A fire detection method and system based on power data analysis

Abstract

The present invention provides a fire detection method and system based on power data analysis. After obtaining a set of power equipment to be analyzed and a historical abnormal monitoring data set of each power equipment to be analyzed in the set of power equipment to be analyzed, one or more abnormal operation events are determined in the historical abnormal monitoring data set. According to the historical abnormal monitoring data set, the power equipment to be analyzed and the abnormal operation events are used as graph vertices to generate a power analysis graph. The power analysis graph is mined for characterization information to obtain a power equipment characterization vector of the power equipment to be analyzed. Based on the power equipment characterization vector, the power equipment to be analyzed is grouped to obtain one or more pending power equipment groups, and a target power equipment group is determined in the pending power equipment groups. The present application can improve the accuracy of the determined power equipment groups, help focus on monitoring the equipment in the group during fire monitoring, and make the fire monitoring more targeted.

Description

Technical Field

[0001] This application relates to the technical field of data processing, and in particular, to a fire detection method and system based on power data analysis. Background Art

[0002] In modern society, the stable operation of the power system is crucial for the normal operation of various fields. As the core component of the power system, the operating state of power equipment directly affects the reliability and safety of power supply. However, during the long-term operation of power equipment, due to various factors such as equipment aging, environmental changes, and improper operation, abnormal operating states may occur. These abnormal operating states not only affect the normal service life of power equipment but may also lead to safety accidents, especially fire accidents, posing a serious threat to people's lives and property safety.

[0003] Existing fire detection methods mainly rely on traditional monitoring means such as smoke sensors and temperature sensors. These sensors can only issue alarms after or when a fire is about to occur, and cannot effectively monitor and warn of potential abnormalities in power equipment in advance. Moreover, traditional monitoring means often can only monitor individual equipment and it is difficult to comprehensively and systematically analyze and evaluate power equipment in the entire power system.

[0004] Some fire detection methods based on data analysis usually simply perform statistics and analysis on the operating data of power equipment, ignoring the correlation between power equipment and the mutual influence between abnormal operating events. This monotonous analysis method is prone to misjudging the abnormal state of power equipment and cannot accurately determine the group of power equipment that needs to be monitored key, thus reducing the pertinence and effectiveness of fire detection. At the same time, due to the lack of comprehensive consideration of the structural characteristics of the abnormal operation map of power equipment, existing methods often need to process a large amount of redundant data during the data analysis process, increasing the cost and difficulty of data analysis. Summary of the Invention

[0005] In view of this, this application provides a fire detection method and system based on power data analysis.

[0006] The technical solution of this application is realized as follows:

[0007] On the one hand, the present application provides a fire detection method based on power data analysis, which is applied to a computer system. The method includes: obtaining a set of power devices to be analyzed and a historical abnormal monitoring data set of each power device to be analyzed in the set of power devices to be analyzed; determining one or more abnormal operation events in the historical abnormal monitoring data set, where the abnormal operation event is an abnormal classification corresponding to the abnormal state of the power device to be analyzed; generating a power analysis graph by using the power device to be analyzed and the abnormal operation event as graph vertices according to the historical abnormal monitoring data set, where the power analysis graph is used to indicate the triggering connection between the power device to be analyzed and the abnormal operation event; mining characterization information of the power analysis graph to obtain a power device characterization vector of the power device to be analyzed; grouping the power devices to be analyzed according to the power device characterization vector to obtain one or more pending power device groups, and determining a target power device group in the pending power device groups, so as to perform fire detection on the power devices in the target power device group.

[0008] On the other hand, the present application provides a computer system, including a memory and a processor. The memory stores a computer program that can run on the processor, and the processor implements the steps in the above method when executing the program.

[0009] After obtaining the set of power devices to be analyzed and the historical abnormal monitoring data set of each power device to be analyzed in the set of power devices to be analyzed, the present invention determines one or more abnormal operation events in the historical abnormal monitoring data set, and generates a power analysis graph by using the power device to be analyzed and the abnormal operation event as graph vertices according to the historical abnormal monitoring data set. Then, it mines the characterization information of the power analysis graph to obtain the power device characterization vector of the power device to be analyzed, and groups the power devices to be analyzed according to the power device characterization vector to obtain one or more pending power device groups, and determines a target power device group in the pending power device groups. Since the present application can determine one or more abnormal operation events in the historical abnormal monitoring data set and generate a power analysis graph based on the abnormal operation events, the graph structure features of the abnormal operation and the features of the graph vertices can be combined. At the same time, based on the power analysis graph, the error of determining the target power device group according to the monotonous connection can be prevented. In other words, the present application can improve the accuracy of the determined power device group, help to focus on monitoring the devices in the group during fire monitoring, make the fire monitoring more targeted, and reduce the cost of data analysis.

[0010] It should be understood that the above general description and the following detailed description are only exemplary and explanatory, and do not limit the technical solution of the present application. Description of the Drawings

[0011] Figure 1 This is a schematic diagram of the implementation process of a fire detection method based on power data analysis provided by an embodiment of the present application.

[0012] Figure 2 This is a schematic diagram of the hardware entity of a computer system provided by an embodiment of the present application. Detailed implementation manners

[0013] An embodiment of the present application provides a fire detection method based on power data analysis, and this method can be executed by a processor of a computer system. Among them, the computer system may refer to devices with data processing capabilities such as servers, laptops, tablets, desktop computers, mobile devices, etc.

[0014] Figure 1 This is a schematic diagram of the implementation process of a fire detection method based on power data analysis provided by an embodiment of the present application. As Figure 1 shown, this method includes:

[0015] Step 100: Obtain a set of power devices to be analyzed and a historical abnormal monitoring data set of each power device to be analyzed in the set of power devices to be analyzed.

[0016] In step 100, obtain a set of power devices to be analyzed and a historical abnormal monitoring data set of each power device to be analyzed in this set. The set of power devices to be analyzed is a combination of a series of power devices to be deeply analyzed. These power devices can be different types of devices in the same substation, such as transformers, circuit breakers, capacitors, etc., or power devices distributed in different regions but having similar functions or operating characteristics.

[0017] The historical anomaly monitoring dataset is a dataset that records information related to abnormal conditions for each power device to be analyzed over a past period of time. Such information typically includes the time when the anomaly occurred, the type of anomaly, the operating parameters of the power device when the anomaly occurred, etc. For example, for a transformer, the historical anomaly monitoring dataset may contain records of multiple over-temperature anomalies, where each record details the specific time when the over-temperature occurred, the oil temperature, winding temperature, load current, etc. of the transformer at that time. It can be screened from the asset database of the power system. The asset database stores the basic information of all power devices owned by the power company, including the type, location, and commissioning time of the devices. The set of power devices to be analyzed can be determined by setting screening conditions according to the purpose of the analysis. If analyzing the operating conditions of power devices in a specific area, all power devices in that area can be screened based on the geographical location information of the devices; if analyzing the abnormal conditions of a certain type of power device, all power devices of that type can be screened based on the type information of the devices.

[0018] When obtaining the historical anomaly monitoring dataset for each power device to be analyzed, data can be extracted from the monitoring system of the power device. The power device is equipped with various sensors for real-time monitoring of the device's operating status, and the data monitored by these sensors is transmitted to the monitoring system for storage. By establishing a data connection with the monitoring system and according to the identification information of the device, the historical anomaly monitoring data of the corresponding power device can be extracted from the monitoring system.

[0019] Step 200: Determine one or more abnormal operation events in the historical anomaly monitoring dataset. An abnormal operation event is the abnormal classification corresponding to the abnormal state of the power device to be analyzed.

[0020] In step 200, determining one or more abnormal operation events in the historical anomaly monitoring dataset, where the abnormal operation event is the abnormal classification corresponding to the abnormal state of the power device to be analyzed, can help extract key abnormal information from the vast amount of historical anomaly monitoring data. The historical anomaly monitoring dataset contains detailed records of abnormal conditions of the power device to be analyzed over a past period of time, and these records may involve various operating parameters and time information of the device.

[0021] To determine abnormal operation events from this data, a data classification algorithm can be adopted to classify historical abnormal monitoring data according to the characteristics and types of abnormalities. For the abnormal data of transformers, it can be classified according to different types such as temperature abnormality, current abnormality, etc.; for the abnormal data of circuit breakers, it can be classified according to abnormal opening and closing time, abnormal contact wear, etc. Through this classification, similar abnormal data can be grouped into one category, thus determining different abnormal operation events. Machine learning algorithms can also be used to identify abnormal operation events. By learning a large amount of historical abnormal monitoring data, an abnormal recognition model can be established, which can automatically identify new abnormal data and classify it into the corresponding abnormal operation events. The support vector machine (SVM) algorithm can be adopted, using historical abnormal monitoring data as training samples to train an SVM model that can accurately identify abnormal operation events.

[0022] In practical applications, the severity of abnormalities can be judged by setting thresholds. For the abnormal oil temperature of transformers, if the oil temperature exceeds a certain proportion of the normal operation range, it is considered a severe abnormality; if the oil temperature only slightly exceeds the normal range, it is considered a minor abnormality. Corresponding early warning strategies can be formulated according to the classification of abnormal operation events to notify relevant personnel for handling in a timely manner. Association analysis can also be carried out on abnormal operation events to find out the potential connections between different abnormal operation events. In a power system, the abnormal temperature of a transformer may be related to the abnormal load current. Through the association analysis of these abnormal operation events, the operation status of power equipment can be deeply understood and potential safety hazards can be discovered. The association rule mining algorithm is adopted to analyze historical abnormal monitoring data to find out the association rules between different abnormal operation events. Through the Apriori algorithm, frequently occurring combinations of abnormal events can be mined from historical abnormal monitoring data, thus discovering the association relationships between different abnormal operation events.

[0023] Step 300: According to the historical abnormal monitoring data set, generate a power analysis graph with the power equipment to be analyzed and the abnormal operation events as graph vertices. The power analysis graph is used to indicate the triggering relationship between the power equipment to be analyzed and the abnormal operation events.

[0024] In step 300, according to the historical abnormal monitoring data set, a power analysis graph is generated with the power equipment to be analyzed and the abnormal operation events as graph vertices. This graph is used to indicate the triggering relationship between the power equipment to be analyzed and the abnormal operation events, which is crucial for deeply understanding the operation status and potential risks of power equipment. The historical abnormal monitoring data set relied on contains detailed records of the abnormal situations that occurred during the past operation of the power equipment to be analyzed, and these records cover various operation parameters of the equipment, the time of abnormality occurrence, etc.

[0025] To clarify the triggering relationship between the power equipment to be analyzed and abnormal operation events, relevant information needs to be extracted from the historical abnormal monitoring dataset. For example, information such as the number of times, time, and frequency of each power equipment triggering specific abnormal operation events is counted. For a transformer, the number of times of over-temperature anomalies triggered in the past year, the specific time of each over-temperature anomaly occurrence, and the frequency of over-temperature anomaly occurrence can be counted. By analyzing these data, the degree of association between power equipment and abnormal operation events can be understood.

[0026] To more accurately determine the triggering relationship, data analysis algorithms can be used. For example, association rule mining algorithms such as the Apriori algorithm can be used to mine the frequent association rules between power equipment and abnormal operation events from the historical abnormal monitoring data. This algorithm calculates the frequency of item sets in the dataset to find association rules with higher support and confidence. Support represents the frequency of an item set appearing in the dataset, and confidence represents the probability of another item set appearing under the condition that an item set appears. Let the support of power equipment A triggering abnormal operation event B be , where represents the number of times power equipment A triggers abnormal operation event B, and N represents the total number of records; the confidence is , where Count(A) represents the number of abnormal records of power equipment A. The abnormal classification information of abnormal operation events can also be considered. Different types of abnormal operation events may have different impacts on power equipment. For transformers, the severity and potential risks of over-temperature anomalies and over-current anomalies may be different. The triggering relationship can be weighted according to the abnormal classification information to more accurately reflect the relationship between power equipment and abnormal operation events. After determining the triggering relationship, the power equipment to be analyzed and the abnormal operation events are used as graph vertices to generate a power analysis graph. Graph vertices are the basic elements in the graph structure and represent power equipment and abnormal operation events here. A graph database can be used to store and manage the power analysis graph. The graph database can efficiently process graph structure data and support fast graph query and analysis operations. When generating the power analysis graph, edges can be added between graph vertices according to the triggering relationship. The attributes of the edges can include information such as the number of triggers, time, and frequency. For the two graph vertices of the transformer and over-temperature anomaly, if the transformer triggers the over-temperature anomaly 10 times in the past year, an edge can be added between these two graph vertices, and the attribute of the edge can be set to the number of triggers 10 times.

[0027] Step 400: Mine the characterization information of the power analysis graph to obtain the power equipment characterization vector of the power equipment to be analyzed.

[0028] In step 400, characterization information mining is performed on the power analysis graph to obtain the power equipment characterization vector of the power equipment to be analyzed. This step is an important link to extract key information from the power analysis graph for more accurate analysis of the power equipment. The power analysis graph is generated based on the historical anomaly monitoring data set, which contains the power equipment to be analyzed and the abnormal operation events as graph vertices, as well as the triggering relationships between them. For example, in the power system of a factory, the power analysis graph may show the associations between multiple motors and abnormal operation events such as overload anomalies and overheat anomalies.

[0029] To perform characterization information mining on the power analysis graph, a characterization vector needs to be assigned to each graph vertex in the graph to represent the feature information of that vertex. These characterization vectors can be constructed based on the attribute information of the graph vertices, such as the type and rated power of the power equipment, and the type and severity of the abnormal operation events. For a transformer, the characterization vector of its graph vertex may include information such as the capacity, voltage level, and historical fault times of the transformer; for the abnormal operation event of over-temperature anomaly, the characterization vector of its graph vertex may include information such as the frequency and duration of the anomaly occurrence. To make the characterization vectors of different graph vertices comparable, these graph vertex characterization vectors need to be projected onto a unified characterization domain to obtain the basic graph vertex characterization vectors. Since the original characterization vectors of different graph vertices may have different dimensions and scales, it is difficult to directly compare and analyze them. A linear transformation method can be used. By determining a transformation matrix, the graph vertex characterization vector is projected onto the graph vertex characterization vector corresponding to the preset unified projection standard dimension. Let the preset unified projection standard dimension be m, the actual vector dimension of the graph vertex characterization vector be n, the transformation matrix be T, and the graph vertex characterization vector be X. Then the projected basic graph vertex characterization vector Y = T×X, where T is an m×n matrix.

[0030] After obtaining the basic graph vertex characterization vectors, these vectors need to be integrated to obtain the power equipment characterization vector of the power equipment to be analyzed. The operating state of the power equipment is not only related to its own characteristics but also related to the surrounding abnormal operation events and the states of other relevant power equipment. The connection relationships between the graph vertices need to be considered, and the characterization vectors of the target graph vertex (i.e., the graph vertex corresponding to the power equipment to be analyzed) and its connected graph vertices (adjacent graph vertices) are integrated. The method of graph neural network (GNN) can be used for vector integration. The graph neural network can effectively process graph-structured data. Through the message passing mechanism, the information of the target graph vertex and its connected graph vertices is aggregated. The target graph vertex can receive information from its connected graph vertices and update its own characterization vector based on this information. Different weights can be assigned to the information transmission of the connected graph vertices according to the attributes of the edges between the graph vertices, such as the triggering times and time, to more accurately reflect the degree of association between them.

[0031] The characterization vectors of power equipment can comprehensively reflect the operating status and characteristics of power equipment. By analyzing the characterization vectors of power equipment, the similarities and differences between power equipment can be discovered, providing a basis for subsequent grouping of power equipment. If the characterization vectors of two power equipment have a high degree of similarity, it indicates that they are similar in terms of operating status and abnormal triggering conditions, etc., and may belong to the same type of equipment group.

[0032] The characterization vectors of power equipment can also be used for fault prediction and risk assessment. By establishing a fault prediction model, using the characterization vectors of power equipment as input, the probability and type of faults occurring in power equipment can be predicted. Machine learning algorithms such as support vector machines and random forests can be used to train a large amount of historical data to obtain an accurate fault prediction model. The potential risks of power equipment can also be evaluated based on the characterization vectors of power equipment, providing a reference for fire detection. For power equipment with high risks, early warnings can be issued in a timely manner to remind relevant personnel to conduct inspections and maintenance to prevent the occurrence of safety accidents such as fires.

[0033] Step 500: Group the power equipment to be analyzed according to the characterization vectors of power equipment, obtain one or more pending power equipment groups, and determine the target power equipment group from the pending power equipment groups, so as to perform fire detection on the power equipment in the target power equipment group.

[0034] In step 500, group the power equipment to be analyzed according to the characterization vectors of power equipment, obtain one or more pending power equipment groups, and determine the target power equipment group from the pending power equipment groups, so as to perform fire detection on the power equipment in the target power equipment group. The characterization vectors of power equipment are obtained after mining the characterization information of the power analysis atlas, and it comprehensively reflects the operating status and characteristics of power equipment.

[0035] Determine the power equipment commonality metric value between the power equipment to be analyzed in the set of power equipment to be analyzed. This process is achieved by calculating the characterization vectors of power equipment. The power equipment commonality metric value is used to measure the similarity degree between different power equipment. The higher the similarity degree, the closer these power equipment are in terms of operating status, abnormal triggering conditions, etc. The cosine similarity algorithm can be used to calculate the power equipment commonality metric value. Suppose the characterization vectors of two power equipment are A and B respectively, then their cosine similarity , where is the dot product of vectors A and B, and are the norms of vectors A and B respectively. The closer the value of the cosine similarity is to 1, the more similar the characterization vectors of the two power equipment are, that is, the higher their commonality metric value.

[0036] Based on the calculated commonality metric values of power equipment, the power equipment to be analyzed is used as graph vertices to generate an interconnected graph. The interconnected graph is an undirected graph, where each graph vertex represents a power equipment, and the edges between vertices represent the commonality metric values between power equipment. The weight of the edge can be set as the commonality metric value of the power equipment. The larger the weight, the higher the similarity between the two power equipment. In the power system of an industrial park, the interconnected graph can visually display the similarity relationships between different power equipment.

[0037] Next, the graph vertices in the interconnected graph are grouped to obtain one or more groups of power equipment to be determined. The purpose of grouping is to divide similar power equipment into the same group for more targeted analysis and management. Community discovery algorithms, such as the Louvain algorithm, can be used to group the interconnected graph. The Louvain algorithm is a community discovery algorithm based on modularity optimization. It continuously moves graph vertices to different communities to maximize the grouping effectiveness of the entire graph. The grouping effectiveness Q is an index to measure the quality of community division in the graph, and its calculation formula is , where m is the total number of edges in the graph, is the weight of the edge between vertices i and j in the graph, and are the degrees of vertices i and j respectively, and are the communities to which vertices i and j belong respectively, is an indicator function, which is 1 when , and 0 otherwise.

[0038] After obtaining the groups of power equipment to be determined, the target groups of power equipment are determined among these groups. The target groups of power equipment are composed of those power equipment that are more likely to have fire hazards. The groups of power equipment to be determined can be evaluated and screened according to factors such as the historical abnormal data and operating environment of the power equipment. Once the target groups of power equipment are determined, fire detection can be performed on the power equipment in these groups. Fire detection can include real-time monitoring of parameters such as the temperature, current, and voltage of the power equipment, and inspection of the insulation performance and grounding conditions of the power equipment. Relevant parameters can be obtained through sensors installed on the power equipment, and these parameters are compared with preset safety thresholds. Once it is found that the parameters exceed the thresholds, early warning signals are immediately sent out to remind relevant personnel to handle them. Through the above methods, it is possible to effectively group the power equipment to be analyzed based on the power equipment characterization vectors, determine the target groups of power equipment, and perform fire detection on them, which helps to improve the safety and reliability of the power system, timely discover and eliminate potential fire hazards, and ensure the normal operation of power equipment and the safety of the lives and property of personnel.

[0039] In one implementation, in step 200, one or more abnormal operation events are determined from the historical abnormal monitoring dataset, including:

[0040] Step 210: Extract the abnormal category data of the power equipment to be analyzed from the historical abnormal monitoring dataset;

[0041] Step 220: Determine one or more abnormal operation events corresponding to the power equipment to be analyzed from the abnormal category data.

[0042] In step 210, the abnormal category data of the power equipment to be analyzed is extracted from the historical abnormal monitoring dataset. The historical abnormal monitoring dataset contains detailed records of the abnormal conditions that occurred during the past operation of the power equipment to be analyzed. These records cover various operation parameters of the equipment, the time when the abnormality occurred, and other information.

[0043] To extract the abnormal category data from the historical abnormal monitoring dataset, methods of data screening and classification can be used. The historical abnormal monitoring data can be screened according to the preset abnormal type labels, and the data that meets the specific abnormal type can be extracted. For over-temperature abnormalities, a temperature threshold can be set. When the monitored equipment temperature exceeds this threshold, it is marked as over-temperature abnormal data. Data classification algorithms, such as decision tree algorithms, can also be used to classify the historical abnormal monitoring data and distinguish different types of abnormal data. The decision tree algorithm constructs a decision tree model and classifies according to the characteristics of the data, and can effectively identify different types of abnormal data.

[0044] After the abnormal category data is extracted, in step 220, one or more abnormal operation events corresponding to the power equipment to be analyzed are determined from the abnormal category data. The abnormal operation event is the abnormal classification corresponding to the abnormal state of the power equipment to be analyzed, and it is a further abstraction and induction of the abnormal situation. For the over-temperature abnormality of the transformer, the abnormal operation event can be divided into different levels of abnormal classifications such as mild over-temperature, moderate over-temperature, and severe over-temperature.

[0045] To determine abnormal operation events in abnormal category data, rule matching and machine learning methods can be adopted. Rule matching means matching the abnormal category data according to preset rules to determine the corresponding abnormal operation events. For the over-temperature abnormality of a transformer, different temperature ranges can be set. When the monitored temperature falls within different ranges, it is classified as over-temperature abnormal operation events of different levels. The machine learning method is to establish an abnormal operation event recognition model through learning a large amount of historical abnormal category data. This model can automatically identify the abnormal operation events corresponding to new abnormal category data. The support vector machine (SVM) algorithm can be used. Taking the historical abnormal category data as training samples, an SVM model that can accurately identify abnormal operation events is trained. In practical applications, the severity of an abnormality can be judged by setting a threshold. For the over-current abnormality of a transformer, if the current exceeds a certain proportion of the rated current, it is considered a serious abnormality; if the current only slightly exceeds the rated current, it is considered a minor abnormality. Corresponding early warning strategies can be formulated according to the classification of abnormal operation events to notify relevant personnel for handling in a timely manner. Association analysis can also be performed on abnormal operation events to find the potential connections between different abnormal operation events. In a power system, the over-temperature abnormality of a transformer may be related to the over-current abnormality. Through the association analysis of these abnormal operation events, the operating status of power equipment can be deeply understood and potential safety hazards can be discovered.

[0046] The association rule mining algorithm can be used to analyze the abnormal category data to find the association rules between different abnormal operation events. Through the Apriori algorithm, frequently occurring combinations of abnormal events can be mined from the abnormal category data, thereby discovering the association relationships between different abnormal operation events.

[0047] In one implementation, step 220 of determining one or more abnormal operation events corresponding to the power equipment to be analyzed in the abnormal category data includes:

[0048] Step 221: Extract the abnormal occurrence time and abnormal electrical nodes of the power equipment to be analyzed from the abnormal category data;

[0049] Step 222: Determine one or more abnormal operation events corresponding to the power equipment to be analyzed based on the abnormal occurrence time and abnormal electrical nodes.

[0050] In step 221, the abnormal occurrence time and abnormal electrical nodes of the power equipment to be analyzed are extracted from the abnormal category data. The abnormal category data is the data related to the abnormal types extracted from the historical abnormal monitoring dataset, which contains various information when the power equipment to be analyzed has an abnormality. Taking the power system of a large factory as an example, the abnormal category data may cover different abnormal situation records of equipment such as transformers and motors. The abnormal occurrence time refers to the specific moment when the power equipment is in an abnormal state, which can reflect the time pattern and trend of the abnormality occurrence. For the over-temperature abnormality of a transformer, the abnormal occurrence time can be accurate to the specific year, month, day, hour, minute, and second. By analyzing these time information, it can be found whether the over-temperature abnormality occurs frequently in a specific time period, such as during the high-temperature period in summer or when the equipment is operating at high load. The abnormal electrical node refers to the specific electrical part where the abnormality occurs in the power equipment, which can help locate the specific position where the abnormality occurs. In a complex power network, a transformer may have multiple electrical nodes, such as the high-voltage side winding and the low-voltage side winding. The abnormal electrical node can clearly indicate which winding has an abnormality.

[0051] To extract the abnormal occurrence time and abnormal electrical nodes from the abnormal category data, data parsing and screening methods can be adopted. According to the storage format of the abnormal category data, the corresponding parsing algorithm can be used to extract the time information and electrical node information therein. If the abnormal category data is stored in a table form, the columns where the abnormal occurrence time and abnormal electrical nodes are located can be located through column names or data positions, and then these data can be extracted. Screening conditions can also be set to only extract the abnormal occurrence time and abnormal electrical node data related to the power equipment to be analyzed, so as to improve the accuracy and pertinence of the data.

[0052] After the abnormal occurrence time and abnormal electrical nodes are extracted, in step 222, one or more abnormal operation events corresponding to the power equipment to be analyzed are determined based on this information. The abnormal operation event is the classification and induction of the abnormal state of the power equipment, which can help to understand the abnormal situation of the power equipment more clearly. For a transformer, the abnormal operation events can include over-temperature abnormality, over-current abnormality, short-circuit abnormality, etc.

[0053] When determining abnormal operation events based on the abnormal occurrence time and abnormal electrical nodes, rule matching and machine learning methods can be used. Rule matching means matching the abnormal occurrence time and abnormal electrical nodes according to preset rules to determine the corresponding abnormal operation events. For the over-temperature abnormality of a transformer, if the abnormal occurrence time is during the high-temperature period in summer and the abnormal electrical node is the high-voltage side winding of the transformer, it can be classified as an over-temperature abnormal operation event of the high-voltage side winding caused by high temperature in summer according to the preset rules. The machine learning method is to establish an abnormal operation event recognition model through learning a large amount of historical abnormal data. This model can automatically identify the corresponding abnormal operation event according to the input abnormal occurrence time and abnormal electrical node information. The neural network algorithm can be used. Taking the abnormal occurrence time and abnormal electrical nodes in the historical abnormal data as inputs and the abnormal operation event as the output, a neural network model that can accurately identify abnormal operation events can be trained. When determining abnormal operation events, factors such as the frequency and duration of the abnormal occurrence can also be considered. If an abnormal electrical node frequently has abnormalities in a short period of time, it indicates that there may be serious potential fault hazards at this node, and the corresponding abnormal operation event can be marked as a high-risk event. The duration of the abnormality can be calculated based on the abnormal occurrence time. If the abnormal duration is relatively long, it will also increase the risk level of the abnormal operation event. Association analysis can also be performed on abnormal operation events. Combining the abnormal occurrence time and abnormal electrical nodes, the potential connections between different abnormal operation events can be found. In a power system, the over-temperature abnormality and over-current abnormality of a transformer may occur simultaneously and the abnormal electrical nodes are close. Through the association analysis of these abnormal operation events, the operating status of power equipment can be deeply understood and potential safety hazards can be discovered.

[0054] In one implementation, step 300, based on the historical abnormal monitoring data set, generate a power analysis graph with the power equipment to be analyzed and the abnormal operation event as graph vertices, including:

[0055] Step 310: Determine the trigger statistical information of the power equipment to be analyzed and the abnormal operation event in the historical abnormal monitoring data set;

[0056] Step 320: Determine the trigger connection between the power equipment to be analyzed and the abnormal operation event according to the abnormal classification information and trigger statistical information of the abnormal operation event;

[0057] Step 330: Generate a power analysis graph with the power equipment to be analyzed and the abnormal operation event as graph vertices according to the trigger connection.

[0058] In step 310, trigger statistical information of the power equipment to be analyzed and abnormal operation events is determined from the historical abnormal monitoring dataset. The historical abnormal monitoring dataset contains detailed records of abnormal situations that occurred during the past operation of the power equipment to be analyzed. These records cover various operating parameters of the equipment, the time when the abnormality occurred, and other information. Taking a power system as an example, the historical abnormal monitoring dataset may contain voltage and current abnormal records of equipment such as transformers, distribution cabinets, and elevator control cabinets, as well as the specific moments when these abnormalities occurred. The trigger statistical information refers to the relevant statistical data of the power equipment to be analyzed triggering abnormal operation events, such as the number of triggers, time, frequency, etc. These information can reflect the degree of association between the power equipment and abnormal operation events. For a transformer, the trigger statistical information can include the number of times it triggered overheating abnormalities in the past year, the specific time of each overheating abnormality occurrence, and the frequency of overheating abnormality occurrence.

[0059] After determining the trigger statistical information, in step 320, based on the abnormal classification information and trigger statistical information of the abnormal operation event, the trigger connection between the power equipment to be analyzed and the abnormal operation event is determined. The abnormal classification information refers to the information for classifying abnormal operation events. Different types of abnormal operation events may have different impacts on power equipment. For transformers, the severity and potential risks of overheating abnormalities and overcurrent abnormalities may be different. The trigger connection refers to the association relationship between the power equipment to be analyzed and the abnormal operation event, which can reflect the possibility and intensity of the power equipment triggering the abnormal operation event. A weighted calculation method can be used to determine the trigger connection based on the abnormal classification information and trigger statistical information. For different types of abnormal operation events, different weights can be set according to their severity and potential risks. For overheating abnormalities, since they may cause transformer insulation aging and even lead to fires, a higher weight can be set; for minor voltage fluctuation abnormalities, a lower weight can be set. The trigger statistical information can be weighted with the corresponding weights to obtain the trigger connection value between the power equipment to be analyzed and the abnormal operation event. Suppose the weight of the abnormal operation event is , the number of times the power equipment D j triggers the abnormal operation event E i is n ij , then the trigger connection value C j between the power equipment D i and the abnormal operation event E ij = w i × n ij .

[0060] Factors such as trigger time and frequency can also be considered to more accurately evaluate the trigger relationship. If a power device frequently triggers a certain abnormal operation event within a short period of time, it indicates a strong trigger relationship between the device and the abnormal operation event. A time decay factor can be introduced to attenuate the trigger events with a relatively distant trigger time, so as to highlight the impact of recent trigger events. Let the time decay factor be , be the power device triggering the abnormal operation event E i . If the current time is t0, then the trigger relationship value after considering time decay is .

[0061] After determining the trigger relationship, in step 330, according to the trigger relationship, the power device to be analyzed and the abnormal operation event are used as graph vertices to generate a power analysis graph. A graph vertex is a basic element in a graph structure, which represents the power device to be analyzed and the abnormal operation event here. The power analysis graph is a graph structure that can intuitively display the trigger relationship between the power device to be analyzed and the abnormal operation event. A graph database can be used to store and manage the power analysis graph. A graph database is a database specifically used to process graph-structured data, which can efficiently store and query information about graph vertices and edges. The power device to be analyzed and the abnormal operation event can be stored as graph vertices in the graph database, and edges are added between the graph vertices according to the trigger relationship. The attribute of the edge can be set as the trigger relationship value, and the greater the weight of the edge, the stronger the trigger relationship between the power device to be analyzed and the abnormal operation event. The power analysis graph can be dynamically updated. As new historical abnormal monitoring data is continuously generated, the information of graph vertices and edges in the graph can be updated in real time to reflect the latest changes in the operating state of the power device. New data can be periodically extracted from the historical abnormal monitoring dataset, the trigger statistical information and trigger relationship are recalculated, and then the power analysis graph is updated.

[0062] Risk assessment can be performed based on the power analysis graph. For each power device to be analyzed, its risk score can be calculated according to the trigger relationship between it and the abnormal operation event. The higher the risk score, the greater the likelihood of the device failing.

[0063] Through the above methods, the trigger statistical information can be accurately determined from the historical abnormal monitoring dataset, the trigger relationship is determined based on the abnormal classification information and trigger statistical information, and the power analysis graph is generated according to the trigger relationship. The power analysis graph can intuitively display the relationship between the power device to be analyzed and the abnormal operation event, providing an important basis for subsequent steps such as characterization information mining, power device grouping, and fire detection, helping to improve the safety and reliability of the power system and reduce potential fire risks.

[0064] In one implementation, in step 400, perform characterization information mining on the power analysis graph spectrum to obtain the power equipment characterization vector of the power equipment to be analyzed, including:

[0065] Step 410: Extract the graph vertex characterization vector of each graph vertex in the power analysis graph spectrum;

[0066] Step 420: Project the graph vertex characterization vector into a unified characterization domain to obtain the basic graph vertex characterization vector;

[0067] Step 430: Integrate the basic graph vertex characterization vectors to obtain the power equipment characterization vector of the power equipment to be analyzed.

[0068] In step 410, when extracting the graph vertex characterization vector in the power analysis graph spectrum, the power analysis graph spectrum is generated based on the historical anomaly monitoring data set, which includes the power equipment to be analyzed and the abnormal operation events as graph vertices, as well as the triggering relationships between them. The graph vertex characterization vector is a vector used to represent the characteristics of the graph vertex, which can comprehensively reflect various attribute information of the power equipment or abnormal operation event represented by the graph vertex. Taking the power analysis graph spectrum of the power supply network in a city as an example, for the graph vertex representing the transformer, its graph vertex characterization vector may include information such as the capacity, voltage level, operation years, and historical fault times of the transformer; for the graph vertex representing the over-temperature anomaly, its graph vertex characterization vector may include information such as the frequency, duration, and influence range of the anomaly occurrence.

[0069] In order to extract the graph vertex characterization vector in the power analysis graph spectrum, it can be constructed based on the attribute information of the graph vertex. If the power analysis graph spectrum is stored in the graph database, the attribute data of each graph vertex can be obtained by querying the graph database. These attribute data can be converted into vector form according to the pre-defined feature dimensions. For the graph vertex of the transformer, the attribute data such as capacity, voltage level, operation years, and historical fault times can be arranged in a certain order to form a vector. The method of feature engineering can also be used to process and transform the attribute data to improve the quality of the graph vertex characterization vector. For some continuous attribute data, normalization processing can be performed to map it to the interval of [0, 1] to eliminate the dimensional difference between different attribute data.

[0070] After obtaining the graph vertex representation vectors of each graph vertex, in step 420, project the graph vertex representation vectors into a unified representation domain to obtain the basic graph vertex representation vectors. Since the original representation vectors of different graph vertices may have different dimensions and scales, it is difficult to directly compare and analyze them. The unified representation domain refers to a vector space with the same dimension and scale. After projecting the graph vertex representation vectors into this space, they can be compared and analyzed under the same standard. A linear transformation method can be used. By determining a transformation matrix, project the graph vertex representation vectors into the graph vertex representation vectors corresponding to the preset unified projection standard dimension. Let the preset unified projection standard dimension be m, the actual vector dimension of the graph vertex representation vector be n, the transformation matrix be T, and the graph vertex representation vector be X. Then the projected basic graph vertex representation vector Y = T×X, where T is an m×n matrix. The transformation matrix T can be determined by methods such as singular value decomposition (SVD). Singular value decomposition is a matrix decomposition method that can decompose a matrix into the product of three matrices. By processing these matrices, a suitable transformation matrix T can be obtained.

[0071] After projecting the graph vertex representation vectors into a unified representation domain, the representation vectors of different graph vertices become comparable, making it more convenient to perform subsequent analysis and processing. By calculating the similarity between the basic graph vertex representation vectors, the similarity relationships between graph vertices can be found. For example, which power equipment is more similar in operation characteristics, and which abnormal operation events have similar characteristics. After obtaining the basic graph vertex representation vectors, in step 430, integrate the basic graph vertex representation vectors to obtain the power equipment representation vector of the power equipment to be analyzed. The operating state of a power equipment is not only related to its own characteristics but also related to the abnormal operation events around it and the states of other relevant power equipment. It is necessary to consider the connection relationships between graph vertices and integrate the representation vectors of the target graph vertex (i.e., the graph vertex corresponding to the power equipment to be analyzed) and its connected graph vertices (adjacent graph vertices).

[0072] The method of graph neural network (GNN) can be used for vector integration. A graph neural network is a neural network specifically designed to process graph-structured data, which can effectively capture the dependence relationships between graph vertices. Through the message passing mechanism, a graph neural network can aggregate the information of the target graph vertex and its connected graph vertices. The target graph vertex can receive information from its connected graph vertices and update its own representation vector according to this information. Different weights can be assigned to the information transfer of connected graph vertices according to the attributes of the edges between graph vertices, such as the trigger times, time, etc., to more accurately reflect the degree of association between them.

[0073] The attention mechanism can also be used to enhance the effect of vector integration. The attention mechanism can automatically assign different attention weights to different vertices of the connection graph, enabling the target graph vertex to pay more attention to the connection graph vertices with stronger relevance to itself. Let the target graph vertex be and its connection graph vertices be . The attention weight can be calculated by the following formula: , where represents the similarity between the target graph vertex and the connection graph vertex , and represents the set of connection graph vertices of the target graph vertex v t . Through the attention mechanism, the basic graph vertex representation vectors can be integrated more flexibly, improving the quality of the power equipment representation vectors.

[0074] The power equipment representation vector can comprehensively reflect the operating state and characteristics of the power equipment. By analyzing the power equipment representation vectors, the similarities and differences between power equipment can be discovered, providing a basis for subsequent grouping of power equipment. If the representation vectors of two power equipment have a high similarity, it indicates that they have similarities in aspects such as operating state and abnormal triggering conditions and may belong to the same equipment group.

[0075] The power equipment representation vector can be used for fault prediction and risk assessment. By establishing a fault prediction model, taking the power equipment representation vector as the input, the probability and type of power equipment failure can be predicted. Machine learning algorithms such as support vector machines and random forests can be used to train a large amount of historical data to obtain an accurate fault prediction model. Let the power equipment representation vector be x and the fault prediction model be f(x), then the predicted fault probability P = f(x).

[0076] The potential risk of power equipment can also be evaluated based on the power equipment representation vector, providing a reference for fire detection. For power equipment with a high risk, early warnings can be issued in a timely manner to remind relevant personnel to conduct inspections and maintenance to prevent the occurrence of safety accidents such as fires. The risk score of power equipment can be calculated based on certain characteristics in the power equipment representation vector, such as the frequency and severity of anomalies. Let the power equipment representation vector be x and the risk assessment function be g(x), then the risk score R of the power equipment = g(x).

[0077] Through the above methods, effective characterization information mining of the power analysis graph can be carried out, successively completing the extraction of graph vertex representation vectors, projection into a unified representation domain, and integration of basic graph vertex representation vectors, and finally obtaining the power equipment representation vector of the power equipment to be analyzed.

[0078] In one implementation, step 420 projects the graph vertex representation vectors into a unified representation domain to obtain the basic graph vertex representation vectors, including:

[0079] Step 421: Obtain the preset unified projection standard dimension and the actual vector dimension of each graph vertex representation vector;

[0080] Step 422: Determine the transformation matrix corresponding to the graph vertex representation vector according to the preset unified projection standard dimension and the actual vector dimension;

[0081] Step 423: Project the graph vertex representation vector onto the graph vertex representation vector corresponding to the preset unified projection standard dimension according to the transformation matrix to obtain the basic graph vertex representation vector.

[0082] In step 421, the preset unified projection standard dimension and the actual vector dimension of each graph vertex representation vector are obtained. The preset unified projection standard dimension is a fixed dimension set in advance, which is used to unify different graph vertex representation vectors into the vector space of this dimension for subsequent comparison and analysis. The actual vector dimension of each graph vertex representation vector refers to the dimension that each graph vertex representation vector has before the projection operation. Taking a power analysis graph containing various power equipment and abnormal operation events as an example, for the graph vertex representing a transformer, its graph vertex representation vector may contain information such as the capacity, voltage level, operation years, and historical fault times of the transformer, and the actual vector dimension may be 4; while for the graph vertex representing an overcurrent anomaly, its graph vertex representation vector may contain information such as the frequency, duration, and influence range of the anomaly occurrence, and the actual vector dimension may be 3.

[0083] To obtain the preset unified projection standard dimension, it can be set according to specific analysis requirements and subsequent processing requirements. If a specific machine learning algorithm needs to be used for analysis later, and this algorithm has certain requirements for the dimension of the input vector, the preset unified projection standard dimension will be set to the dimension required by this algorithm. It is also possible to select a suitable preset unified projection standard dimension through experiments and optimization methods to improve the effect of subsequent analysis.

[0084] While obtaining the preset unified projection standard dimension, the actual vector dimension of each graph vertex representation vector is obtained. The dimension information can be directly obtained by querying the data structure of the graph vertex representation vector. If the graph vertex representation vector is stored in the form of an array, its actual vector dimension can be determined by obtaining the length of the array.

[0085] After obtaining the preset unified projection standard dimension and the actual vector dimension of each graph vertex representation vector, in step 422, a transformation matrix corresponding to the graph vertex representation vector is determined based on this information. The transformation matrix is a matrix used to project the graph vertex representation vector from its actual vector dimension to the preset unified projection standard dimension. Let the preset unified projection standard dimension be m, the actual vector dimension of the graph vertex representation vector be n, and the transformation matrix be T. Then T is an m×n matrix. The singular value decomposition (SVD) method can be used to determine the transformation matrix T. Singular value decomposition is a matrix decomposition method that can decompose a matrix A into the product of three matrices, namely , where U is an m×m orthogonal matrix, is an m×n diagonal matrix, and V T is an n×n orthogonal matrix. The matrix composed of the graph vertex representation vectors can be subjected to singular value decomposition, and then appropriate parts can be selected from the decomposed matrices to form the transformation matrix T according to the preset unified projection standard dimension and the actual vector dimension. In addition to singular value decomposition, the principal component analysis (PCA) method can also be used to determine the transformation matrix.

[0086] After determining the transformation matrix T, in step 423, according to the transformation matrix, the graph vertex representation vector is projected onto the graph vertex representation vector corresponding to the preset unified projection standard dimension to obtain the basic graph vertex representation vector. Let the graph vertex representation vector be X and the transformation matrix be T. Then the projected basic graph vertex representation vector Y can be calculated by the formula Y = T×X.

[0087] When performing the projection operation, the same processing is carried out on each graph vertex representation vector. For each graph vertex in the power analysis graph spectrum, the corresponding graph vertex representation vector X is multiplied by the transformation matrix T through matrix multiplication to obtain the projected basic graph vertex representation vector Y. In this way, all graph vertex representation vectors can be unified into a vector space with a preset unified projection standard dimension. The original representation vectors of different graph vertices may have different dimensions and scales, and it is difficult to directly compare and analyze them. By projecting into a unified representation domain, the differences in dimensions and scales can be eliminated, enabling the representation vectors of different graph vertices to be compared and analyzed under the same standard. This helps to more accurately discover the similarity relationships and differences between graph vertices, providing a more reliable data basis for subsequent analysis such as power equipment grouping and fault prediction. After obtaining the basic graph vertex representation vectors, these vectors can be further analyzed and processed. The similarity between the basic graph vertex representation vectors can be calculated to discover the similarities between power equipment or abnormal operation events. Commonly used similarity calculation methods include cosine similarity, Euclidean distance, etc. Cluster analysis can also be performed using the basic graph vertex representation vectors to group similar graph vertices into the same category. Commonly used clustering algorithms include K - means clustering, hierarchical clustering, etc. Through cluster analysis, power equipment or abnormal operation events can be classified for more targeted management and processing.

[0088] In addition, the basic graph vertex representation vectors can be used as inputs to construct machine learning models for fault prediction and risk assessment. For the fault prediction of power equipment, machine learning algorithms such as support vector machines and neural networks can be used. Taking the basic graph vertex representation vectors as input features, a model capable of accurately predicting the faults of power equipment can be trained. Let the basic graph vertex representation vector be Y and the fault prediction model be f(Y), then the predicted fault probability P = f(Y).

[0089] The potential risks of power equipment can be evaluated based on the basic graph vertex representation vectors. For power equipment with high risks, early warnings can be issued in a timely manner to remind relevant personnel to conduct inspections and maintenance to prevent safety accidents such as fires. The risk score of power equipment can be calculated according to certain features in the basic graph vertex representation vectors, such as the frequency and severity of abnormal occurrences. Let the basic graph vertex representation vector be Y and the risk assessment function be g(Y), then the risk score R of the power equipment is R = g(Y).

[0090] In the above way, the transformation matrix can be determined based on the preset unified projection standard dimension and the actual vector dimension of the graph vertex representation vectors, and the graph vertex representation vectors can be projected into a unified representation domain to obtain the basic graph vertex representation vectors.

[0091] In one implementation, step 430 is to integrate the basic graph vertex representation vectors to obtain the power equipment representation vector of the power equipment to be analyzed, including:

[0092] Step 431: Determine the graph vertex corresponding to the power equipment to be analyzed in the graph vertices to obtain the target graph vertex, and sample one or more connected graph vertices of the target graph vertex in the power analysis graph spectrum;

[0093] Step 432: Determine the target graph vertex representation vector corresponding to the target graph vertex and the connected graph vertex representation vectors corresponding to the connected graph vertices in the basic graph vertex representation vectors;

[0094] Step 433: Integrate the target graph vertex representation vector and the connected graph vertex representation vectors to obtain the power equipment representation vector of the power equipment to be analyzed.

[0095] In step 431, determine the graph vertex corresponding to the power equipment to be analyzed in the graph vertices to obtain the target graph vertex, and sample one or more connected graph vertices of the target graph vertex in the power analysis graph spectrum. In the power analysis graph spectrum, each graph vertex represents a power equipment or an abnormal operation event, and the graph vertex corresponding to the power equipment to be analyzed is the target graph vertex. Taking the power analysis graph spectrum of the power system in a large industrial park as an example, if it is necessary to analyze the operation status of a certain transformer, then the graph vertex representing this transformer is the target graph vertex. The connected graph vertices are the graph vertices that have a connection relationship with the target graph vertex in the graph spectrum, and they may represent abnormal operation events related to this power equipment or other related power equipment. For example, if there is an over-temperature abnormality in this transformer, then the graph vertex representing the over-temperature abnormality is the connected graph vertex of the target graph vertex.

[0096] To determine the target graph vertex, it is necessary to search in the power analysis graph spectrum according to the identification information of the power equipment to be analyzed. If the power analysis graph spectrum is stored in a graph database, the corresponding graph vertex can be located through a query statement according to the identification information such as the equipment number and name. After determining the target graph vertex, it is necessary to sample one or more of its connected graph vertices. The purpose of sampling is to reduce the calculation amount and improve the processing efficiency while ensuring the integrity of information. Sampling methods such as random sampling and stratified sampling can be used. Random sampling means randomly selecting a certain number of graph vertices from all the connected graph vertices of the target graph vertex; stratified sampling is to stratify according to certain characteristics of the connected graph vertices and then select a certain number of graph vertices from each layer.

[0097] After obtaining the target graph vertices and connection graph vertices, in step 432, the target graph vertex representation vector corresponding to the target graph vertex and the connection graph vertex representation vector corresponding to the connection graph vertex are determined in the basic graph vertex representation vectors. The basic graph vertex representation vectors are vectors obtained by projecting the graph vertex representation vectors into a unified representation domain, and they have the same dimension, which is convenient for comparison and analysis. The target graph vertex representation vector and the connection graph vertex representation vector respectively represent the characteristic information of the target graph vertex and the connection graph vertex. For example, for the transformer represented by the target graph vertex, its target graph vertex representation vector may include characteristic information such as the capacity, voltage level, and operating years of the transformer; for the over-temperature anomaly represented by the connection graph vertex, its connection graph vertex representation vector may include characteristic information such as the frequency and duration of the anomaly occurrence.

[0098] The corresponding vector can be found in the basic graph vertex representation vector set according to the identification information of the graph vertex. If the basic graph vertex representation vectors are stored in the form of an array or a list, and each element corresponds to the representation vector of a graph vertex, the corresponding element can be found through the index or identification information of the graph vertex.

[0099] After determining the target graph vertex representation vector and the connection graph vertex representation vector, in step 433, these vectors are integrated to obtain the power equipment representation vector of the power equipment to be analyzed. The operating state of the power equipment is not only related to its own characteristics, but also related to the abnormal operating events around it and the states of other related power equipment. Therefore, integrating the target graph vertex representation vector and the connection graph vertex representation vector can more comprehensively reflect the operating state of the power equipment. For example, the method of weighted summation can be used for vector integration, or the attention mechanism can also be used for vector integration. Through the attention mechanism, different attention weights are automatically assigned to different connection graph vertices, so that the target graph vertex can pay more attention to the connection graph vertices with stronger relevance to itself. Let the target graph vertex be , and its connection graph vertex be , and the attention weight can be calculated by the following formula: , where represents the similarity between the target graph vertex representation vector and the connection graph vertex representation vector , such as cosine similarity. Then, the integrated power equipment representation vector V can be calculated by the formula .

[0100] The power equipment characterization vector can comprehensively reflect the operating status and characteristics of power equipment. By analyzing the power equipment characterization vector, the similarities and differences between power equipment can be discovered, providing a basis for subsequent grouping of power equipment. If the characterization vectors of two power equipment have a high similarity, it indicates that they are similar in terms of operating status and abnormal triggering conditions, etc., and may belong to the same equipment group. Methods such as cosine similarity and Euclidean distance can be used to calculate the similarity between power equipment characterization vectors.

[0101] Through the above method, the target graph vertex and the connected graph vertex can be determined in the graph vertices, the corresponding characterization vectors can be obtained, and they can be integrated to finally obtain the power equipment characterization vector of the power equipment to be analyzed.

[0102] In one implementation, in step 431, sampling one or more connected graph vertices of the target graph vertex in the power analysis graph includes:

[0103] Step 4311: Determine the number of graph vertices of the peripheral graph vertices adjacent to the target graph vertex in the power analysis graph;

[0104] Step 4312: Determine the sampling strategy for the connected graph vertices of the target graph vertex according to the number of graph vertices. For example, determining the sampling strategy for the connected graph vertices of the target graph vertex according to the number of graph vertices includes: comparing the number of graph vertices with a reference value. If the number of graph vertices is greater than the reference value, then determine the sampling strategy as a non-repetitive sampling strategy; if the number of graph vertices is less than or equal to the reference value, then determine the sampling strategy as a reusable sampling strategy; if the sampling strategy is a non-repetitive sampling strategy, directly sample the connected graph vertices corresponding to the reference value in the peripheral graph vertices. If the sampling strategy is a reusable sampling strategy, sample the connected graph vertices corresponding to the reference value in the peripheral graph vertices using the reusable sampling strategy.

[0105] Step 4313: Sample one or more connected graph vertices in the peripheral graph vertices according to the sampling strategy.

[0106] In step 4311, determining the number of graph vertices of the peripheral graph vertices adjacent to the target graph vertex in the power analysis graph. In the power analysis graph, the target graph vertex represents the power equipment to be analyzed, and the peripheral graph vertices are the graph vertices that have a direct connection relationship with the target graph vertex. These peripheral graph vertices may represent abnormal operation events related to the power equipment or other related power equipment. To determine the number of graph vertices of the peripheral graph vertices, it can be achieved by querying the graph structure information of the power analysis graph. If the power analysis graph is stored in a graph database, the query language provided by the graph database can be used to find the graph vertices directly connected to the target graph vertex according to the identification information of the target graph vertex and count the number of these graph vertices.

[0107] After determining the number of graph vertices of the peripheral graph vertices, in step 4312, according to the number of graph vertices, determine the sampling strategy for the connecting graph vertices of the target graph vertex. The sampling strategy refers to the method used when selecting connecting graph vertices from the peripheral graph vertices. Different numbers of graph vertices may require different sampling strategies to ensure the effectiveness and representativeness of the sampling results. Usually, the number of graph vertices is compared with a reference value, which is a pre-set threshold used to determine which sampling strategy to adopt.

[0108] If the number of graph vertices is greater than the reference value, it indicates that the number of peripheral graph vertices is relatively large. At this time, the sampling strategy can be determined as a non-repetitive sampling strategy. The non-repetitive sampling strategy means that during the sampling process, each peripheral graph vertex can only be selected once, which can avoid repeatedly selecting the same graph vertex and ensure the randomness and diversity of the sampling. For example, if the reference value is set to 10 and the number of peripheral graph vertices of the target graph vertex is 15, a non-repetitive sampling strategy can be adopted to randomly select a certain number of graph vertices from these 15 peripheral graph vertices as connecting graph vertices.

[0109] If the number of graph vertices is less than or equal to the reference value, it indicates that the number of peripheral graph vertices is relatively small. At this time, the sampling strategy can be determined as a reusable sampling strategy. The reusable sampling strategy means that during the sampling process, the same peripheral graph vertex can be selected multiple times, which can ensure that enough connecting graph vertices can be extracted to meet the needs of subsequent analysis. For example, if the reference value is set to 10 and the number of peripheral graph vertices of the target graph vertex is 5, a reusable sampling strategy can be adopted to extract 10 graph vertices from these 5 peripheral graph vertices as connecting graph vertices, and some graph vertices may be selected repeatedly.

[0110] After determining the sampling strategy, in step 4313, according to the sampling strategy, sample one or more connecting graph vertices from the peripheral graph vertices. If the sampling strategy is a non-repetitive sampling strategy, a random number generator can be used to assign a random number to each peripheral graph vertex, and then sort the peripheral graph vertices according to the size of the random number, and select a certain number of graph vertices with the top ranking as connecting graph vertices. Let the set of peripheral graph vertices be , generate a random number r i for each graph vertex v i , sort the graph vertices in ascending order according to the random number. If k connecting graph vertices need to be sampled, then select the first k graph vertices after sorting.

[0111] If the sampling strategy is a reusable sampling strategy, random numbers can be continuously generated, and graph vertices are selected from the set of peripheral graph vertices according to the indexes corresponding to the random numbers until enough connecting graph vertices are sampled. Let the set of peripheral graph vertices be , each time a random integer \(j\) in the range \([1, n]\) is generated, and the graph vertex \(v\) is selected j As the connected graph vertices, repeat this process \(k\) times to obtain \(k\) connected graph vertices.

[0112] The representation vectors of the sampled connected graph vertices contain information about abnormal operation events related to the target graph vertex or other relevant power equipment. Integrating them with the representation vector of the target graph vertex can more comprehensively reflect the operating state of the power equipment represented by the target graph vertex. After obtaining the connected graph vertices, vector integration is performed based on the representation vectors of these connected graph vertices and the representation vector of the target graph vertex to generate the power equipment representation vector. The power equipment representation vector can comprehensively reflect the operating state and characteristics of the power equipment. By analyzing the power equipment representation vector, the similarities and differences between power equipment can be found, providing a basis for subsequent grouping of power equipment. If the representation vectors of two power equipment are highly similar, it indicates that they are similar in terms of operating state and abnormal triggering conditions, etc., and may belong to the same equipment group. Methods such as cosine similarity and Euclidean distance can be used to calculate the similarity between power equipment representation vectors.

[0113] In addition, the power equipment representation vector can also be used for fault prediction and risk assessment. By establishing a fault prediction model, using the power equipment representation vector as the input, the probability and type of power equipment failure can be predicted. Machine learning algorithms such as support vector machines and random forests can be used to train a large amount of historical data to obtain an accurate fault prediction model. Let the power equipment representation vector be \(V\) and the fault prediction model be \(f(V)\), then the predicted fault probability \(P = f(V)\). The potential risk of power equipment can also be evaluated based on the power equipment representation vector, providing a reference for fire detection. For power equipment with a high risk, early warnings can be issued in a timely manner to remind relevant personnel to conduct inspections and maintenance to prevent the occurrence of safety accidents such as fires. The risk score of power equipment can be calculated based on certain characteristics in the power equipment representation vector, such as the frequency and severity of anomalies. Let the power equipment representation vector be \(V\) and the risk assessment function be \(g(V)\), then the risk score \(R\) of the power equipment is \(R = g(V)\).

[0114] In one implementation, in step 433, the target graph vertex representation vector and the connected graph vertex representation vector are integrated to obtain the power equipment representation vector of the power equipment to be analyzed, including:

[0115] Step 4331: Determine the graph vertex hop count between the obtained connected graph vertex and the target graph vertex in the power analysis graph;

[0116] Step 4332: Group the connected graph vertices by the graph vertex hop count to obtain the graph vertex group corresponding to each connected graph vertex, and determine the target connected graph vertex among the connected graph vertices according to the graph vertex group.

[0117] Step 4333: Integrate the connected graph vertex representation vector corresponding to the target connected graph vertex with the target graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed.

[0118] In Step 4331, the graph vertex hop count between the connected graph vertex and the target graph vertex is determined in the power analysis graph. The graph vertex hop count refers to the minimum number of edges passed from one graph vertex to another in the power analysis graph, which reflects the distance and the degree of association between the two graph vertices. Taking a power analysis graph as an example, if the target graph vertex represents the transformer of a certain substation and the connected graph vertex represents the over-temperature abnormal event related to the transformer, if only one edge is passed from the transformer graph vertex to the over-temperature abnormal event graph vertex, then the graph vertex hop count is 1; if two edges need to be passed to reach, then the graph vertex hop count is 2.

[0119] To determine the graph vertex hop count between the connected graph vertex and the target graph vertex, the breadth-first search (BFS) algorithm can be used. The breadth-first search algorithm is an algorithm for traversing or searching a tree or a graph. It starts from the starting vertex and visits the vertices in the graph layer by layer. Taking the target graph vertex as the starting vertex, use the breadth-first search algorithm to traverse the power analysis graph, and record the layer number when each connected graph vertex is visited. This layer number is the graph vertex hop count between the connected graph vertex and the target graph vertex.

[0120] After determining the graph vertex hop count, in Step 4332, group the connected graph vertices by the graph vertex hop count to obtain the graph vertex group corresponding to each connected graph vertex, and determine the target connected graph vertex among the connected graph vertices according to the graph vertex group. The hop grouping means dividing the connected graph vertices into different groups according to the graph vertex hop count, similar to dividing into levels one by one. The target graph vertex is the first group (the first layer), and those with a hop count of 2 are in the second group (the second layer), and so on. Through the hop grouping, the distance relationship between the connected graph vertices and the target graph vertex can be understood more clearly, providing a basis for subsequent vector integration.

[0121] A hash table or an array can be used to store the graph vertex group information of each connected graph vertex. For each connected graph vertex, according to its graph vertex hop count, it is assigned to the corresponding group. Let the set of connected graph vertices be and the set of graph vertex hop counts be Create an array G, where the index of the array represents the grouping number, and each element is a set used to store the connected graph vertices in that group. For each connected graph vertex vci , add it to G[h i ]middle.

[0122] After completing the jump grouping, the graph vertices are grouped according to the graph vertices, and the target graph vertices are determined from the graph vertices. Target graph vertices are those that play an important role in integrating the target graph vertex representation vectors. Target graph vertices can be determined based on factors such as the grouping level and the importance of the graph vertices. For groups with lower levels (fewer hops), the graph vertices in those groups are more closely associated with the target graph vertices, and the graph vertices in these groups can be prioritized as target graph vertices. Graph vertices can also be sorted based on the importance of their features, such as the frequency and severity of anomalies, with the top-ranked graph vertices selected as target graph vertices.

[0123] After determining the target connection graph vertex, in step 4333, the connection graph vertex representation vector corresponding to the target connection graph vertex is integrated with the target graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed. The power equipment representation vector can comprehensively reflect the operating status and characteristics of the power equipment. Integrating the information of the target connection graph vertex into the target graph vertex representation vector allows the power equipment representation vector to more comprehensively reflect the actual situation of the power equipment. Vector integration can be performed using a weighted summation method or an attention mechanism.

[0124] In one implementation, step 4333 integrates the connection graph vertex representation vector corresponding to the target connection graph vertex with the target graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed, including:

[0125] Step 43331: Integrate the connection graph vertex representation vector corresponding to the target connection graph vertex with the target graph vertex representation vector to obtain the integrated graph vertex representation vector corresponding to the current graph vertex group;

[0126] Step 43332: If the level of the current graph vertex grouping is at the maximum number of layers of the preset graph vertex grouping, the integrated graph vertex representation vector is used as the target graph vertex representation vector, and the process jumps to the step of determining the target connected graph vertex from the connected graph vertices based on the graph vertex grouping until the current graph vertex grouping reaches the maximum number of layers of the preset graph vertex grouping, thereby obtaining the target integrated graph vertex representation vector of the power equipment to be analyzed;

[0127] Step 43333: perform a normalization operation on the target integrated graph vertex representation vector to obtain a compressed graph vertex representation vector, and use the compressed graph vertex representation vector as the power equipment representation vector of the power equipment to be analyzed.

[0128] In step 43331, the connection graph vertex representation vectors corresponding to the target connection graph vertices are integrated with the target graph vertex representation vectors to obtain the integrated graph vertex representation vectors corresponding to the current graph vertex group. In the power analysis graph spectrum, the target graph vertex represents the power equipment to be analyzed, and the target connection graph vertex is an important connection graph vertex that is closely associated with the target graph vertex and determined after grouping according to the graph vertex hop count. Each target connection graph vertex has its corresponding connection graph vertex representation vector, and these vectors contain the characteristic information of the abnormal operation event or related power equipment represented by the connection graph vertex.

[0129] Taking the power analysis graph spectrum of the power system of a large factory as an example, if the target graph vertex represents a key transformer, the target connection graph vertices may represent events such as over-temperature abnormality and over-current abnormality of the transformer. The target graph vertex representation vector may contain information such as the rated power and operating duration of the transformer, while the target connection graph vertex representation vectors respectively contain information such as the occurrence frequency of over-temperature abnormality and the current peak value of over-current abnormality. The method of weighted summation can be used for vector integration.

[0130] After obtaining the integrated graph vertex representation vectors corresponding to the current graph vertex group, in step 43332, it is judged whether the level of the current graph vertex group is at the maximum number of layers of the preset graph vertex group. The maximum number of layers of the preset graph vertex group is a threshold set in advance to control the depth of vector integration. If the level of the current graph vertex group is at the maximum number of layers of the preset graph vertex group, it means that the maximum depth of vector integration has been reached. At this time, the integrated graph vertex representation vector is used as the target graph vertex representation vector. At the same time, jump to execute the step of determining the target connection graph vertex among the connection graph vertices according to the graph vertex group until the current graph vertex group reaches the maximum number of layers of the preset graph vertex group, and the target integrated graph vertex representation vector of the power equipment to be analyzed is obtained.

[0131] Continuing with the example of the factory transformer, assume that the maximum number of layers of the preset graph vertex group is 3. When the vector integration of the first layer (the layer where the target graph vertex is located) and the target connection graph vertices of the second layer is completed, the integrated graph vertex representation vectors corresponding to the current graph vertex group (the second layer) are obtained. At this time, check whether the second layer is the maximum number of layers of the preset graph vertex group. If not, continue to determine the target connection graph vertices of the third layer and integrate their connection graph vertex representation vectors with the current integrated graph vertex representation vectors to obtain new integrated graph vertex representation vectors. Repeat this process until the maximum number of layers of the preset graph vertex group is reached to obtain the target integrated graph vertex representation vector.

[0132] In step 43333, perform a standardization operation on the target integrated graph vertex representation vector to obtain a compressed graph vertex representation vector, and use it as the power equipment representation vector of the power equipment to be analyzed. The standardization operation is to eliminate the dimensional differences of different features in the vector, make the various features in the vector comparable, and also help the subsequent machine learning algorithms to run more stably. The Z-score standardization method can be adopted. Let the target integrated graph vertex representation vector be where the j-th feature is The mean of this feature is and the standard deviation is Then the standardized j-th feature can be calculated by the formula The standardized vector is the compressed graph vertex representation vector.

[0133] In the above way, the connection graph vertex representation vector corresponding to the target connection graph vertex and the target graph vertex representation vector can be integrated. After multiple iterations until the preset maximum number of graph vertex grouping layers is reached, the target integrated graph vertex representation vector is obtained, and then a standardization operation is performed on it to finally obtain the power equipment representation vector of the power equipment to be analyzed.

[0134] In one implementation, in step 500, group the power equipment to be analyzed according to the power equipment representation vector to obtain one or more pending power equipment groups, including:

[0135] Step 510: Determine the power equipment commonality metric values between the power equipment to be analyzed in the set of power equipment to be analyzed according to the power equipment representation vector;

[0136] Step 520: Generate an interconnected graph spectrum with the power equipment to be analyzed as graph vertices according to the power equipment commonality metric values;

[0137] Step 530: Group the graph vertices in the interconnected graph spectrum to obtain one or more pending power equipment groups.

[0138] In step 510, according to the power equipment representation vector, determine the power equipment commonality metric values between the power equipment to be analyzed in the set of power equipment to be analyzed. The power equipment representation vector is obtained by mining the representation information of the power analysis graph spectrum, which comprehensively reflects the operating state and characteristics of the power equipment. The power equipment commonality metric value is used to measure the similarity degree between different power equipment. The higher the similarity degree, the closer these power equipment are in terms of operating state, abnormal trigger situation, etc.

[0139] Taking a power system as an example, the set of power equipment to be analyzed may include multiple transformers, distribution cabinets, motors, etc. Each power equipment has its corresponding power equipment characterization vector, and these vectors contain various characteristic information of the equipment, such as power, voltage, operation duration, abnormal occurrence frequency, etc. Methods such as cosine similarity algorithm, Euclidean distance, and Manhattan distance can be used to calculate the commonality metric value of power equipment.

[0140] After determining the commonality metric value of power equipment, in step 520, according to the commonality metric value of power equipment, the power equipment to be analyzed is used as graph vertices to generate an interconnected graph. The interconnected graph is an undirected graph, where each graph vertex represents a power equipment, and the edges between vertices represent the commonality metric value between power equipment. The weight of the edge can be set as the commonality metric value of power equipment. The larger the weight, the higher the similarity between the two power equipment. A graph database can be used to store and manage the interconnected graph, and the graph database can efficiently process graph-structured data and support fast graph query and analysis operations.

[0141] In step 530, the graph vertices in the interconnected graph are grouped to obtain one or more pending power equipment groups. The purpose of grouping is to divide similar power equipment into the same group for more targeted analysis and management. Community discovery algorithms, such as the Louvain algorithm, can be used to group the interconnected graph. By continuously moving graph vertices to different communities, the grouping effectiveness of the entire graph, that is, modularity, is maximized. For the formula, please refer to the previous introduction and will not be elaborated here. When using the Louvain algorithm for grouping, first, each graph vertex is assigned to a separate community, and then the current grouping effectiveness is calculated. Then, each graph vertex is traversed, and it is attempted to move it to an adjacent community and calculate the grouping effectiveness after the move. If the grouping effectiveness after the move increases, the vertex is moved to the new community. This process is repeated continuously until the grouping effectiveness no longer increases. After completing a round of vertex movement, each community is regarded as a new vertex, and the interconnected graph is reconstructed, and the above process is repeated until the grouping effectiveness reaches the maximum value. Through the Louvain algorithm, the graph vertices in the interconnected graph can be divided into different communities, and each community corresponds to a pending power equipment group. The power equipment in these pending power equipment groups has high similarity, and they may have common characteristics in terms of operating status, abnormal trigger conditions, etc. For example, in the power system of a commercial building, all transformers may be divided into one pending power equipment group, and all motors may be divided into another pending power equipment group.

[0142] For each group of power equipment to be determined, unified monitoring and analysis can be carried out to predict the failure probability and risk level of the power equipment within the group. Targeted maintenance plans and emergency response plans can be formulated based on the common characteristics of the power equipment within the group. If the power equipment in a certain group of power equipment to be determined often shows over-temperature anomalies, the temperature monitoring of the equipment in this group can be strengthened, and heat dissipation equipment and maintenance tools can be prepared in advance.

[0143] The group of power equipment to be determined can be dynamically updated. As new historical anomaly monitoring data continues to be generated, the operating status of the power equipment will change, and the similarity relationship between the power equipment will also change accordingly. It is necessary to regularly recalculate the commonality metric values of the power equipment, update the interconnection graph, and perform grouping operations again to ensure that the group of power equipment to be determined can accurately reflect the latest status of the power equipment. The group of power equipment to be determined can also be used for knowledge discovery and decision support. By analyzing the common characteristics and anomaly patterns of the power equipment in the group of power equipment to be determined, potential fault hazards and operating rules can be discovered. Based on these discoveries, decision-making basis can be provided for the optimization and upgrade of the power system, such as reasonably adjusting the operating parameters of the equipment, planning the replacement and expansion of the equipment, etc. The information of the group of power equipment to be determined can be combined with fire detection. For some groups of power equipment to be determined with high risks, the fire monitoring of the power equipment in this group can be strengthened, and more fire-fighting equipment such as temperature sensors and smoke alarms can be installed to detect and handle potential fire hazards in a timely manner. Different levels of fire warning strategies can be formulated according to the failure probability and risk level of the power equipment within the group to improve the pertinence and effectiveness of fire detection.

[0144] Through the above methods, the commonality metric value of the power equipment can be determined based on the characterization vector of the power equipment, the interconnection graph can be generated according to the commonality metric value, and the graph vertices in the interconnection graph can be grouped to obtain one or more groups of power equipment to be determined.

[0145] In one implementation, step 530, grouping the graph vertices in the interconnection graph to obtain one or more groups of power equipment to be determined, includes:

[0146] Step 531: Aggregate each graph vertex in the interconnection graph into the corresponding tight graph vertex group, and determine the basic grouping effectiveness of the tight graph vertex group;

[0147] Step 532: Determine the adjacent graph vertices corresponding to each graph vertex in the interconnection graph, and aggregate the graph vertices into the target tight graph vertex group corresponding to the adjacent graph vertices;

[0148] Step 533: Determine the current grouping validity of the target tight graph vertex group, and based on the basic grouping validity and the current grouping validity, determine one or more power equipment to be analyzed corresponding to the target tight graph vertex group in the set of power equipment to be analyzed, so as to obtain a group of power equipment to be determined.

[0149] In step 531, each graph vertex in the interconnected graph spectrum is grouped into the corresponding tight graph vertex group, and the basic grouping validity of the tight graph vertex group is determined. The tight graph vertex group can be understood as a Node Community, which is a set of vertices in the interconnected graph spectrum. The connections between these vertices are relatively tight, representing power equipment with similar characteristics or relationships.

[0150] In order to group the graph vertices into the corresponding tight graph vertex groups, a clustering algorithm such as spectral clustering algorithm can be used. The spectral clustering algorithm is based on the spectral theory of graphs. By performing eigen-decomposition on the Laplacian matrix of the graph, the vertices are mapped into a low-dimensional space, and then clustering is performed in the low-dimensional space. First, calculate the Laplacian matrix L = D - A of the interconnected graph spectrum, where D is the degree matrix and A is the adjacency matrix. Then, perform eigen-decomposition on the Laplacian matrix to obtain the eigenvectors. Select the eigenvectors corresponding to the first k smallest non-zero eigenvalues to form a matrix U, and use the row vectors of each vertex in U as its representation in the low-dimensional space. Finally, use the K-means clustering algorithm to cluster these low-dimensional vectors, and divide the vertices into different tight graph vertex groups. To determine the basic grouping validity of the tight graph vertex group, the concept of grouping validity (Modularity) is required. Grouping validity is an index to measure the quality of community division in the graph spectrum, and its calculation formula is referred to the introduction in the previous step 500. The closer the value of the grouping validity is to 1, the more reasonable the grouping is, the closer the connections within the tight graph vertex group are, and the sparser the connections between the groups are. According to the current grouping situation, substitute the corresponding values into the grouping validity formula for calculation to obtain the basic grouping validity of the tight graph vertex group.

[0151] After completing the grouping of the graph vertices and the determination of the basic grouping validity, in step 532, determine the adjacent graph vertices corresponding to each graph vertex in the interconnected graph spectrum, and group the graph vertices into the target tight graph vertex group corresponding to the adjacent graph vertices. The adjacent graph vertices refer to other graph vertices directly connected to the current graph vertex. In the interconnected graph spectrum of the power supply network, if a graph vertex represents a transformer, then the graph vertices corresponding to other power equipment directly connected to the transformer in the circuit are its adjacent graph vertices.

[0152] The adjacent graph vertices corresponding to each graph vertex can be determined by querying the adjacency matrix of the interconnected graph spectrum. For each graph vertex i, check the i-th row of the adjacency matrix A. If , it indicates that vertex j is an adjacent graph vertex of vertex i. Then, attempt to group graph vertex i into the target tight graph vertex group corresponding to its adjacent graph vertex j. During the grouping process, it is necessary to recalculate the grouping effectiveness and evaluate whether the overall grouping effectiveness is improved after moving vertex i to the new group.

[0153] In step 533, determine the current grouping effectiveness of the target tight graph vertex group, and based on the basic grouping effectiveness and the current grouping effectiveness, determine one or more power equipment to be analyzed corresponding to the target tight graph vertex group in the set of power equipment to be analyzed, and obtain the power equipment group to be determined. After moving the graph vertex to the target tight graph vertex group corresponding to the adjacent graph vertex, use the grouping effectiveness formula again to calculate the current grouping effectiveness of the target tight graph vertex group.

[0154] By comparing the basic grouping effectiveness and the current grouping effectiveness, determine whether the grouping is optimized. If the current grouping effectiveness is greater than the basic grouping effectiveness, it indicates that after moving the graph vertex to the new group, the grouping is more reasonable, and this move will be retained; if the current grouping effectiveness is less than the basic grouping effectiveness, it indicates that this move reduces the quality of the grouping, and the graph vertex will be restored to the original group.

[0155] In the process of continuously moving the graph vertex to the target tight graph vertex group corresponding to the adjacent graph vertex and comparing the grouping effectiveness, multiple iterations will be performed until the grouping effectiveness no longer improves. In this process, the composition of the target tight graph vertex group will be continuously updated, making the connections between the graph vertices within each group tighter and the connections between groups sparser.

[0156] After the grouping effectiveness reaches a stable state, determine one or more power equipment to be analyzed corresponding to the target tight graph vertex group in the set of power equipment to be analyzed, and obtain the power equipment group to be determined. The power equipment represented by the graph vertices within each target tight graph vertex group has high similarity, and they may have common characteristics in terms of operating status, abnormal trigger conditions, etc. In the power supply network, a target tight graph vertex group may include power equipment of the same type, in the same area, or with similar fault patterns.

[0157] Fault prediction and risk assessment can be carried out using a group of power equipment to be determined. By analyzing the common characteristics and historical fault data of the power equipment in the group, a fault prediction model can be established to predict the probability and type of faults occurring in the power equipment in the group. For a group of power equipment to be determined that often experiences overcurrent anomalies, it can be predicted that other power equipment in the group may also experience overcurrent faults, and early warnings can be issued in a timely manner. The information of the group of power equipment to be determined can also be combined with fire detection. For some high-risk groups of power equipment to be determined, the fire monitoring of the power equipment in the group can be strengthened, and more fire-fighting equipment such as temperature sensors and smoke alarms can be installed to detect and handle potential fire hazards in a timely manner. Different levels of fire warning strategies can be formulated according to the fault probability and risk level of the power equipment in the group to improve the pertinence and effectiveness of fire detection.

[0158] In one implementation, step 533, based on the basic grouping effectiveness and the current grouping effectiveness, determine one or more power equipment to be analyzed corresponding to the target tight graph vertex group in the set of power equipment to be analyzed, and obtain a group of power equipment to be determined, including:

[0159] Step 5331: Determine the error between the basic grouping effectiveness and the current grouping effectiveness of the target tight graph vertex group to obtain the grouping effectiveness error;

[0160] Step 5332: Iterate the graph vertices in the target tight graph vertex group according to the grouping effectiveness error to obtain the iterated tight graph vertex group, and verify the group strength of the iterated tight graph vertex group;

[0161] Step 5333: If the iterated tight graph vertex group meets the strength requirements, use the iterated tight graph vertex group as the graph vertices to generate an interconnected graph, obtain the current interconnected graph, and based on the current interconnected graph, determine one or more power equipment to be analyzed corresponding to the iterated tight graph vertex group in the set of power equipment to be analyzed, and obtain a group of power equipment to be determined. For example, if all the graph vertices are collected into the target tight graph vertex group and the grouping effectiveness reaches the maximum value at the same time, it is determined that the iterated tight graph vertex group meets the strength requirements; otherwise, jump to the step of collecting the graph vertices into the target tight graph vertex group corresponding to the adjacent graph vertices until the iterated tight graph vertex group meets the strength requirements; when the iterated tight graph vertex group meets the strength requirements, the method of using the iterated tight graph vertex group as the graph vertices to generate an interconnected graph is, for example, to merge all the graph vertices in the iterated tight graph vertex group into one graph vertex, that is, each iterated tight graph vertex group is regarded as one graph vertex, and then, the edges between the tight graph vertex groups remain unchanged to generate the current interconnected graph.

[0162] In step 5331, determine the error between the basic grouping validity and the current grouping validity of the target tight graph vertex group to obtain the grouping validity error. The basic grouping validity is the grouping validity value calculated using the grouping validity formula (refer to the foregoing content) after initially grouping the graph vertices into the tight graph vertex group, and the current grouping validity is the grouping validity value calculated again using the same grouping validity formula after moving the graph vertices to the corresponding target tight graph vertex group of the adjacent graph vertices.

[0163] Taking an interconnected graph of a power supply network as an example, during the initial grouping, group the graph vertices representing power equipment in different substations into different tight graph vertex groups, and calculate the basic grouping validity Q1. Then, attempt to move some graph vertices to the corresponding target tight graph vertex groups of their adjacent graph vertices, and calculate the grouping validity again to obtain the current grouping validity Q2. The grouping validity error can be calculated by the formula The grouping validity error reflects the degree of change in the grouping validity after the grouping adjustment. The larger the error, the greater the impact of the grouping adjustment on the overall grouping quality.

[0164] After obtaining the grouping validity error, in step 5332, based on the grouping validity error, iterate the graph vertices within the target tight graph vertex group to obtain the iterated tight graph vertex group and perform group strength verification on the iterated tight graph vertex group. Determine whether it is necessary to continue moving and adjusting the grouping of the graph vertices according to the grouping validity error. If the grouping validity error is large, it indicates that there is still room for optimizing the grouping, and the graph vertices will continue to be moved to the corresponding target tight graph vertex groups of the adjacent graph vertices, and the grouping validity will be recalculated until the grouping validity error is less than a preset threshold.

[0165] During each iteration, traverse each graph vertex within the target tight graph vertex group, attempt to move it to different groups, and compare the grouping validity before and after the move. If the grouping validity after the move increases, retain this move; otherwise, restore the graph vertex to the original group. Through continuous iteration, the grouping can be gradually optimized so that the connections between the graph vertices within each tight graph vertex group are closer and the connections between groups are sparser. After obtaining the iterated tight graph vertex group, perform group strength verification on it. The group strength verification is to ensure that the iterated tight graph vertex group has sufficient stability and rationality. Verify by checking indicators such as the average connection strength between the graph vertices within the group and the size of the group. If the average connection strength between the graph vertices within the group is low, it indicates that the relationship between the vertices within the group is not close enough, and this group may need further adjustment; if the size of the group is too large or too small, it may also not meet the actual grouping requirements.

[0166] In step 5333, it is determined whether the vertex group of the post-iteration dense graph meets the strength requirement. If it meets the strength requirement, the vertex group of the post-iteration dense graph is used as the graph vertices to generate an interconnected graph, obtaining the current interconnected graph. And based on the current interconnected graph, one or more power equipment to be analyzed corresponding to the vertex groups of the post-iteration dense graph are determined in the set of power equipment to be analyzed, obtaining the group of power equipment to be determined. Determining whether the vertex group of the post-iteration dense graph meets the strength requirement can be carried out according to multiple conditions. If all the graph vertices are grouped into the target vertex group of the dense graph and the grouping effectiveness reaches the maximum value, that is, the grouping effectiveness no longer increases, it can be determined that the vertex group of the post-iteration dense graph meets the strength requirement.

[0167] When the vertex group of the post-iteration dense graph meets the strength requirement, the vertex group of the post-iteration dense graph is used as the graph vertices to generate an interconnected graph. The specific method is to merge all the graph vertices in the vertex group of the post-iteration dense graph into one graph vertex, that is, each vertex group of the post-iteration dense graph is regarded as one graph vertex. Then, the edges between the vertex groups of the dense graph remain unchanged to generate the current interconnected graph. In the interconnected graph of the power supply network, if after iteration, several vertex groups of the dense graph representing partial power equipment in different substations meet the strength requirement, these groups will be merged into one graph vertex respectively, and the weights of the edges between these new graph vertices are determined according to the original connection situation between the groups.

[0168] Based on the current interconnected graph, one or more power equipment to be analyzed corresponding to the vertex groups of the post-iteration dense graph are determined in the set of power equipment to be analyzed, obtaining the group of power equipment to be determined. The corresponding power equipment can be found in the set of power equipment to be analyzed according to the information of the vertex groups of the post-iteration dense graph represented by each graph vertex in the current interconnected graph. If a graph vertex represents a vertex group of the post-iteration dense graph containing multiple transformers, these transformers will be found in the set of power equipment to be analyzed and taken as a group of power equipment to be determined.

[0169] These groups of power equipment to be determined are of great significance for the management and maintenance of power equipment. Personalized monitoring and maintenance strategies can be formulated for each group of power equipment to be determined. For a group of power equipment to be determined containing multiple old transformers, the monitoring of parameters such as the temperature and oil quality of the transformers in the group can be strengthened, and repair and replacement plans can be formulated in advance.

[0170] Fault prediction and risk assessment can also be performed using a group of to-be-determined power equipment. By analyzing the common characteristics and historical fault data of the power equipment within the group, a fault prediction model can be established to predict the probability and type of faults occurring in the power equipment within the group. For a group of to-be-determined power equipment that often experiences overcurrent anomalies, it can be predicted that other power equipment within the group may also experience overcurrent faults, and early warnings can be issued in a timely manner.

[0171] In one implementation, in step 5333, based on the current interconnection map, one or more power equipment to be analyzed corresponding to the tightly connected graph vertex groups after iteration are determined from the set of power equipment to be analyzed, and a group of to-be-determined power equipment is obtained, including:

[0172] Step 53331: Determine the current interconnection map as the interconnection map, and jump to the step of aggregating each graph vertex in the interconnection map into the corresponding tightly connected graph vertex group for execution until the current interconnection map stops changing, and obtain the target interconnection map;

[0173] Step 53332: Determine the group label of the current tightly connected graph vertex group corresponding to each graph vertex in the target interconnection map;

[0174] Step 53333: Based on the group label, one or more power equipment to be analyzed within the current tightly connected graph vertex group are determined from the set of power equipment to be analyzed, and a group of to-be-determined power equipment is obtained. For example, if the current interconnection map stops changing, it means that the tightly connected graph vertex groups in the current interconnection map have stabilized. At this time, the group label of the current tightly connected graph vertex group corresponding to each graph vertex can be determined in the obtained target interconnection map. Here, the group label can be the graph vertex label of the graph vertices included in the current tightly connected graph vertex group. Based on this graph vertex label, the power equipment to be analyzed included in the current tightly connected graph vertex group can be determined in the set of power equipment to be analyzed, and the power equipment to be analyzed within the current tightly connected graph vertex group can be used as a group of to-be-determined power equipment, so as to obtain the group of to-be-determined power equipment corresponding to each current tightly connected graph vertex group in the target interconnection map.

[0175] In step 53331, the current interconnection map is determined as the interconnection map, and the step of aggregating each graph vertex in the interconnection map into the corresponding tightly connected graph vertex group is jumped to for execution until the current interconnection map stops changing, and the target interconnection map is obtained. The current interconnection map is a new interconnection map generated by using the iteratively tightly connected graph vertex group as graph vertices when the iteratively tightly connected graph vertex group meets the strength requirement. In this new interconnection map, each graph vertex represents an iteratively tightly connected graph vertex group, and the edges between the vertices represent the connection relationships between the groups.

[0176] During the process of continuously repeating the grouping operation, the grouping validity is continuously calculated, and the change in the grouping validity after each grouping adjustment is compared. If, after multiple adjustments, the grouping validity no longer increases and the grouping situation of the graph vertices no longer changes, that is, the current interconnected graph stops changing, then the target interconnected graph is obtained at this time. The target interconnected graph is a graph in which the grouping reaches a stable state after multiple optimizations, and it can more accurately reflect the similarity relationships and group structures among power equipment.

[0177] After obtaining the target interconnected graph, in step 53332, the group label of the current dense graph vertex group corresponding to each graph vertex in the target interconnected graph is determined. The group label is information used to identify each dense graph vertex group, and it can be the graph vertex label of the graph vertices included in the current dense graph vertex group. In the target interconnected graph of the power system in the industrial park, a graph vertex represents a dense graph vertex group composed of multiple transformers, and the group label of this group can be the set of the numbers of these transformers. By traversing each graph vertex in the target interconnected graph, the graph vertex labels of the graph vertices within the dense graph vertex group it represents can be obtained, and these labels can be combined as the group label of this group.

[0178] In step 53333, based on the group label, one or more power equipment to be analyzed within the current dense graph vertex group are determined from the set of power equipment to be analyzed, and a group of power equipment to be determined is obtained. The set of power equipment to be analyzed is a series of power equipment initially obtained that needs to be analyzed.

[0179] Through the above method, based on the current interconnected graph, the target interconnected graph can be obtained through multiple iterations and optimizations, the group label can be determined, and the corresponding group of power equipment to be determined can be found in the set of power equipment to be analyzed.

[0180] Figure 2 The following is a schematic diagram of the hardware entity of a computer system provided by an embodiment of this application, as Figure 2 shown. The hardware entity of the computer system 1000 includes: a processor 1001 and a memory 1002. Among them, the memory 1002 stores a computer program that can run on the processor 1001, and when the processor 1001 executes the program, it implements the steps in the method of any of the above embodiments.

[0181] The memory 1002 stores a computer program that can run on a processor. The memory 1002 is configured to store instructions and applications executable by the processor 1001, and can also cache data to be processed or already processed by the processor 1001 and each module in the computer system 1000 (such as, image data, audio data, voice communication data, and video communication data), and can be implemented by flash memory (FLASH) or random access memory (RAM).

[0182] When the processor 1001 executes the program, it implements the steps of the fire detection method based on power data analysis in any one of the above. The processor 1001 generally controls the overall operation of the computer system 1000.

[0183] As described above, it is only the implementation mode of this application, but the protection scope of this application is not limited thereto. Any person skilled in the art in the technical field disclosed in this application can easily think of changes or substitutions within the technical scope disclosed in this application, and all should be covered by the protection scope of this application.

Claims

1. A fire detection method based on power data analysis, characterized in that, Applied to a computer system, the method includes: Obtain a set of power equipment to be analyzed and a historical abnormal monitoring data set of each piece of power equipment to be analyzed in the set of power equipment to be analyzed; Determine one or more abnormal operation events in the historical abnormal monitoring data set, where the abnormal operation event is an abnormal classification corresponding to the abnormal state of the power equipment to be analyzed; According to the historical abnormal monitoring data set, generate a power analysis graph with the power equipment to be analyzed and the abnormal operation events as graph vertices, where the power analysis graph is used to indicate the triggering relationship between the power equipment to be analyzed and the abnormal operation events; Mine characterization information from the power analysis graph to obtain a power equipment characterization vector of the power equipment to be analyzed; Based on the power equipment characterization vector, determine the power equipment commonality metric values between the power equipment to be analyzed in the set of power equipment to be analyzed; According to the power equipment commonality metric values, generate an interconnected graph with the power equipment to be analyzed as graph vertices; Group each graph vertex in the interconnected graph into a corresponding tight graph vertex group, and determine the basic grouping effectiveness of the tight graph vertex group; Determine the adjacent graph vertices corresponding to each graph vertex in the interconnected graph, and group the graph vertices into the target tight graph vertex group corresponding to the adjacent graph vertices; Determine the current grouping effectiveness of the target tight graph vertex group, and determine the error between the basic grouping effectiveness and the current grouping effectiveness of the target tight graph vertex group to obtain the grouping effectiveness error; Based on the grouping effectiveness error, iterate the graph vertices in the target tight graph vertex group to obtain an iterated tight graph vertex group, and verify the group strength of the iterated tight graph vertex group; If the iterated tight graph vertex group meets the strength requirement, then generate an interconnected graph with the iterated tight graph vertex group as graph vertices to obtain the current interconnected graph, and based on the current interconnected graph, determine one or more pieces of power equipment corresponding to the iterated tight graph vertex group in the set of power equipment to be analyzed to obtain a pending power equipment group, and determine a target power equipment group in the pending power equipment group so as to perform fire detection on the power equipment in the target power equipment group.

2. The fire detection method based on power data analysis according to claim 1, wherein The determining one or more abnormal operation events in the historical abnormal monitoring data set includes: Extract the abnormal category data of the power equipment to be analyzed from the historical abnormal monitoring data set; Extract the abnormal occurrence time and abnormal electrical node of the power equipment to be analyzed from the abnormal category data; Based on the abnormal occurrence time and abnormal electrical node, determine one or more abnormal operation events corresponding to the power equipment to be analyzed; The generating a power analysis graph with the power equipment to be analyzed and the abnormal operation events as graph vertices according to the historical abnormal monitoring data set includes: Determine the trigger statistical information of the power equipment to be analyzed and the abnormal operation event in the historical abnormal monitoring dataset; Determine the trigger relationship between the power equipment to be analyzed and the abnormal operation event according to the abnormal classification information of the abnormal operation event and the trigger statistical information; Generate a power analysis graph with the power equipment to be analyzed and the abnormal operation event as graph vertices according to the trigger relationship; 3. The fire detection method based on power data analysis according to claim 1 or 2, characterized in that Mining the characterization information of the power analysis graph to obtain the power equipment characterization vector of the power equipment to be analyzed, including: Extract the graph vertex characterization vector of each graph vertex in the power analysis graph; Project the graph vertex characterization vector into a unified characterization domain to obtain a basic graph vertex characterization vector; Integrate the basic graph vertex characterization vectors to obtain the power equipment characterization vector of the power equipment to be analyzed; 4. The fire detection method based on power data analysis according to claim 3, wherein, The step of projecting the graph vertex characterization vector into a unified characterization domain to obtain a basic graph vertex characterization vector includes: Obtain the preset unified projection standard dimension and the actual vector dimension of each graph vertex characterization vector; Determine the transformation matrix corresponding to the graph vertex characterization vector according to the preset unified projection standard dimension and the actual vector dimension; Project the graph vertex characterization vector into the graph vertex characterization vector corresponding to the preset unified projection standard dimension according to the transformation matrix to obtain a basic graph vertex characterization vector; The step of integrating the basic graph vertex characterization vectors to obtain the power equipment characterization vector of the power equipment to be analyzed includes: Determine the graph vertex corresponding to the power equipment to be analyzed in the graph vertices to obtain the target graph vertex, and sample one or more connected graph vertices of the target graph vertex in the power analysis graph; Determine the target graph vertex characterization vector corresponding to the target graph vertex and the connected graph vertex characterization vector corresponding to the connected graph vertex in the basic graph vertex characterization vectors; Integrate the target graph vertex characterization vector and the connected graph vertex characterization vector to obtain the power equipment characterization vector of the power equipment to be analyzed; 5. The fire detection method based on power data analysis according to claim 4, wherein The step of sampling one or more connected graph vertices of the target graph vertex in the power analysis graph includes: Determine the number of graph vertices of the surrounding graph vertices adjacent to the target graph vertex in the power analysis graph; Determine the sampling strategy for the connected graph vertices of the target graph vertex according to the number of graph vertices; Sample one or more connected graph vertices from the surrounding graph vertices according to the sampling strategy; Among them, the step of determining the sampling strategy for the connected graph vertices of the target graph vertex according to the number of graph vertices includes: Compare the number of graph vertices with a reference value; if the number of graph vertices is greater than the reference value, determine that the sampling strategy is a non-repetitive sampling strategy; if the number of graph vertices is less than or equal to the reference value, determine that the sampling strategy is a reusable sampling strategy; if the sampling strategy is a non-repetitive sampling strategy, directly perform non-repetitive sampling among the surrounding graph vertices to obtain the connected graph vertices corresponding to the reference value; if the sampling strategy is a reusable sampling strategy, perform repetitive sampling among the surrounding graph vertices to obtain the connected graph vertices corresponding to the reference value; The step of integrating the target graph vertex representation vector and the connection graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed includes: Determining the number of graph vertex hops between the connection graph vertex and the target graph vertex in the power analysis graph; According to the number of graph vertex hops, the connection graph vertices are hop-grouped to obtain a graph vertex group corresponding to each of the connection graph vertices, and according to the graph vertex group, a target connection graph vertex is determined among the connection graph vertices; The connection graph vertex representation vector corresponding to the target connection graph vertex is integrated with the target graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed.

6. The fire detection method based on power data analysis according to claim 5, wherein, The step of integrating the connection graph vertex representation vector corresponding to the target connection graph vertex with the target graph vertex representation vector to obtain the power equipment representation vector of the power equipment to be analyzed includes: Integrate the connection graph vertex representation vector corresponding to the target connection graph vertex with the target graph vertex representation vector to obtain an integrated graph vertex representation vector corresponding to the current graph vertex group; If the level of the current graph vertex grouping is at the maximum number of layers of the preset graph vertex grouping, the integrated graph vertex representation vector is used as the target graph vertex representation vector, and the step of determining the target connected graph vertex from the connected graph vertices based on the graph vertex grouping is skipped and executed until the current graph vertex grouping reaches the maximum number of layers of the preset graph vertex grouping, thereby obtaining the target integrated graph vertex representation vector of the power equipment to be analyzed; A normalization operation is performed on the target integrated graph vertex representation vector to obtain a compressed graph vertex representation vector, and the compressed graph vertex representation vector is used as the power equipment representation vector of the power equipment to be analyzed.

7. The fire detection method based on power data analysis according to claim 6, wherein The step of determining, based on the current interconnection graph, one or more power devices to be analyzed corresponding to vertex groups of the iterative dense graph from the set of power devices to be analyzed, and obtaining the pending power device group, includes: Determine the current interconnection graph as the interconnection graph, and jump to the step of grouping each graph vertex in the interconnection graph into a corresponding dense graph vertex group until the current interconnection graph stops changing, thereby obtaining a target interconnection graph; Determining in the target interconnected graph a group label of a current dense graph vertex group corresponding to each graph vertex; Based on the group tag, one or more power devices to be analyzed within the current tight graph vertex group are determined from the set of power devices to be analyzed, and the group of power devices to be determined is obtained.

8. A computer system, comprising a memory and a processor, the memory storing a computer program executable on the processor, characterized in that, When the processor executes the program, the steps in the method according to any one of claims 1 to 7 are implemented.

Images