Electric power Internet of Things equipment data analysis method and system based on artificial intelligence

Abstract

The invention provides an electric power Internet of Things equipment data analysis method and system based on artificial intelligence, and aims to realize accurate fault positioning through deep analysis of equipment operation event data. The method comprises the following steps: firstly, converting running event data in an equipment log into a graph convolution vector set by utilizing graph convolution processing; and then, by adopting a heuristic fault path prediction algorithm and combining a graph embedding vector corresponding to the predefined fault knowledge graph, analyzing the graph convolution vector set, and generating heuristic fault path data. In the process, feature derivation is carried out on a graph convolution vector set by calculating a feature distance between priori fault path prediction knowledge and the graph convolution vector set, and a strengthened vector set is formed. And finally, the priori knowledge is iteratively optimized by using the enhanced vector set and the target map embedded vector, so that a fault path is accurately predicted, the accuracy of power Internet of Things equipment fault positioning is remarkably improved, and stable operation of a power system is facilitated.

Description

Technical Field

[0001] The present application relates to the field of artificial intelligence technology, and more specifically, to an artificial intelligence-based power Internet of Things device data analysis method and system. Background Art

[0002] With the rapid development of the power Internet of Things, a large number of devices are connected to the network, forming a large and complex system. These devices generate a large amount of log data during operation, including information on various device operation events. How to effectively analyze and use this data to achieve fault prediction and location of power Internet of Things devices has become an important challenge facing the current industry.

[0003] Traditional data analysis methods mainly rely on manual analysis and simple statistical techniques, which make it difficult to deeply explore the hidden information and associations in log data. These methods are often inefficient and their accuracy cannot be guaranteed when processing large-scale, high-dimensional power IoT data. Especially in terms of fault path prediction, traditional methods cannot fully utilize the complex associations between historical data and devices, resulting in inaccurate prediction results and difficulty in timely and effective fault location. Summary of the invention

[0004] In view of the above-mentioned problems, an embodiment of the present application provides a method for analyzing data of power Internet of Things devices based on artificial intelligence, the method comprising: Perform graph convolution processing on the operation event data of each device included in the target power IoT device log to generate a corresponding graph convolution vector set; Perform heuristic fault path prediction on the graph convolution vector set to generate corresponding heuristic fault path data, and generate a fault location result of the target power Internet of Things device log based on the predicted heuristic fault path data; The step of performing heuristic fault path prediction on the graph convolution vector set to generate corresponding heuristic fault path data includes: In each round of fault path prediction, a target graph embedding vector corresponding to the heuristic fault path data generated by the previous round of prediction is generated according to the graph embedding vectors corresponding to each predefined fault knowledge graph; According to the feature distance between the prior fault path prediction knowledge generated in the previous round of fault path prediction and the graph convolution vector set, feature derivation is performed on the graph convolution vector set to generate a reinforcement vector set corresponding to the current round of fault path prediction, wherein the prior fault path prediction knowledge is used to represent: past fault path prediction information of the graph convolution vector set; The priori fault path prediction knowledge is iteratively optimized according to the set of reinforcement vectors and the target graph embedding vector, and fault path prediction is performed according to the iteratively optimized priori fault path prediction knowledge to generate heuristic fault path data for this round of fault path prediction.

[0005] For example, in a possible implementation of the first aspect, the step of generating a target graph embedding vector corresponding to the heuristic fault path data generated by the previous round of prediction according to the graph embedding vectors corresponding to each predefined fault knowledge graph includes: Loading predefined fault knowledge graphs from a preset knowledge base, each of the predefined fault knowledge graphs represents a specific fault mode path, and each of the predefined fault knowledge graphs is associated with a corresponding graph embedding vector, which is learned by graph embedding technology in a training phase; Parsing the heuristic fault path data generated by the previous round of prediction, identifying key nodes and edges in the heuristic fault path data, wherein the key nodes and edges represent key steps and transfers in the fault path; Matching the key nodes and edges with nodes and edges in a predefined fault knowledge graph, and retrieving a candidate graph embedding vector corresponding to the matched target fault knowledge graph from a predefined graph embedding vector library according to the matching result, wherein if there are multiple matching target fault knowledge graphs, selecting the best matching candidate graph embedding vector according to a similarity or confidence score; If the retrieved candidate graph embedding vector completely matches the heuristic fault path data, directly using the candidate graph embedding vector as the target graph embedding vector; If the retrieved candidate graph embedding vector partially matches or does not fully match the heuristic fault path data, the retrieved candidate graph embedding vector is adjusted using interpolation, extrapolation or a rule-based method to generate a target graph embedding vector that better matches the heuristic fault path data.

[0006] In a possible implementation of the first aspect, the method is implemented by completing a parameter-optimized power Internet of Things fault prediction model, and the method further includes: Parameter optimization is performed on a graph convolution parameter array included in a graph convolution network through multiple fault path description templates, and when a model convergence condition is met, a graph convolution parameter array that has completed parameter optimization is generated, each fault path description template is a description sequence composed of fault paths corresponding to multiple device state description units, and the graph convolution network is used to map the fault path description template to a corresponding graph convolution vector set using the graph convolution parameter array; Using the graph convolution parameter array that has completed parameter optimization, the graph convolution parameter array in the power Internet of Things fault prediction model is guided to generate a target graph convolution parameter array, wherein the target graph convolution parameter array includes graph embedding vectors corresponding to each of the predefined fault knowledge graphs; Optimizing the parameters of the guided power Internet of Things fault prediction model according to multiple fault prediction training data, and generating the power Internet of Things fault prediction model with the optimized parameters when the model convergence condition is met, each fault prediction training data includes: a template power Internet of Things device log and its corresponding heuristic fault path label data; In a possible implementation of the first aspect, performing parameter optimization on a graph convolution parameter array included in a graph convolution network by using multiple fault path description templates includes: Utilizing the multiple fault path description templates to perform cyclic parameter optimization on the graph convolution network until the graph convolution network meets the network convergence condition, and in each round of parameter optimization, performing feature shielding processing on each fault path description template loaded in this round to shield the device state description unit of the shielded node in each fault path description template; According to the graph convolution parameter array called in this round, the graph convolution vector distributions corresponding to each fault path description template after feature masking are generated respectively; According to the distribution of each generated graph convolution vector, the device state description unit corresponding to the shielded node in each fault path description template is predicted respectively; Optimizing the graph convolution parameter array according to the difference parameters between the device state description units actually shielded by the respective fault path description templates and the predicted device state description units; For example, in a possible implementation of the first aspect, the step of performing feature shielding processing on each fault path description template loaded in this round to shield the device state description unit of the shielded node in each fault path description template includes: Load the fault path description template to be processed in this round, parse the structure of the fault path description template, and identify the nodes in the fault path description template and their corresponding device status description units; Determine the nodes to be shielded according to preset rules, create a shielded node list, and record the node information of the nodes to be shielded, wherein the preset rules include rules based on the type, attribute, and location of the nodes; Traverse each node in the fault path description template and check whether the node is in the shielded node list; For a node in the shielded node list, performing a shielding operation, wherein the shielding operation includes deleting, replacing or hiding a device state description unit of the node; After the feature masking process is completed, the fault path description template is updated.

[0007] In a possible implementation of the first aspect, graph convolution processing is performed on each device operation event data included in the target power Internet of Things device log to generate a corresponding graph convolution vector set, including: Performing device state migration node parsing on the target electric power IoT device log to generate a corresponding device state migration node set, wherein each device state migration node in the device state migration node set corresponds to a device operation event data; Performing a serialized graph convolution process on the device state transition node set until a graph convolution vector corresponding to each device state transition node is generated. In each round of graph convolution process, if the device operation event data of this round is the first device operation event data, then generating a graph convolution vector of the device operation event data of this round according to the device state transition node of the device operation event data of this round and a defined initialization vector; If the device operation event data of this round is not the first device operation event data, then the graph convolution vector of the device operation event data of this round is generated according to the graph convolution vector extracted in the previous round and the device state transition node of the device operation event data of this round; Generating the graph convolution vector set according to each generated graph convolution vector; For example, in a possible implementation of the first aspect, the device state migration node is parsed for the target power Internet of Things device log to generate a corresponding device state migration node set, including: Extracting keyword features related to the device status from the target power IoT device log using a natural language processing algorithm; According to the extracted keyword features, the record units of each time period in the target power Internet of Things device log are classified into different device states, and a corresponding timestamp is marked for each identified device state; Arranging the identified device states into a state sequence according to the order of the marked timestamps, and detecting a transition point from one state to another state in the state sequence, wherein the transition point is the device state migration node; Determine the migration type of each device state migration node, and encapsulate the node information of each device state migration node into a data object according to the migration type of each device state migration node, combine all the encapsulated data objects into a device state migration node set, and associate a corresponding device operation event data for each device state migration node in the device state migration node set; In a possible implementation of the first aspect, the method further includes: In the generated device state migration node set, remove duplicate or invalid device state migration nodes; Alternatively, a time window is set, and only the device state migration nodes within the time window are retained in the device state migration node set.

[0008] On the other hand, an embodiment of the present application also provides an artificial intelligence-based power Internet of Things device data analysis system, including a processor and a machine-readable storage medium, the machine-readable storage medium is connected to the processor, the machine-readable storage medium is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the machine-readable storage medium to implement the above method.

[0009] Based on the above aspects, the embodiment of the present application can effectively extract the deep features of the equipment operation event data from the complex power Internet of Things equipment log through graph convolution processing, generate a graph convolution vector set, and provide an accurate data basis for subsequent fault path prediction. The introduction of the heuristic fault path prediction mechanism can intelligently predict possible fault paths based on the graph convolution vector set, greatly improving the accuracy and efficiency of fault location. By combining the graph embedding vector corresponding to the predefined fault knowledge graph and the prediction results of the previous round in each round of fault path prediction, the target graph embedding vector is generated, and the dynamic adjustment and optimization of the fault path prediction is realized. The feature distance between the prior fault path prediction knowledge and the graph convolution vector set is used to derive features and generate a reinforcement vector set, which further enhances the precision and sensitivity of the fault path prediction. Through the iterative optimization of the prior fault path prediction knowledge, not only the accuracy of the fault path prediction is improved, but also the prediction system has the ability of self-learning and improvement, so that it can continuously adapt to the complex changes in the power Internet of Things environment. As a result, the accuracy, efficiency and intelligence level of the fault location of the power Internet of Things equipment are significantly improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0010] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present application and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other corresponding drawings can be obtained based on these drawings without creative work.

[0011] Figure 1 It is a schematic diagram of the execution flow of the artificial intelligence-based power Internet of Things device data analysis method provided in an embodiment of the present application.

[0012] Figure 2 This is a schematic diagram of the hardware architecture of an artificial intelligence-based power Internet of Things device data analysis system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0013] The following description is intended to enable one of ordinary skill in the art to implement and utilize the present application, and the description is provided in the context of a specific application scenario and its requirements. It will be apparent to one of ordinary skill in the art that various changes may be made to the disclosed embodiments, and that the general principles defined in the present application may be applied to other embodiments and application scenarios without departing from the principles and scope of the present application. Therefore, the present application is not limited to the described embodiments, but should be given the broadest scope consistent with the claims.

[0014] The present application will be described in detail below with reference to the accompanying drawings. Figure 1 This is a flow chart of an artificial intelligence-based power Internet of Things device data analysis method provided by an embodiment of the present application. The artificial intelligence-based power Internet of Things device data analysis method is introduced in detail below.

[0015] Step S110, performing graph convolution processing on the operation event data of each device included in the target power Internet of Things device log to generate a corresponding graph convolution vector set.

[0016] In this embodiment, the power IoT device data analysis based on artificial intelligence acts as a server. After receiving the power IoT device logs, it first pre-processes these power IoT device logs and extracts the device operation event data. Next, the server uses a graph convolutional network to process these device operation event data. Among them, the power IoT device log records the operating status and events of each power IoT device in the power grid in detail.

[0017] During the graph convolution process, the server regards each power IoT device as a node in the graph and establishes the connection relationship between the nodes based on the data path association and time sequence between the power IoT devices. Then, the server converts these nodes and their connection relationships into a series of graph convolution vectors through graph convolution operations. These graph convolution vectors not only contain the characteristic information of a single power IoT device, but also incorporate the information of other power IoT devices associated with it, thus forming a comprehensive description of the operating status of the power IoT device.

[0018] For example, in a more detailed explanation, the target power IoT device log refers to a record file collected from the power IoT device, which contains the device's operating status, events, errors, warnings and other information, which are usually recorded in order of timestamps and can be used for device monitoring, fault prediction, performance analysis, etc. For example, suppose a smart meter records its readings, voltage, current and other data at regular intervals and saves this information in log files. These log files are the logs of the power IoT device.

[0019] The equipment operation event data refers to the data of specific events or state changes that occur during the operation of the equipment. For example, the start, stop, failure, warning, etc. of the equipment can be regarded as operation events. For example, if a transformer triggers a warning due to overheating, this warning event and its related data (such as trigger time, temperature reading, etc.) will be recorded in the equipment operation event data.

[0020] Graph convolution is a mathematical operation that works on graph data (a data structure consisting of nodes and edges). In graph convolution, the feature value of each node is updated based on the feature values ​​of its neighboring nodes, thereby capturing the global structural information of the graph. In machine learning and deep learning, graph convolutional networks (GCNs) are often used to process graph data. For example, suppose that the nodes in the power network graph represent devices (such as generators, transformers, etc.) and the edges represent the connections between devices. Through graph convolution, the relationship between each device and its neighboring devices can be considered to more accurately predict or analyze the operating status of the device. After graph convolution, each node (in this context, the power IoT device) generates a new feature vector (also called vector), which captures the structure and attribute information of the node in the graph. When the feature vectors of all nodes are combined together, a graph convolution vector set is formed. For example, suppose there are 10 device nodes in the power network graph. After graph convolution, 10 feature vectors will be obtained, each vector represents the new feature of a device node after considering its neighboring nodes, and the set of these 10 feature vectors is the graph convolution vector set.

[0021] That is, in this embodiment, the device operation event data is first extracted from the log of the power Internet of Things device; then, a graph is constructed to represent the relationship between these devices; then, graph convolution processing is used to update the feature value of each device so that it contains the relationship information with its adjacent devices; finally, a graph convolution vector set is obtained, and this graph convolution vector set can be used for subsequent tasks such as equipment status analysis and fault prediction.

[0022] Step S120, performing heuristic fault path prediction on the graph convolution vector set, generating corresponding heuristic fault path data, and generating a fault location result of the target power Internet of Things device log based on the predicted heuristic fault path data.

[0023] In detail, heuristic fault path prediction refers to the use of set rules, connection relationships between devices, historical fault data and other information to predict possible fault propagation paths. This method can quickly give possible fault paths without the need for complex calculations or modeling. For example, if a transformer fails in a power network, heuristic fault path prediction can quickly predict which cables or switches the fault may propagate along, and thus which devices it may affect, based on historical data and data feature relationships between devices.

[0024] Heuristic fault path data refers to data generated through heuristic fault path prediction, including information such as predicted fault paths, potentially affected devices, and the probability of fault propagation. These heuristic fault path data can help operation and maintenance personnel quickly locate and resolve faults. For example, if it is predicted that a switch may fail due to overload, and this fault may propagate along a specific cable to several key devices, then these predicted fault paths and propagation information are heuristic fault path data.

[0025] The fault location result refers to the specific fault point or fault range obtained through analysis and calculation based on the heuristic fault path data. The fault location result can help operation and maintenance personnel quickly find and repair the fault, reducing power outage time and losses. For example, if a specific switch or cable is determined to be the fault point through analysis of the heuristic fault path data, then this conclusion is the fault location result. The operation and maintenance personnel can go directly to the site to repair it based on this result.

[0026] That is, in this embodiment, a graph convolution vector set is used to perform heuristic fault path prediction to generate possible fault paths and propagation information; then, heuristic fault path data is generated based on this information; finally, by analyzing these heuristic fault path data, specific fault location results are obtained, thereby guiding operation and maintenance personnel to perform rapid repairs.

[0027] Wherein, step S120 may include: Step S121, during each round of fault path prediction, a target graph embedding vector corresponding to the heuristic fault path data generated in the previous round of prediction is generated according to the graph embedding vectors corresponding to each predefined fault knowledge graph.

[0028] In detail, the predefined fault knowledge graph is a graph model built based on historical fault data, expert knowledge and experience, etc., which is used to represent the possible fault relationships and patterns between power IoT devices, and can help understand and predict the propagation path and impact range of faults. For example, according to historical data, the failure of a certain device often leads to the failure of another device, and this relationship can be represented in the predefined fault knowledge graph.

[0029] The graph embedding vector converts the nodes in the graph or the entire graph into a low-dimensional vector representation, while retaining the structural information of the graph and the similarity between nodes. This embedding can be used for various machine learning tasks, such as classification, clustering, and visualization. For example, assuming there is a power network graph containing multiple devices, the graph embedding vector can convert each device or the entire network into a numerical vector that captures the properties of the device and the relationship between the devices.

[0030] The target graph embedding vector refers to a graph embedding vector generated based on specific fault path data (such as heuristic fault path data) and is used to represent the characteristics of the fault path. For example, if a specific fault path is predicted, a target graph embedding vector can be generated based on the devices and relationships on this path for subsequent analysis and prediction.

[0031] Step S122, based on the feature distance between the prior fault path prediction knowledge generated in the previous round of fault path prediction and the graph convolution vector set, feature derivation is performed on the graph convolution vector set to generate a reinforcement vector set corresponding to this round of fault path prediction, and the prior fault path prediction knowledge is used to represent: the past fault path prediction information of the graph convolution vector set.

[0032] In detail, the feature distance is a measure of the similarity or difference between two feature vectors. In this embodiment, it can be used to compare the difference between the graph convolution vector and the prior knowledge of fault path prediction. For example, if there are two graph convolution vectors representing the state of a device, the feature distance can help determine whether the two states are similar or how different they are.

[0033] The prior fault path prediction knowledge refers to the knowledge or information that has been obtained before the current prediction is made. For example, the prior fault path prediction knowledge refers to the fault path prediction results and information of the previous round, which can be used to guide the current round of prediction. For example, if the previous round of prediction shows that a certain device is prone to failure, then this information can be used as prior fault path prediction knowledge and taken into account in the current round of prediction.

[0034] The enhanced vector set is a graph convolution vector set derived from features that enhances specific features (such as fault features). The enhanced vector set places more emphasis on features related to faults to improve the accuracy of fault path prediction. For example, assuming there is a set of graph convolution vectors, by enhancing features related to faults (such as voltage fluctuations, temperature anomalies, etc.), an enhanced vector set can be obtained, which may be more effective in fault path prediction.

[0035] Step S123, iteratively optimizing the priori fault path prediction knowledge according to the set of reinforcement vectors and the target graph embedding vector, and performing fault path prediction according to the iteratively optimized priori fault path prediction knowledge to generate heuristic fault path data for this round of fault path prediction.

[0036] In detail, iterative optimization is a process of gradually improving the results by repeatedly adjusting parameters or strategies, which can be used to optimize the prior knowledge of fault path prediction to improve the accuracy of fault path prediction. For example, if it is found that the current fault path prediction result deviates from the actual result, the prediction model or parameters can be adjusted through iterative optimization methods in the hope of obtaining more accurate results in the next prediction.

[0037] Therefore, in each round of fault path prediction, the target graph embedding vector is first generated according to the predefined fault knowledge graph; then the enhanced vector set is generated using the prior fault path prediction knowledge and the graph convolution vector set; then the prior fault path prediction knowledge is improved through iterative optimization; finally, the optimized knowledge is used to perform fault path prediction.

[0038] That is, in this embodiment, when performing each round of fault path prediction, the server first refers to the predefined fault knowledge graphs, which are developed by experts in the power field based on historical fault data and experience, and describe different types of faults and their possible propagation paths. The server converts these predefined fault knowledge graphs into graph embedding vectors through graph embedding technology, and then selects the target graph embedding vector that best matches the heuristic fault path data generated by the previous round of prediction as a reference.

[0039] Next, the server can evaluate the feature distance between the prior fault path prediction knowledge generated in the previous round of fault path prediction and the current graph convolution vector set, which reflects the correlation between past prediction information and the current device status. Based on this feature distance, the server performs feature derivation on the graph convolution vector set, strengthening those vectors that are highly correlated with past prediction information, while weakening or ignoring those irrelevant or contradictory vectors. In this way, the server obtains a strengthened vector set that is more focused on possible fault paths.

[0040] Finally, the server uses the reinforcement vector set and the target graph embedding vector to iteratively optimize the prior fault path prediction knowledge. In this process, the server continuously adjusts and optimizes the prediction results of the fault path to make it more consistent with the current equipment status and historical fault mode. After multiple rounds of iterative optimization, the server finally generates heuristic fault path data for this round of fault path prediction. These data not only reveal possible fault sources and propagation paths, but also provide important basis for subsequent fault location and repair. As a result, the server can accurately identify potential fault paths from complex power IoT device logs and take corresponding measures in a timely manner to prevent the spread of faults and affect the stable operation of the power grid.

[0041] Based on the above steps, the embodiment of the present application can effectively extract the deep features of the equipment operation event data from the complex power Internet of Things equipment log through graph convolution processing, generate a graph convolution vector set, and provide an accurate data basis for subsequent fault path prediction. The introduction of the heuristic fault path prediction mechanism can intelligently predict possible fault paths based on the graph convolution vector set, greatly improving the accuracy and efficiency of fault location. By combining the graph embedding vector corresponding to the predefined fault knowledge graph and the prediction results of the previous round in each round of fault path prediction, the target graph embedding vector is generated, and the dynamic adjustment and optimization of the fault path prediction is realized. The feature distance between the prior fault path prediction knowledge and the graph convolution vector set is used to derive features and generate a reinforcement vector set, which further enhances the precision and sensitivity of the fault path prediction. Through the iterative optimization of the prior fault path prediction knowledge, not only the accuracy of the fault path prediction is improved, but also the prediction system has the ability of self-learning and improvement, so that it can continuously adapt to the complex changes in the power Internet of Things environment. As a result, the accuracy, efficiency and intelligence level of the fault location of the power Internet of Things equipment are significantly improved.

[0042] In a possible implementation, step S121 may include: Step S1211, loading predefined fault knowledge graphs from a preset knowledge base, each of the predefined fault knowledge graphs represents a specific fault mode path, and each predefined fault knowledge graph is associated with a corresponding graph embedding vector, which is learned by graph embedding technology during the training phase.

[0043] In this embodiment, the server first accesses its built-in preset knowledge base, which stores a variety of predefined fault knowledge graphs. Each predefined fault knowledge graph is constructed based on historical fault data, expert knowledge and experience, and represents a specific fault mode path. For example, one of the predefined fault knowledge graphs may describe the fault mode of "when the transformer is overheated, it can cause aging of the insulation of the connected cable, thereby causing a short circuit". These predefined fault knowledge graphs have been associated with corresponding graph embedding vectors during the training phase. These graph embedding vectors are learned through graph embedding technology and can effectively represent the structural information of the graph and the similarity between nodes.

[0044] Step S1212, parsing the heuristic fault path data generated by the previous round of prediction, identifying key nodes and edges in the heuristic fault path data, wherein the key nodes and edges represent key steps and transitions in the fault path.

[0045] In this embodiment, the server obtains the heuristic fault path data generated by the previous round of prediction, which contains the fault path information that may occur in the power network. The server begins to parse the heuristic fault path data and identify the key nodes and edges therein. For example, in a heuristic fault path data, the key node may be a faulty transformer, and the key edge may be the line connecting the transformer and a section of cable. These key nodes and edges represent the key steps and transfers in the fault path.

[0046] Step S1213, matching the key nodes and edges with the nodes and edges in the predefined fault knowledge graph, and based on the matching result, retrieving the candidate graph embedding vector corresponding to the matched target fault knowledge graph from the predefined graph embedding vector library, wherein, if there are multiple matching target fault knowledge graphs, the best matching candidate graph embedding vector is selected based on the similarity or confidence score.

[0047] In this embodiment, the server matches the parsed key nodes and edges with the nodes and edges in the predefined fault knowledge graph. For example, the server finds that the faulty transformer in the previous round of prediction matches the "transformer overheating" node in a predefined fault knowledge graph, and the cable connected to the faulty transformer also matches the "cable insulation aging" edge in the knowledge graph. Based on this matching result, the server retrieves the candidate graph embedding vector corresponding to the matching target fault knowledge graph from the predefined graph embedding vector library. If there are multiple matching target fault knowledge graphs, the server can select the best matching candidate graph embedding vector based on the similarity or confidence score.

[0048] Step S1214: If the retrieved candidate graph embedding vector completely matches the heuristic fault path data, the candidate graph embedding vector is directly used as the target graph embedding vector.

[0049] In this embodiment, if the retrieved candidate graph embedding vector completely matches the heuristic fault path data generated by the previous round of prediction, for example, all key nodes and edges in the fault path can find completely corresponding graph embedding vectors in the predefined fault knowledge graph, then the server can directly use the candidate graph embedding vector as the target graph embedding vector.

[0050] Step S1215: If the retrieved candidate graph embedding vector partially matches or does not completely match the heuristic fault path data, the retrieved candidate graph embedding vector is adjusted by interpolation, extrapolation or a rule-based method to generate a target graph embedding vector that better matches the heuristic fault path data.

[0051] In this embodiment, if the retrieved candidate graph embedding vector partially matches or does not completely match the heuristic fault path data, for example, some key nodes or edges in the fault path do not have a directly corresponding graph embedding vector in the predefined fault knowledge graph, then the server can use interpolation, extrapolation or rule-based methods to adjust the retrieved candidate graph embedding vector. For example, the server can generate a new graph embedding vector based on the existing graph embedding vector and rules to more accurately represent a specific part of the fault path, thereby generating a target graph embedding vector that better matches the heuristic fault path data.

[0052] In a possible implementation manner, the method is implemented by completing a parameter-optimized power Internet of Things fault prediction model, and the method further includes: Step S101, performing parameter optimization on a graph convolution parameter array included in a graph convolution network through multiple fault path description templates, and generating a graph convolution parameter array that has completed parameter optimization when the model convergence conditions are met, each fault path description template is a description sequence consisting of fault paths corresponding to multiple device status description units, and the graph convolution network is used to map the fault path description template into a corresponding graph convolution vector set using the graph convolution parameter array.

[0053] In this embodiment, the server first obtains multiple fault path description templates, which are description sequences composed of fault paths corresponding to multiple device status description units. For example, a fault path description template may describe a fault path from "device A abnormality" to "device B overload" and then to "device C failure". The server uses these fault path description templates to perform parameter optimization on the graph convolution parameter array of the graph convolution network. During the optimization process, the graph convolution network uses the graph convolution parameter array to map these fault path description templates to corresponding graph convolution vector sets. When the model meets the convergence conditions, such as the loss function reaches a preset minimum value or the number of iterations reaches an upper limit, the server generates a graph convolution parameter array that completes the parameter optimization.

[0054] Step S102, using the graph convolution parameter array that has completed parameter optimization to guide the graph convolution parameter array in the power Internet of Things fault prediction model, to generate a target graph convolution parameter array, wherein the target graph convolution parameter array includes graph embedding vectors corresponding to each of the predefined fault knowledge graphs.

[0055] In this embodiment, the server then adjusts the graph convolution parameter array in the power Internet of Things fault prediction model with the graph convolution parameter array that has completed parameter optimization as a guide. In this process, the server can take into account various predefined fault knowledge graphs and generate corresponding graph embedding vectors for each fault knowledge graph. These graph embedding vectors can effectively represent the structure and characteristics of the fault knowledge graph. Finally, the server generates a target graph convolution parameter array, which contains the convolution parameters of the graph embedding vectors corresponding to each predefined fault knowledge graph.

[0056] Step S103, optimizing the parameters of the guided electric power Internet of Things fault prediction model according to multiple fault prediction training data, and generating the electric power Internet of Things fault prediction model with optimized parameters when the model convergence conditions are met, and each fault prediction training data includes: a template electric power Internet of Things device log and its corresponding heuristic fault path label data.

[0057] In this embodiment, the server obtains multiple fault prediction training data, each of which includes a template power Internet of Things device log and its corresponding heuristic fault path label data. For example, a training data may contain a device log that describes the operating status of the device within a certain period of time, as well as the corresponding fault path label, such as "from device D abnormality to device E shutdown". The server uses these training data to further optimize the parameters of the previously guided power Internet of Things fault prediction model. During this process, the model will learn how to predict possible fault paths from the device log. When the model meets the convergence conditions, such as the accuracy on the validation set reaches the preset value or the training round reaches the upper limit, the server generates a power Internet of Things fault prediction model with optimized parameters. The power Internet of Things fault prediction model can accurately predict possible fault paths based on the device log, thereby helping operation and maintenance personnel to troubleshoot and repair faults in a timely manner.

[0058] In a possible implementation, step S101 may include: Step S1011, use the multiple fault path description templates to perform cyclic parameter optimization on the graph convolutional network until the graph convolutional network meets the network convergence conditions. In each round of parameter optimization, feature shielding processing is performed on each fault path description template loaded in this round to shield the device status description unit of the shielded node in each fault path description template.

[0059] In this embodiment, the server starts to use multiple fault path description templates to perform cyclic parameter optimization on the graph convolutional network. This process will continue until the graph convolutional network meets the network convergence condition, that is, the performance of the model tends to be stable and no longer has significant improvement. In each round of parameter optimization, the server can perform feature masking processing on each fault path description template loaded in this round, which means that the server can deliberately mask the device status description unit of certain nodes to test the prediction ability of the model in the absence of some information.

[0060] Step S1012, based on the graph convolution parameter array called in this round, respectively generate graph convolution vector distributions corresponding to each fault path description template after feature masking processing.

[0061] In this embodiment, after feature masking processing, the server can generate graph convolution vector distributions corresponding to each fault path description template after feature masking processing based on the graph convolution parameter array called in this round. These graph convolution vector distributions are actually a mathematical representation of the fault path, capturing the relationship and characteristics between each node in the path.

[0062] Step S1013, predicting the device state description units corresponding to the shielded nodes in each fault path description template respectively according to the generated graph convolution vector distributions.

[0063] The server will then use the generated graph convolution vector distribution to predict the device status description unit corresponding to the shielded node in each fault path description template. This is actually testing the recovery ability of the model, that is, whether the model can infer the status of the shielded part based on the existing information.

[0064] Step S1014, optimizing the graph convolution parameter array according to the difference parameters between the device state description units actually shielded by each fault path description template and the predicted device state description units.

[0065] Finally, the server can optimize the graph convolution parameter array based on the difference parameters between the real shielded device state description unit and the device state description unit predicted by the model in each fault path description template. The difference parameter reflects the accuracy of the model prediction. If the prediction is significantly different from the actual situation, then the graph convolution parameter array needs to be adjusted accordingly to improve the prediction ability of the model. The optimization process is achieved through the back propagation algorithm, that is, the parameters of the model are updated according to the prediction error, so that the model can more accurately restore the shielded device state description unit in the next round of prediction. This process will be repeated until the performance of the model reaches a satisfactory level.

[0066] In a possible implementation, step S1011 includes: Step S1011 - 1 , loading the fault path description template to be processed in this round, parsing the structure of the fault path description template, and identifying the nodes in the fault path description template and their corresponding device status description units.

[0067] In this embodiment, the server first loads the fault path description templates that need to be processed in this round. These fault path description templates are stored in a specific format and may contain multiple nodes and their relationships. The server uses a special parser to read these fault path description templates and identify each node and the corresponding device status description unit. For example, a fault path description template may describe a fault path in a power system from "generator failure" to "transmission line interruption" to "substation outage", and each node contains a specific device status description, such as the voltage, current and other parameters of the generator.

[0068] Step S1011 - 2 , determining the nodes to be shielded according to preset rules, and creating a shielded node list to record the node information of the nodes to be shielded, wherein the preset rules include rules based on the type, attribute, and location of the nodes.

[0069] In this embodiment, the server determines which nodes need to be blocked according to preset rules. These rules may be based on the type of node (such as blocking only devices of a specific type), attribute (such as blocking devices with abnormal status), or location (such as blocking nodes at a specific location in the path). According to these rules, the server creates a blocked node list, which records the detailed information of all nodes to be blocked, such as the node identifier, type, etc.

[0070] Step S1011 - 3 traverses each node in the fault path description template and checks whether the node is in the shielded node list.

[0071] In this embodiment, the server starts to traverse each node in the fault path description template. For each node, the server checks whether it is in the previously created blocked node list, which may require comparing the node's identifier or other unique identification information. If a node finds a match in the blocked list, it will be considered as a node to be blocked.

[0072] Step S1011 - 4 : performing a shielding operation on a node in the shielded node list, wherein the shielding operation includes deleting, replacing or hiding a device status description unit of the node.

[0073] In this embodiment, for the nodes in the blocked node list, the server performs a blocking operation, which may include deleting the device status description unit of the node, or replacing it with a specific placeholder, or even simply hiding this information so that it is not considered in subsequent processing. The specific blocking method depends on the configuration and requirements of the server.

[0074] Step S1011 - 5 , after completing the feature masking process, updating the fault path description template.

[0075] In this embodiment, after completing the shielding operation of all nodes, the server can update the original fault path description template, which may involve writing the shielded node information back to the template, or creating a new template version to store these changes. The updated template will be used in the subsequent graph convolutional network parameter optimization process to ensure that the model can take into account the impact of these shielding operations when processing similar fault paths.

[0076] In a possible implementation, before step S101, the method further includes: Step A110 , extracting the fault operation vectors contained in each fault operation log from the fault operation log sequences with multiple set category attributes, and generating a reference fault operation vector sequence.

[0077] In this embodiment, the server first starts working from a sequence of fault operation logs with multiple set category attributes. These fault operation logs may record the operation status, fault codes, environmental parameters and other information of various devices at different time points. The server uses data extraction techniques, such as regular expressions, data cleaning and conversion tools, to extract key fault operation vectors from these fault operation logs. These vectors are numerical features that represent the operation status of the equipment, such as equipment temperature, pressure, current value, etc.

[0078] After the extraction is completed, the server arranges the fault operation vectors in chronological order or in the order in which the faults occurred, and generates a reference fault operation vector sequence, which provides basic data for subsequent analysis and template generation.

[0079] Step A120 , removing the fault operation vectors that do not meet the set template requirement from the reference fault operation vector sequence to generate a target fault operation vector sequence.

[0080] After generating the reference fault operation vector sequence, the server starts to filter the sequence. The server removes the fault operation vectors that do not meet the preset template requirements, such as only including specific types of faults, only considering faults occurring in a specific time period, etc.

[0081] This process may involve complex logical judgment and data processing techniques, such as using SQL query language for data screening, or using machine learning algorithms for outlier detection, etc. After removing vectors that do not meet the requirements, the server generates a new, more refined target fault operation vector sequence.

[0082] Step A130, for each fault operation vector in the target fault operation vector sequence, fault path mapping is performed through a fault path knowledge graph to generate the multiple fault path description templates, wherein the fault path knowledge graph contains the fault path corresponding to each device state description unit.

[0083] The server then uses a pre-built fault path knowledge graph to map the fault paths. The knowledge graph is a complex network of device state description units and possible fault paths between them. Each device state description unit is a node in the graph, and the lines between the nodes represent possible fault propagation paths.

[0084] The server matches each vector in the target fault operation vector sequence with the nodes in the knowledge graph to find the most likely fault path, a process that may involve complex graph search and reasoning algorithms, such as the shortest path algorithm, graph traversal algorithm, etc. Through mapping, the server is able to generate multiple fault path description templates that describe the complete path from the initial fault state to the final fault state.

[0085] Finally, the server obtained multiple accurate fault path description templates, which will be used in the subsequent graph convolutional network parameter optimization process to improve the model's prediction and diagnosis capabilities for similar faults.

[0086] In a possible implementation, step S122 may include: Step S1221, performing a correlation analysis on the prior fault path prediction knowledge and the graph convolution vector set, generating a correlation cost parameter corresponding to each graph convolution vector in the graph convolution vector set, each correlation cost parameter is used to represent: a characteristic distance between the prior fault path prediction knowledge and the corresponding graph convolution vector.

[0087] In this embodiment, the server first obtains the prior fault path prediction knowledge generated in the previous round of fault path prediction, which is usually a data set containing historical fault paths, device state changes and their correlations. At the same time, the server also obtains the graph convolution vector set currently generated by the graph convolution network, which reflects the state characteristics of each node in the fault path.

[0088] Next, the server starts the cross-correlation analysis, which involves comparing the features of the prior fault path prediction knowledge and the current graph convolution vector to find the similarities and differences between them. The server can use various algorithms, such as cosine similarity, Euclidean distance, etc., to quantify this feature distance.

[0089] After completing the correlation analysis, the server can generate a set of cross-correlation cost parameters. Each graph convolution vector will have a corresponding cross-correlation cost parameter, which represents the characteristic distance between the prior fault path prediction knowledge and the graph convolution vector. Simply put, if a graph convolution vector is highly similar to the prior fault path prediction knowledge, then its cross-correlation cost parameter will be relatively low; conversely, if the difference is large, the cost parameter will be relatively high.

[0090] Step S1222: According to the mutual correlation cost parameters corresponding to each of the graph convolution vectors, feature enhancement processing based on mutual correlation cost is performed on the graph convolution vector set to generate the enhanced vector set.

[0091] In this embodiment, after obtaining the cross-correlation cost parameters of each graph convolution vector, the server will use these parameters to perform feature enhancement processing on the graph convolution vector set. This process aims to adjust the weight of the current graph convolution vector based on the prior fault path prediction knowledge, so as to more accurately reflect the actual situation in the subsequent fault path prediction.

[0092] Specifically, the server can give greater weight to those graph convolution vectors that are highly similar to the prior fault path prediction knowledge, because they are more likely to represent the true fault path characteristics. For those vectors that are very different from the prior fault path prediction knowledge, the server can reduce their weights to reduce their influence in the prediction.

[0093] Through this feature enhancement processing based on cross-correlation cost, the server will eventually generate a new set of enhanced vectors. The vectors in this set not only retain the feature information of the original graph convolution vectors, but also incorporate the influencing factors of prior fault path prediction knowledge, which is expected to improve the accuracy and reliability of fault path prediction.

[0094] In a possible implementation, step S1221 may include: Step S1221 - 1 , projecting the priori fault path prediction knowledge and each graph convolution vector into a predefined feature representation domain respectively.

[0095] In this embodiment, the server first projects the prior fault path prediction knowledge and each graph convolution vector in the graph convolution vector set into a predefined feature representation domain. The predefined feature representation domain can be regarded as a multi-dimensional space, and each dimension represents a specific fault or equipment status feature.

[0096] Specifically, the server can use a mapping function or algorithm to transform the prior fault path prediction knowledge and graph convolution vectors from their original data space into this predefined feature representation domain. This process is similar to converting data from one coordinate system to another for easier comparison and analysis.

[0097] For example, assuming that the prior knowledge of fault path prediction includes the temperature changes of devices in a historical fault path, and a graph convolution vector also contains temperature-related information, the server can map these temperature data to the corresponding dimensions in the feature representation domain for subsequent comparison.

[0098] Step S1221-2, based on the mapped prior fault path prediction knowledge and the deviation state value of each graph convolution vector in the predefined feature representation domain, generate the mutual correlation cost parameters corresponding to each graph convolution vector respectively, the larger the deviation state value, the larger the mutual correlation cost parameter.

[0099] After projecting the data into the predefined feature representation domain, the server may start to calculate the deviation state value between the mapped prior fault path prediction knowledge and each graph convolution vector in the feature representation domain, and the deviation state value reflects the similarity or difference between the two.

[0100] To calculate the deviation state value, the server can use various metrics, such as Euclidean distance, cosine similarity, etc., to measure the distance or angle between the prior fault path prediction knowledge and the graph convolution vector in the feature representation domain. The larger the deviation state value, the greater the difference between the two.

[0101] Next, the server can generate the mutual correlation cost parameter corresponding to each graph convolution vector according to these deviation state values. This parameter is actually a weight value used to indicate the degree of correlation between the prior fault path prediction knowledge and the graph convolution vector. When the deviation state value is large, it means that the correlation between the prior fault path prediction knowledge and the graph convolution vector is weak, so the mutual correlation cost parameter will increase accordingly; conversely, if the deviation state value is small, it means that the correlation between the two is strong, and the mutual correlation cost parameter will decrease.

[0102] In this way, the server can quantify the correlation between the prior fault path prediction knowledge and the graph convolution vector, providing an important basis for subsequent feature enhancement processing and fault path prediction.

[0103] In a possible implementation, step S1221 may further include: Step S1221 - 3 , determining the target device operation event data whose inter-correlation cost parameter is less than the preset cost parameter, or the target device operation event data whose inter-correlation cost parameter is the smallest, from the inter-correlation cost parameters generated in the previous round of fault path prediction.

[0104] In this embodiment, the server first traces back the mutual correlation cost parameters generated in the previous round of fault path prediction, which reflect the degree of correlation between the prior fault path prediction knowledge and the graph convolution vector. The server can set a preset cost parameter as a threshold for screening the mutual correlation cost parameters.

[0105] The server can search for device operation event data with an inter-correlation cost parameter less than the preset cost parameter. Such data is considered to be highly relevant to the prior fault path prediction knowledge and is referred to as "target device operation event data". If there are multiple such data, the server can select the one with the smallest inter-correlation cost parameter, i.e., the event most closely related to the prior fault path prediction knowledge.

[0106] Step S1221-4: split the graph convolution vector set according to the target device operation event data to remove the device operation event data before the target device operation event data from the graph convolution vector set.

[0107] In this embodiment, after determining the target device operation event data, the server can segment the current graph convolution vector set. The purpose of segmentation is to remove all device operation event data before the target device operation event data, because these data may have a low correlation with the current fault path prediction.

[0108] For example, if the target device operation event data is a data recording abnormal device temperature, then all other device operation event data such as temperature records and pressure records before this data will be removed, and only the target device operation event data and the data after it will be retained.

[0109] Step S1221-5, performing a correlation analysis on the priori fault path prediction knowledge and the segmented graph convolution vector set, and generating a correlation cost parameter corresponding to each graph convolution vector in the segmented graph convolution vector set.

[0110] After the segmentation is completed, the server will perform a correlation analysis on the remaining graph convolution vector set and the prior fault path prediction knowledge. This process is similar to the previous correlation analysis, but this time the analysis is on the filtered and segmented data.

[0111] The server can calculate the deviation state value between each graph convolution vector and the prior fault path prediction knowledge, which reflects the similarity or difference between them. Based on these deviation state values, the server can generate new correlation cost parameters, which will be used for subsequent feature enhancement processing.

[0112] Then step S1222 may include: performing feature enhancement processing based on the mutual correlation cost on the segmented graph convolution vector set according to the generated mutual correlation cost parameters, to generate the enhanced vector set.

[0113] In this embodiment, the server can perform feature enhancement processing on the segmented graph convolution vector set according to the newly generated mutual correlation cost parameter. The purpose of this processing is to increase the weight of graph convolution vectors with high correlation with prior fault path prediction knowledge, while reducing the influence of vectors with low correlation.

[0114] Specifically, for graph convolution vectors with smaller mutual correlation cost parameters (i.e., vectors with high correlation with prior fault path prediction knowledge), the server can increase their weight in subsequent fault path prediction; while for vectors with larger mutual correlation cost parameters (i.e., vectors with low correlation), their weight will be reduced.

[0115] Through this feature enhancement processing, the server can generate a set of enhanced vectors, in which the vectors are more focused on features that are highly correlated with prior fault path prediction knowledge, thereby improving the accuracy and reliability of fault path prediction.

[0116] In a possible implementation, step S1222 may include: Step S1222-1, performing weighted fusion on the graph convolution vector set according to the mutual correlation cost parameters corresponding to each graph convolution vector, to generate a fused vector set.

[0117] In this embodiment, after obtaining the mutual correlation cost parameters corresponding to each graph convolution vector, the server will use these parameters to perform weighted fusion on the graph convolution vector set. The purpose of weighted fusion is to readjust the weights of each graph convolution vector in the fusion process according to the correlation between each graph convolution vector and the prior fault path prediction knowledge.

[0118] Specifically, the server can traverse each vector in the graph convolution vector set. For each vector, the server can determine a weight value according to its corresponding mutual correlation cost parameter. Generally, vectors with smaller mutual correlation cost parameters (i.e., more correlated with prior fault path prediction knowledge) are assigned larger weights, while vectors with larger mutual correlation cost parameters (i.e., less correlated) are assigned smaller weights.

[0119] Then, the server can fuse these weighted graph convolution vectors. Fusion can be a simple weighted average or a more complex fusion algorithm, such as fusion based on the attention mechanism. Through weighted fusion, the server can obtain a new vector set that fuses all graph convolution vector information, called the "fused vector set". The vectors in this set not only retain the information of the original graph convolution vectors, but also reflect their relevance to the prior fault path prediction knowledge.

[0120] Step S1222-2: Integrate the fused vector set with the graph convolution vector set to generate the enhanced vector set.

[0121] After generating the fused vector set, the server can integrate it with the original graph convolution vector set to generate an enhanced vector set. The purpose of the integration is to combine the weighted fused information with the original information to obtain an enhanced feature set that contains both the original features and reflects the relevance to the prior fault path prediction knowledge.

[0122] There are many ways to integrate, such as concatenating, adding or transforming the fused vector set with the original graph convolution vector set. The specific method depends on the specific application scenario and requirements.

[0123] For example, the server may choose to concatenate each vector in the fused vector set with the vector at the corresponding position in the original graph convolution vector set to form a longer vector, which contains both the information of the original graph convolution vector and the information after weighted fusion, and thus has stronger feature expression capabilities.

[0124] In this way, the server will eventually generate a set of enhanced vectors, the vectors in which not only retain the feature information of the original graph convolution vectors, but also incorporate the correlation with the prior fault path prediction knowledge, which is expected to improve the accuracy and reliability of subsequent fault path predictions.

[0125] In a possible implementation, step S123 may include: Step S1231, integrating the enhanced vector set and the target graph embedding vector to generate an integrated vector sequence for this round of fault path prediction.

[0126] In this embodiment, after the server has the set of reinforcement vectors and the target graph embedding vector, it will perform an integration operation. The set of reinforcement vectors is obtained through the feature reinforcement processing based on the mutual correlation cost in the previous step, and contains feature information that is highly correlated with prior knowledge. The target graph embedding vector represents the characteristics of the target node or key device in the fault path.

[0127] During the integration process, the server can combine each vector in the reinforcement vector set with the target graph embedding vector. This combination can be a simple concatenation or a combination after some transformation (such as linear transformation, nonlinear activation function, etc.), depending on the design of the model and the training objectives.

[0128] For example, if each vector in the set of reinforced vectors represents the state characteristics of a device failure, and the target graph embedding vector represents the key devices in the failure path, then the integrated vector sequence will contain both the state characteristics of the device and the identification information of the key devices.

[0129] Through this integration step, the server obtains an integrated vector sequence that combines the enhanced features and the target features, which will serve as the basis for the subsequent iterative optimization process.

[0130] Step S1232, based on the first importance coefficient defined for the integrated vector sequence and the second importance coefficient defined for the priori fault path prediction knowledge, the integrated vector sequence and the priori fault path prediction knowledge are merged to generate the iteratively optimized priori fault path prediction knowledge.

[0131] The first importance coefficient and the second importance coefficient are learned in a network parameter optimization phase.

[0132] In this embodiment, after obtaining the integrated vector sequence, the server may fuse it with the priori fault path prediction knowledge. The fusion process is to integrate new feature information into the priori knowledge, thereby achieving iterative optimization.

[0133] Before fusion, the server may define a first importance coefficient and a second importance coefficient for the integrated vector sequence and the priori fault path prediction knowledge, respectively. These two coefficients are obtained through learning in the network parameter optimization phase, and they reflect their respective weights in the fusion process.

[0134] Specifically, the first importance coefficient may represent the importance or credibility of the features in the integrated vector sequence, while the second importance coefficient represents the reliability or influence of the prior fault path prediction knowledge. The specific values ​​of these two coefficients will be adjusted according to the training process of the model and the characteristics of the data.

[0135] During fusion, the server can weight the integrated vector sequence and prior knowledge according to the two importance coefficients. For example, if the first importance coefficient is higher, the features in the integrated vector sequence will have a larger proportion in the fusion result; conversely, if the second importance coefficient is higher, the influence of prior knowledge will be more significant.

[0136] Through this importance coefficient-based fusion method, the server is able to generate an iteratively optimized prior fault path prediction knowledge, which not only retains the useful information of the original prior knowledge, but also incorporates new feature information, which is expected to improve the accuracy and robustness of subsequent fault path predictions.

[0137] In a possible implementation, step S110 may include: Step S111, performing device state migration node parsing on the target electric power IoT device log to generate a corresponding device state migration node set, wherein each device state migration node in the device state migration node set corresponds to a device operation event data.

[0138] In this embodiment, the server first carefully analyzes the logs of the target power IoT devices, which record various device operation event data, such as device startup, shutdown, failure, recovery, etc. The server converts these device operation event data into device state migration nodes through a specific parsing algorithm.

[0139] For example, the target power IoT device log may record that a transformer has migrated from the "normal operation" state to the "overload" state, and then migrated back to the "normal operation" state. The server can use these two state migrations as two device state migration nodes, each of which corresponds to a specific device operation event data.

[0140] In this way, the server can generate a set of device state migration nodes, each of which represents a change in the device state.

[0141] Step S112, performing serialized graph convolution processing on the set of device state transition nodes until a graph convolution vector corresponding to each device state transition node is generated. In each round of graph convolution processing, if the device operation event data of this round is the first device operation event data, then based on the device state transition node of the device operation event data of this round and the defined initialization vector, a graph convolution vector of the device operation event data of this round is generated.

[0142] In this embodiment, next, the server may perform a serialized graph convolution process on the device state migration node set. Graph convolution is a convolution operation performed on graph data, which can effectively extract feature information in the graph.

[0143] During the processing, the server can perform graph convolution on the device state transition nodes one by one in the order of the device operation event data. For the first device operation event data in the set, since there is no previous data for reference, the server can use a predefined initialization vector to perform graph convolution operation with the device state transition node of this round of device operation event data to generate a graph convolution vector for the event.

[0144] For example, if the first event is the startup of a transformer, the server can combine the initialization vector and the device state transition node of the startup event, and obtain a graph convolution vector representing the startup characteristics of the transformer through graph convolution processing.

[0145] Step S113: if the device operation event data of this round is not the first device operation event data, the graph convolution vector of the device operation event data of this round is generated according to the graph convolution vector extracted in the previous round and the device state transition node of the device operation event data of this round.

[0146] In this embodiment, for device operation event data other than the first one, the server can use the graph convolution vector extracted in the previous round and the device state migration node of the current round of device operation event data to perform graph convolution processing. This can capture information about how the device state changes over time, making the generated graph convolution vector more time-series and context-related.

[0147] For example, if the second event is a transformer overload, the server can perform a graph convolution operation on the graph convolution vector of the first event (transformer startup) with the device state transition node of the overload event to obtain a graph convolution vector that combines the startup and overload features.

[0148] Step S114: Generate the graph convolution vector set based on the generated graph convolution vectors.

[0149] In this embodiment, as the server performs graph convolution processing on each device operation event data, a set of graph convolution vectors corresponding to each device state transition node is gradually accumulated. This set is the graph convolution vector set, which captures the characteristic information of all important events in the device log.

[0150] Finally, the server can use this graph convolution vector set for subsequent analysis and prediction, such as fault warning, performance optimization, etc. In this way, the server can more accurately understand the operating status and performance of power IoT devices, providing strong support for the stable operation of the power system.

[0151] In a possible implementation, step S111 includes: Step S1111, using a natural language processing algorithm to extract keyword features related to the device status from the target power Internet of Things device log.

[0152] In this embodiment, in order to extract useful information from these complex target power Internet of Things device logs, the server can use a natural language processing algorithm for processing.

[0153] The natural language processing algorithm will first segment and tag the log data of the target power IoT device, and then extract the keyword features related to the device status. For example, the log of the target power IoT device may contain words such as "startup success", "temperature is too high", "voltage abnormality" and so on that describe the device status, which will be recognized and extracted by the algorithm.

[0154] Step S1112: Classify the record units of each time period in the target power Internet of Things device log into different device states according to the extracted keyword features, and mark the corresponding timestamp for each identified device state.

[0155] In this embodiment, after extracting the keyword features, the server can classify the record units of each time period in the log into different device states according to these keyword features. For example, if the keyword "temperature is too high" frequently appears in the log of a certain time period, the device state in this time period can be classified as "overheated state".

[0156] At the same time, the server will also mark the corresponding timestamp for each identified device status so that the change process of the device status can be clearly tracked later.

[0157] Step S1113, arranging the identified device states into a state sequence according to the order of the marked timestamps, and detecting a transition point from one state to another in the state sequence, wherein the transition point is the device state migration node.

[0158] In this embodiment, the server can arrange the identified device states into a state sequence according to the marked timestamp order. Then, the server can detect the transition points from one state to another in the state sequence. These transition points are device state migration nodes, which represent important changes in the device state.

[0159] For example, in the state sequence, the server may find that the device suddenly changes from the "normal operation state" to the "overload state", and then this transition point will be identified as a device state transition node.

[0160] Step S1114, determine the migration type of each device state migration node, and encapsulate the node information of each device state migration node into a data object according to the migration type of each device state migration node, then combine all encapsulated data objects into a device state migration node set, and associate a corresponding device operation event data with each device state migration node in the device state migration node set.

[0161] After determining the device state migration node, the server can further determine the migration type of each node, such as migration from "normal operation" to "fault state", or migration from "shutdown state" to "operation state", etc.

[0162] Then, the server can encapsulate the relevant information of each device state migration node (such as migration time, migration type, device state before and after migration, etc.) into a data object according to the migration type of each device state migration node. These data objects not only contain the change information of the device state, but also facilitate subsequent data processing and analysis.

[0163] Finally, the server can combine all the encapsulated data objects into a device state migration node set. In this set, each device state migration node is associated with a corresponding device operation event data, so as to fully reflect the operation state and change process of the device.

[0164] In a possible implementation, the method further includes: Step S1115: remove duplicate or invalid device state transition nodes from the generated device state transition node set.

[0165] In this embodiment, after the server completes the generation of the device state transition node set, the device state transition node set may be further processed to remove duplicate or invalid device state transition nodes.

[0166] First, the server can traverse the entire device state migration node set to check whether there are duplicate nodes. Duplicate nodes may be caused by redundant information in log records or errors in data processing. For example, if the server detects the same device state migration (such as migration from "running state" to "stopped state") at two consecutive time points, and the interval between these two time points is very short, then these two nodes may be duplicates.

[0167] In order to remove these duplicate nodes, the server can set an appropriate time threshold. If the time difference between two adjacent device state migration nodes is less than this time threshold, and the migration type and the device state before and after the migration are exactly the same, then the server will determine that the two nodes are duplicates and delete one of the nodes from the set.

[0168] In addition, the server will check whether the device state migration node is valid. Invalid nodes may be caused by errors or incomplete log data. For example, if the migration type or device state information before and after the migration of a device state migration node is missing or unclear, then the node will be considered invalid. The server can remove these invalid nodes from the collection to ensure the accuracy of subsequent analysis.

[0169] Alternatively, in step S1116, a time window is set, and only the device state transition nodes within the time window are retained in the device state transition node set.

[0170] After processing duplicate and invalid nodes, the server can also set a time window according to actual needs to further filter the data in the device state migration node set.

[0171] This time window can be a specific time period, such as "the past 24 hours", "this week" or "last month". The server can filter out qualified data from the device state migration node set based on the setting of this time window. Only the device state migration nodes that fall within this time window will be retained for subsequent analysis and processing.

[0172] For example, if the time window set by the server is "the past 24 hours", it will traverse the entire device state migration node set and delete those nodes whose timestamps are not within the past 24 hours. This can help the server focus on the device state changes in the recent period of time, so as to discover and solve problems more promptly. At the same time, by adjusting the size of the time window, the server can also flexibly control the granularity and scope of the analysis.

[0173] Figure 2 The hardware structure of the power Internet of Things device data analysis system 100 based on artificial intelligence for implementing the above-mentioned power Internet of Things device data analysis method based on artificial intelligence provided in the embodiment of the present application is shown as follows: Figure 2 As shown, the artificial intelligence-based power Internet of Things device data analysis system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

[0174] In one possible design, the power Internet of Things device data analysis system 100 based on artificial intelligence can be a single server or a server group. The server group can be centralized or distributed (for example, the power Internet of Things device data analysis system 100 based on artificial intelligence can be a distributed system). In some embodiments, the power Internet of Things device data analysis system 100 based on artificial intelligence can be local or remote. For example, the power Internet of Things device data analysis system 100 based on artificial intelligence can access information and / or data stored in the machine-readable storage medium 120 via a network. For another example, the power Internet of Things device data analysis system 100 based on artificial intelligence can be directly connected to the machine-readable storage medium 120 to access the stored information and / or data. In some embodiments, the power Internet of Things device data analysis system 100 based on artificial intelligence can be implemented on the power Internet of Things device data analysis system based on artificial intelligence. By way of example only, the power Internet of Things device data analysis system based on artificial intelligence can include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, etc. or any combination thereof.

[0175] The machine-readable storage medium 120 may store data and / or instructions. In some embodiments, the machine-readable storage medium 120 may store data acquired from an external terminal. In some embodiments, the machine-readable storage medium 120 may store data and / or instructions that the artificial intelligence-based power IoT device data analysis system 100 uses to execute or use to complete the exemplary method described in this application.

[0176] During the specific implementation process, at least one processor 110 executes computer executable instructions stored in the machine-readable storage medium 120, so that the processor 110 can execute the artificial intelligence-based power Internet of Things device data analysis method as described in the above method embodiment. The processor 110, the machine-readable storage medium 120 and the communication unit 140 are connected via the bus 130, and the processor 110 can be used to control the sending and receiving actions of the communication unit 140.

[0177] The specific implementation process of the processor 110 can refer to the various method embodiments executed by the above-mentioned artificial intelligence-based power Internet of Things device data analysis system 100. The implementation principles and technical effects are similar, and this embodiment will not be repeated here.

[0178] In addition, an embodiment of the present application also provides a readable storage medium, in which computer executable instructions are preset. When the processor executes the computer executable instructions, the above-mentioned artificial intelligence-based power Internet of Things device data analysis method is implemented.

[0179] It should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof. Similarly, it should be noted that in order to simplify the description disclosed in this application and thus help understand one or more embodiments of the invention, in the foregoing description of the embodiments of the present application, multiple features are sometimes combined into one embodiment, drawings, or descriptions thereof.

Claims

1. A method for analyzing data of power Internet of Things devices based on artificial intelligence, characterized in that: The method comprises: Perform graph convolution processing on the operation event data of each device included in the target power IoT device log to generate a corresponding graph convolution vector set; Perform heuristic fault path prediction on the graph convolution vector set to generate corresponding heuristic fault path data, and generate a fault location result of the target power Internet of Things device log based on the predicted heuristic fault path data; The step of performing heuristic fault path prediction on the graph convolution vector set to generate corresponding heuristic fault path data includes: In each round of fault path prediction, a target graph embedding vector corresponding to the heuristic fault path data generated by the previous round of prediction is generated according to the graph embedding vectors corresponding to each predefined fault knowledge graph; According to the feature distance between the prior fault path prediction knowledge generated in the previous round of fault path prediction and the graph convolution vector set, feature derivation is performed on the graph convolution vector set to generate a reinforcement vector set corresponding to the current round of fault path prediction, wherein the prior fault path prediction knowledge is used to represent: past fault path prediction information of the graph convolution vector set; The priori fault path prediction knowledge is iteratively optimized according to the set of reinforcement vectors and the target graph embedding vector, and fault path prediction is performed according to the iteratively optimized priori fault path prediction knowledge to generate heuristic fault path data for this round of fault path prediction.

2. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to claim 1 is characterized in that: The method is implemented by completing a parameter-optimized power Internet of Things fault prediction model, and the method also includes: Parameter optimization is performed on a graph convolution parameter array included in a graph convolution network through multiple fault path description templates, and when a model convergence condition is met, a graph convolution parameter array that has completed parameter optimization is generated, each fault path description template is a description sequence composed of fault paths corresponding to multiple device state description units, and the graph convolution network is used to map the fault path description template to a corresponding graph convolution vector set using the graph convolution parameter array; Using the graph convolution parameter array that has completed parameter optimization, the graph convolution parameter array in the power Internet of Things fault prediction model is guided to generate a target graph convolution parameter array, wherein the target graph convolution parameter array includes graph embedding vectors corresponding to each of the predefined fault knowledge graphs; Optimizing the parameters of the guided power Internet of Things fault prediction model according to multiple fault prediction training data, and generating the power Internet of Things fault prediction model with the optimized parameters when the model convergence condition is met, each fault prediction training data includes: a template power Internet of Things device log and its corresponding heuristic fault path label data; The graph convolution parameter array included in the graph convolution network is optimized through multiple fault path description templates, including: Utilizing the multiple fault path description templates to perform cyclic parameter optimization on the graph convolution network until the graph convolution network meets the network convergence condition, and in each round of parameter optimization, performing feature shielding processing on each fault path description template loaded in this round to shield the device state description unit of the shielded node in each fault path description template; According to the graph convolution parameter array called in this round, the graph convolution vector distribution corresponding to each fault path description template after feature masking processing is generated respectively; According to the distribution of each generated graph convolution vector, the device state description unit corresponding to the shielded node in each fault path description template is predicted respectively; The graph convolution parameter array is optimized according to the difference parameters between the device state description units actually shielded by each fault path description template and the predicted device state description units.

3. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to claim 2 is characterized in that: Before performing parameter optimization on the graph convolution parameter array included in the graph convolution network by using a plurality of fault path description templates, the method further includes: Extracting the fault operation vectors contained in each fault operation log from a plurality of fault operation log sequences with set category attributes, and generating a reference fault operation vector sequence; Remove the fault operation vectors that do not meet the set template requirements from the reference fault operation vector sequence to generate a target fault operation vector sequence; For each fault operation vector in the target fault operation vector sequence, fault path mapping is performed through a fault path knowledge graph to generate the multiple fault path description templates, wherein the fault path knowledge graph contains a fault path corresponding to each device state description unit.

4. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to any one of claims 1 to 3, characterized in that: The feature distance between the priori fault path prediction knowledge generated in the previous round of fault path prediction and the graph convolution vector set is used to derive features from the graph convolution vector set to generate a reinforcement vector set corresponding to the current round of fault path prediction, including: Performing a correlation analysis on the priori fault path prediction knowledge and the graph convolution vector set to generate a correlation cost parameter corresponding to each graph convolution vector in the graph convolution vector set, wherein each correlation cost parameter is used to represent: a characteristic distance between the priori fault path prediction knowledge and the corresponding graph convolution vector; According to the mutual correlation cost parameters respectively corresponding to the respective graph convolution vectors, feature enhancement processing based on mutual correlation cost is performed on the graph convolution vector set to generate the enhanced vector set.

5. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to claim 4 is characterized in that: Performing a correlation analysis on the priori fault path prediction knowledge and the graph convolution vector set to generate a correlation cost parameter corresponding to each graph convolution vector in the graph convolution vector set, including: Projecting the priori fault path prediction knowledge and each graph convolution vector into a predefined feature representation domain respectively; According to the mapped prior fault path prediction knowledge and the deviation state value of each graph convolution vector in the predefined feature representation domain, the mutual correlation cost parameters corresponding to each graph convolution vector are generated respectively. The larger the deviation state value, the larger the mutual correlation cost parameter.

6. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to claim 4 is characterized in that: Performing a correlation analysis on the priori fault path prediction knowledge and the graph convolution vector set to generate a correlation cost parameter corresponding to each graph convolution vector in the graph convolution vector set, including: Determine, from the inter-correlation cost parameters generated in the previous round of fault path prediction, target device operation event data whose inter-correlation cost parameters are less than the preset cost parameters, or target device operation event data whose inter-correlation cost parameters are the smallest; According to the target device operation event data, the graph convolution vector set is segmented to remove the device operation event data before the target device operation event data from the graph convolution vector set; Performing a correlation analysis on the priori fault path prediction knowledge and the segmented graph convolution vector set to generate a correlation cost parameter corresponding to each graph convolution vector in the segmented graph convolution vector set; Then, according to the mutual correlation cost parameters respectively corresponding to the respective graph convolution vectors, the graph convolution vector set is subjected to feature enhancement processing based on the mutual correlation cost to generate the enhanced vector set, including: According to each generated mutual correlation cost parameter, the segmented graph convolution vector set is subjected to feature enhancement processing based on mutual correlation cost to generate the enhanced vector set.

7. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to claim 4 is characterized in that: According to the mutual correlation cost parameters respectively corresponding to the respective graph convolution vectors, the graph convolution vector set is subjected to feature enhancement processing based on the mutual correlation cost to generate the enhanced vector set, including: According to the mutual correlation cost parameters corresponding to the respective graph convolution vectors, the graph convolution vector set is weightedly fused to generate a fused vector set; The fused vector set is integrated with the graph convolution vector set to generate the enhanced vector set.

8. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to any one of claims 1 to 3, characterized in that: According to the set of reinforcement vectors and the target graph embedding vector, the priori fault path prediction knowledge is iteratively optimized, including: Integrating the set of enhanced vectors and the target graph embedding vector to generate an integrated vector sequence for this round of fault path prediction; According to a first importance coefficient defined for the integrated vector sequence and a second importance coefficient defined for the priori fault path prediction knowledge, the integrated vector sequence is merged with the priori fault path prediction knowledge to generate the iteratively optimized priori fault path prediction knowledge; The first importance coefficient and the second importance coefficient are learned in a network parameter optimization phase.

9. The method for analyzing data of power Internet of Things devices based on artificial intelligence according to any one of claims 1 to 3, characterized in that: Perform graph convolution processing on the operation event data of each device included in the target power IoT device log to generate a corresponding graph convolution vector set, including: Performing device state migration node parsing on the target electric power IoT device log to generate a corresponding device state migration node set, wherein each device state migration node in the device state migration node set corresponds to a device operation event data; Performing a serialized graph convolution process on the device state transition node set until a graph convolution vector corresponding to each device state transition node is generated. In each round of graph convolution process, if the device operation event data of this round is the first device operation event data, then generating a graph convolution vector of the device operation event data of this round according to the device state transition node of the device operation event data of this round and a defined initialization vector; If the device operation event data of this round is not the first device operation event data, then the graph convolution vector of the device operation event data of this round is generated according to the graph convolution vector extracted in the previous round and the device state transition node of the device operation event data of this round; The graph convolution vector set is generated according to each generated graph convolution vector.

10. An artificial intelligence-based power Internet of Things equipment data analysis system, characterized in that: The artificial intelligence-based power Internet of Things device data analysis system includes a processor and a memory, the memory is connected to the processor, the memory is used to store programs, instructions or codes, and the processor is used to execute the programs, instructions or codes in the memory to implement the artificial intelligence-based power Internet of Things device data analysis method described in any one of claims 1 to 9 above.

Images