Substation equipment operation detection method and device

Abstract

The invention belongs to the technical field of transformer substation intelligent operation inspection, and particularly relates to a transformer substation equipment operation detection method and device. According to the method, knowledge maps of different time periods with equipment as nodes are established based on structured data of each time period, a knowledge map encoder is constructed and extracted based on an attention mechanism and a GRU model, and the knowledge map encoder sets corresponding processing units according to the number of the time periods and is used for processing knowledge map data in the corresponding time periods. Each processing unit comprises a GRU model and two layers of attention mechanisms and is used for processing data in a corresponding time period, and text feature vectors which carry time information and are extracted from equipment maintenance records are introduced into the attention mechanisms of the processing units; the structural data of the equipment and the unstructured data of the corresponding time are fused and then feature extraction of the structured data is performed, so that the structured feature vector can be obtained more accurately, and the accuracy of equipment fault detection in operation and inspection of the transformer substation is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent operation and inspection of substations, and in particular relates to a method and device for detecting operation of substation equipment. Background Art

[0002] There are a large number of communication devices in smart substations and the power network structure is complex. The problem of being unable to quickly and accurately locate equipment failures is becoming increasingly prominent. In operation detection, deep learning models are generally used to analyze equipment data to achieve automatic identification and early warning of equipment failures. However, the data type based on deep learning models is often relatively single, and structured fault data is very scarce compared to normal operation data, which easily leads to the prediction results of the model being biased towards the normal state, making it difficult to capture the complex relationship between devices, resulting in high errors, and large deep learning models may not meet the requirements of low latency. To this end, the prior art proposes to establish a knowledge graph containing structured data and unstructured data, where structured data is information such as equipment sensor data, and unstructured data includes maintenance records and image information, etc., and then feature extraction is performed on structured data and unstructured data respectively, and the extracted feature vector data is fused to obtain equipment detection results to achieve intelligent operation and inspection. Although this method can improve the efficiency of fault location and solve the problem of high errors caused by single data, the structured feature vector used in the fusion of each feature vector is only extracted from the structured data of the device, and the dependency and association information between the device and the unstructured data are not considered, resulting in a low accuracy rate of fault detection. Summary of the invention

[0003] The object of the present invention is to provide a substation equipment operation detection method and device to solve the problem of low detection accuracy of existing substation equipment fault detection methods.

[0004] In order to solve the above technical problems, the present invention provides a substation equipment operation detection method, comprising: Acquire structured and unstructured data of equipment. Structured data includes equipment sensor data and equipment operation and maintenance data. Unstructured data includes text data consisting of equipment maintenance records. Using the structured data of each time period, a knowledge graph of substation structured data of the corresponding time period is established with the equipment as the node; Extracting feature vectors of text data as feature vectors of unstructured data; The obtained substation structured data knowledge graph of each time period is input into the constructed knowledge graph encoder for feature extraction to obtain an attention feature vector; the knowledge graph encoder includes N processing units, N is the number of time periods of the input substation structured data knowledge graph; each processing unit includes a GRU model and a two-layer attention mechanism, wherein the first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, the second-layer attention mechanism is used to process the feature vector of the text data and the output result of the first-layer attention mechanism, the GRU model is used to process according to the output result of the GRU model in the processing unit corresponding to the previous time period and the output results of each layer of the attention mechanism in the processing unit, and the feature vector output by the GRU model corresponding to the last time period is the feature vector of the structured data; The detection result of the device is obtained based on the feature vectors of the extracted structured data and unstructured data.

[0005] Furthermore, the first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, and output the results of each device in the corresponding time period after weighting by the importance and time of different fault types.

[0006] Furthermore, the unstructured data also includes image data, and a feature vector of the image data is extracted as the feature vector of the unstructured data; the image data includes at least one of a substation image, a device appearance picture, and a device surrounding environment.

[0007] Furthermore, the operation and maintenance data includes at least one of start / stop status data, abnormal data and maintenance data.

[0008] Furthermore, the detection result of the equipment is obtained by inputting the feature vectors of the extracted structured data and unstructured data into a multimodal model; the multimodal model includes at least 2 perception units and a classification output layer, each perception unit includes a fully connected layer, a batch normalization layer and an activation function layer, and after the feature vectors of each structured data and unstructured data are input into the perception unit for processing, the fault type of the equipment is output through the classification output layer.

[0009] Furthermore, a Dropout layer is provided between the perception units to prevent overfitting.

[0010] Furthermore, the method also includes scoring the degree of fault for the obtained fault type; the method for scoring the degree of fault is to determine the feature set of the device under the fault type according to the fault type of the device, perform a similarity comparison between the current classification features obtained by the classification output layer and the feature set of the device under the fault type, and determine the degree of fault score based on the similarity comparison result.

[0011] Furthermore, the feature vector of the text data is extracted using a text encoder consisting of Transformer blocks.

[0012] Furthermore, the feature vector of the image data is extracted using an image encoder, and the image encoder is a YOLOv11 model.

[0013] The beneficial effects of the above technical solution are as follows: the present invention is a pioneering invention, which establishes a knowledge graph of different time periods with equipment as nodes based on the structured data of each time period, and constructs an extracted knowledge graph encoder based on the attention mechanism and the GRU model. The knowledge graph encoder sets corresponding processing units according to the number of time periods to process the knowledge graph data in the corresponding time period. The processing units include a GRU model and a two-layer attention mechanism to process the data in the corresponding time period, wherein the attention mechanism of the processing unit introduces a feature vector with time information extracted from the maintenance record, and the structural data of the equipment and the unstructured data of the corresponding time are fused to extract the features of the structured data, so that the correlation and time correlation of each modal data of the equipment are also considered when extracting the structured features, so that the structured feature vector can be obtained more accurately, thereby improving the accuracy of equipment fault detection in substation operation and inspection.

[0014] In order to solve the above technical problems, the present invention also provides a substation equipment operation detection device, including a processor, wherein the processor is used to execute computer program instructions to implement the steps in the substation equipment operation detection method introduced above.

[0015] The beneficial effects of the above technical solution are as follows: the present invention is a pioneering invention, which establishes a knowledge graph of different time periods with equipment as nodes based on the structured data of each time period, and constructs an extracted knowledge graph encoder based on the attention mechanism and the GRU model. The knowledge graph encoder sets corresponding processing units according to the number of time periods to process the knowledge graph data in the corresponding time period. The processing units include a GRU model and a two-layer attention mechanism to process the data in the corresponding time period, wherein the attention mechanism of the processing unit introduces a feature vector with time information extracted from the maintenance record, and the structural data of the equipment and the unstructured data of the corresponding time are fused to extract the features of the structured data, so that the correlation and time correlation of each modal data of the equipment are also considered when extracting the structured features, so that the structured feature vector can be obtained more accurately, thereby improving the accuracy of equipment fault detection in substation operation and inspection. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a schematic diagram of the device structure of an embodiment of the device of the present invention; Figure 2 It is a flow chart of equipment operation detection of an embodiment of the device of the present invention; Figure 3 It is a software architecture diagram of a detection device of an embodiment of the device of the present invention; Figure 4 is a structural diagram of a text encoder of an embodiment of the device of the present invention; Figure 5 is a structural diagram of an image encoder according to an embodiment of the device of the present invention; Figure 6 is a structural diagram of a knowledge graph encoder of an embodiment of the device of the present invention; Figure 7 is a multimodal model structure diagram of an embodiment of the device of the present invention; Figure 8 It is a schematic diagram of the knowledge graph of an embodiment of the device of the present invention. DETAILED DESCRIPTION

[0017] In order to make the purpose, technical solutions and advantages of the present invention more clear, the specific implementation methods of the present invention are further described below in conjunction with the accompanying drawings.

[0018] The present invention establishes knowledge graphs of different time periods with devices as nodes based on structured data of each time period, constructs an extracted knowledge graph encoder based on the attention mechanism and the GRU model, and the knowledge graph encoder sets corresponding processing units according to the number of time periods to process the knowledge graph data in the corresponding time period. The processing units all include a GRU model and a two-layer attention mechanism to process the data in the corresponding time period, wherein the attention mechanism of the processing unit introduces a feature vector with time information extracted from the maintenance record, and the structured data of the equipment and the unstructured data of the corresponding time are fused to extract features of the structured data, so that the correlation and time correlation of each modal data of the equipment are also considered when extracting the structured features, so that the structured feature vector can be obtained more accurately, thereby improving the accuracy of equipment fault detection in substation operation and inspection.

[0019] Device Embodiment The present invention provides a substation equipment operation detection device, the overall structure of which is as follows Figure 1 As shown, it includes a processor, which is set in the operation and maintenance control platform of the substation and is used to execute computer program instructions to implement a substation equipment operation detection method. The specific process of the method is as follows Figure 2 As shown, the detailed description is given below.

[0020] 1. Obtain the structured data and unstructured data of the equipment. The structured data includes equipment sensor data and equipment operation and maintenance data. The unstructured data includes text data, and the text data includes equipment maintenance records.

[0021] Specifically, the operation and maintenance personnel deploy a variety of sensors on the equipment of the substation, such as temperature sensors, humidity sensors, vibration sensors, current transformers and voltage transformers, etc. The above sensors are connected to the substation operation and maintenance management platform, which collects and records the data collected by the sensors in real time and stores them in the database. The operation and maintenance data includes at least one of the start and stop status data, abnormal data and maintenance data. The operation and maintenance management system also collects information such as the start and stop status, abnormalities and maintenance records of electrical equipment to form a structured database.

[0022] The operation and maintenance personnel collect the maintenance records of substation equipment every day and upload them to the substation database in text file format. In order to further improve the detection accuracy, the unstructured data also includes image data, including at least one of the substation image, equipment appearance picture and equipment surrounding environment. The operation and maintenance personnel deploy multiple drones in the substation and install multiple high-definition cameras around the substation equipment to capture the overall layout of the substation, the appearance of the equipment and the surrounding environment of the equipment. The unstructured database is generated based on the image data and equipment maintenance records.

[0023] 2. Preprocess the data by denoising and normalizing.

[0024] Among them, the preprocessing of structured data includes data cleaning, data conversion and data integration. Data cleaning includes deleting duplicate records, processing missing values ​​and correcting erroneous data. Data conversion includes numerical standardization and category coding. Data integration is used to solve the inconsistency problem of data from different sources, including unifying data formats and units, and merging related data from different tables.

[0025] The preprocessing of text data in unstructured data includes the removal of stop words, correction of spelling errors and format correction. Specifically, it removes irrelevant information in the text, removes stop words that are common in the language but have little significance for analysis, corrects spelling errors in the text, and adjusts the data to the correct format, such as date and time formatting, numerical type formatting, etc., and finally uploads it to the substation database.

[0026] The preprocessing of image data in unstructured data includes image standardization and data enhancement. First, blurry, poorly exposed or damaged images are removed and converted to standard size and format; then the image richness is increased through data enhancement technology; finally, basic features are extracted using image processing technology to ensure the quality and consistency of image data before uploading to the substation database.

[0027] 3. Utilize the structured data of each time period and establish a knowledge graph of substation structured data for the corresponding time period with equipment as nodes.

[0028] According to the structured data of substation equipment in each time period, the substation equipment list is counted, the equipment in the knowledge graph is defined as an entity, the connection relationship between equipment and equipment, the association relationship between equipment and faults, etc. are collected, data is extracted according to the entity relationship, and the knowledge representation result is obtained. The knowledge graph of structured data of substations in each time period is constructed with equipment as a node, such as Figure 8 shown.

[0029] 4. Extract feature vectors.

[0030] like Figure 3 As shown, the feature extraction model includes a knowledge graph encoder, a text encoder, and an image encoder.

[0031] The text encoder is used to extract the feature vector of text data in unstructured data. The text encoder takes the preprocessed substation equipment maintenance record as text input and converts the input text into a text feature vector for output. The text encoder consists of Transformer blocks. Figure 4 As shown, in this embodiment, the text encoder consists of 6 layers of Transformer blocks, each layer contains a multi-head self-attention and a feedforward neural network, and also includes normalization and residual connection layers for preventing gradient disappearance.

[0032] The image encoder is used to extract feature vectors of image data from unstructured data. The image encoder may use a YOLO model or a graph neural network model. Preferably, in order to improve the efficiency of image feature extraction, the image encoder uses a YOLOv11 model, such as Figure 5 As shown in the figure, the preprocessed image data is input into the YOLOv11 model and the image feature vector is output.

[0033] The knowledge graph encoder is used to extract the feature vector of structured data. The knowledge graph of structured data of the substation in each time period is input into the constructed knowledge graph encoder for feature extraction to obtain the feature vector of structured data, that is, the attention feature vector.

[0034] In order to better capture the attention feature vector, the knowledge graph encoder combines the attention mechanism with the gated recurrent unit to form a temporal attention module to extract the structured features of devices in different time periods. Figure 6As shown, the knowledge graph encoder includes N processing units, where N is the number of time periods of the input substation structured data knowledge graph; each processing unit includes a GRU model and a two-layer attention mechanism, wherein the first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, and the second-layer attention mechanism is used to perform attention calculation processing on the text feature vector and the output result of the first-layer attention mechanism, and the GRU model is used to process according to the output result of the GRU model in the processing unit corresponding to the previous time period and the output result of each layer of the attention mechanism in the processing unit, and the GRU model corresponding to the last time period outputs the feature vector of the structured data, which is recorded as the attention feature vector; wherein the first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, and output the result of each device in the corresponding time period after being weighted by the importance and time of different fault types.

[0035] Assume that the time period of the input knowledge graph is T1~T N , then T1~T N The text feature vector corresponding to the equipment maintenance record within the time range is input into the knowledge graph encoder. The text feature vector carries time information. The processing unit automatically filters out the text feature vector belonging to the corresponding time period and introduces the selected text feature vector into the second-layer attention mechanism for processing. When there is no equipment maintenance record in all time periods, the input text feature vector is None.

[0036] 5. Obtain the detection results of the equipment based on the feature vectors of the extracted structured data and unstructured data.

[0037] The feature vectors of different modes extracted previously are fused at the decision level to generate multimodal features, and the detection result of the device is obtained based on the generated multimodal features. The present invention obtains the detection result of the device by inputting the feature vectors of the extracted structured data and unstructured data into the multimodal model. Figure 7 As shown in the figure, the multimodal model refers to the structure of MLP (multi-layer perceptron) to independently extract text feature vectors, image feature vectors and attention feature vectors. The model includes at least 2 perception units and a classification output layer. Each perception unit includes a fully connected layer, a batch normalization layer and an activation function layer ( Figure 7 The parameterized ReLU in reduces the risk of neurons becoming inactive and helps the model better adapt to the data). After each feature vector is input to the perception unit for processing, it passes through the classification output layer ( Figure 7 The SoftMax layer in the output device detection results, including the fault type and whether it is a fault. In order to prevent overfitting, a Dropout layer is also set between the perception units.

[0038] The model also assigns a fault severity score F to the obtained fault type. The fault severity score method is to determine the feature set of the equipment under the fault type according to the fault type of the equipment, compare the current classification features obtained by the classification output layer with the feature set of the equipment under the fault type, and determine the fault severity score based on the similarity comparison result, so as to know the equipment operation status. The operation and maintenance control platform will display the obtained fault severity score, and the staff will judge the degree of equipment failure through the information displayed on the operation and maintenance control platform, so as to reasonably arrange maintenance time and resources and avoid idle or trampling operation and maintenance resources.

[0039] Method Embodiment The specific implementation process of the substation equipment operation detection method of the present invention has been described in detail in the device embodiment and will not be repeated here.

[0040] The present invention extracts feature vectors from the knowledge graph of each time period through the designed knowledge graph encoder, thereby improving the accuracy of equipment fault detection in substation operation and inspection. By integrating multimodal data from multiple data sources of the substation to judge the operating status of substation equipment, the type and degree of equipment fault can be judged more comprehensively and accurately. The operation detection method based on knowledge graph and multimodal technology can quickly retrieve and associate relevant information of substation equipment, including equipment model, historical fault records, maintenance records, etc., combined with efficient image processing methods to achieve rapid positioning of faults.

Claims

1. A substation equipment operation detection method, characterized in that: include: Acquire structured and unstructured data of equipment. Structured data includes equipment sensor data and equipment operation and maintenance data. Unstructured data includes text data consisting of equipment maintenance records. Using the structured data of each time period, a knowledge graph of substation structured data of the corresponding time period is established with the equipment as the node; Extracting feature vectors of text data as feature vectors of unstructured data; The obtained substation structured data knowledge graph of each time period is input into the constructed knowledge graph encoder for feature extraction to obtain an attention feature vector; the knowledge graph encoder includes N processing units, N is the number of time periods of the input substation structured data knowledge graph; each processing unit includes a GRU model and a two-layer attention mechanism, wherein the first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, the second-layer attention mechanism is used to process the feature vector of the text data and the output result of the first-layer attention mechanism, the GRU model is used to process according to the output result of the GRU model in the processing unit corresponding to the previous time period and the output results of each layer of the attention mechanism in the processing unit, and the feature vector output by the GRU model corresponding to the last time period is the feature vector of the structured data; The detection result of the device is obtained based on the feature vectors of the extracted structured data and unstructured data.

2. The substation equipment operation detection method according to claim 1, characterized in that: The first-layer attention mechanism is used to process the input substation structured data knowledge graph of the corresponding time period, and output the results of each device in the corresponding time period after weighting by the importance and time of different fault types.

3. The substation equipment operation detection method according to claim 1, characterized in that: The unstructured data also includes image data, and a feature vector of the image data is extracted as the feature vector of the unstructured data; the image data includes at least one of a substation image, a device appearance picture, and a device surrounding environment.

4. The substation equipment operation detection method according to claim 1, characterized in that: The operation and maintenance data includes at least one of start / stop status data, abnormal data and maintenance data.

5. The substation equipment operation detection method according to claim 3, characterized in that: The detection result of the equipment is obtained by inputting the feature vectors of the extracted structured data and unstructured data into a multimodal model; the multimodal model includes at least two perception units and a classification output layer, each perception unit includes a fully connected layer, a batch normalization layer and an activation function layer, and after the feature vectors of each structured data and unstructured data are input into the perception unit for processing, the fault type of the equipment is output through the classification output layer.

6. The substation equipment operation detection method according to claim 5, characterized in that: A Dropout layer is also provided between the perception units to prevent overfitting.

7. The substation equipment operation detection method according to claim 5 or 6, characterized in that: The method also includes scoring the degree of fault for the obtained fault type; the method for scoring the degree of fault is to determine the feature set of the device under the fault type according to the fault type of the device, perform a similarity comparison between the current classification features obtained by the classification output layer and the feature set of the device under the fault type, and determine the degree of fault score based on the similarity comparison result.

8. The substation equipment operation detection method according to claim 1, characterized in that: The feature vectors of text data are extracted using a text encoder consisting of Transformer blocks.

9. The substation equipment operation detection method according to claim 3, characterized in that: The feature vector of the image data is extracted using the image encoder, which is the YOLOv11 model.

10. A substation equipment operation detection device, comprising a processor, characterized in that: The processor is used to execute computer program instructions to implement the steps in the substation equipment operation detection method according to any one of claims 1 to 9.

Images