Method for detecting abnormal operation state of electric power metering equipment

Abstract

The invention relates to the technical field of equipment anomaly detection, in particular to an electric power metering equipment abnormal operation state detection method, which comprises the steps that operation state detection information of electric power metering equipment uploaded by a worker is acquired, and the operation state detection information comprises an operation state image and an operation state detection text; and performing state detection on the running state image based on a pre-trained state detection model to obtain a running state detection result of the running state image, and calculating the similarity between the running state detection text and the running state detection result. According to the method, image detection and text analysis are combined, the state of the electric power metering equipment can be effectively monitored in a multi-level and all-around mode, the running state image and the detection text are combined, misjudgment or missed judgment caused by insufficient information of a single source is reduced, and when image state detection is not accurate, the detection accuracy of the electric power metering equipment is improved. And further accuracy guarantee is provided through monitoring of the sensor group.

Description

Technical Field

[0001] The present invention relates to the technical field of device anomaly detection, and particularly to a method for detecting the abnormal operation state of power metering devices. Background Art

[0002] Power metering devices are instruments used to measure and record power consumption, and they are commonly used in power systems to ensure accurate metering of power usage.

[0003] Traditional methods usually rely on a single detection method, such as only performing fault detection through sensor data or image recognition. This may lead to misjudgment or missed judgment caused by inaccurate or incomplete data from a certain data source, reducing the accuracy of fault identification. Moreover, traditional methods often adopt a fixed monitoring mode and it is difficult to dynamically adjust the detection strategy to adapt to the operation changes of the device at different time periods. The performance of the device may be different under different working states, and traditional methods cannot handle these changes well. In addition, traditional methods rely on manual state analysis and anomaly detection. Manual inspection not only has low efficiency but is also easily limited by the work intensity and energy of people, which may lead to judgment errors or detection delays. And traditional methods have weak potential fault prediction ability for devices, often responding only after a fault occurs and unable to achieve preventive maintenance. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to overcome the above-mentioned disadvantages of the prior art and provide a method for detecting the abnormal operation state of power metering devices.

[0005] The technical solution adopted to solve the above technical problem is: A method for detecting the abnormal operation state of power metering devices, comprising:

[0006] Obtaining the operation state detection information of the power metering device uploaded by the staff, wherein the operation state detection information includes an operation state image and an operation state detection text;

[0007] Performing state detection on the operation state image based on a pre-trained state detection model to obtain the operation state detection result of the operation state image, and calculating the similarity between the operation state detection text and the operation state detection result;

[0008] Judging whether the operation state detection information is accurate based on the similarity. If the operation state detection information is inaccurate, performing state monitoring on the power metering device based on a sensor group to obtain the state data sequence of the power metering device;

[0009] Sample the state data sequence based on a preset sliding window to obtain multiple time windows, and perform anomaly detection on the multiple time windows based on a pre-trained anomaly detection model to obtain the anomaly score of the power metering device;

[0010] Compare the anomaly score of the power metering device with a preset score threshold to obtain the operation status detection label of the power metering device.

[0011] Preferably, the state detection model uses an improved YOLOv3 network, and improves the backbone network of the YOLOv3 network based on a hybrid attention model.

[0012] Preferably, the attention module includes a coordinate information unit, a channel information unit, and a spatial attention unit. Among them, the expression of the coordinate information unit is as follows:

[0013]

[0014] Among them, represents the feature map output by the coordinate information unit, Conv represents the convolution operation, Concat represents the channel concatenation operation, X i1 (i, j) represents the first input feature map, x coord and y coord represent the coordinate information of the X and Y axis channels in the coordinate information unit;

[0015] The expression of the channel information unit is as follows:

[0016]

[0017] Among them, represents the feature map output by the channel information unit, σ represents the Sigmoid activation function, ω1 and b1 represent the weight vector and bias vector of the channel information unit, H and W represent the height and width of the feature map;

[0018] The expression of the spatial attention unit is as follows:

[0019]

[0020] Among them, represents the feature map output by the spatial attention unit, GN represents group normalization, ω2 and b2 represent the weight vector and bias vector of the spatial attention unit, X i2 (i, j) represents the second input feature map.

[0021] Preferably, calculating the similarity between the operation status detection text and the operation status detection result includes:

[0022] Perform word segmentation on the operation state detection result and the operation state detection text to obtain a label word segmentation sequence corresponding to the operation state detection result and a problem word segmentation sequence corresponding to the operation state detection text;

[0023] Based on the Word2vec model, sequence the label word segmentation sequence and the problem word segmentation sequence to obtain a first word vector matrix of the label word segmentation sequence and a second word vector matrix of the problem word segmentation sequence;

[0024] Interact the first word vector matrix and the second word vector matrix to obtain a first interactive attention matrix corresponding to the first word vector matrix and a second interactive attention matrix corresponding to the second word vector matrix;

[0025] Concatenate the first word vector matrix and the first interactive attention matrix to obtain a first concatenated matrix, and concatenate the second word vector matrix and the second interactive attention matrix to obtain a second concatenated matrix.

[0026] Preferably, calculating the similarity between the operation state detection text and the operation state detection result further includes:

[0027] Input the first concatenated matrix and the second concatenated matrix into the Transformer model respectively, and output a first text feature of the first concatenated matrix and a second text feature of the second concatenated matrix based on the Transformer model;

[0028] Based on a fully connected layer, one-dimensionalize the first text feature and the second text feature to obtain a first semantic feature and a second semantic feature;

[0029] Calculate the difference and product of the first semantic feature and the second semantic feature, and concatenate the difference and the product to obtain a fusion feature;

[0030] Process the fusion feature based on a two-layer fully connected network to obtain the similarity between the operation state detection result and the operation state detection text, where the first layer of the fully connected network uses the ReLU activation function and the second layer of the connected network uses the Softmax normalization function.

[0031] Preferably, the state data includes current data, voltage data, vibration data and temperature data, and the expressions of the multiple time windows are as follows:

[0032] S = {S1, S2,..., S t ,..., S n};

[0033] Among them, S represents multiple time windows, and S t represents the t-th time window among the multiple time windows, and S t = [x t-w , x t-w+1 ,..., x t-1 , and x t-1 represents the state data at the (t - 1)-th moment.

[0034] Preferably, the anomaly detection model includes a data reconstruction module, a graph prediction module, and an anomaly scoring module. Among them, the data reconstruction module includes an encoder and a decoder. Among them, the encoder is used to calculate the mean, standard deviation, and Gaussian distribution of the multiple time windows, and the decoder is used to input the Gaussian distribution into the encoder again to obtain a reconstruction window and map the reconstruction window to the reconstructed multivariate time series data. The graph prediction module is used to fuse the reconstructed multivariate time series data with multiple time windows into an embedding vector, generate a time series state graph based on the embedding vector, and predict the mean, standard deviation, and Gaussian distribution of the sensor group at future times based on the time series state graph according to a graph neural network. The anomaly scoring module is used to generate an anomaly score of the power metering device based on the mean, standard deviation, and Gaussian distribution of the sensor group at future times.

[0035] Preferably, the calculation formulas for the mean, standard deviation, and Gaussian distribution of the multiple time windows are as follows:

[0036]

[0037] Among them, μ(t), σ(t), and Z(t) represent the mean, standard deviation, and Gaussian distribution of the multiple time windows, W μ and W σ represent weight matrices for linear transformation, LSTM represents a long short-term memory network, Reshape represents a reshaping function, b μ and b σ represent bias terms, and ∈ represents a preset weight coefficient;

[0038] The expression of the reconstruction window is as follows:

[0039]

[0040] Among them, represents the t-th reconstruction window.

[0041] Preferably, generating a time series state graph based on the embedding vector and predicting the mean, standard deviation, and Gaussian distribution of the sensor group at future times based on the time series state graph according to a graph neural network includes:

[0042] Regarding each sensor in the sensor group as a graph node in the time series state graph;

[0043] Calculate the cosine similarity between each graph node in the temporal state graph based on the embedding vectors. If the cosine similarity is greater than a preset similarity threshold, connect the graph nodes to obtain an edge between the two graph nodes.

[0044] Calculate the correlation coefficient between each graph node in the temporal state graph. The calculation formula for the correlation coefficient is as follows:

[0045]

[0046] where α i，j represents the correlation coefficient between graph node v i and graph node v j , a represents the parameter vector to be learned, LeakyReLU represents the activation function, and represent the features of graph node v i and graph node v j at time t, N(i) represents the neighbor graph nodes of graph node v i .

[0047] Perform graph node aggregation and update on each graph node in the temporal state graph to obtain node aggregation features. The formula for graph node aggregation and update is as follows:

[0048]

[0049] where α i，j represents the node aggregation feature of graph node v i at time t, W o represents the weight matrix of the linear transformation for graph node aggregation and update, represents the concatenated feature of graph node v i at time t, and represents the state data of graph node v i at time t, represents the Gaussian distribution of graph node v i at time t;

[0050] Predict the mean, standard deviation, and Gaussian distribution of the sensor group at future times based on the node aggregation features of each node in the temporal state graph. The calculation formulas for the mean, standard deviation, and Gaussian distribution of the sensor group at future times are as follows:

[0051]

[0052] where, and represent graph node vi The mean, standard deviation, and Gaussian distribution at time t are the mean, standard deviation, and Gaussian distribution of the i-th sensor in the sensor group at time t. γ and δ represent the parameters to be learned, and f Φ represents a stacked fully connected layer.

[0053] Preferably, the calculation formula for the anomaly score of the power metering device is as follows:

[0054]

[0055] Among them, L abn represents the anomaly score of the power metering device, and N represents the number of sensor groups.

[0056] The beneficial effects of the present invention are as follows: (1) By combining image detection and text analysis, the method of the present invention can effectively monitor the status of power metering devices in a multi-level and all-round manner. The combination of operation status images and detection texts reduces misjudgment or missed judgment caused by insufficient information from a single source. When the image status detection is inaccurate, the monitoring by the sensor group provides further accuracy assurance. By comprehensively considering various detection means, the ability to identify equipment failures can be greatly improved; (2) The sliding window sampling technology of the present invention enables the method to dynamically adjust the status data sequence of power metering devices, adapt to the operation changes of the devices in different time periods. By performing anomaly detection on multiple time windows, it ensures that the monitoring of the device status is not affected by time fluctuations. And based on the preset anomaly detection model, the method can automatically identify abnormal states without manual intervention. This automation ability not only improves the detection efficiency but also significantly reduces the workload of manual inspection; (3) By accurately judging the operation status of the device, potential failures can be discovered in a timely manner, avoiding the device being in an abnormal operation state for a long time, reducing downtime and maintenance costs. And by long-term monitoring of the data of power metering devices, potential failure modes can be identified, potential failures of the device can be predicted in advance, and preventive measures can be taken to avoid sudden failures. Through intelligent analysis of different types of status data (images, texts, sensor data), the effect of cross-domain integration is achieved, promoting the intelligent process of power metering device detection. Description of the Drawings

[0057] Figure 1 It is a schematic diagram of the step flow of the overall method in an embodiment proposed by the present invention. Detailed Embodiments

[0058] Embodiment 1, as Figure 1 shown, a method for detecting abnormal operation status of a power metering device proposed by the present invention includes:

[0059] S1. Obtain the operation status detection information of the power metering device uploaded by the staff, where the operation status detection information includes the operation status image and the operation status detection text;

[0060] S2. Perform status detection on the operation status image based on a pre-trained status detection model to obtain the operation status detection result of the operation status image, and calculate the similarity between the operation status detection text and the operation status detection result;

[0061] S3. Judge whether the operation status detection information is accurate based on the similarity. If the operation status detection information is inaccurate, perform status monitoring on the power metering device based on the sensor group to obtain the status data sequence of the power metering device;

[0062] S4. Sample the status data sequence based on a preset sliding window to obtain multiple time windows, and perform anomaly detection on the multiple time windows based on a pre-trained anomaly detection model to obtain the anomaly score of the power metering device;

[0063] S5. Compare the anomaly score of the power metering device with a preset score threshold to obtain the operation status detection label of the power metering device.

[0064] In the present invention, the operation status image refers to an image captured by a camera, sensor or other monitoring device of the power metering device, which is used to reflect the working status of the power metering device, and can be the appearance image of the device or the key picture during the working process; the operation status detection text refers to the text information describing the current operation status of the power metering device, usually provided by the staff or an automated system, and may include whether the device is normal, the type of fault, the operation mode, etc.; the sensor group refers to multiple sensors used to detect and monitor the operation status of the device, and these sensors may collect various data such as temperature, voltage, current, vibration, etc. for real-time monitoring of the device status; the sliding window is a data processing technology, usually used to segment time series data. In this case, the sliding window refers to intercepting data of multiple time periods from the status data sequence at a certain time interval, and each time period is a "window". Through such processing, the data can be analyzed step by step to capture the subtle changes during the operation of the device.

[0065] Embodiment 2. A method for detecting the abnormal operation status of a power metering device proposed by the present invention. Compared with Embodiment 1, this embodiment further includes: the status detection model adopts an improved YOLOv3 network, and the backbone network of the YOLOv3 network is improved based on a hybrid attention model.

[0066] It should be noted that YOLO is a popular object detection algorithm designed to quickly detect multiple objects in an image through a single forward pass. YOLOv3 is the third version of the YOLO series of algorithms. It processes images through a convolutional neural network (CNN), capable of improving the detection speed while maintaining a relatively high detection accuracy. The attention mechanism is a technique widely used in deep learning, aiming to enable the model to focus on more important parts when processing information, thereby enhancing the model's performance. The hybrid attention model is a model that combines multiple attention mechanisms, usually including channel attention, spatial attention, etc.

[0067] In an optional embodiment, the attention module includes a coordinate information unit, a channel information unit, and a spatial attention unit. Among them, the expression of the coordinate information unit is as follows:

[0068]

[0069] Among them, represents the feature map output by the coordinate information unit, Conv represents the convolution operation, Concat represents the channel concatenation operation, X i1 (i, j) represents the first input feature map, x coord and y coord represent the coordinate information of the X and Y axis channels in the coordinate information unit;

[0070] The expression of the channel information unit is as follows:

[0071]

[0072] Among them, represents the feature map output by the channel information unit, σ represents the Sigmoid activation function, ω1 and b1 represent the weight vector and bias vector of the channel information unit, H and W represent the height and width of the feature map;

[0073] The expression of the spatial attention unit is as follows:

[0074]

[0075] Among them, represents the feature map output by the spatial attention unit, GN represents group normalization, ω2 and b2 represent the weight vector and bias vector of the spatial attention unit, X i2 (i, j) represents the second input feature map.

[0076] In an optional embodiment, calculating the similarity between the running state detection text and the running state detection result includes:

[0077] A1. Perform word segmentation on the operation status detection result and the operation status detection text to obtain the labeled word segmentation sequence corresponding to the operation status detection result and the problem word segmentation sequence corresponding to the operation status detection text;

[0078] A2. Align the labeled word segmentation sequence and the problem word segmentation sequence based on the Word2vec model to obtain the first word vector matrix of the labeled word segmentation sequence and the second word vector matrix of the problem word segmentation sequence;

[0079] A3. Interact the first word vector matrix and the second word vector matrix to obtain the first interactive attention matrix corresponding to the first word vector matrix and the second interactive attention matrix corresponding to the second word vector matrix;

[0080] A4. Concatenate the first word vector matrix and the first interactive attention matrix to obtain the first concatenated matrix, and concatenate the second word vector matrix and the second interactive attention matrix to obtain the second concatenated matrix.

[0081] It should be noted that word segmentation is a basic technique in natural language processing for splitting text strings into individual words or phrases; Word2Vec is a popular word embedding model for converting words or phrases into vectors. Word2Vec learns the semantic relationships of words in the context by training a large-scale text dataset, enabling word vectors to capture the similarity between words; the interactive attention matrix is a matrix generated by calculating the similarity or correlation between two word vector matrices. Each element in the matrix represents the influence or attention of one word vector on another, which helps the model to weight the relationships between two input sequences during processing; matrix concatenation refers to combining multiple matrices into a new matrix according to certain rules (usually by columns or rows). The concatenation operation combines the word vector matrix and the interactive attention matrix, thereby enhancing the model's learning ability for this information.

[0082] In an optional embodiment, calculating the similarity between the operation status detection text and the operation status detection result further includes:

[0083] A5. Input the first concatenated matrix and the second concatenated matrix into the Transformer model respectively, and output the first text feature of the first concatenated matrix and the second text feature of the second concatenated matrix based on the Transformer model;

[0084] A6. One-dimensionalize the first text feature and the second text feature based on the fully connected layer to obtain the first semantic feature and the second semantic feature;

[0085] A7. Calculate the difference and product of the first semantic feature and the second semantic feature, and splice the difference and the product to obtain a fusion feature;

[0086] Process the fusion feature based on a two-layer fully connected network to obtain the similarity between the operating state detection result and the operating state detection text. Among them, the first-layer fully connected network uses the ReLU activation function, and the second-layer connection network uses the Softmax normalization function.

[0087] It should be noted that Transformer is a deep learning architecture widely used in natural language processing tasks (such as machine translation, text generation, etc.).

[0088] In an alternative embodiment, the state data includes current data, voltage data, vibration data, and temperature data. The expressions of multiple time windows are as follows:

[0089] S = {S1, S2,..., S t ,..., S n};

[0090] Among them, S represents multiple time windows, and S t represents the t-th time window among multiple time windows, and S t = [x t-w , x t-w+1 ,..., x t-1 , and x t-1 represents the state data at the (t - 1)-th moment.

[0091] In an alternative embodiment, the anomaly detection model includes a data reconstruction module, a graph prediction module, and an anomaly scoring module. Among them, the data reconstruction module includes an encoder and a decoder. The encoder is used to calculate the mean, standard deviation, and Gaussian distribution of multiple time windows. The decoder is used to input the Gaussian distribution into the encoder again to obtain a reconstructed window, and map the reconstructed window to the reconstructed multivariate time series data. The graph prediction module is used to fuse the reconstructed multivariate time series data with multiple time windows into an embedding vector, generate a time series state graph based on the embedding vector, and predict the mean, standard deviation, and Gaussian distribution of the sensor group at future times based on the time series state graph according to the graph neural network. The anomaly scoring module is used to generate an anomaly score of the power metering device based on the mean, standard deviation, and Gaussian distribution of the sensor group at future times.

[0092] It should be noted that the goal of the data reconstruction module is to restore or reconstruct the original input information through the process of encoding and decoding the input data. In the time series anomaly detection task, through data reconstruction, abnormal data that does not conform to the normal pattern can be identified; the role of the encoder is to transform the original input data into a low-dimensional representation or "encoding". The encoder extracts features from the input data by calculating the mean, standard deviation, and Gaussian distribution of multiple time windows. These features can capture the statistical properties of the data and help the model better understand the overall structure and distribution of the time series data; the role of the decoder is to transform the output information of the encoder (such as the parameters of the Gaussian distribution) back into the form of the original data; multivariate time series data represents a combination of multiple sensors or multiple time series, and each time point has multiple feature values. This data usually comes from multiple sensors and reflects the behavior of a system over time; the time series state diagram is a graphical representation based on time series data. In this diagram, nodes represent different time windows or different sensor groups, and edges represent the relationships between them. By constructing the time series state diagram, the time-varying characteristics of the time series data and their dependencies can be better understood; graph neural network is a type of neural network model used to process graph data, and it can learn the dependencies between nodes through the topological structure of the graph; anomaly score is an indicator used to evaluate whether a data point is an anomaly. The higher the score, the more likely the data point is an anomaly. In this model, the anomaly score is calculated based on the predicted future time series data distribution. Specifically, the model will calculate the deviation between the actual observed data and the predicted data according to the predicted mean, standard deviation, and Gaussian distribution, and use this deviation to score each data point. If the deviation of the data point is large, it means that it is very likely to be an anomaly.

[0093] In an optional embodiment, the calculation formulas for the mean, standard deviation, and Gaussian distribution of multiple time windows are as follows:

[0094]

[0095] Where, μ(t), σ(t), and Z(t) represent the mean, standard deviation, and Gaussian distribution of multiple time windows, W μ and W σ represent the weight matrices for linear transformation, LSTM represents the long short-term memory network, Reshape represents the reshaping function, b μ and b σ represent the bias terms, and ∈ represents the preset weight coefficient;

[0096] The expression of the reconstruction window is as follows:

[0097]

[0098] Where, Denote the t-th reconstruction window.

[0099] In an optional embodiment, a temporal state graph is generated based on the embedding vectors, and the mean, standard deviation, and Gaussian distribution of the sensor group at future times are predicted based on the temporal state graph according to a graph neural network, including:

[0100] B1. Take each sensor in the sensor group as a graph node in the temporal state graph;

[0101] B2. Calculate the cosine similarity between each pair of graph nodes in the temporal state graph. If the cosine similarity is greater than a preset similarity threshold, connect the graph nodes to obtain an edge between the two graph nodes;

[0102] B3. Calculate the correlation coefficient between each pair of graph nodes in the temporal state graph. The calculation formula of the correlation coefficient is as follows:

[0103]

[0104] where α i，j represents the correlation coefficient between graph node v i and graph node v j , a represents a parameter vector to be learned, LeakyReLU represents an activation function, and represent the features of graph node v i and graph node v j at time t, and N(i) represents the neighbor graph nodes of graph node v i ;

[0105] B4. Aggregate and update each graph node in the temporal state graph to obtain a node aggregation feature. The formula for graph node aggregation and update is as follows:

[0106]

[0107] where α i，j represents the node aggregation feature of graph node v i at time t, W o represents the weight matrix of the linear transformation for graph node aggregation and update, represents the concatenated feature of graph node v i at time t, and represents the state data of graph node v i at time t, represents the Gaussian distribution of graph node v i at time t;

[0108] B5. Predict the mean, standard deviation, and Gaussian distribution of the sensor group at future times based on the aggregated features of each node in the timing state diagram. The calculation formulas for the mean, standard deviation, and Gaussian distribution of the sensor group at future times are as follows:

[0109]

[0110] Among them, and represent the mean, standard deviation, and Gaussian distribution of the graph node v i at time t, which is the mean, standard deviation, and Gaussian distribution of the i-th sensor in the sensor group at time t. γ and δ represent the parameters to be learned, and f Φ represents the stacked fully connected layers.

[0111] It should be noted that the graph node aggregation update in the graph neural network refers to updating the features of a node by integrating the features of its neighbor nodes. In this model, the features of the graph nodes (such as the features of sensors) will be aggregated through the information of the neighbor nodes to generate new node features. This aggregation process helps the model capture the dependencies between nodes.

[0112] In an optional embodiment, the calculation formula for the anomaly score of the power metering device is as follows:

[0113]

[0114] Among them, L abn represents the anomaly score of the power metering device, and N represents the number of sensor groups.

[0115] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited thereto. Various changes can be made without departing from the spirit of the present invention within the scope of knowledge possessed by those skilled in the art.

Claims

1. A method for detecting abnormal operation status of electric power metering equipment, characterized in that: include: Obtain the operating status detection information of the power metering device uploaded by the staff, wherein the operating status detection information includes an operating status image and an operating status detection text: Performing state detection on the running state image based on a pre-trained state detection model to obtain a running state detection result of the running state image, and calculating the similarity between the running state detection text and the running state detection result; Based on the similarity, judging whether the running state detection information is accurate; if the running state detection information is inaccurate, performing state monitoring on the power metering device based on the sensor group to obtain a state data sequence of the power metering device; Sampling the state data sequence based on a preset sliding window to obtain a plurality of time windows, and performing anomaly detection on the plurality of time windows based on a pre-trained anomaly detection model to obtain an anomaly score of the power metering device; The abnormality score of the power metering device is compared with a preset score threshold to obtain an operation status detection label of the power metering device.

2. A method for detecting abnormal operation status of electric power metering equipment according to claim 1, characterized in that: The state detection model adopts an improved YOLOv3 network, and improves the backbone network of the YOLOv3 network based on a hybrid attention model.

3. A method for detecting abnormal operation status of electric power metering equipment according to claim 2, characterized in that: The attention module includes a coordinate information unit, a channel information unit and a spatial attention unit, wherein the expression of the coordinate information unit is as follows: in, Represents the feature map output by the coordinate information unit, Conv represents the convolution operation, Concat represents the channel cascade operation, X i1 (i, j) represents the first input feature map, x coord and coord Represents the coordinate information of the X and Y axis channels in the coordinate information unit; The expression of the channel information unit is as follows: in, represents the feature map output by the channel information unit, σ represents the Sigmoid activation function, ω1 and b1 represent the weight vector and bias vector of the channel information unit, and H and W represent the height and width of the feature map; The expression of the spatial attention unit is as follows: in, represents the feature map output by the spatial attention unit, GN represents group normalization, ω2 and b2 represent the weight vector and bias vector of the spatial attention unit, X i2 (i, j) represents the second input feature map.

4. A method for detecting abnormal operation status of electric power metering equipment according to claim 1, characterized in that: Calculating the similarity between the running status detection text and the running status detection result includes: Performing a word segmentation operation on the running status detection result and the running status detection text to obtain a label word segmentation sequence corresponding to the running status detection result and a question word segmentation sequence corresponding to the running status detection text; Sequencing the label word segmentation sequence and the question word segmentation sequence based on the Word2vec model to obtain a first word vector matrix of the label word segmentation sequence and a second word vector matrix of the question word segmentation sequence; Interacting the first word vector matrix with the second word vector matrix to obtain a first interactive attention matrix corresponding to the first word vector matrix and a second interactive attention matrix corresponding to the second word vector matrix; The first word vector matrix and the first interaction attention matrix are matrix-concatenated to obtain a first concatenated matrix, and the second word vector matrix and the second interaction attention matrix are matrix-concatenated to obtain a second concatenated matrix.

5. A method for detecting abnormal operation status of electric power metering equipment according to claim 4, characterized in that: Calculating the similarity between the running status detection text and the running status detection result also includes: Inputting the first splicing matrix and the second splicing matrix into a Transformer model respectively, and outputting a first text feature of the first splicing matrix and a second text feature of the second splicing matrix based on the Transformer model; One-dimensionalizing the first text feature and the second text feature based on a fully connected layer to obtain a first semantic feature and a second semantic feature; Calculating a difference and a product of the first semantic feature and the second semantic feature, and concatenating the difference and the product to obtain a fusion feature; The fusion features are processed based on a two-layer fully connected network to obtain the similarity between the running status detection result and the running status detection text, wherein the first layer of the fully connected network adopts the ReLU activation function and the second layer of the fully connected network adopts the Softmax normalization function.

6. A method for detecting abnormal operation status of electric power metering equipment according to claim 1, characterized in that: The state data includes current data, voltage data, vibration data and temperature data. The expressions of the multiple time windows are as follows: S={S1,S2,…,S t ,…,S n }; Among them, S represents multiple time windows, S t represents the tth time window among multiple time windows, and S t =[x t-w ,x t-w+1 , ..., x t-1 ], x t-1 Indicates the status data at time t-1.

7. A method for detecting abnormal operation status of electric power metering equipment according to claim 6, characterized in that: The anomaly detection model includes a data reconstruction module, a graph prediction module and an anomaly scoring module, wherein the data reconstruction module includes an encoder and a decoder, wherein the encoder is used to calculate the mean, standard deviation and Gaussian distribution of the multiple time windows, and the decoder is used to input the Gaussian distribution into the encoder again to obtain a reconstruction window, and map the reconstruction window to the reconstructed multivariate time series data, the graph prediction module is used to fuse the reconstructed multivariate time series data with the multiple time windows into an embedding vector, generate a time series state diagram based on the embedding vector, and predict the mean, standard deviation and Gaussian distribution of the sensor group at a future time based on the time series state diagram based on a graph neural network, and the anomaly scoring module is used to generate an anomaly score for the power metering device based on the mean, standard deviation and Gaussian distribution of the sensor group at a future time.

8. A method for detecting abnormal operation status of electric power metering equipment according to claim 7, characterized in that: The calculation formulas for the mean, standard deviation and Gaussian distribution of the multiple time windows are as follows: Among them, μ(t), σ(t) and Z(t) represent the mean, standard deviation and Gaussian distribution of multiple time windows, W μ and W σ represents the weight matrix for linear transformation, LSTM represents the long short-term memory network, Reshape represents the reshaping function, and b μ and b σ represents the bias term, ∈ represents the preset weight coefficient; The expression of the reconstruction window is as follows: in, represents the tth reconstruction window.

9. A method for detecting abnormal operation status of electric power metering equipment according to claim 8, characterized in that: Generate a time series state diagram based on the embedding vector, and predict the mean, standard deviation and Gaussian distribution of the sensor group at a future time based on the time series state diagram based on a graph neural network, including: Using each sensor in the sensor group as a graph node in the timing state graph; Calculating the cosine similarity between each graph node in the time series state graph based on the embedding vector, and if the cosine similarity is greater than a preset similarity threshold, connecting the graph nodes to obtain a connecting edge between two graph nodes; Calculate the correlation coefficient between each graph node in the time series state graph, wherein the calculation formula of the correlation coefficient is as follows: Among them, α i，j Represents a graph node v i and graph node v j The correlation coefficient between them, a represents the parameter vector to be learned, LeakyReLU represents the activation function, and Represents a graph node v i and graph node v j The features at time t, N(i) represents the graph node v i Neighbor graph nodes of Perform graph node aggregation update on each graph node in the time series state graph to obtain node aggregation features, wherein the graph node aggregation update formula is as follows: Among them, α i，j Represents a graph node v i The node aggregation feature at time t, W o represents the weight matrix of the linear transformation used for the aggregate update of the graph nodes, Represents a graph node v i The splicing features at time t, and Represents a graph node v i The state data at time t, Represents a graph node v i Gaussian distribution at time t; The mean, standard deviation and Gaussian distribution of the sensor group at a future time are predicted based on the aggregated features of each node in the time series state diagram, wherein the calculation formulas for the mean, standard deviation and Gaussian distribution of the sensor group at a future time are as follows: in, and Represents a graph node v i The mean, standard deviation and Gaussian distribution at time t are the mean, standard deviation and Gaussian distribution of the i-th sensor in the sensor group at time t. γ and δ represent the parameters to be learned. f Φ Represents a stack of fully connected layers.

10. A method for detecting abnormal operation status of electric power metering equipment according to claim 9, characterized in that: The calculation formula of the abnormal score of the power metering equipment is as follows: Among them, L abn represents the anomaly score of the power metering device, and N represents the number of sensor groups.

Images