A method and apparatus for determining the opening and closing status of a disconnector switch based on an event camera.

Abstract

This invention provides a method and apparatus for determining the opening and closing status of a disconnector switch based on an event camera. The method includes: Step 1: acquiring event stream data output by the event camera; Step 2: preprocessing the event stream data acquired in Step 1 to remove noise; Step 3: performing spatiotemporal point cloud encoding on the preprocessed event stream data in Step 2 to form point cloud data with temporal and spatial dimensions; Step 4: analyzing the point cloud data obtained in Step 3 using a deep learning model to determine the opening and closing status of the disconnector switch. This invention effectively solves the influence of strong light and nighttime light on status determination, is easy to install, is unaffected by high voltage and strong magnetic fields, and greatly improves the accuracy and applicability of opening and closing status determination.

Description

Technical Field

[0001] This invention relates to the field of power equipment monitoring technology, specifically a method and device for determining the opening and closing status of disconnecting switches based on event cameras. Background Technology

[0002] Disconnect switches are used on high-voltage equipment of various levels in substations and power plants to change circuit connections, connecting or isolating lines or equipment from the power source. Because disconnect switches operate outdoors for extended periods and their operation frequency is not very high, the contact surfaces and transmission mechanisms are prone to contamination and corrosion. This can cause contact jamming and high resistance during opening and closing operations, ultimately leading to incomplete disconnection and closing, thus affecting the safe and stable operation of the entire substation.

[0003] To improve safety and work efficiency, one-click sequential control has emerged. With the advancement of one-click sequential control, numerous substations have installed disconnector switch opening / closing status determination devices. Existing disconnector switch opening / closing status determination devices mainly include microswitches, image detection, attitude sensing, and pressure monitoring types. Actual operation has revealed that while image-based intelligent recognition technology offers intuitive and highly automated identification, its accuracy is low and it is significantly affected by weather and lighting conditions. Attitude sensor-based methods suffer from signal transmission problems in high-voltage, strong magnetic environments and difficulties in powering the attitude sensor. Auxiliary contact-based methods also face challenges in signal transmission and auxiliary switch functionality under high-voltage, strong magnetic environments. These disconnector switch opening / closing status determination devices have numerous issues regarding reliability and accuracy, hindering the comprehensive and in-depth application of one-click sequential control technology.

[0004] With the development of machine vision, event cameras, also known as dynamic vision sensors, have attracted increasing attention. Event cameras mimic the human retina, responding to pixel pulses that indicate brightness changes due to motion. Therefore, they can capture scene brightness changes at extremely high frame rates, recording events at specific times and locations within the image, forming an event stream. Compared to traditional standard cameras, event cameras offer advantages such as high dynamic range, high temporal resolution, and no dynamic blur. Therefore, event camera-based solutions provide a new research approach for determining the status of high-voltage disconnect switches.

[0005] Therefore, in order to solve the above problems, a method and device for judging the opening and closing status of disconnecting switches based on event cameras are designed to improve the operational reliability and judgment accuracy of the disconnecting switch opening and closing status judgment device, and improve the application efficiency of one-click sequential control. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides a method and apparatus for determining the opening and closing status of disconnecting switches based on an event camera. This invention effectively solves the problem of strong light and nighttime light affecting status determination, is easy to install, is unaffected by high voltage and strong magnetic fields, and greatly improves the accuracy and applicability of opening and closing status determination.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] In a first aspect, the present invention provides a method for determining the opening and closing status of a disconnector switch based on an event camera, comprising the following steps:

[0009] Step 1: Obtain the event stream data output by the event camera;

[0010] Step 2: Preprocess the event stream data obtained in Step 1 to remove noise from the event stream;

[0011] Step 3: Perform spatiotemporal point cloud encoding on the preprocessed event stream data from Step 2 to form point cloud data with temporal and spatial dimensions;

[0012] Step 4: Use a deep learning model to analyze the point cloud data obtained in Step 3 to determine the open and closed status of the disconnector switch.

[0013] Furthermore, the event stream data specifically includes: event trigger coordinates, event trigger timestamps, and event trigger polarity, wherein the event trigger polarity includes: event trigger positive polarity and event trigger negative polarity. Event trigger positive polarity indicates that the light intensity brightens beyond the trigger threshold, and event trigger negative polarity indicates that the light intensity dims beyond the trigger threshold.

[0014] Furthermore, step 2 preprocesses the event stream data obtained in step 1 to remove noise from the event stream, specifically including:

[0015] First, Gaussian filtering is used for noise reduction: for each event, K points are taken in its vicinity, which are called neighbor points. Then, the average distance between the event and the K neighbor points is calculated and filtered by a pre-set standard deviation.

[0016] Then, outlier filtering is performed using the DBSCAN clustering algorithm. To perform outlier filtering using the DBSCAN clustering algorithm, the neighborhood radius Eps and the threshold MinPts for the number of data objects in the neighborhood need to be set in advance. The specific steps are as follows:

[0017] (1) Select any event point p from the event data;

[0018] (2) Using the selected event point p as the core point, find the event points that the density of p can reach, and form a cluster;

[0019] (3) If the selected event point p is an edge point, select another event point;

[0020] (4) Repeat steps (2) and (3) until all points have been processed;

[0021] Finally, by using random downsampling, some data are randomly selected from the majority class samples to represent a relatively complete action with fewer events.

[0022] Further, step 4: Analyze the point cloud data obtained in step 3 using a deep learning model to determine the opening and closing status of the disconnector switch, specifically including:

[0023] Step 4.1: Create a dataset for the opening and closing of disconnecting switches;

[0024] Step 4.2: Establish a model for judging the opening and closing status of disconnecting switches;

[0025] Step 4.3: Train the disconnector switch opening and closing status discrimination model established in Step 4.2 based on the disconnector switch opening and closing dataset created in Step 4.1;

[0026] Step 4.4: Use the disconnector switch opening and closing status discrimination model trained in Step 4.3 to identify the disconnector switch opening and closing status in real time.

[0027] Furthermore, in step 4.1, the action categories of the disconnector switch opening and closing data are divided into four categories: normal disconnector switch closing, normal disconnector switch opening, abnormal disconnector switch closing, and abnormal disconnector switch opening.

[0028] Furthermore, the disconnector switch opening and closing status discrimination model established in step 4.2 is a PointConv point cloud classification model, and the specific establishment steps include:

[0029] First, a shared encoding module PointConv_1 is established to convolve the input 3D spatiotemporal event point cloud data. PointConv_1 is composed of MLP and BatchNorm. PointConv_1 is upgraded to 64, 64 and 128 dimensions respectively. Then, PointConv_2 and PointConv_3 modules are established in sequence for convolution. After these three modules, enough features are obtained and the data feature dimension is increased to 1024 dimensions.

[0030] Then, a Linear join module is built to reduce the data dimensionality to 512. The Dropout module is used to deactivate the data, and then the Linear join module is used again to reduce the data dimensionality to 256.

[0031] Finally, it performs state discrimination with a four-class fully connected model;

[0032] When a three-dimensional spatiotemporal event point cloud data is input for model inference, the PointConv network outputs the probability values ​​of four categories: normal closing of the disconnector switch, normal opening of the disconnector switch, abnormal closing of the disconnector switch, and abnormal opening of the disconnector switch. The one with the highest probability is taken as the final discrimination result.

[0033] Further, step 4.3: Based on the disconnector switch opening and closing dataset created in step 4.1, the disconnector switch opening and closing status discrimination model established in step 4.2 is trained, specifically including:

[0034] First, the dataset is randomly divided into training, test, and validation sets in a ratio of 7:2:1. During training, the PointConv model is used, and data augmentation is performed on the dataset, including adding Gaussian noise, random displacement, random rotation, and random scaling.

[0035] The PointConv model uses the cross-entropy loss function and SGD stochastic gradient descent during training. Cross-entropy can measure the degree of difference between two different probability distributions in the same random variable. As training progresses, the weights are continuously adjusted, and the loss value continues to decrease until the optimal model is obtained.

[0036] Further, step 4.4: Using the disconnector switch opening and closing status discrimination model trained in step 4.3, the opening and closing status of the disconnector switch is identified in real time, specifically including:

[0037] First, the disconnector switch opening / closing status discrimination model is deployed in the intelligent computing hardware device. Step 1 receives real-time event stream data, step 2 processes the event stream data, step 3 encodes the event stream data, and finally inputs it into the trained disconnector switch opening / closing status discrimination model. The intelligent computing hardware calls the AI ​​chip to complete the model inference and outputs the disconnector switch opening / closing status.

[0038] A method for determining the opening and closing status of a disconnector switch based on an event camera, comprising an event camera and a core board;

[0039] The event camera is used to monitor the opening and closing status of the high-voltage disconnector in real time and obtain event stream data.

[0040] The core board is used to preprocess the event stream data acquired by the event camera, remove noise from the event stream, perform spatiotemporal point cloud encoding on the preprocessed event stream data to form point cloud data in time and space dimensions, and use a deep learning model to analyze the obtained point cloud data to realize the determination of the opening and closing status of the disconnecting switch.

[0041] Furthermore, it also includes the network port and power supply that connect to the core board;

[0042] The network port is used to send and receive data over the network;

[0043] The power supply is responsible for powering the core board and the event camera.

[0044] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects:

[0045] 1. In open-air substation scenarios, strong light interference and blurred images at night can impair edge extraction in visible light images, leading to a higher misjudgment rate in disconnection and closing status identification based on standard visible light cameras, making it difficult to meet actual business needs. This invention provides a method for identifying the disconnection and closing status of disconnectors based on event cameras. Using an event camera instead of a standard camera, the method preprocesses the event stream through Gaussian filtering, DBSCAN clustering, and random downsampling. This solves the problem of strong light and nighttime lighting affecting status capture and removes noise points through filtering and clustering algorithms. Regarding the recognition algorithm, the method innovatively uses the PointConv point cloud recognition network for disconnection and closing status identification. Preprocessing and spatiotemporal point cloud encoding of the input event stream data achieves an accuracy of 98.5% on the validation set, significantly improving the accuracy of disconnection and closing status identification.

[0046] 2. In addition, the disconnection switch opening and closing status discrimination device based on event camera is easy to install, is not affected by high voltage and strong magnetic field, is safe and reliable to use, and has high applicability. Attached Figure Description

[0047] Figure 1 This is a flowchart of the method for determining the opening and closing status of a disconnecting switch based on an event camera, provided in an embodiment of the present invention.

[0048] Figure 2 This is a network structure diagram of the disconnection switch opening and closing status discrimination algorithm based on event camera provided in an embodiment of the present invention;

[0049] Figure 3 This is a comparison chart of model indicators provided in the embodiments of the present invention;

[0050] Figure 4 This is an example of the opening and closing of a disconnector switch based on an event flow provided in this embodiment of the invention;

[0051] Figure 5 This is a schematic diagram of the structure of the disconnection switch opening and closing status discrimination device based on an event camera provided in an embodiment of the present invention;

[0052] Figure 6 This is an application diagram of the disconnection switch opening and closing status discrimination device based on an event camera provided in an embodiment of the present invention. Detailed Implementation

[0053] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] This invention provides a method for determining the opening and closing status of a disconnector switch based on an event camera. The method is based on the event stream data (also known as asynchronous spatiotemporal pulse signal) output by the event camera. First, noise in the event stream is removed through preprocessing. Then, the event stream data is encoded to form point cloud data with time and spatial dimensions. Finally, a deep learning model is used to analyze the point cloud data to determine the opening and closing status of the disconnector switch.

[0055] Specifically, such as Figure 1 As shown, the method includes the following steps:

[0056] Step 1: Obtain the event stream data output by the event camera.

[0057] The opening and closing status of the high-voltage disconnector is monitored in real time by an event camera to obtain event stream data. The event stream data output by the event camera specifically includes: event trigger coordinates, event trigger timestamp, and event trigger polarity. The event trigger polarity includes: event trigger positive polarity and event trigger negative polarity. Event trigger positive polarity indicates that the light intensity has increased beyond the trigger threshold, and event trigger negative polarity indicates that the light intensity has decreased beyond the trigger threshold.

[0058] Step 2: Preprocess the event stream data obtained in Step 1 to remove noise from the event stream.

[0059] Event cameras are highly sensitive to changes in ambient light intensity and are inevitably susceptible to the influence of optical hardware circuitry and environmental factors. This results in a significant amount of noise being mixed in with the output motion information, impacting the accuracy of subsequent neural network training. Preprocessing of event stream data specifically includes:

[0060] First, Gaussian filtering is used for denoising. The specific steps for denoising event stream data based on Gaussian filtering are as follows: take each event as the center, take K points in its vicinity, these points are called neighbor points, then calculate the average distance between the event and the K neighbor points, and filter it by the pre-set standard deviation.

[0061] Then, outlier filtering is performed using the DBSCAN clustering algorithm. To perform outlier filtering based on the DBSCAN clustering algorithm, the neighborhood radius Eps and the threshold value MinPts for the number of data objects in the neighborhood need to be set in advance. The specific steps are as follows:

[0062] (1) Select any event point p from the event data;

[0063] (2) Using the selected event point p as the core point, find the event points that the density of p can reach, and form a cluster;

[0064] (3) If the selected event point p is an edge point, select another event point;

[0065] (4) Repeat steps (2) and (3) until all points have been processed.

[0066] Finally, random downsampling is used. The idea is simple: randomly select some data from the majority class samples. This method can represent a relatively complete action with fewer events, providing support for the subsequent process.

[0067] Step 3: Perform spatiotemporal point cloud encoding on the preprocessed event stream data from Step 2 to form point cloud data with temporal and spatial dimensions.

[0068] The purpose of spatiotemporal point cloud encoding is to eliminate the adverse effects caused by singular sample data. First, the spatial coordinates are normalized, and then the temporal coordinates are encoded in sequence to obtain the encoded three-dimensional spatiotemporal event point cloud data.

[0069] Suppose an event is defined as ei(xi,yi,ti,pi), where xi andyi represent spatial coordinates, ti represents a timestamp, and pi represents a polarity value. Then, it is spatiotemporally encoded, and the encoded event is represented as Ei(Xi,Yi,Zi,Pi). Assuming there are N events in this time period, the width of the frame image is W, and the height of the frame image is H, the detailed encoding rules are as follows:

[0070]

[0071] X i =x i / W (2)

[0072] Y i =(Hy i ) / H (3)

[0073]

[0074] Step 4: Analyze the point cloud data obtained in Step 3 using a deep learning model to determine the open / closed state of the disconnector switch. This includes the following steps:

[0075] Step 4.1: Create a dataset for disconnecting switch opening and closing.

[0076] To assess the actual operational status of disconnectors, the actions in the dataset are categorized into four types: normal closing, normal opening, abnormal closing, and abnormal opening (each action lasting approximately 6 seconds). Considering the impact of the real-world substation environment on the disconnector identification algorithm, different on-site environments were set up for data collection. This process mainly consisted of two parts. First, data was collected under ideal conditions: normal lighting, a simple background, and the data acquisition point was located in front of the disconnector, primarily to test the algorithm's effectiveness. Second, data was collected under complex conditions: single or multiple adverse conditions were set, such as insufficient lighting at night, direct sunlight during the day, a background with numerous electrical devices, the data acquisition point being located to the side of the disconnector, or a slightly farther distance, etc., primarily to test the algorithm's reliability.

[0077] like Figure 4 As shown, the action example images obtained by the event camera through the above steps 1, 2, and 3 are shown. The two-dimensional image is a cumulative image of the frequency of this event, and the three-dimensional image is a three-dimensional display of all events that occur in the (x,y,t) coordinate system composed of spatial coordinates and time axis. Red dots represent polarity 1, and green dots represent polarity 0.

[0078] Step 4.2: Establish a model for judging the opening and closing status of disconnecting switches.

[0079] Because the event stream data output by the event camera is similar to the point cloud in terms of its 3D spatial data structure distribution, and is sparser in its temporal distribution, this embodiment of the invention uses a point cloud network-based structure to process the event stream data. For example... Figure 1 The flowchart of the method for determining the opening and closing status of disconnecting switches based on event cameras shows that after preprocessing the event stream data and encoding the spatiotemporal point cloud, point cloud data with time and space dimensions is formed, which is a set of points in three-dimensional space.

[0080] At this point, the 3D spatiotemporal point cloud data retains the original spatial geometry and can well represent the surface features and depth information of objects. However, it also has the characteristics of uneven spatial distribution and scattered, disordered data. Therefore, a point cloud classification model can be established to determine the opening and closing status of disconnecting switches.

[0081] Currently, among point cloud classification models, the PointConv network can efficiently utilize convolutional networks in point clouds with non-uniform sampling, while simultaneously ensuring the invariance of point cloud arrangement and transformation. Therefore, this embodiment of the invention innovatively selects the PointConv point cloud classification model to determine the opening and closing states of disconnecting switches.

[0082] The model is based on Figure 2 The PointConv network structure is established by first creating a shared encoding module, PointConv_1, to convolve the input 3D spatiotemporal event point cloud data. The PointConv_1 module is composed of MLP and BatchNorm. Here, PointConv_1 will perform dimensionality upscaling operations, upscaling to 64, 64, and 128 respectively. Then, PointConv_2 and PointConv_3 modules are created sequentially for convolution. After these three modules, a sufficient number of features can be obtained. Here, the data feature dimension is increased to 1024 dimensions.

[0083] Then, a Linear Connection module is built to reduce the data dimensionality to 512. The Dropout module is used to deactivate the data, and then the Linear Connection module is used again to reduce the data dimensionality to 256. It can be seen from the created disconnector switch opening and closing dataset that the disconnector switch opening and closing action categories are currently divided into four categories: normal disconnector switch closing, normal disconnector switch opening, abnormal disconnector switch closing, and abnormal disconnector switch opening. Therefore, the last step in building the model is to perform state discrimination with a four-category fully connected model.

[0084] When a three-dimensional spatiotemporal event point cloud data is input for model inference, the PointConv network will output the probability values ​​of four categories: normal closing of disconnector switch, normal opening of disconnector switch, abnormal closing of disconnector switch, and abnormal opening of disconnector switch. The one with the highest probability is taken as the final discrimination result.

[0085] The PointConv module in the model is a novel convolution operation, as shown in Equation (5):

[0086]

[0087] In the formula: F is the feature of a point in a local region G surrounding a point P(x,y,z); S is a density function, which takes the coordinates of each point as input and outputs the inverse density coefficient of each point; W is the weight of the feature of that point, which is learned by MLP and returns the weight coefficient of the point by taking the coordinates of the point as input.

[0088] The final layer of the model is the output layer, which uses the SoftMax classifier to obtain the classification results of the four actions of opening and closing the disconnecting switch. The SoftMax function is shown in formula (6):

[0089]

[0090] In the formula: Oi represents the output value corresponding to the i-th node, and Cz is the total number of categories. In this embodiment, Cz = 4.

[0091] Step 4.3: Train the disconnector switch opening and closing status discrimination model established in Step 4.2 based on the disconnector switch opening and closing dataset created in Step 4.1.

[0092] First, the dataset is randomly divided into training, test, and validation sets in a 7:2:1 ratio. Since both data quality and quantity are essential for training high-performance deep neural networks, especially in 3D deep learning, data augmentation is performed during training to improve the model's generalization ability. Specific data augmentation methods are shown below:

[0093] (1) Add Gaussian noise

[0094] In point cloud data, Gaussian noise of the same dimension is first generated and then added to the original data.

[0095] (2) Random displacement

[0096] The point cloud is randomly moved a certain distance in the x, y, and z directions. For each point in the point cloud, three random numbers are generated to represent the random displacement in the x, y, and z directions. Then, the previously generated random displacements are added to the x, y, and z coordinates of all points in the same point cloud. After this operation, the point cloud can be randomly displaced in any direction.

[0097] (3) Random rotation

[0098] Similar to displacement, rotation also occurs around the x, y, and z directions. For each point cloud dataset, this paper first randomly generates random rotation angles in the three directions, and then rotates each point in the point cloud according to the randomly generated angles.

[0099] (4) Random scaling

[0100] For a point cloud scaling operation, first, a scaling center is randomly generated, then a scaling ratio is randomly generated, and finally, the distance from each point in the point cloud to the scaling center is scaled according to this ratio to obtain new point cloud data.

[0101] The PointConv model employs a cross-entropy loss function and stochastic gradient descent (SGD) during training. Cross-entropy measures the difference between two different probability distributions of the same random variable, while SGD is simple, efficient, and can escape local optima. As training progresses, the weights are continuously adjusted, and the loss value decreases until the optimal model is obtained. Figure 3 Based on the disconnector switch opening and closing dataset created in step 4, PointNet, PointNet++, and the PointConv model used in this invention were trained respectively. Verification showed that PointNet achieved an overall classification accuracy of 97.2%, PointNet++ achieved 98%, and the PointConv model used in this application achieved 98.5%. This demonstrates that PointConv efficiently utilizes convolutional networks while simultaneously ensuring invariance to point cloud arrangement and transformation. Compared to other classification models in the table, this paper's classification model exhibits high accuracy and a simple processing method.

[0102] Step 4.4: Use the disconnector switch opening and closing status discrimination model trained in Step 4.3 to identify the disconnector switch opening and closing status in real time.

[0103] First, the disconnector switch opening / closing status discrimination model is deployed in the intelligent computing hardware device. Step 1 receives real-time event stream data, step 2 processes the event stream data, step 3 encodes the event stream data, and finally inputs it into the trained disconnector switch opening / closing status discrimination model. The intelligent computing hardware calls the AI ​​chip to complete the model inference and outputs the disconnector switch opening / closing status.

[0104] Preferably, in this embodiment, the event camera used is the DAVIS346, which combines DVS and a traditional camera and introduces an active pixel sensor (APS), so it can output RGB images while outputting event stream data.

[0105] like Figure 5 As shown, this embodiment also provides a device for determining the opening and closing status of a disconnecting switch based on an event camera, including an event camera, a core board, a network port, and a power supply.

[0106] The event camera is used to monitor the opening and closing status of the high-voltage disconnector in real time and obtain event stream data.

[0107] The core board is used to preprocess the event stream data acquired by the event camera, remove noise from the event stream, perform spatiotemporal point cloud encoding on the preprocessed event stream data to form point cloud data in time and space dimensions, and use a deep learning model to analyze the obtained point cloud data to realize the determination of the opening and closing status of the disconnecting switch.

[0108] The network port is used to send and receive data over the network;

[0109] The power supply is responsible for powering the core board and the event camera.

[0110] Existing research on the detection of the opening and closing status of disconnecting switches relies on visible light standard cameras. The advantages of this approach are its in-depth research, mature technology, intuitive judgment process, and high degree of automation. However, its disadvantages include low accuracy, particularly susceptible to weather and lighting conditions. In open-air substation scenarios, strong light interference and blurred images at night can impair edge extraction from visible light images, leading to a higher misjudgment rate for opening and closing status detection, making it difficult to meet practical operational needs. In contrast, using an event camera to detect the opening and closing status based on event streams can eliminate the impact of lighting on image recognition.

[0111] To verify the superiority of the event sensor, an image frame-based algorithm for detecting the opening and closing status of disconnecting switches was constructed. Based on the RGB image data output by a standard camera, the rotated_faster_rcnn rotating target detection model was used to detect the disconnecting switches.

[0112] During the acquisition of RGB image data, the same method as in step 4 for creating the disconnector switch opening and closing dataset was used. Different on-site environments were set up for data acquisition. First, data was acquired under ideal conditions: normal lighting, simple background, and the acquisition location was in front of the disconnector switch. This was primarily used to test the effectiveness of the algorithm. Second, data was acquired under complex conditions, including single or multiple adverse conditions such as insufficient lighting at night, direct sunlight on the camera during the day, numerous electrical devices in the background, the acquisition location being on the side of the disconnector switch, or a slightly farther distance. This was primarily used to test the reliability of the algorithm.

[0113] After training, rotated_faster_rcnn achieved an accuracy of 96.1% at the optimal threshold of 0.8, while the PointConv algorithm based on event cameras in this invention achieved an accuracy of 98.5%. It can be concluded that the PointConv algorithm model using event cameras has higher accuracy and stronger generalization ability than the detection algorithm using standard cameras, as shown in Table 1.

[0114] Table 1

[0115] plan accuracy characteristic Based on event stream 98.5％ Highly adaptable Based on image frames 96.1％ Susceptible to strong light interference; images are blurry at night.

[0116] The disconnector switch opening / closing status discrimination device based on event camera provided in this embodiment, such as Figure 6As shown in the application diagram, the device acts as a terminal device to monitor the high-voltage disconnecting switch of the substation and collects real-time opening and closing data when the disconnecting switch is working normally.

[0117] The device is equipped with a pre-trained disconnector switch opening and closing status discrimination model. Through model reasoning, the disconnector switch opening and closing status is output and pushed to the WEB terminal or mobile terminal in real time, which facilitates the inspection or replacement of disconnectors with defects, saves labor costs and improves detection efficiency.

[0118] Preferably, in this embodiment, the core board selects NVIDIA Jetson NANO as the intelligent computing hardware, with 128 CUDA cores and 4G memory, providing excellent speed and energy efficiency for AI computing.

[0119] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for discriminating the on-off state of an isolating switch based on an event camera, characterized in that: Comprising the following steps: Step 1: obtaining event stream data output by an event camera; Step 2: preprocessing the event stream data obtained in step 1 to remove noise in the event stream; Step 3: spatiotemporal point cloud coding on the event stream data preprocessed in step 2 to form point cloud data in time and space dimensions; Step 4: analyzing the point cloud data obtained in step 3 using a deep learning model to realize discrimination of the on-off state of the disconnecting switch, specifically comprising: Step 4.1: making a disconnecting switch on-off data set; Step 4.2: establishing a disconnecting switch on-off state discrimination model; Step 4.3: training the disconnecting switch on-off state discrimination model established in step 4.2 based on the disconnecting switch on-off data set made in step 4.1; Step 4.4: using the disconnecting switch on-off state discrimination model trained in step 4.3 to realize real-time recognition of the on-off state of the disconnecting switch; The disconnecting switch on-off state discrimination model established in step 4.2 is a PointConv point cloud classification model, and the specific establishment steps comprise: First, a shared coding module PointConv_1 is established to perform convolution on the input three-dimensional spatiotemporal event point cloud data, PointConv_1 is composed of MLP and BatchNorm, PointConv_1 performs dimension lifting operation to 64, 64, 128 respectively, and then PointConv_2 and PointConv_3 modules are established in sequence for convolution, sufficient features are obtained after passing through these three modules, and the data feature dimension is lifted to 1024 dimensions; Then, a Linear connection module is established to reduce the data to 512 dimensions, a Dropout module is used for inactivation operation, and then a Linear connection module is used to reduce the data to 256 dimensions; Finally, a four-class full connection model is used for state discrimination; When a segment of three-dimensional spatiotemporal event point cloud data is input for model inference, the PointConv network outputs probability values of four categories of disconnecting switch normal closing, disconnecting switch normal opening, disconnecting switch abnormal closing, and disconnecting switch abnormal opening, and the maximum probability is taken as the final discrimination result.

2. The method for event camera based isolator opening and closing state discrimination of claim 1, wherein: The event stream data specifically comprises: event trigger coordinates, event trigger timestamps, and event trigger polarity, wherein the event trigger polarity comprises: event trigger positive polarity and event trigger negative polarity, the event trigger positive polarity indicates that the light intensity becomes brighter than the trigger threshold, and the event trigger negative polarity indicates that the light intensity becomes darker than the trigger threshold.

3. The method for event camera based isolator opening and closing state discrimination of claim 1, wherein: Step 2 preprocesses the event stream data obtained in step 1 to remove noise in the event stream, specifically comprising: First, Gaussian filtering is used for denoising: taking each event as the center, K points near the event are taken as neighbor points, then the average distance between the event and the K neighbor points is calculated, and filtering is performed through the pre-set standard deviation; Based on the DBSCAN clustering algorithm, the neighborhood radius Eps and the threshold MinPts of the number of data objects in the neighborhood are set in advance, and the specific steps are as follows: (1) randomly select an event point p from the event data; (2) take the selected event point p as the core point, find the event points with reachable density p, and form a cluster; (3) if the selected event point p is an edge point, select another event point; (4) repeat steps (2) and (3) until all points are processed; Finally, through random down-sampling, some data are randomly selected from the majority class samples to represent a relatively complete action with fewer events.

4. The method for event camera based isolator open-close state discrimination of claim 1, wherein: The action category of the disconnector opening and closing data set in step 4.1 is divided into four categories: disconnector normal closing, disconnector normal opening, disconnector abnormal closing and disconnector abnormal opening.

5. The method for event camera based isolator open-close state discrimination of claim 1, wherein: Step 4.3: Train the disconnector opening and closing state discrimination model established in step 4.2 based on the disconnector opening and closing data set made in step 4.1, which includes: First, randomly divide the data set into training set, test set and validation set according to the ratio of 7:2:1, use PointConv model in training, and perform data enhancement on the data set, including adding Gaussian noise, random displacement, random rotation and random scaling; PointConv model uses cross-entropy loss function and SGD random gradient descent in the training process. Cross-entropy can measure the difference between two different probability distributions of the same random variable. With the training, the weights are constantly adjusted, and the loss value is constantly reduced until the optimal model is obtained.

6. The method for event camera based isolator open-close state discrimination of claim 1, wherein: Step 4.4: Real-time identification of disconnector opening and closing state using the disconnector opening and closing state discrimination model trained in step 4.3, which includes: First, deploy the disconnector opening and closing state discrimination model to the intelligent computing hardware device, receive real-time event stream data through step 1, process event stream data through step 2, encode event stream data through step 3, and finally input into the trained disconnector opening and closing state discrimination model. Through intelligent computing hardware, AI chips are called to complete model inference, and the opening and closing state of the disconnector is output.

7. An apparatus for event camera based discrimination of open / close states of a disconnector, for implementing the method of any one of claims 1-6, characterized by: The device includes an event camera and a core board; The event camera is used to monitor the opening and closing state of the high-voltage disconnector in real time and obtain event stream data; The core board is used to preprocess the event stream data obtained by the event camera, remove noise in the event stream, and perform spatio-temporal point cloud coding on the preprocessed event stream data to form point cloud data in time and space dimensions. Deep learning model is used to analyze the obtained point cloud data to realize discrimination of the disconnector opening and closing state.

8. The apparatus for event camera based isolator opening and closing state discrimination of claim 7, wherein: It also includes a network port and a power supply connected to the core board; The network port is used for data sending and receiving through the network; The power supply is responsible for power supply of the core board and the event camera.

Images