CN116363411A - Method and device for judging opening and closing states of isolating switch based on event camera - Google Patents
Method and device for judging opening and closing states of isolating switch based on event camera Download PDFInfo
- Publication number
- CN116363411A CN116363411A CN202310120669.3A CN202310120669A CN116363411A CN 116363411 A CN116363411 A CN 116363411A CN 202310120669 A CN202310120669 A CN 202310120669A CN 116363411 A CN116363411 A CN 116363411A
- Authority
- CN
- China
- Prior art keywords
- event
- opening
- isolating switch
- data
- closing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 46
- 238000007781 pre-processing Methods 0.000 claims abstract description 12
- 238000013136 deep learning model Methods 0.000 claims abstract description 10
- 238000012549 training Methods 0.000 claims description 20
- 230000002159 abnormal effect Effects 0.000 claims description 18
- 230000009471 action Effects 0.000 claims description 11
- 230000008569 process Effects 0.000 claims description 11
- 238000001914 filtration Methods 0.000 claims description 9
- 238000012545 processing Methods 0.000 claims description 8
- 238000012360 testing method Methods 0.000 claims description 7
- 238000006073 displacement reaction Methods 0.000 claims description 6
- 238000009826 distribution Methods 0.000 claims description 6
- 230000006870 function Effects 0.000 claims description 6
- 238000012544 monitoring process Methods 0.000 claims description 6
- 238000012216 screening Methods 0.000 claims description 6
- 230000001960 triggered effect Effects 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 5
- 230000005540 biological transmission Effects 0.000 claims description 4
- 238000004519 manufacturing process Methods 0.000 claims description 4
- 238000005282 brightening Methods 0.000 claims description 3
- 230000002779 inactivation Effects 0.000 claims description 2
- 230000005415 magnetization Effects 0.000 abstract 1
- 238000013145 classification model Methods 0.000 description 6
- 238000001514 detection method Methods 0.000 description 6
- 238000010586 diagram Methods 0.000 description 5
- 238000005286 illumination Methods 0.000 description 5
- 230000007547 defect Effects 0.000 description 4
- 238000005516 engineering process Methods 0.000 description 4
- 230000000694 effects Effects 0.000 description 3
- 238000011160 research Methods 0.000 description 3
- 238000013528 artificial neural network Methods 0.000 description 2
- 238000006243 chemical reaction Methods 0.000 description 2
- 238000013480 data collection Methods 0.000 description 2
- 238000000605 extraction Methods 0.000 description 2
- 230000005389 magnetism Effects 0.000 description 2
- 230000000007 visual effect Effects 0.000 description 2
- HPTJABJPZMULFH-UHFFFAOYSA-N 12-[(Cyclohexylcarbamoyl)amino]dodecanoic acid Chemical group OC(=O)CCCCCCCCCCCNC(=O)NC1CCCCC1 HPTJABJPZMULFH-UHFFFAOYSA-N 0.000 description 1
- 230000002411 adverse Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 230000007797 corrosion Effects 0.000 description 1
- 238000005260 corrosion Methods 0.000 description 1
- 230000009849 deactivation Effects 0.000 description 1
- 238000013135 deep learning Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 238000002955 isolation Methods 0.000 description 1
- 230000007246 mechanism Effects 0.000 description 1
- 230000003287 optical effect Effects 0.000 description 1
- 210000001525 retina Anatomy 0.000 description 1
- 239000004065 semiconductor Substances 0.000 description 1
- 230000008054 signal transmission Effects 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/764—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using classification, e.g. of video objects
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/762—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using clustering, e.g. of similar faces in social networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06V—IMAGE OR VIDEO RECOGNITION OR UNDERSTANDING
- G06V10/00—Arrangements for image or video recognition or understanding
- G06V10/70—Arrangements for image or video recognition or understanding using pattern recognition or machine learning
- G06V10/82—Arrangements for image or video recognition or understanding using pattern recognition or machine learning using neural networks
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Evolutionary Computation (AREA)
- Physics & Mathematics (AREA)
- Artificial Intelligence (AREA)
- General Physics & Mathematics (AREA)
- Computing Systems (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Software Systems (AREA)
- Databases & Information Systems (AREA)
- Medical Informatics (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Multimedia (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Molecular Biology (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Image Analysis (AREA)
Abstract
The invention provides a method and a device for judging the opening and closing states of an isolating switch based on an event camera, wherein the method comprises the following steps: step 1: acquiring event stream data output by an event camera; step 2: preprocessing the event stream data obtained in the step 1, and removing noise in the event stream; step 3: carrying out space-time point cloud coding on the event stream data preprocessed in the step 2 to form point cloud data of time and space dimensions; step 4: and (3) analyzing the point cloud data obtained in the step (3) by using a deep learning model, and judging the opening and closing states of the isolating switch. The invention can effectively solve the problem of influence of strong light and light at night on state discrimination, is convenient to install, is not influenced by high-voltage magnetization, and greatly improves the accuracy and applicability of state discrimination of opening and closing.
Description
Technical Field
The invention relates to the technical field of power equipment monitoring, in particular to a method and a device for judging the opening and closing states of an isolating switch based on an event camera.
Background
The isolating switch is used for various high-voltage equipment in a transformer substation and a power station, and is used for changing circuit connection to enable lines or equipment to be communicated with or isolated from a power supply. Because the isolating switch operates in an outdoor environment for a long time, and the action frequency of equipment is not very high, the phenomena of pollution falling, corrosion and the like of the contact surface and a transmission mechanism of the isolating switch are easy to occur, the phenomena of contact clamping stagnation, high switching resistance and the like occur when the isolating switch is switched on and off, and finally the fault that the switching on and off of the isolating switch are not in place is caused, so that the safe and stable operation of the whole transformer substation is influenced.
In order to improve the safety level and the working efficiency, one-key sequential control is generated. Along with the promotion of the one-key sequential control work, a large number of power stations are provided with isolating switch opening and closing state distinguishing devices. The existing disconnecting switch opening and closing state judging device mainly comprises a micro switch, image detection, gesture sensing, pressure monitoring and other types. Actual operation discovers that the disconnecting switch opening and closing position distinguishing method based on the image intelligent recognition technology is visual in distinguishing process and high in automation degree, but is low in recognition accuracy and greatly influenced by weather and illumination. The isolating switch position judging method based on the attitude sensor technology has the defect that signals cannot be transmitted under a high-pressure magneto environment and the attitude sensor is difficult to supply power. The isolating switch position judging mode based on the auxiliary contact principle has the defect of signal transmission and auxiliary switch function under a high-voltage magnetic environment. The disconnecting switch opening and closing state judging device has more problems in reliability, judging accuracy and the like, and prevents the comprehensive and deep application of the one-key sequential control technology.
With the development of the machine vision field, event cameras, also called dynamic vision sensors, draw more and more attention. The event camera simulates the retina of a human being, responds to pixel point pulses that produce brightness changes due to motion, so it can capture the brightness changes of a scene at a very high frame rate, record events at specific points in time and at specific locations in an image, and form an event stream. Compared with the traditional standard camera, the event camera has the advantages of high dynamic range, high time resolution, no dynamic blur and the like, so that a new research thought is provided for judging the state of the high-voltage isolating switch based on the scheme of the event camera.
Therefore, in order to solve the problems, a method and a device for judging the opening and closing state of the isolating switch based on the event camera are designed, so that the operation reliability and the judgment accuracy of the device for judging the opening and closing state of the isolating switch are improved, and the application quality of one-key sequential control work is improved.
Disclosure of Invention
Aiming at the defects in the prior art, the invention provides the method and the device for judging the opening and closing state of the isolating switch based on the event camera, which can effectively solve the problem that strong light and light at night affect the state judgment, are convenient to install, are not affected by high-pressure magnetism, and greatly improve the accuracy and the applicability of the opening and closing state judgment.
In order to achieve the above object, the present invention adopts the following technical scheme:
in a first aspect, the present invention provides a method for discriminating the opening and closing states of an isolating switch based on an event camera, comprising the following steps:
step 1: acquiring event stream data output by an event camera;
step 2: preprocessing the event stream data obtained in the step 1, and removing noise in the event stream;
step 3: carrying out space-time point cloud coding on the event stream data preprocessed in the step 2 to form point cloud data of time and space dimensions;
step 4: and (3) analyzing the point cloud data obtained in the step (3) by using a deep learning model, and judging the opening and closing states of the isolating switch.
Further, the event stream data specifically includes: event trigger coordinates, event trigger time stamp, and event trigger polarity, wherein the event trigger polarity comprises: an event-triggered positive polarity, which indicates that the light intensity is brightening beyond the trigger threshold, and an event-triggered negative polarity, which indicates that the light intensity is darkening beyond the trigger threshold.
Further, step 2 pre-processes the event stream data obtained in step 1 to remove noise in the event stream, and specifically includes:
firstly, denoising processing is carried out by using Gaussian filtering: taking each event as a center, taking K nearby points, which are called neighbor points, calculating the average distance between the event and the K neighbor points, and filtering through the standard deviation set in advance;
then, abnormal point screening is carried out through a DBSCAN clustering algorithm, abnormal point screening is carried out based on the DBSCAN clustering algorithm, and firstly, a neighborhood radius Eps and a neighborhood data object number threshold MinPts are required to be set in advance, and the specific steps are as follows:
(1) An event point p is arbitrarily selected from the event data;
(2) Using the selected event point p as a core point, finding out event points with reachable p density, and forming a cluster;
(3) If the selected event point p is an edge point, selecting another event point;
(4) Repeating the steps (2) and (3) until all points are processed;
finally, through random downsampling, some data are randomly selected from a plurality of types of samples, so that a relatively complete action is represented by a smaller event number.
Further, step 4: and 3, analyzing the point cloud data obtained in the step 3 by using a deep learning model to judge the opening and closing state of the isolating switch, wherein the method specifically comprises the following steps of:
step 4.1: manufacturing an isolating switch opening and closing data set;
step 4.2: establishing an isolating switch opening and closing state judging model;
step 4.3: training the disconnecting switch opening and closing state judging model established in the step 4.2 based on the disconnecting switch opening and closing data set manufactured in the step 4.1;
step 4.4: and (3) identifying the opening and closing state of the isolating switch in real time by using the trained distinguishing model of the opening and closing state of the isolating switch in the step (4.3).
Further, the action categories of the data set number of the opening and closing of the isolating switch in the step 4.1 are divided into four categories of normal opening of the isolating switch, abnormal opening of the isolating switch and abnormal opening of the isolating switch.
Further, the disconnecting switch opening and closing state discrimination model established in the step 4.2 is a PointConv point cloud classification model, and the specific establishing steps include:
firstly, establishing a sharing coding module PointConv_1, convolving input three-dimensional space-time event point cloud data, wherein the PointConv_1 is formed by combining MLP and BatchNorm, the PointConv_1 is subjected to dimension increasing operation to 64, 64 and 128 respectively, and then sequentially establishing a PointConv_2 module and a PointConv_3 module for convolving, so that enough characteristics are obtained after passing through the three modules, and the dimension of the data characteristics is increased to 1024 dimensions;
then establishing a Linear connection module, reducing the data to 512, performing inactivation operation through a Dropout module, and reducing the data to 256 by using the Linear connection module;
finally, carrying out state discrimination with a four-class full-connection model;
when a section of three-dimensional space-time event point cloud data is input for model reasoning, the PointConv network outputs probability values of four categories, namely normal closing of the isolating switch, normal opening of the isolating switch, abnormal closing of the isolating switch and abnormal opening of the isolating switch, and the probability value is the largest and is taken as a final judging result.
Further, step 4.3: training the disconnecting switch opening and closing state judging model established in the step 4.2 based on the disconnecting switch opening and closing data set manufactured in the step 4.1 specifically comprises the following steps:
firstly, the data set is according to the training set, the testing set and the verification set proportion 7:2:1, adopting a PointConv model during training, and carrying out data enhancement on a data set, wherein the data enhancement comprises the following steps: adding Gaussian noise, random displacement, random rotation and random scaling;
the PointConv model adopts a cross entropy loss function and adopts SGD random gradient descent in the training process, cross entropy can measure the difference degree of two different probability distributions in the same random variable, the weight is continuously adjusted along with the training, and the loss value is continuously reduced until an optimal model is obtained.
Further, step 4.4: the identification of the opening and closing state of the isolating switch is carried out in real time by utilizing the trained distinguishing model of the opening and closing state of the isolating switch in the step 4.3, and the method specifically comprises the following steps:
firstly, deploying an isolating switch opening and closing state judging model into an intelligent computing hardware device, receiving real-time event stream data through a step 1, carrying out data processing on the event stream data through a step 2, carrying out data encoding on the event stream data through a step 3, finally inputting the event stream data into a trained isolating switch opening and closing state judging model, calling an AI chip through the intelligent computing hardware to complete model reasoning, and outputting the opening and closing state of the isolating switch.
A method for judging the opening and closing state of an isolating switch based on an event camera comprises the event camera and a core board;
the event camera is used for monitoring the opening and closing state of the high-voltage isolating switch in real time to obtain event stream data;
the core board is used for preprocessing event stream data acquired by the event camera, removing noise in the event stream, performing space-time point cloud coding on the preprocessed event stream data to form point cloud data of time and space dimensions, and analyzing the obtained point cloud data by using a deep learning model to judge the opening and closing state of the isolating switch.
Further, the device also comprises a network port and a power supply which are connected with the core board;
the network port is used for completing data transmission and data reception through a network;
the power supply is responsible for supplying power to the core board and the event camera.
In general, the above technical solutions conceived by the present invention, compared with the prior art, enable the following beneficial effects to be obtained:
1. under an open-air transformer substation scene, strong light interference and night image blurring can cause poor extraction effect of the edges of the visible light images, so that the judgment error judgment rate based on the opening and closing state of the visible light standard camera is increased, and the actual service requirements are difficult to meet. The invention provides a method for judging the opening and closing state of an isolating switch based on an event camera, which uses the event camera to replace a standard camera, and in the aspect of event data processing, preprocessing of event streams is completed through Gaussian filtering, DBSCAN clustering and random downsampling, so that the influence of strong light and light at night on state capture is solved, and the influence of noise points is removed through filtering and clustering algorithms. In the aspect of an identification algorithm, the PointConv point cloud identification network is used for judging the opening and closing state, the input event stream data is preprocessed and space-time point cloud coded, and finally the accuracy of 98.5% is achieved on a verification set, so that the accuracy of judging the opening and closing state of the isolating switch is remarkably improved;
2. in addition, the disconnecting switch opening and closing state judging device based on the event camera is convenient to install, is not influenced by high-voltage magnetism, is safe and reliable to use and has high applicability.
Drawings
Fig. 1 is a general flow chart of a method for judging the opening and closing states of an isolating switch based on an event camera according to an embodiment of the present invention;
fig. 2 is a network structure diagram of an event camera-based disconnecting switch opening and closing state discrimination algorithm according to an embodiment of the present invention;
FIG. 3 is a graph of model index comparisons provided by embodiments of the present invention;
fig. 4 is an example of opening and closing of an isolating switch based on event flow according to an embodiment of the present invention;
fig. 5 is a schematic structural diagram of an isolating switch on-off state discriminating device based on an event camera according to an embodiment of the present invention;
fig. 6 is an application schematic diagram of an isolating switch opening and closing state discriminating device based on an event camera according to an embodiment of the present invention.
Detailed Description
For the purpose of making the objects, technical solutions and advantages of the embodiments of the present invention more apparent, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention, and it is apparent that the described embodiments are some embodiments of the present invention, but not all embodiments of the present invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.
The embodiment of the invention provides a method for judging the opening and closing state of an isolating switch based on an event camera, which is based on event stream data (also called asynchronous space-time pulse signals) output by the event camera, and firstly noise in the event stream is removed through preprocessing; then encoding the event stream data to form point cloud data of time dimension and space dimension; and finally, analyzing the point cloud data by using a deep learning model to judge the opening and closing states of the isolating switch.
Specifically, as shown in fig. 1, the method includes the following steps:
step 1: and acquiring event stream data output by the event camera.
The on-off state of the high-voltage isolating switch is monitored in real time through the event camera to obtain event stream data, wherein the event stream data output by the event camera specifically comprises: event trigger coordinates, event trigger time stamp, and event trigger polarity, wherein the event trigger polarity comprises: an event-triggered positive polarity, which indicates that the light intensity is brightening beyond the trigger threshold, and an event-triggered negative polarity, which indicates that the light intensity is darkening beyond the trigger threshold.
Step 2: preprocessing the event stream data obtained in the step 1, and removing noise in the event stream.
The event camera is very sensitive to the light intensity change in the environment, is inevitably easy to be influenced by an optical hardware circuit and environmental factors, and is mixed with a large amount of noise while outputting action information, so that the subsequent neural network training precision is influenced. The preprocessing of the event stream data specifically comprises:
firstly, denoising processing is carried out by using Gaussian filtering, and the specific steps of denoising event stream data based on Gaussian filtering are as follows: taking each event as a center, taking K nearby points, which are called neighbor points, and then calculating the average distance between the event and the K neighbor points, and filtering through standard deviation set in advance.
Then, abnormal point screening is carried out through a DBSCAN clustering algorithm, abnormal point screening is carried out based on the DBSCAN clustering algorithm, firstly, the neighborhood radius Eps needs to be set in advance, and the threshold MinPts of the number of data objects in the neighborhood is set as follows:
(1) An event point p is arbitrarily selected from the event data;
(2) Using the selected event point p as a core point, finding out event points with reachable p density, and forming a cluster;
(3) If the selected event point p is an edge point, selecting another event point;
(4) Repeating the steps (2) and (3) until all points are processed.
Finally, through random downsampling, the idea is simpler, namely, some data are randomly selected from a plurality of types of samples, and in this way, a relatively complete action can be represented by a smaller event number, so that support is provided for the following flow.
Step 3: and (3) carrying out space-time point cloud coding on the event stream data preprocessed in the step (2) to form point cloud data of time and space dimensions.
The purpose of space-time point cloud coding is to eliminate adverse effects caused by singular sample data, firstly, normalize space coordinates, then progressively encode time coordinates according to sequence, and finally obtain coded three-dimensional space-time event point cloud data.
Let an event be defined as Ei (Xi, yi, ti, pi), where Xi, yi represent spatial coordinates, ti represent time stamps, pi represent polarity values, then space-time coded, the coded event is denoted as Ei (Xi, yi, zi, pi), let N number of events in this period of time, W width of frame image, H height of frame image, the detailed coding rules are as follows:
X i ＝x i /W (2)
Y i ＝(H-y i )/H (3)
step 4: and (3) analyzing the point cloud data obtained in the step (3) by using a deep learning model, and judging the opening and closing states of the isolating switch. The method specifically comprises the following steps:
step 4.1: making a data set of opening and closing of an isolating switch
Aiming at the operation state of the isolating switch in reality, the action categories in the data set are divided into four categories (the duration of the action process is about 6 s) of normal closing of the isolating switch, normal opening of the isolating switch and abnormal closing of the isolating switch. In consideration of the influence of the real environment of the transformer substation on the disconnecting switch judging algorithm, different field environments are set up to be respectively collected in the data collection process, wherein the data collection process mainly comprises two parts, namely, data under the ideal condition is collected, and for the conditions of normal light and simple background, the collection position is positioned on the front face of the disconnecting switch and is mainly used for testing the effectiveness of the algorithm. Secondly, collecting data under the complex scene condition, setting data under single or multiple bad conditions, such as insufficient illumination at night, direct sunlight camera at daytime, more power equipment on the background of a picture, collecting positions on the side face of a disconnecting switch, a slightly far distance and the like, and mainly testing the reliability of an algorithm.
As shown in fig. 4, the motion example image captured by the event camera obtained in the steps 1, 2 and 3 above, wherein the two-dimensional graph is a frequency accumulated graph of the event, and the three-dimensional graph is a three-dimensional display of all events of which one motion occurs under the (x, y, t) condition consisting of the spatial coordinates and the time axis, wherein the red color point is displayed to represent the polarity as 1, and the green color point is displayed to represent the polarity as 0.
Step 4.2: establishing a distinguishing model of the opening and closing states of the isolating switch
Because the event stream data output by the event camera is similar to the point cloud in three-dimensional space data structure distribution and is sparser in time domain distribution, the embodiment of the invention adopts a structure based on a point cloud network to process the event stream data. As shown in a flowchart of a disconnecting switch opening and closing state judging method based on an event camera in fig. 1, after preprocessing event stream data and space-time point cloud coding, point cloud data of time dimension and space dimension, namely a set of three-dimensional space points, are formed.
The three-dimensional space-time event point cloud data at the moment reserves the original space geometric structure, can well express the surface characteristics and depth information of objects, and has the characteristics of uneven space distribution and random data. Therefore, the judgment of the opening and closing states of the isolating switch can be completed by establishing a point cloud classification model.
In the current point cloud classification model, the PointConv network can efficiently use a convolution network in non-uniformly sampled point clouds, and meanwhile, the arrangement invariance and conversion invariance of the point clouds can be guaranteed, so that the PointConv point cloud classification model is innovatively selected to judge the opening and closing states of the isolating switch.
The model is built according to the PointConv network structure diagram in FIG. 2, firstly, a shared coding module PointConv_1 is built to convolve the input three-dimensional space-time event point cloud data, the PointConv_1 module is formed by combining MLP and BatchNorm, the PointConv_1 can perform dimension increasing operation, the dimension increasing operation is respectively carried out to 64, 64 and 128, then, pointConv_2 and PointConv_3 modules are built in sequence for convolving, and enough characteristics can be obtained after the three modules are used, and the dimension of the data characteristics is increased to 1024 dimensions.
Then, a Linear connection module is established to reduce the data dimension to 512, the Dropout module is used for carrying out the deactivation operation, the Linear connection module is used for reducing the data dimension to 256, the manufactured isolating switch opening and closing data set can be seen, and the isolating switch opening and closing action categories are divided into four categories of normal closing of the isolating switch, normal opening of the isolating switch, abnormal closing of the isolating switch and abnormal opening of the isolating switch at present, so that the last step of model establishment is to carry out state discrimination with a four-category full connection model.
When a section of three-dimensional space-time event point cloud data is input for model reasoning, the PointConv network outputs probability values of four categories, namely normal closing of the isolating switch, normal opening of the isolating switch, abnormal closing of the isolating switch and abnormal opening of the isolating switch, and the probability value is the largest and is taken as a final judging result.
The PointConv module in the model is a new convolution operation, as shown in equation (5):
wherein: f is the characteristic of a point within the local area G around a point P (x, y, z), S is a function of density, the coordinates of each point are input, the output is the inverse density coefficient corresponding to each point, W is the weight corresponding to the characteristic of the point, and the weight coefficient of the point is returned by the input point through the coordinates of the point learned by MLP.
And finally, an output layer of the model adopts a softMax classifier to obtain classification results of four actions of opening and closing of the isolating switch, and a function of the softMax is shown in a formula (6):
wherein: oi represents the output value corresponding to the i-th node, cz is the total category number, and cz=4 in this embodiment.
Step 4.3: and training the disconnecting switch opening and closing state judging model established in the step 4.2 based on the disconnecting switch opening and closing data set manufactured in the step 4.1.
Firstly, the data set is according to the training set, the testing set and the verification set proportion 7:2:1, performing random division in a mode of 1. Since both the quality and the quantity of data are necessary conditions for training a high-performance deep neural network, the method is more important especially in three-dimensional deep learning. Therefore, the data of the data set is enhanced during training, and the generalization capability of the model is further improved. The data enhancements specifically used are shown below:
(1) Adding Gaussian noise
In the point cloud data, gaussian noise of the same dimension is first generated and then added to the original data.
(2) Random displacement
The point cloud is randomly moved for a certain distance in the x, y and z directions. For each point cloud, three random numbers are first generated, representing the random displacements of the point cloud in the x, y and z directions. All points in the same point cloud are then given the previously generated random displacements in the x, y, z coordinates. The point cloud after the operation can be randomly displaced in any direction.
(3) Randomly rotate
Like displacement, rotation is also performed in three directions, x, y, and z. For each point cloud data, random rotation angles in three directions are first randomly generated, and then each point in the point cloud is rotated according to the randomly generated angles.
(4) Random scaling
For the scaling operation of a point cloud, a scaling center is randomly generated, then a scaling scale is randomly generated, and finally the distance from each point in the point cloud to the scaling center is scaled according to the scaling scale, so that new point cloud data are obtained.
The PointConv model adopts a cross entropy loss function and adopts SGD random gradient descent in the training process, cross entropy can measure the difference degree of two different probability distributions in the same random variable, and SGD is simple and efficient and can jump out of a local optimal solution. And continuously adjusting the weight along with training, and continuously reducing the loss value until an optimal model is obtained. The isolation switch opening and closing data set manufactured according to the step 4 is shown in fig. 3, and is trained by using PointNet, pointNet ++ and the PointConv model used by the invention. Through verification, the overall classification accuracy of PointNet is 97.2%, the overall classification accuracy of PointNet++ is 98%, the PointConv model accuracy adopted by the method is 98.5%, the convolution network is effectively utilized by the PointConv, meanwhile, the arrangement invariance and conversion invariance of point clouds are guaranteed, and a conclusion that the classification model accuracy is high and the processing mode is concise can be obtained relative to other classification models in a table.
Step 4.4: and (3) identifying the opening and closing state of the isolating switch in real time by using the trained distinguishing model of the opening and closing state of the isolating switch in the step (4.3).
Firstly, deploying an isolating switch opening and closing state judging model into an intelligent computing hardware device, receiving real-time event stream data through a step 1, carrying out data processing on the event stream data through a step 2, carrying out data encoding on the event stream data through a step 3, finally inputting the event stream data into a trained isolating switch opening and closing state judging model, calling an AI chip through the intelligent computing hardware to complete model reasoning, and outputting the opening and closing state of the isolating switch.
Preferably, in the present embodiment, the event camera used is a DAVIS346, and the combination of the camera DVS and the conventional camera incorporates an Active Pixel Sensor (APS), so that RGB images can be output while event stream data is output.
As shown in fig. 5, this embodiment further provides an isolating switch opening and closing state determining device based on an event camera, which includes the event camera, a core board, a network port and a power supply.
The event camera is used for monitoring the opening and closing state of the high-voltage isolating switch in real time to obtain event stream data;
the core board is used for preprocessing event stream data acquired by the event camera, removing noise in the event stream, performing space-time point cloud coding on the preprocessed event stream data to form point cloud data of time and space dimensions, and analyzing the obtained point cloud data by using a deep learning model to judge the opening and closing state of the isolating switch.
The network port is used for completing data transmission and data reception through a network;
the power supply is responsible for supplying power to the core board and the event camera.
In the existing research, the detection of the opening and closing state of the isolating switch is based on a visible light standard camera, and the method has the advantages of deep research, mature technology, visual judging process, high automation degree, low identification accuracy and large influence of weather and illumination. Under an open-air transformer substation scene, strong light interference and night image blurring can cause poor extraction effect of the edges of the visible light images, so that the judgment misjudgment rate of the opening and closing state is increased, and the actual service requirements are difficult to meet. And the event camera is used for detecting the opening and closing state based on event stream, so that the influence of illumination on image recognition can be ignored.
In order to verify the superiority of the event sensor, an isolating switch opening and closing state detection algorithm based on image frames is constructed, and the isolating switch is detected by using a rotated_master_rcnn rotating target detection model based on RGB image data output by a standard camera.
In the acquisition process of RGB image data, different field environments are set up to acquire respectively in the same mode as the mode of manufacturing the opening and closing data set of the isolating switch in the step 4, firstly, the data under the ideal condition is acquired, the light is normal, the background is simple, the acquisition position is positioned on the front face of the isolating switch, and the method is mainly used for testing the effectiveness of an algorithm. Secondly, collecting data under the complex scene condition, setting data under single or multiple bad conditions, such as insufficient illumination at night, direct sunlight camera at daytime, more power equipment on the background of a picture, collecting positions on the side face of a disconnecting switch, a slightly far distance and the like, and mainly testing the reliability of an algorithm.
Through training, the accuracy rate of the rotated_master_rcnn is 96.1% on the basis of the optimal threshold value of 0.8, and the accuracy rate of the PointConv algorithm based on the event camera is 98.5%, so that a conclusion that the PointConv algorithm model adopting the event camera is high in accuracy and strong in generalization capability compared with the detection algorithm adopting the standard camera can be obtained, as shown in a table 1.
TABLE 1
| Scheme for the production of a semiconductor device | Accuracy rate of | Characteristics of |
| Event stream based | 98.5％ | High adaptability |
| Based on image frames | 96.1％ | Is easy to be interfered by strong light and has blurred images at night |
The disconnecting switch opening and closing state judging device based on the event camera provided by the embodiment is used as terminal equipment for monitoring the high-voltage disconnecting switch of the transformer substation, and performs real-time opening and closing data acquisition when the disconnecting switch works normally, as shown in an application schematic diagram of fig. 6.
The trained disconnecting switch opening and closing state judging model is deployed in the device, the opening and closing state of the disconnecting switch is output through model reasoning, and the opening and closing state of the disconnecting switch is pushed to the WEB end or the mobile end in real time, so that the disconnecting switch with defects is convenient to overhaul or replace, the labor cost is saved, and the detection efficiency is improved.
Preferably, in this embodiment, the core board selects NVIDIA Jetson NANO as the intelligent computing hardware, 128 CUDA cores, 4G memory, to provide excellent speed and energy efficiency for AI computation.
The foregoing is merely illustrative embodiments of the present invention, and the present invention is not limited thereto, and any changes or substitutions that may be easily contemplated by those skilled in the art within the scope of the present invention should be included in the scope of the present invention. Therefore, the protection scope of the present invention should be subject to the protection scope of the claims.
Claims (10)
1. A method for judging the opening and closing state of an isolating switch based on an event camera is characterized by comprising the following steps of: the method comprises the following steps:
step 1: acquiring event stream data output by an event camera;
step 2: preprocessing the event stream data obtained in the step 1, and removing noise in the event stream;
step 3: carrying out space-time point cloud coding on the event stream data preprocessed in the step 2 to form point cloud data of time and space dimensions;
step 4: and (3) analyzing the point cloud data obtained in the step (3) by using a deep learning model, and judging the opening and closing states of the isolating switch.
2. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 1, wherein: the event stream data specifically includes: event trigger coordinates, event trigger time stamp, and event trigger polarity, wherein the event trigger polarity comprises: an event-triggered positive polarity, which indicates that the light intensity is brightening beyond the trigger threshold, and an event-triggered negative polarity, which indicates that the light intensity is darkening beyond the trigger threshold.
3. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 1, wherein: step 2, preprocessing the event stream data acquired in the step 1 to remove noise in the event stream, and specifically includes:
firstly, denoising processing is carried out by using Gaussian filtering: taking each event as a center, taking K nearby points, which are called neighbor points, calculating the average distance between the event and the K neighbor points, and filtering through the standard deviation set in advance;
then, abnormal point screening is carried out through a DBSCAN clustering algorithm, abnormal point screening is carried out based on the DBSCAN clustering algorithm, and firstly, a neighborhood radius Eps and a neighborhood data object number threshold MinPts are required to be set in advance, and the specific steps are as follows:
(1) An event point p is arbitrarily selected from the event data;
(2) Using the selected event point p as a core point, finding out event points with reachable p density, and forming a cluster;
(3) If the selected event point p is an edge point, selecting another event point;
(4) Repeating the steps (2) and (3) until all points are processed;
finally, through random downsampling, some data are randomly selected from a plurality of types of samples, so that a relatively complete action is represented by a smaller event number.
4. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 1, wherein: step 4: and 3, analyzing the point cloud data obtained in the step 3 by using a deep learning model to judge the opening and closing state of the isolating switch, wherein the method specifically comprises the following steps of:
step 4.1: manufacturing an isolating switch opening and closing data set;
step 4.2: establishing an isolating switch opening and closing state judging model;
step 4.3: training the disconnecting switch opening and closing state judging model established in the step 4.2 based on the disconnecting switch opening and closing data set manufactured in the step 4.1;
step 4.4: and (3) identifying the opening and closing state of the isolating switch in real time by using the trained distinguishing model of the opening and closing state of the isolating switch in the step (4.3).
5. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 4, wherein: and 4.1, dividing the action categories of the data set number of the opening and closing of the isolating switch into four categories of normal opening and closing of the isolating switch, normal opening and closing of the isolating switch and abnormal opening and closing of the isolating switch.
6. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 4, wherein: the disconnecting switch opening and closing state judging model established in the step 4.2 is a PointConv point cloud classifying model, and the specific establishing steps comprise:
firstly, establishing a sharing coding module PointConv_1, convolving input three-dimensional space-time event point cloud data, wherein the PointConv_1 is formed by combining MLP and BatchNorm, the PointConv_1 is subjected to dimension increasing operation to 64, 64 and 128 respectively, and then sequentially establishing a PointConv_2 module and a PointConv_3 module for convolving, so that enough characteristics are obtained after passing through the three modules, and the dimension of the data characteristics is increased to 1024 dimensions;
then establishing a Linear connection module, reducing the data to 512, performing inactivation operation through a Dropout module, and reducing the data to 256 by using the Linear connection module;
finally, carrying out state discrimination with a four-class full-connection model;
when a section of three-dimensional space-time event point cloud data is input for model reasoning, the PointConv network outputs probability values of four categories, namely normal closing of the isolating switch, normal opening of the isolating switch, abnormal closing of the isolating switch and abnormal opening of the isolating switch, and the probability value is the largest and is taken as a final judging result.
7. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 4, wherein: step 4.3: training the disconnecting switch opening and closing state judging model established in the step 4.2 based on the disconnecting switch opening and closing data set manufactured in the step 4.1 specifically comprises the following steps:
firstly, the data set is according to the training set, the testing set and the verification set proportion 7:2:1, adopting a PointConv model during training, and carrying out data enhancement on a data set, wherein the data enhancement comprises the following steps: adding Gaussian noise, random displacement, random rotation and random scaling;
the PointConv model adopts a cross entropy loss function and adopts SGD random gradient descent in the training process, cross entropy can measure the difference degree of two different probability distributions in the same random variable, the weight is continuously adjusted along with the training, and the loss value is continuously reduced until an optimal model is obtained.
8. The method for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 4, wherein: step 4.4: the identification of the opening and closing state of the isolating switch is carried out in real time by utilizing the trained distinguishing model of the opening and closing state of the isolating switch in the step 4.3, and the method specifically comprises the following steps:
firstly, deploying an isolating switch opening and closing state judging model into an intelligent computing hardware device, receiving real-time event stream data through a step 1, carrying out data processing on the event stream data through a step 2, carrying out data encoding on the event stream data through a step 3, finally inputting the event stream data into a trained isolating switch opening and closing state judging model, calling an AI chip through the intelligent computing hardware to complete model reasoning, and outputting the opening and closing state of the isolating switch.
9. A method for judging the opening and closing state of an isolating switch based on an event camera is characterized by comprising the following steps of: the system comprises an event camera and a core board;
the event camera is used for monitoring the opening and closing state of the high-voltage isolating switch in real time to obtain event stream data;
the core board is used for preprocessing event stream data acquired by the event camera, removing noise in the event stream, performing space-time point cloud coding on the preprocessed event stream data to form point cloud data of time and space dimensions, and analyzing the obtained point cloud data by using a deep learning model to judge the opening and closing state of the isolating switch.
10. The device for distinguishing the opening and closing states of the isolating switch based on the event camera as claimed in claim 9, wherein: the device also comprises a network port and a power supply which are connected with the core board;
the network port is used for completing data transmission and data reception through a network;
the power supply is responsible for supplying power to the core board and the event camera.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310120669.3A CN116363411A (en) | 2023-02-16 | 2023-02-16 | Method and device for judging opening and closing states of isolating switch based on event camera |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310120669.3A CN116363411A (en) | 2023-02-16 | 2023-02-16 | Method and device for judging opening and closing states of isolating switch based on event camera |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN116363411A true CN116363411A (en) | 2023-06-30 |
Family
ID=86930741
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310120669.3A Pending CN116363411A (en) | 2023-02-16 | 2023-02-16 | Method and device for judging opening and closing states of isolating switch based on event camera |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116363411A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117155705A (en) * | 2023-10-27 | 2023-12-01 | 三峡高科信息技术有限责任公司 | Data transmission system, method, equipment and storage medium based on internet of things gateway |
-
2023
- 2023-02-16 CN CN202310120669.3A patent/CN116363411A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117155705A (en) * | 2023-10-27 | 2023-12-01 | 三峡高科信息技术有限责任公司 | Data transmission system, method, equipment and storage medium based on internet of things gateway |
| CN117155705B (en) * | 2023-10-27 | 2024-02-02 | 三峡高科信息技术有限责任公司 | Data transmission system, method, equipment and storage medium based on internet of things gateway |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108109385B (en) | System and method for identifying and judging dangerous behaviors of power transmission line anti-external damage vehicle | |
| CN106356757B (en) | A kind of power circuit unmanned plane method for inspecting based on human-eye visual characteristic | |
| CN109635744A (en) | A kind of method for detecting lane lines based on depth segmentation network | |
| CN108596169B (en) | Block signal conversion and target detection method and device based on video stream image | |
| CN110378288A (en) | A kind of multistage spatiotemporal motion object detection method based on deep learning | |
| CN106407903A (en) | Multiple dimensioned convolution neural network-based real time human body abnormal behavior identification method | |
| CN110070530A (en) | A kind of powerline ice-covering detection method based on deep neural network | |
| CN109636758B (en) | Water surface floater detection method based on space-time dynamic operator | |
| CN106096603A (en) | A kind of dynamic flame detection method merging multiple features and device | |
| CN102831393A (en) | Rapid image recognizing method of power tower pole outline | |
| CN107463931A (en) | A kind of real-time pointer instrument reading method and device based on ARM platforms | |
| CN114842363B (en) | Identification method and system for key power equipment in digital twin platform area | |
| CN112183472A (en) | Method for detecting whether test field personnel wear work clothes or not based on improved RetinaNet | |
| CN116363411A (en) | Method and device for judging opening and closing states of isolating switch based on event camera | |
| CN112232356A (en) | Event camera denoising method based on cluster degree and boundary characteristics | |
| CN111145222A (en) | Fire detection method combining smoke movement trend and textural features | |
| CN103489012A (en) | Crowd density detecting method and system based on support vector machine | |
| CN105825219A (en) | Machine vision detection system | |
| CN110880184A (en) | Method and device for carrying out automatic camera inspection based on optical flow field | |
| CN111767826A (en) | Timing fixed-point scene abnormity detection method | |
| CN114241310B (en) | Improved YOLO model-based intelligent identification method for piping dangerous case of dike | |
| CN111931577A (en) | Intelligent inspection method for specific foreign matters of power grid line | |
| CN111582076A (en) | Picture freezing detection method based on pixel motion intelligent perception | |
| CN110197131A (en) | A method of remotely monitoring lightning arrester discharge action frequency based on machine vision | |
| CN114926456A (en) | Rail foreign matter detection method based on semi-automatic labeling and improved deep learning |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |