A method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning

Abstract

The present invention discloses a method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning, and the specific steps are as follows: Step S1: Set auxiliary markers and collect a high-speed image sequence; Step S2: Input the high-speed image sequence into a trained deep learning model to track the motion trajectory of the auxiliary markers; Step S3: Calculate the stroke-time curve of the moving contact of the circuit breaker according to the trajectory of the auxiliary markers. By adopting the above method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning, non-contact measurement is used, and the motion characteristics of the moving contact of the circuit breaker are measured by tracking the components that move synchronously with the moving contact of the circuit breaker, and the problems of target size change, background color interference, and camera jitter can be effectively addressed to achieve accurate tracking.

Description

Technical Field

[0001] The present invention relates to the technical field of detection of moving contacts of circuit breakers, and more particularly to a method for measuring the motion characteristics of moving contacts of circuit breakers based on deep learning. Background Art

[0002] In the power system, high-voltage circuit breakers are crucial. Once a failure occurs, it may cause potential hazards to power grid operations and even threaten the lives of relevant personnel. The reliability assessment of high-voltage circuit breakers by the International Council on Large Electric Systems shows that 64.8% of the main failures of circuit breakers are operating mechanical failures. Measuring the motion characteristic parameters of the moving contacts of circuit breakers can evaluate the operating performance of circuit breakers and is an important technical means for detecting mechanical failures of circuit breakers.

[0003] At present, most of the developed and mature measuring tools for the motion characteristic parameters of the moving contacts of circuit breakers adopt contact measurement methods, including: electromagnetic oscillators, drum meters, slide wire rheostats combined with light oscilloscopes, grating displacement sensors (grating rulers), linear sensors (slide wire rheostats), angle sensors (rotary potentiometers), etc. Different circuit breaker models require different sizes of sensor fixtures, the sensors are difficult to install, and the debugging process is complex. Moreover, the external dimensions of vacuum circuit breakers are continuously shrinking, and the space available for installing displacement sensors near the moving contacts is becoming more and more limited, making it increasingly difficult to install traditional sensors on the operating mechanism and the main shaft.

[0004] The non-contact measurement methods in the prior art, such as semiconductor laser displacement sensors, infrared ranging sensors, and ultrasonic displacement sensors, etc., but currently lack professional instruments supporting the high-voltage switch field. The non-contact measurement combined with computer vision technology in the prior art is limited by the performance of the camera, and too few motion images are collected, and the stroke of the circuit breaker is not fully analyzed. The traditional correlation matching algorithm is used to track the target, and the selected tracking punctuation is similar to the background color, resulting in large matching errors and slow speeds. In addition, using the gray pixel values of the image for image matching has the problems of large computational complexity and slow speed, and needs to be improved in terms of matching accuracy and speed. Summary of the Invention

[0005] The object of the present invention is to provide a method for measuring the motion characteristics of moving contacts of circuit breakers based on deep learning, which adopts non-contact measurement, tracks the components that move synchronously with the moving contacts of the circuit breaker to achieve the measurement of the motion characteristics of the moving contacts of the circuit breaker, and can effectively cope with the problems of target size change, background color interference, and camera jitter, and achieve accurate tracking.

[0006] To achieve the above object, the present invention provides a method for measuring the motion characteristics of moving contacts of circuit breakers based on deep learning, and the specific steps are as follows:

[0007] Step S1: Set up auxiliary markers and collect a high-speed image sequence;

[0008] Step S2: Input the high-speed image sequence into the trained deep learning model to track the motion trajectory of the auxiliary markers;

[0009] Step S3: Calculate the travel time curve of the moving contact of the circuit breaker based on the trajectory of the auxiliary markers.

[0010] Preferably, step S1 is specifically as follows:

[0011] Step S11: Respectively set a first auxiliary marker and a second auxiliary marker on the insulating pull rod and the rotating shaft that move synchronously with the moving contact of the circuit breaker. There is one said first auxiliary marker on the linearly moving insulating pull rod, and two second auxiliary markers are set on the rotating shaft with rotational motion;

[0012] Step S12: Adjust the position and field of view of the high-speed camera to collect the motion images of the auxiliary markers and generate a high-speed image sequence.

[0013] Preferably, step S2 is specifically as follows:

[0014] Step S21: Feature extraction, input the high-speed image sequence into the improved ResNet50 network to extract multi-layer feature maps;

[0015] Step S22: Feature fusion, fuse the multi-layer feature maps extracted through the pyramid feature enhancement network to fuse multi-layer features;

[0016] Step S23: Feature enhancement, input the fused multi-layer features into the feature enhancement network to obtain a feature-enhanced fused vector;

[0017] Step S24: Input the fused vector into the prediction head to obtain the motion trajectory of the tracked auxiliary markers.

[0018] Preferably, the improved ResNet50 network replaces the 7*7 convolution in the first layer of the input part of the traditional ResNet50 network with three 3*3 convolutions;

[0019] The pyramid feature enhancement network is a double-layer pyramid structure. The double-layer pyramid structure adds a path from bottom to top and then from top to bottom in the traditional pyramid structure, and uses lateral connections to fuse the features of each layer;

[0020] The feature enhancement network includes a CECA / CCFA structure and a CCFA structure. The CECA / CCFA structure includes two CECA structures and two CCFA structures. The fused multi-layer features are input into the CECA / CCFA structure and repeated four times, and then the two output branches are fused through the CCFA structure;

[0021] The prediction head includes a classification branch for predicting whether each position contains an auxiliary marker and a regression branch for predicting the regression box. Both the classification branch and the regression branch are composed of three linear layers and the ReLU activation function. The classification branch is used to predict whether the tracking target exists at each position, and the regression branch is used to predict the regression box.

[0022] Preferably, a secondary sampling layer is provided in both the CECA structure and the CCFA structure. The process of secondary sampling is as follows:

[0023] First, reshape the one-dimensional vector into a two-dimensional feature map;

[0024] Then, use depthwise separable convolution to perform secondary sampling on the feature map. The depthwise separable convolution is specifically divided into: traversing a 3×3 convolution kernel in each channel to obtain an output with the same number of channels as the original feature map, and using a 1×1×number of channels point convolution to expand the depth.

[0025] Preferably, in step S3,

[0026] When tracking the movement of the insulating pull rod, the first auxiliary marker moves in a straight line, and the change in the centroid coordinates is calculated using the Euclidean distance. The calculation formula is as follows:

[0027]

[0028] where D is the distance from the initial position of the centroid of the first auxiliary marker to the end position, (x1, y1) is the coordinate of the centroid of the first auxiliary marker at the initial position, (x2, y2) is the coordinate of the centroid of the first auxiliary marker at the end position. Combining with the camera shooting rate, a travel time curve graph of the moving contact of the circuit breaker is generated;

[0029] When tracking the movement of the rotating shaft, the rotation angle of the line connecting the centroids of the two second auxiliary markers is the rotation angle of the rotating shaft. Line a is the line connecting the centroids of the two second auxiliary markers at the initial position, and the slope of line a is k1. Line b is the line connecting the centroids of the two second auxiliary markers at the end position, and the slope of line b is k2. The included angle between line a and line b is θ, θ ∈ [0, π), and θ is the angular displacement of the moving contact of the circuit breaker.

[0030] When both k1 and k2 exist, the calculation formula is as follows:

[0031]

[0032] When both k1 and k2 do not exist, line a is parallel to line b, and θ = 0;

[0033] When k1 does not exist and the included angle α2 between line b and the horizontal axis is less than π / 2, θ = α2 + π / 2;

[0034] When k1 does not exist and the included angle α2 between the line b and the horizontal axis is greater than π / 2, θ = α2 - π / 2;

[0035] When k2 does not exist and the included angle α1 between the line a and the horizontal axis is less than π / 2, θ = α1 - π / 2;

[0036] When k2 does not exist and the included angle α1 between the line a and the horizontal axis is greater than π / 2, θ = α1 + π / 2.

[0037] Therefore, the present invention adopts the above-mentioned method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning, and has the following beneficial effects:

[0038] (1) Non-contact measurement is carried out by using deep learning technology. An auxiliary marker is set, and the components moving synchronously with the moving contact of the circuit breaker are tracked to realize the measurement of the motion characteristics of the moving contact of the circuit breaker. A high-speed image sequence is collected. The auxiliary marker is a sticker with distinct color, which can effectively cope with the problems of target size change and background color interference. At the same time, a high-speed image sequence is collected to increase the number of collected images, and the stroke analysis is more sufficient.

[0039] (2) An improved ResNet50 network is used to extract multi-layer feature maps, enhancing the learning ability of features while reducing the computational amount.

[0040] (3) A pyramid feature enhancement network is used to fuse multi-layer features, enhancing the adaptability to size.

[0041] (4) The feature enhancement network consists of a CECA / CCFA structure and a CCFA structure. In the secondary sampling layer, depthwise separable convolution is used to replace the original position linear projection of multi-head self-attention, enhancing the modeling of local spatial context, while compressing the computational amount and compensating for the information loss caused by the reduction of resolution.

[0042] The technical solution of the present invention will be further described in detail below with reference to the drawings and embodiments. Description of the Drawings

[0043] Figure 1 It is a block diagram of a method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning according to the present invention;

[0044] Figure 2 It is a position diagram of the auxiliary marker and the high-speed camera set according to the present invention;

[0045] Figure 3 It is a schematic structural diagram of the original position linear projection layer;

[0046] Figure 4 It is a schematic structural diagram of the secondary sampling layer;

[0047] Figure 5It is a schematic diagram of the CECA structure;

[0048] Figure 6 It is a schematic diagram of the CCFA structure.

[0049] Reference numerals

[0050] 1. Second auxiliary marker; 2. First auxiliary marker; 3. High-speed camera; 4. Insulating pull rod; 5. Rotating shaft. Detailed implementation manners

[0051] Embodiment

[0052] Figure 1 It is a block diagram of a method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning. As shown in the figure, a method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning includes the following specific steps:

[0053] Step S1: Set auxiliary markers and collect a high-speed image sequence.

[0054] Step S11: Respectively set a first auxiliary marker 2 and a second auxiliary marker 1 on the insulating pull rod 4 and the rotating shaft 5 that move synchronously with the moving contact of the circuit breaker. As Figure 2 shown, Figure 2 It is a diagram showing the setting positions of the auxiliary markers and the high-speed camera of the present invention. A first auxiliary marker 2 is set on the linearly moving insulating pull rod 4, and two second auxiliary markers 1 are set on the rotating shaft 5 that rotates.

[0055] Step S12: Adjust the position and field of view of the high-speed camera 3 to collect the motion images of the auxiliary markers and generate a high-speed image sequence.

[0056] Step S2: Input the high-speed image sequence into the trained deep learning model to track the motion trajectories of the auxiliary markers.

[0057] Step S21: Feature extraction. Input the high-speed image sequence into the improved ResNet50 network to extract multi-layer feature maps. The improved ResNet50 network replaces the 7*7 convolution in the first layer of the input part of the traditional ResNet50 network with three 3*3 convolutions, enhancing the feature learning ability and reducing the computational amount at the same time.

[0058] Step S22: Feature fusion. The feature maps extracted from the second layer to the fifth layer are fused with multi-layer features through the pyramid feature enhancement network. The pyramid feature enhancement network is a double-layer pyramid structure. The double-layer pyramid structure adds a path from bottom to top and then from top to bottom in the traditional pyramid structure, and uses lateral connections to fuse the features of each layer to enhance the adaptability to sizes.

[0059] Step S23: Feature enhancement. The fused multi-layer features are input into a feature enhancement network to obtain a feature-enhanced fused vector. The feature enhancement network includes a CECA / CCFA structure and a CCFA structure. The CECA / CCFA structure includes two CECA structures and two CCFA structures. The original position linear projection of the multi-head self-attention (MHSA) is replaced by depth-wise separable convolutions to enhance the modeling of local spatial context and compress the computational amount, resulting in the CECA / CCFA structure. The fused multi-layer features are input into the CECA / CCFA structure and repeated four times, and then the two output branches are fused through the CCFA structure.

[0060] Both the CECA structure and the CCFA structure are provided with a secondary sampling layer. The process of secondary sampling is as follows:

[0061] First, reshape the one-dimensional vector into a two-dimensional feature map. Then, use depth-wise separable convolutions to perform secondary sampling on the feature map. The depth-wise separable convolution is specifically divided into: traversing a 3×3 convolution kernel in each channel to obtain an output with the same number of channels as the original feature map, and using a 1×1×number of channels point convolution to expand the depth. The convolution stride is 2, which further extracts local spatial context information and compensates for the information loss caused by the resolution reduction.

[0062] Step S24: Input the fused vector into the prediction head to obtain the motion trajectory of the tracking auxiliary marker. The prediction head includes a classification branch for predicting whether each position contains an auxiliary marker and a regression branch for predicting the regression box. Both the classification branch and the regression branch are composed of 3 linear layers and the RELU activation function. The classification branch is used to predict whether the tracking target exists at each position, and the regression branch is used to predict the regression box.

[0063] Step S3: Calculate the travel time curve of the moving contact of the circuit breaker according to the trajectory of the auxiliary marker.

[0064] When tracking the movement of the insulating rod, the first auxiliary marker moves in a straight line, and the change in the centroid coordinates is calculated using the Euclidean distance. The calculation formula is as follows:

[0065]

[0066] where D is the distance from the initial position of the centroid of the first auxiliary marker to the end position, (x1, y1) is the coordinate of the centroid of the first auxiliary marker at the initial position, and (x2, y2) is the coordinate of the centroid of the first auxiliary marker at the end position. Combining with the camera shooting rate, a travel time curve graph of the moving contact of the circuit breaker is generated;

[0067] When tracking the motion of the rotating shaft, the rotation angle of the connecting line of the centroids of the two second auxiliary markers is the rotation angle of the rotating shaft. The straight line a is the connecting line of the centroids of the two second auxiliary markers at the initial position, the slope of the straight line a is k1, the straight line b is the connecting line of the centroids of the two second auxiliary markers at the end position, the slope of the straight line b is k2, and the included angle between the straight line a and the straight line b is θ, θ ∈ [0, π), and θ is the angular displacement of the moving contact of the circuit breaker.

[0068] When both k1 and k2 exist, the calculation formula is as follows:

[0069]

[0070] When both k1 and k2 do not exist, the straight line a is parallel to the straight line b, and θ = 0;

[0071] When k1 does not exist and the included angle α2 between the straight line b and the horizontal axis is < π / 2, θ = α2 + π / 2;

[0072] When k1 does not exist and the included angle α2 between the straight line b and the horizontal axis is > π / 2, θ = α2 - π / 2;

[0073] When k2 does not exist and the included angle α1 between the straight line a and the horizontal axis is < π / 2, θ = α1 - π / 2;

[0074] When k2 does not exist and the included angle α1 between the straight line a and the horizontal axis is > π / 2, θ = α1 + π / 2.

[0075] Therefore, the present invention adopts the above-mentioned method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning, uses non-contact measurement, and tracks the components that move synchronously with the moving contact of the circuit breaker to measure the motion characteristics of the moving contact of the circuit breaker, and can effectively cope with the problems of target size change, background color interference, and camera jitter, and achieve accurate tracking.

[0076] Finally, it should be noted that: the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that: they can still modify or equivalently replace the technical solutions of the present invention, and these modifications or equivalent replacements cannot make the modified technical solutions deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning, characterized in that, The specific steps are as follows: Step S1: Set up auxiliary markers and collect a high-speed image sequence; Step S2: Input the high-speed image sequence into the trained deep learning model to track the motion trajectory of the auxiliary markers. Specifically, Step S2 is as follows: Step S21: Feature extraction. Input the high-speed image sequence into the improved ResNet50 network to extract multi-layer feature maps; Step S22: Feature fusion. Fuse the multi-layer feature maps extracted through the pyramid feature enhancement network; Step S23: Feature enhancement. Input the fused multi-layer features into the feature enhancement network to obtain a feature-enhanced fusion vector; Step S24: Input the fusion vector into the prediction head to obtain the motion trajectory of the tracked auxiliary markers; The improved ResNet50 network replaces the 7*7 convolution in the first layer of the input part of the traditional ResNet50 network with three 3*3 convolutions; The pyramid feature enhancement network is a double-layer pyramid structure. The double-layer pyramid structure adds a bottom-up and then top-down path to the traditional pyramid structure and uses lateral connections to fuse the features of each layer; The feature enhancement network includes a CECA / CCFA structure and a CCFA structure. The CECA / CCFA structure includes two CECA structures and two CCFA structures. The fused multi-layer features are input into the CECA / CCFA structure and repeated four times, and then the two output branches are fused through the CCFA structure; The prediction head includes a classification branch for predicting whether each position contains an auxiliary marker and a regression branch for predicting the regression box. Both the classification branch and the regression branch are composed of three linear layers and the RELU activation function. The classification branch is used to predict whether the tracking target exists at each position, and the regression branch is used to predict the regression box; Step S3: Calculate the travel time curve of the moving contact of the circuit breaker based on the trajectory of the auxiliary markers.

2. The method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning according to claim 1, characterized in that, Specifically, Step S1 is as follows: Step S11: Set a first auxiliary marker and a second auxiliary marker on the insulating rod and the rotating shaft that move synchronously with the moving contact of the circuit breaker. One first auxiliary marker is set on the linearly moving insulating rod, and two second auxiliary markers are set on the rotating shaft with rotational motion; Step S12: Adjust the position and field of view of the high-speed camera to collect the motion images of the auxiliary markers and generate a high-speed image sequence.

3. The method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning according to claim 2, characterized in that: Secondary sampling layers are set in both the CECA structure and the CCFA structure. The process of secondary sampling is as follows: First, reshape the one-dimensional vector into a two-dimensional feature map; Then, use depthwise separable convolution to perform secondary sampling on the feature map. The depthwise separable convolution is specifically divided into: traversing a 3*3 convolution kernel in each channel to obtain an output with the same number of channels as the original feature map, and using a 1*1*number of channels point convolution to expand the depth.

4. The method for measuring the motion characteristics of the moving contact of a circuit breaker based on deep learning according to claim 3, characterized in that: In Step S3, When tracking the motion of the insulating rod, the first auxiliary marker moves linearly, and the change in the centroid coordinates is calculated using the Euclidean distance. The calculation formula is as follows: Among them, D is the distance from the initial position of the centroid of the first auxiliary mark to the end position, (x1, y1) are the coordinates of the centroid of the first auxiliary mark at the initial position, (x2, y2) are the coordinates of the centroid of the first auxiliary mark at the end position. Combining with the camera shooting rate, a travel time curve graph of the moving contact of the circuit breaker is generated; When tracking the movement of the rotating shaft, the rotation angle of the connecting line of the centroids of the two second auxiliary marks is the rotation angle of the rotating shaft. Line a is the connecting line of the centroids of the two second auxiliary marks at the initial position, the slope of line a is k1, line b is the connecting line of the centroids of the two second auxiliary marks at the end position, the slope of line b is k2, and the included angle between line a and line b is θ, θ ∈ [0, π), and θ is the angular displacement of the moving contact of the circuit breaker. When both k1 and k2 exist, the calculation formula is as follows: When both k1 and k2 do not exist, line a is parallel to line b, and θ = 0; When k1 does not exist and the included angle α2 between line b and the horizontal axis is < π / 2, θ = α2 + π / 2; When k1 does not exist and the included angle α2 between line b and the horizontal axis is > π / 2, θ = α2 - π / 2; When k2 does not exist and the included angle α1 between line a and the horizontal axis is < π / 2, θ = α1 - π / 2; When k2 does not exist and the included angle α1 between line a and the horizontal axis is > π / 2, θ = α1 + π / 2.

Images