Double-confirmation discrimination method for opening and closing states of 10kV circuit breaker and load switch

Abstract

The invention relates to a double-confirmation judgment method for opening and closing states of a 10kV circuit breaker and a load switch. The double-confirmation judgment method comprises the following steps: realizing image recognition of a switch indication sign and comprehensive acquisition and analysis of an auxiliary point switch state through a visible light module, a switching value acquisition module and a calculation module; identification of a sign image is realized and formation and structure of the sign image and the state of a switch node are realized, that is, target detection and identification are carried out on an indication board for displaying the opening and closing states of a circuit breaker and a load switch through a trained and optimized YOLOv8 network model, a green indication board represents an opening state, and a red indication board represents a closing state; the switching value acquisition module acquires opening and closing state signals of the auxiliary contact and then transmits the opening and closing state signals to the calculation module, and the calculation module judges whether the switch is closed or opened according to the signals; according to the method, the identification of the mark image is realized, and the accuracy and the reliability of judgment are improved and the risk of misjudgment is reduced through double confirmation with the state formation and the structure of the switch node.

Description

Technical Field

[0001] The present invention relates to the technical field of switch state discrimination, and particularly to a double-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and a load switch. Background Art

[0002] In recent years, the State Grid Corporation has accelerated the implementation progress of the double-confirmation work for disconnectors and circuit breakers, and there is a large demand for the transformation of a large number of 10 kV circuit breakers and load switches. At present, double-confirmation mostly adopts attitude sensors, video linkage and micro-switch methods, which have problems such as large transformation difficulty and complex construction. Among them, video linkage is a relatively reliable detection method, but it has the problem of reduced efficiency due to human participation. The present invention upgrades video linkage to short-region switch indication image recognition and combines it with the auxiliary point switch state detection method to realize the double-confirmation discrimination method. Summary of the Invention

[0003] The present invention provides a double-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and a load switch, realizes the recognition of a marker image, forms a structure with the switch node state, and improves the accuracy and reliability of discrimination by double-confirmation, reducing the risk of misjudgment.

[0004] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0005] A double-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and a load switch includes the following steps:

[0006] S1. A visible light module collects an indication plate image on the circuit breaker and the load switch;

[0007] S2. Target detection and recognition are performed by training a YOLOv8 network model: an attention mechanism CBAM is introduced into the YOLOv8 network model to make the model focus on the color and shape features of the indication plate and ignore background interference. A deformable convolution DCNv4 is introduced into the YOLOv8 network model to adapt to the diversity and deformation features of the switch indication plates in the transmission line. The improved YOLOv8 network model introduced with the attention mechanism CBAM and the deformable convolution DCNv4 is trained, and an adaptive broadband algorithm is used to optimize the hyperparameters of the training, and then target detection is performed on the collected indication plate image;

[0008] S3. A switch quantity acquisition module collects the auxiliary contact state signals of the circuit breaker and the load switch;

[0009] S4. A calculation module receives the switch quantity signals of the circuit breaker and the load switch, performs logical judgment, and determines the switch states of the circuit breaker and the load switch;

[0010] S5. Hybrid Dual Confirmation Discrimination: Perform a logical AND operation on the image recognition result in step S2 and the switching signal analysis result in step S4. When the results are consistent, confirm this result as the final state of the switch.

[0011] S5. Output the final discrimination result.

[0012] Furthermore, the improved YOLOv8 network model includes an input end Input layer, a network backbone module Backbone layer, a feature enhancement module Neck layer, and an output end Head layer. A CBAM module is introduced after the C2f module in both the network backbone module Backbone layer and the feature enhancement module Neck layer. The C2f module and the CBAM module introduced after it form a C2f-cb module.

[0013] Furthermore, deformable convolution DCNv4 is introduced in the YOLOv8 network model. A dynamic offset is added to each convolution kernel sampling point. The position and size of the convolution kernel are dynamically adjusted according to the outline of the opening and closing state indicator signs of the circuit breaker and load switch. The position encoding and self-attention mechanism are used to optimize the detection performance of the indicator signs in complex backgrounds. The value at a certain sampling point position p0 of the output feature map of the opening and closing state indicator sign picture is calculated using DCNv4.

[0014]

[0015] Among them, is the offset obtained through the position encoding and self-attention mechanism, and is the offset of the kth sampling point in the gth group;

[0016] is the sum over all groups g;

[0017] is the sum over all sampling points k;

[0018] w g is the weight of the gth group;

[0019] m gk is the mask value of the kth sampling point in the gth group;

[0020] is the input feature value at the position in the gth group;

[0021] p n is the position element in the sampling area of the convolution kernel;

[0022] Δp n is the additional offset added to the regular convolution sampling point.

[0023] Furthermore, the improved YOLOv8 network model incorporating the attention mechanism CBAM and deformable convolution DCNv4 is trained, and the hyperparameters of the training are optimized using the adaptive bandwidth algorithm, which specifically includes the following steps:

[0024] S2.1, Data and Preparation: Prepare the photos of the switch opening and closing status indicator signs taken, annotate them, and divide them into a training set and a test set;

[0025] S2.2, Initialize the YOLOv8 model and set the initial hyperparameters;

[0026] S2.3, During model training, use the adaptive bandwidth algorithm to dynamically adjust the learning rate;

[0027] S2.4, Use the adaptive bandwidth algorithm to dynamically adjust the step size;

[0028] S2.5, Iteration and Evaluation;

[0029] S2.6, Output the optimal switch opening and closing status recognition model.

[0030] Furthermore, the dynamic learning rate:

[0031] α t = α0 × λ t (2)

[0032] where α t is the learning rate of the t-th round, α0 is the initial learning rate, and λ t is the adaptive adjustment factor;

[0033] The calculation of the adaptive adjustment factor λ t is as follows:

[0034]

[0035] where L t and L t-1 are the validation set losses of the t-th round and the (t - 1)-th round.

[0036] Compared with the prior art, the beneficial effects of the present invention are:

[0037] 1) Incorporate the CBAM attention mechanism and deformable convolution DCNv4 into the original YOLOv8 network structure to better meet the requirements of the indicator sign detection task, effectively extract the key features of the indicator sign, suppress the background information, and enhance the adaptability of the model to irregular shapes;

[0038] 2) Replace the original two-dimensional conventional convolution with DCNv4. DCNv4 introduces position encoding and self-attention mechanism, enabling the improved YOLOv8 algorithm to better fit the detection targets, improve the detection performance, and enhance the ability to capture spatial offsets. Description of the Drawings

[0039] Figure 1 It is the structural diagram of the traditional YOLOv8 network.

[0040] Figure 2 It is the flowchart of the double-confirmation discrimination method for the opening and closing states of the 10kV circuit breaker and load switch according to the present invention.

[0041] Figure 3 It is the structural diagram of the CBAM attention mechanism according to the present invention.

[0042] Figure 4 It is the structural diagram of the C2f-cb module according to the present invention.

[0043] Figure 5 It is the diagram of the SPPFv4 module with DCNv4 introduced according to the embodiment of the present invention.

[0044] Figure 6 It is the structural diagram of the finally optimized YOLOv8 network according to the embodiment of the present invention. Detailed Embodiments

[0045] The following further describes the detailed embodiments of the present invention with reference to the drawings:

[0046] A double-confirmation discrimination method for the opening and closing states of a 10kV circuit breaker and a load switch according to the present invention realizes the comprehensive acquisition and analysis of the image recognition of the switch indication sign and the state of the auxiliary point switch through a visible light module, a switch quantity acquisition module, and a calculation module; realizes the recognition of the sign image and forms a combination with the state of the switch node, that is, the target detection and recognition of the indication sign showing the opening and closing states of the circuit breaker and the load switch are carried out through the trained and optimized YOLOv8 network model, where the green indication sign represents the open state and the red indication sign represents the closed state; the switch quantity acquisition module will collect the opening and closing state signals of the auxiliary contacts and then transmit them to the calculation module, and the calculation module judges whether the switch is closed or open according to these signals, so that two judgment methods are available to realize the hybrid double-confirmation operation, as shown in Figure 2 , and specifically includes the following steps:

[0047] S1. The visible light module collects the image of the switch state indication sign;

[0048] Use the visible light module to take pictures of the indication signs on the circuit breaker and the load switch. The indication sign being red represents the closed state, and green represents the open state. Take 500 pictures, 450 of which are used as the training set and 50 as the test set.

[0049] S2. Perform object detection and recognition by training the YOLOv8 network model;

[0050] Use the pre-trained and optimized YOLOv8 model to perform object detection on the collected sign images; Through deep learning technology, the YOLOv8 model can accurately identify the color (red or green) and position of the switch opening and closing status signs;

[0051] The detection result outputs the specific status of the sign: If a green sign is detected, it is determined to be in the off state; If a red sign is detected, it is determined to be in the closed state;

[0052] 1) Switch sign recognition method based on improved YOLOv8 model

[0053] YOLO series algorithms are widely used in various object detection and recognition tasks due to their advantages such as high detection speed, strong real-time performance, and simple structure. As the latest YOLO series algorithm, YOLOv8 further optimizes the network structure compared with previous generations and improves the comprehensive detection performance. This invention uses YOLOv8 as the basic model to study a switch indication sign recognition algorithm based on improved YOLOv8; The network structure of YOLOv8 mainly consists of four parts: the input end Input layer, the network backbone module Backbone layer, the feature enhancement module Neck layer, and the output end Head layer. Its structure diagram is as Figure 1 shown;

[0054] When recognizing the switch status signs (red indicates closed, green indicates open) in the transmission line, the environment where the target object (i.e., the sign) is located is complex, and the differences in the characteristics (such as color and shape) of the sign itself and the background significantly affect the detection efficiency of the model. To improve the performance of the model, this invention introduces two mechanisms in the original YOLOv8 network structure: the CBAM attention mechanism and the deformable convolution DCNv4 to better meet the requirements of the sign detection task;

[0055] 1.1. Introduction of the CBAM attention mechanism

[0056] The color and shape features of the switch status signs are crucial in the detection task. However, complex backgrounds and negative samples (such as other irrelevant objects or environmental interferences) often distract the attention of the model and lead to a decrease in detection efficiency. Therefore, this invention introduces the CBAM (Convolutional Block Attention Module) hybrid attention mechanism, see Figure 3, which combines a Channel Attention Module (CAM) and a Spatial Attention Module (SAM), can effectively extract the key features of the sign and suppress background information;

[0057] Specifically, see Figure 4 , the CBAM module is introduced after the C2f module in the Backbone layer and the Neck layer, and is generally represented as the C2f-cb module, enabling the model to focus more on important features such as the color (red or green) and shape of the sign, and ignoring the interference of the background; for example, when the model detects a red sign, CBAM will enhance the feature extraction of the red area while weakening the influence of background pixels, thereby improving the recognition accuracy of the closed state; similarly, the open state of the green sign can also be detected more accurately;

[0058] 1.2. Introduce the deformable convolution DCNv4;

[0059] Traditional convolution operations use rectangular convolution kernels of fixed size, which are difficult to adapt to the diversity and deformation characteristics of switch signs in transmission lines; for example, some signs may present irregular shapes due to angles, lighting, or occlusion, resulting in conventional convolutions being unable to fully extract their key features. Therefore, the present invention introduces the deformable convolution DCNv4 in the Spatial Pyramid Pooling-Fast (SPPF) module of YOLOv8 to enhance the model's adaptability to irregular shapes;

[0060] The principle of deformable convolution DCN (Deformable Convolution) is to add an offset to each sampling point on the basis of traditional convolution. Assuming the convolution kernel size is 3×3, then there are 9 sampling points in this convolution kernel. An offset is assigned to each of these 9 sampling points, so that the position and size of the convolution kernel can be dynamically adjusted according to the contour of the opening and closing state signs of the circuit breaker and load switch, and can fit the size and shape of the opening and closing state signs more closely during feature extraction, optimizing the detection performance of the network model for irregular and severely deformed signs;

[0061] For traditional two-dimensional standard convolution, the formula for the output feature value of a certain sampling point p0 in the photo of the opening and closing state sign is:

[0062]

[0063] R is the sampling area of the convolution kernel, where R = {(-1,-1),(-1,0),…,(0,1),(1,1)}, w(p n ) is at pn The convolution kernel weight value at the position, x is the feature map to be detected, i.e., the feature map of the photo of the opening and closing state indicator signs of the circuit breaker and load switch, and p n is the element at the position in R, and x(p0 + p n ) is the value of the feature map of the input image (the image of the switch opening and closing state indicator sign) at the position (p0 + p n );

[0064] Δp n is the offset added at the conventional convolution sampling points. The sampling point offset is updated and optimized using the interpolation algorithm backpropagation algorithm. After introducing the offset, the pixel x(p0 + p n + Δp n ) at the sampling point position of the switch opening and closing state indicator sign image is calculated using the bilinear interpolation method. The formula is:

[0065] x(p) = Σ q G(q, p)·x(q) (5)

[0066] where p = p0 + p n + Δp n is any position in the region, q represents all the spatial positions in the input feature map, i.e., the four integer points around p, x(q) is the value of the points at all integer positions in the feature map (the feature values of the switch opening and closing state indicator sign image), and G(q, p) is the bilinear interpolation kernel function of a two-dimensional kernel, which can be decomposed into two one-dimensional kernels. The formula is:

[0067] G(q, p) = g(q x , p x )·g(q y , p y ) (6)

[0068] g(a, b) = max{0, 1 - |a - b|} (7)

[0069] By adding a dynamic offset to each convolution kernel sampling point, DCNv4 enables it to automatically adjust the convolution kernel sampling method according to the specific shape and position of the indicator sign. For example, when the switch state indicator sign presents an irregular shape due to occlusion, DCNv4 can dynamically adjust the sampling points to ensure that the model can capture the core features of the indicator sign (such as the red or green boundary), thereby improving the detection accuracy. In addition, DCNv4 also further optimizes the detection performance of the indicator sign in complex backgrounds through position encoding and self-attention mechanisms;

[0070] DCNv4 uses a more advanced position encoding and self-attention mechanism scheme to improve the ability to capture spatial offsets.

[0071]

[0072] Among them, is the offset obtained through the improved positional encoding and self-attention mechanism, y(p0) v4 is the value of the output feature map of the opening and closing state indicator sign picture at position p0, calculated using DCNv4, is the sum over all groups g, is the sum over all sampling points k, w g is the weight of the g-th group, m gk is the mask value of the kk-th sampling point in the g-th group. In the g-th group, at position is the input feature value. is the offset of the k-th sampling point in the g-th group;

[0073] In the original SPPF module of the present invention, DCNv4 is used to replace the original two-dimensional conventional convolution, so that the improved YOLOv8 algorithm can better fit the detection target and improve the detection performance;

[0074] By introducing CBAM and DCNv4 into YOLOv8, the present invention significantly improves the detection ability of the model for switch state indicator signs (red closed, green open) in complex environments;

[0075] The SPPFv4 module introducing deformable convolution is as Figure 5 shown.

[0076] (2) Train the improved YOLOv8 model, and during the training, use the adaptive broadband algorithm to optimize the hyperparameters of the training;

[0077] See Figure 6 , the final optimized YOLOv8 network structure diagram. When training the YOLOv8 model, there are common hyperparameters such as learning rate, step size, and batch size that need to be adjusted manually. The present invention uses the adaptive broadband algorithm to optimize the hyperparameters, and the steps are as follows:

[0078] (1) Data and preparation

[0079] Prepare the photos of the switch opening and closing state indicator signs taken, including the state photos of green and red, and perform annotation, and divide them into a training set and a test set;

[0080] (2) Initial parameter setting

[0081] Initialize the YOLOv8 model and set the initial hyperparameters;

[0082] Initial learning rate (α0): 0.01;

[0083] Batch size (B): 16;

[0084] Initial step size (L0): The step size of the first epoch of training, which can be set to 1;

[0085] Loss function: YOLOv8 usually uses a composite loss function, including classification loss, confidence loss, and bounding box regression loss;

[0086] (3) Dynamic learning rate adjustment

[0087] During model training, an adaptive bandwidth algorithm is used to dynamically adjust the learning rate;

[0088] Dynamic learning rate formula:

[0089] α t = α0 × λ t (9)

[0090] where α t is the learning rate of the t-th round, and λ t is the adaptive adjustment factor;

[0091] Adaptive adjustment factor λ t Calculation:

[0092]

[0093] where L t and L t-1 are the validation set losses of the t-th round and the (t - 1)-th round;

[0094] (4) Dynamic step size adjustment

[0095] Similarly, an adaptive bandwidth algorithm is used to dynamically adjust the step size. Dynamic step size formula:

[0096] L t = L t-1 + ΔL (11)

[0097] where ΔL is the step size increment dynamically adjusted according to the validation set performance. Calculation of step size increment ΔL:

[0098]

[0099] where γ is an adjustable weight parameter, usually taking values from 0.1 to 0.2;

[0100] (5) Iteration and evaluation

[0101] Training set loss: L train (t);

[0102] Validation set loss: L val (t);

[0103] Indicator tracking: After each round of training, save the training set loss L train (t) and the validation set loss L val (t);

[0104] Early stopping method: When the validation set loss has not improved for several consecutive rounds, reduce the learning rate or stop training;

[0105] if L val (t) - L val (t - 1) > ∈ stop then stop (13)

[0106] where ∈ stop is a threshold used to determine whether the change in loss is significant;

[0107] (6) Output the optimal switch opening and closing state recognition model

[0108] When the number of iterations reaches the maximum or the loss value has not decreased for a long time, stop training and output the optimal switch opening and closing state recognition model. At this time, by detecting the switch opening and closing state indicator sign input in real time through the camera, the switch state can be detected as closed or open.

[0109] S3. The digital input module collects the auxiliary contact status signal;

[0110] Use the digital input module to collect the opening and closing status signals of the auxiliary contacts. The auxiliary contacts output the opening and closing status information of the switch through mechanical or electrical contacts. The digital input module converts these signals into digital signals 0 or 1.

[0111] S4. The calculation module processes the digital signal

[0112] The calculation module receives the digital signal and performs logical judgment. The calculation module determines the state of the switch according to the digital signal: if the signal is 1, it is determined to be in the closed state. If the signal is 0, it is determined to be in the open state.

[0113] S5. Hybrid double confirmation discrimination

[0114] Perform a logical AND operation on the image recognition result and the digital signal analysis result;

[0115] Only when the image recognition result is consistent with the digital signal analysis result, the final state of the switch is confirmed; for example, if the image recognition determines the closed state (red indicator sign), and at the same time the digital signal determines the closed state (signal is 1), then the switch is confirmed to be in the closed state; if they are inconsistent, an alarm is triggered or data is re-collected to ensure the reliability of the detection result.

[0116] S6. Output the final discrimination result

[0117] Output the double-confirmation discrimination result for system monitoring or automatic control; the final result can be used for indicator light display, protection logic triggering, or status update of the SCADA system.

[0118] The double-confirmation mechanism improves the accuracy and reliability of discrimination and reduces the risk of misjudgment.

[0119] The above embodiments are implemented on the premise of the technical solution of the present invention, and detailed implementation manners and specific operation processes are given, but the protection scope of the present invention is not limited to the above embodiments. The methods used in the above embodiments are conventional methods unless otherwise specified.

Claims

1. A double - confirmation discrimination method for the opening and closing states of a 10kV circuit breaker and a load switch, characterized in that, It includes the following steps: S1. The visible light module captures the images of the indicator plates on the circuit breaker and the load switch; S2. Perform object detection and recognition by training the YOLOv8 network model: Introduce the attention mechanism CBAM into the YOLOv8 network model to make the model focus on the color and shape features of the indicator plate and ignore background interference. Introduce the deformable convolution DCNv4 into the YOLOv8 network model to adapt to the diversity and deformation features of the switch indicator plates in the transmission line. Train the improved YOLOv8 network model with the attention mechanism CBAM and the deformable convolution DCNv4 introduced, optimize the hyperparameters of the training using the adaptive broadband algorithm, and then perform object detection on the captured indicator plate images; S3. The digital input module captures the status signals of the auxiliary contacts of the circuit breaker and the load switch; S4. The calculation module receives the digital signals of the circuit breaker and the load switch, conducts logical judgment, and determines the switch states of the circuit breaker and the load switch; S5. Hybrid double confirmation discrimination: Perform a logical AND operation on the image recognition result in step S2 and the analysis result of the digital signal in step S4. When the results are consistent, confirm this result as the final state of the switch; S5. Output the final discrimination result.

2. The dual-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and a load switch according to claim 1, wherein, The improved YOLOv8 network model includes an input end Input layer, a network backbone module Backbone layer, a feature enhancement module Neck layer, and an output end Head layer. The CBAM module is introduced after the C2f module in both the network backbone module Backbone layer and the feature enhancement module Neck layer. The C2f module and the CBAM module introduced after it form a C2f-cb module.

3. A dual-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and a load switch according to claim 1, characterized in that, The deformable convolution DCNv4 is introduced into the YOLOv8 network model. A dynamic offset is added to each convolution kernel sampling point. The position and size of the convolution kernel are dynamically adjusted according to the contour of the indicator plate of the opening and closing state of the circuit breaker and the load switch. The position encoding and self-attention mechanism are used to optimize the detection performance of the indicator plate in the complex background. The value at a certain sampling point position p0 of the output feature map of the opening and closing state indicator plate image is calculated using DCNv4. Among them, is the offset obtained through positional encoding and self-attention mechanism, and is the offset of the k-th sampling point in the g-th group; Sum over all groups g; For summing over all sampling points k; w g is the weight for the g-th group; m gk is the mask value of the k-th sampling point in the g-th group; For the input eigenvalue at position in the g-th group; p n is the position element in the sampling area of the convolution kernel; Δp n is the offset added at the conventional convolution sampling points.

4. A dual-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and load switch according to claim 1, characterized in that, The improved YOLOv8 network model with the attention mechanism CBAM and the deformable convolution DCNv4 introduced is trained, and the hyperparameters of the training are optimized using the adaptive broadband algorithm. Specifically, it includes the following steps: S2.

1. Data and preparation: Prepare the photos of the switch opening and closing state indicator plates taken, perform annotation, and divide them into a training set and a test set; S2.

2. Initialize the YOLOv8 model and set the initial hyperparameters; S2.

3. In the model training, use the adaptive broadband algorithm to dynamically adjust the learning rate; S2.

4. Use the adaptive broadband algorithm to dynamically adjust the step size; S2.

5. Iteration and evaluation; S2.

6. Output the optimal switch opening and closing state recognition model.

5. A dual-confirmation discrimination method for the opening and closing states of a 10 kV circuit breaker and load switch according to claim 1, characterized in that The dynamic learning rate: α t = α0×λ t (2) where α t is the learning rate of the t-th round, α0 is the initial learning rate, and λ t is the adaptive adjustment factor; Adaptive adjustment factor λ t Calculation: where L t and L t-1 are the validation set losses in the t-th round and the (t-1)-th round, respectively.

Images