A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual-branch

Abstract

The present invention relates to a fault diagnosis method for power grid fault waveform analysis based on YOLOv5 double branches, including constructing a neural network architecture, denoising traveling wave signal data; converting the denoised traveling wave signal data into a fault waveform image; using the first branch of the YOLOv5 model to detect the traveling wave head and the fault point in the traveling wave signal data; when the traveling wave head is detected, immediately perform feature fusion on the fault waveform image and electrical data features before and after the traveling wave head moment; input the fused feature vector into the second branch of the YOLOv5 model, and output the fault diagnosis result. Using the neural network denoising technology, the key features in the waveform signal can be accurately extracted, and the influence of noise can be greatly reduced, thereby improving the accuracy and stability of fault diagnosis. The dual-branch YOLOv5 model is used to automatically identify the fault time and fault type, which is not only fast and efficient, but also can accurately capture a variety of fault features in a complex environment, thereby improving the intelligent level of system fault diagnosis.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid fault diagnosis equipment, and in particular to a fault diagnosis method for power grid fault waveform analysis based on YOLOv5 double branches. Background Art

[0002] In the field of distribution network fault diagnosis, traditional methods mainly rely on manual inspections and sensor-based signal measurement and analysis. Manual inspections require a lot of human resources and time, and are easily affected by human factors, resulting in low efficiency. Although traditional signal measurement and analysis methods can detect basic power grid anomalies, their processing capabilities and accuracy are significantly limited when faced with complex fault types and diverse power grid structures.

[0003] Traditional fault location methods usually rely on pre-defined thresholds and rules, and lack the ability to handle complex waveforms and noise interference, making it difficult to achieve high-precision fault location and analysis. In addition, although expert systems and model-based prediction methods can use prior knowledge to assist decision-making, they have limited adaptability to new faults and complex situations and cannot achieve automatic learning and optimization. Kalman filtering is a recursive algorithm widely used in signal processing and control systems that can estimate the state of the system in the presence of noise and uncertainty. However, although Kalman filtering has overcome some limitations of traditional methods to a certain extent, its application in fault diagnosis still faces some challenges and limitations.

[0004] The main limitations of existing technologies include, but are not limited to, reliance on manual intervention, low diagnostic accuracy, low processing efficiency, and weak processing capabilities for noise and complex waveforms. These problems limit the application of traditional methods in large-scale power grid systems and cannot meet the needs of intelligent and automated power grid operation and maintenance. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches to overcome the deficiencies in the above-mentioned prior art.

[0006] The technical solution of the present invention to solve the above technical problem is as follows: a fault diagnosis method based on YOLOv5 dual-branch power grid fault waveform analysis, comprising the following steps:

[0007] Step S01: Capture the traveling wave signal data within 1 second;

[0008] Step S02: construct a neural network architecture to perform denoising on the traveling wave signal data;

[0009] Step S03: converting the denoised traveling wave signal data into a fault waveform image;

[0010] Step S04: using the first branch of the YOLOv5 model to detect the traveling wave head and the fault point in the traveling wave signal data;

[0011] Step S05: when the traveling wave head is detected in step S04, the fault waveform image and the electrical data features before and after the traveling wave head are immediately fused;

[0012] Step S06: input the fused feature vector into the second branch of the YOLOv5 model, and output the fault diagnosis result;

[0013] Step S03 specifically includes the following steps:

[0014] Step S31: The traveling wave signal x(t) is used as the time series data x(t1), x(t2), ..., x(t N ) records, dividing the long-term traveling wave signal into time windows T of fixed length, each time window T contains the time from t i to i+T-1 The traveling wave signal data;

[0015] Step S32: Convert the traveling wave signal in each time window T into a fault waveform image I.

[0016] The beneficial effects of the present invention are as follows: the fault diagnosis method based on traveling wave signal conversion to image and YOLOv5 dual-branch target detection proposed in the present invention has obvious advantages. First, by using the neural network denoising technology, the key features in the waveform signal can be accurately extracted, and the influence of noise can be greatly reduced, thereby improving the accuracy and stability of fault diagnosis. Secondly, the dual-branch YOLOv5 model is used to automatically identify the fault time and fault type, which is not only fast and efficient, but also can accurately capture a variety of fault characteristics in complex environments, thereby improving the intelligent level of system fault diagnosis. It overcomes the limitations of traditional technologies in efficiency, accuracy and scope of application, and also provides innovative and reliable solutions for the safe operation and maintenance of power grid systems.

[0017] Based on the above technical solution, the present invention can also be improved as follows.

[0018] Further, step S02 specifically includes the following steps:

[0019] Step S21: collecting noisy traveling wave signal data and corresponding noise-free or low-noise traveling wave signal data;

[0020] Step S22: marking the collected traveling wave signal data, where the marking includes a noise signal and a target noise-free signal;

[0021] Step S23: The neural network architecture includes:

[0022] Input layer: It is used to receive the original traveling wave signal data x = {x1, x2, ..., x T}, where x t is the tth data point in the time series data;

[0023] Hidden layer: It is used to extract features from traveling wave signals using convolutional neural networks and long short-term memory network structures;

[0024] Output layer: It is used to generate denoised traveling wave signal data

[0025] Step S24: using a loss function to measure the difference between the denoised traveling wave signal data and the target noise-free signal.

[0026] Further, the formula of the loss function in step S24 is as follows:

[0027]

[0028] Among them, MSE is the mean square error, y i is the target noise-free signal, is the denoised signal.

[0029] Further, the conversion in step S32 specifically includes the following steps:

[0030] Step S321: Use a normalized method to map the amplitude of the traveling wave signal to a grayscale value or color pixel intensity, and the formula is as follows:

[0031] I(x,y)=A(t i )·255,

[0032] Among them, A(t i ) indicates that at time t i The amplitude value of , and I(x,y) represents the pixel value in the fault waveform image;

[0033] Step S322: Map the time axis information to the horizontal axis of the fault waveform image.

[0034] Further, step S03 also includes the following steps:

[0035] Step S33: Format the fault waveform image I; the formatting steps are as follows:

[0036] Step S331: adjust the image to a fixed size required by the model;

[0037] Step S332: adjusting the boundaries of the fault waveform images as needed so that all input fault waveform images have the same size.

[0038] Furthermore, the feature fusion in step S05 specifically includes the following steps:

[0039] Step S51: Use a convolutional neural network to extract high-dimensional features from the fault waveform image. Let F img Represents the feature vector extracted from the fault waveform image, and the formula is:

[0040] F img =CNN img (waveform_image),

[0041] Step S52: Use a fully connected layer or other neural network structure to extract electrical data features from the electrical data. Let F elec represents the feature vector extracted from electrical data, and the formula is:

[0042] F elec =FC elec (electrical_data),

[0043] Step S53: F img and F elec Perform fusion operation to obtain the fused feature vector F combined , the formula is:

[0044] F combined =concat(F img ,F elec ).

[0045] Further, the electrical data features extracted in step S52 include voltage and current.

[0046] Further, the output fault diagnosis result in step S06 is specifically:

[0047] Based on the fused feature vector, the pre-trained and fine-tuned YOLOv5 model is used to output the fault type, fault location, and diagnosis report;

[0048] The structure of the YOLOv5 model includes:

[0049] Input layer: It is used to receive the fused feature vector F combined ;

[0050] Convolutional layers: These are used to extract spatial features and patterns;

[0051] Detection layer: It is used to use the detection mechanism of YOLO to output the fault type, fault location and diagnosis report. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is the overall flow chart of the present invention;

[0053] Figure 2 This is a waveform diagram for identifying a traveling wave head according to the present invention. DETAILED DESCRIPTION

[0054] The principles and features of the present invention are described below in conjunction with the accompanying drawings. The examples given are only used to explain the present invention and are not used to limit the scope of the present invention.

[0055] like Figure 1-2 As shown, in Example 1, a fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches includes the following steps:

[0056] Step S01: Use high-precision equipment to capture the traveling wave signal data within 1 second; this ensures that we can obtain sufficiently detailed and accurate traveling wave signals for subsequent analysis and processing. In the specific implementation, the waveform data is directly obtained by calling the database API.

[0057] Step S02: Construct a neural network architecture to denoise the traveling wave signal data; compared with traditional filtering operations, neural network denoising technology can effectively reduce the impact of noise and improve the clarity and accuracy of data.

[0058] Step S03: converting the denoised traveling wave signal data into a fault waveform image;

[0059] Step S04: using the first branch of the YOLOv5 model to detect the traveling wave head and the fault point in the traveling wave signal data; the first branch of the YOLOv5 model can quickly and efficiently identify the position of the traveling wave head to ensure the timeliness of fault detection;

[0060] Step S05: When the traveling wave head is detected in step S04, the fault waveform image and electrical data features before and after the traveling wave head are immediately fused; the second branch combines multiple features to automatically identify and classify the fault type. Through feature fusion, the model can accurately capture multiple fault features in a complex environment and improve the intelligence level of fault diagnosis.

[0061] Step S06: Input the fused feature vector into the second branch of the YOLOv5 model and output the fault diagnosis result, which can be used for the maintenance and management of the power grid system and provide timely fault handling solutions.

[0062] The fault diagnosis method based on traveling wave signal conversion to image and YOLOv5 dual-branch target detection proposed in the present invention has obvious advantages. First, by using neural network denoising technology, the key features in the waveform signal can be accurately extracted, and the influence of noise can be greatly reduced, thereby improving the accuracy and stability of fault diagnosis. Secondly, the dual-branch YOLOv5 model is used to automatically identify the fault time and fault type, which is not only fast and efficient, but also can accurately capture a variety of fault characteristics in complex environments, thereby improving the intelligent level of system fault diagnosis. It overcomes the limitations of traditional technologies in efficiency, accuracy and scope of application, and also provides innovative and reliable solutions for the safe operation and maintenance of power grid systems.

[0063] Embodiment 2: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0064] Step S02 specifically includes the following steps:

[0065] Step S21: collecting noisy traveling wave signal data and corresponding noise-free or low-noise traveling wave signal data; these data can be generated through operation records or simulations of actual power systems.

[0066] Step S22: Mark the collected traveling wave signal data, including the noise signal and the target noise-free signal; these marked data will serve as the basis for training the neural network.

[0067] Step S23: The neural network architecture includes:

[0068] Input layer: It is used to receive the original traveling wave signal data x = {x1, x2, ..., x T}, where x t is the tth data point in the time series data;

[0069] Hidden layer: It is used to extract the features of the traveling wave signal using the convolutional neural network (CNN) and the long short-term memory network (LSTM) structure; the convolutional neural network is good at processing local features and is suitable for extracting the detailed features of the waveform signal. Its output feature map is F cnn =f cnn (X); while the long short-term memory network is good at processing sequence data and is suitable for capturing the time correlation in waveform signals. Its hidden state sequence is H lstm =f lstm (X);

[0070] Output layer: It is used to generate denoised traveling wave signal data In order to ensure the quality of the output signal, a fully connected layer can be used Among them, σ is the activation function, W fc and b fcThey are the weights and biases of the fully connected layer; the deconvolution layer Yes deconv Deconvolution operation, F hid is the feature representation of the hidden layer.

[0071] Step S24: using a loss function to measure the difference between the denoised traveling wave signal data and the target noise-free signal.

[0072] In the process of power grid fault diagnosis, traveling wave signals are often interfered by various noises, including environmental noise, electromagnetic interference, and equipment noise. Although traditional denoising methods such as filters can reduce noise to a certain extent, they often cause signal distortion or loss of important features when processing complex waveform signals. To this end, we introduced neural network denoising technology to achieve more accurate and effective traveling wave signal denoising.

[0073] Embodiment 3, this embodiment is a further improvement on the basis of embodiment 2, and its details are as follows:

[0074] The formula of the loss function in step S24 is as follows:

[0075]

[0076] Among them, MSE is the mean square error, y i is the target noise-free signal, is the denoised signal. The mean square error can amplify larger errors and is more suitable for scenarios that require accurate denoising.

[0077] Embodiment 4: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0078] Step S03 specifically includes the following steps:

[0079] Step S31: The traveling wave signal x(t) is used as the time series data x(t1), x(t2), ..., x(t N ) records, dividing the long-term traveling wave signal into time windows T of fixed length, each time window T contains the time from t i to i+T-1 The traveling wave signal data;

[0080] Step S32: Convert the traveling wave signal in each time window T into a fault waveform image I.

[0081] Embodiment 5, this embodiment is a further improvement on the basis of embodiment 4, and its details are as follows:

[0082] The conversion in step S32 specifically includes the following steps:

[0083] Step S321: Use a normalized method to map the amplitude of the traveling wave signal to a grayscale value or color pixel intensity, and the formula is as follows:

[0084] I(x, y) = A(t i )·255,

[0085] Among them, A(t i ) indicates that at time t i The amplitude value of , and I(x,y) represents the pixel value in the fault waveform image;

[0086] Step S322: Map the time axis information to the horizontal axis of the fault waveform image. This ensures that the order and changes of the time series data can be accurately reflected in the fault waveform image. For example, if the time window length is T=100, the traveling wave signal in each time window will be mapped to a fault waveform image of size 100×W, where W is the pixel width and can be adjusted according to the specific situation.

[0087] Embodiment 6: This embodiment is a further improvement on the basis of embodiment 4, and its details are as follows:

[0088] Step S03 also includes the following steps:

[0089] Step S33: formatting the fault waveform image I to meet the input requirements of the neural network model and optimize the image processing effect; the formatting steps are as follows:

[0090] Step S331: adjust the image to a fixed size required by the model; usually a rectangular fault waveform image of H×W, where H is the pixel height and W is the pixel width;

[0091] Step S332: Adjust the boundary of the fault waveform image as needed so that all input fault waveform images have the same size. Simplify the processing of the neural network.

[0092] Embodiment 7, this embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0093] The feature fusion in step S05 specifically includes the following steps:

[0094] Step S51: Use a convolutional neural network (CNN) to extract high-dimensional features from the fault waveform image. Let F img Represents the feature vector extracted from the fault waveform image, and the formula is:

[0095] F img =CNN img (waveform_image),

[0096] Step S52: Use a fully connected layer or other neural network structure to extract electrical data features from the electrical data. Let F elec represents the feature vector extracted from electrical data, and the formula is:

[0097] F elec =FC elec (electrical_data),

[0098] Step S53: F img and F elec Perform fusion operation to obtain the comprehensive feature vector F combined , the formula is:

[0099] F combined =concat(F img ,F elec ).

[0100] The process of feature fusion can combine data from different sources to form a comprehensive input feature vector.

[0101] Embodiment 8, this embodiment is a further improvement on the basis of embodiment 7, and its details are as follows:

[0102] The electrical data features extracted in step S52 include voltage and current.

[0103] Embodiment 9: This embodiment is a further improvement on the basis of embodiment 1, and its details are as follows:

[0104] The output fault diagnosis result in step S06 is specifically:

[0105] Based on the fused feature vector, the pre-trained and fine-tuned YOLOv5 model is used to output the fault type, fault location and diagnosis report; the advantage of the YOLOv5 model lies in its fast and efficient target detection capability, which can accurately identify different types of faults in complex environments.

[0106] The structure of the YOLOv5 model includes:

[0107] Input layer: It is used to receive the fused feature vector F combined ;

[0108] Convolutional layers: These are used to extract spatial features and patterns;

[0109] Detection layer: It is used to use the detection mechanism of YOLO to output the fault type, fault location and diagnosis report.

[0110] Although the embodiments of the present invention have been shown and described above, it is to be understood that the above embodiments are exemplary and are not to be construed as limitations of the present invention. A person skilled in the art may change, modify, replace and vary the above embodiments within the scope of the present invention.

Claims

1. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual-branch, characterized in that: The steps include: Step S01: Capture the traveling wave signal data within 1 second; Step S02: constructing a neural network architecture to perform denoising on the traveling wave signal data; Step S03: converting the denoised traveling wave signal data into a fault waveform image; Step S04: using the first branch of the YOLOv5 model to detect the traveling wave head and the fault point in the traveling wave signal data; Step S05: when the traveling wave head is detected in step S04, the fault waveform image and electrical data features before and after the traveling wave head are immediately fused; Step S06: input the fused feature vector into the second branch of the YOLOv5 model, and output the fault diagnosis result; The step S03 specifically includes the following steps: Step S31: The traveling wave signal x(t) is used as the time series data x(t1), x(t2), ..., x(t N ) records, dividing the long-term traveling wave signal into time windows T of fixed length, each time window T contains the time from t i to i+T-1 The traveling wave signal data; Step S32: Convert the traveling wave signal in each time window T into a fault waveform image I.

2. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 1, characterized in that: The step S02 specifically includes the following steps: Step S21: collecting noisy traveling wave signal data and corresponding noise-free or low-noise traveling wave signal data; Step S22: marking the collected traveling wave signal data, where the marking includes a noise signal and a target noise-free signal; Step S23: The neural network architecture includes: Input layer: It is used to receive the original traveling wave signal data x = {x1, x2, ..., x T }, where x t is the tth data point in the time series data; Hidden layer: It is used to extract features from traveling wave signals using convolutional neural networks and long short-term memory network structures; Output layer: It is used to generate the denoised traveling wave signal data Step S24: using a loss function to measure the difference between the denoised traveling wave signal data and the target noise-free signal.

3. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 2, characterized in that: The formula of the loss function in step S24 is as follows: Among them, MSE is the mean square error, y i is the target noise-free signal, is the denoised signal.

4. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 1, characterized in that: The conversion in step S32 specifically includes the following steps: Step S321: Use a normalized method to map the amplitude of the traveling wave signal to a grayscale value or color pixel intensity, and the formula is as follows: I(x,y)=A(t i )·255, Among them, A(t i ) indicates that at time t i The amplitude value of , and I(x, y) represents the pixel value in the fault waveform image; Step S322: Map the time axis information to the horizontal axis of the fault waveform image.

5. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 1, characterized in that: The step S03 also includes the following steps: Step S33: formatting the fault waveform image I; the formatting steps are as follows: Step S331: adjust the image to a fixed size required by the model; Step S332: adjusting the boundary of the fault waveform image as needed so that all input fault waveform images have the same size.

6. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 1, characterized in that: The feature fusion in step S05 specifically includes the following steps: Step S51: Use a convolutional neural network to extract high-dimensional features from the fault waveform image. Let F img represents the feature vector extracted from the fault waveform image, and the formula is: F img =CNN img (waveform_image), Step S52: Use a fully connected layer or other neural network structure to extract the electrical data features from the electrical data. elec represents the feature vector extracted from the electrical data, and the formula is: F elec =FC elec (electrical_data), Step S53: F img and F elec Perform fusion operation to obtain the fused feature vector F combined , the formula is: F combined =concat(F img ,F elec )。 7. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 6, characterized in that: The electrical data features extracted in step S52 include voltage and current.

8. A fault diagnosis method for power grid fault waveform analysis based on YOLOv5 dual branches according to claim 1, characterized in that: The output fault diagnosis result in step S06 is specifically: Based on the fused feature vector, the pre-trained and fine-tuned YOLOv5 model is used to output the fault type, fault location and diagnosis report; The structure of the YOLOv5 model includes: Input layer: It is used to receive the fused feature vector F combined ; Convolutional layers: These are used to extract spatial features and patterns; Detection layer: It is used to use the detection mechanism of YOLO to output the fault type, fault location and diagnosis report.

Images