A fault wave analysis method for industrial substation closing resistance recording based on YOLOv8 neural network

Abstract

The present invention discloses a method for analyzing fault waves recorded from the closing resistor of an industrial power substation based on a YOLOv8 neural network, belonging to the field of analyzing fault waves recorded from the closing resistor of a circuit breaker. Through the present invention, in the detection of closing resistors in the industrial power substation field, the actual values ​​of the current data to be tested are obtained by parsing a comtrade data set file. Smoothing and Fourier transform fitting are then performed to determine the error type of the current waveform. This method has a high recognition rate and can effectively analyze whether the closing resistor in the industrial power substation is damaged by a rapid increase in current at the moment of closing. By improving the algorithm architecture and optimizing the model parameters, unique neural network parameters are obtained after training the YOLOv8 neural network, thereby improving detection accuracy and inference speed.

Description

Technical Field

[0001] The present invention belongs to the field of circuit breaker closing resistance recording fault wave analysis, and specifically relates to an industrial substation closing resistance recording fault wave analysis method based on a YOLOv8 neural network. Background Art

[0002] Circuit breaker closing resistors are primarily used in high-voltage and ultra-high-voltage power systems, playing a crucial role in high-voltage transmission lines, substations, reactive power compensation equipment, grid interconnection, renewable energy access, and large industrial users. In these applications, circuit breaker closing resistors improve power system stability and reliability by limiting transient overvoltages and inrush currents during closing operations, reducing the risk of damage to power equipment and thus ensuring safe grid operation. As power systems develop and voltage levels increase, the use of circuit breaker closing resistors will become even more important.

[0003] Power waveform recording can reveal relevant characteristic information, such as voltage or current changes. To verify the proper functioning of circuit breaker closing resistors in power systems, power system models collect transient waveforms and accident data. This allows us to determine from the waveforms whether closing resistors in industrial substations are damaged by the sudden increase in current during the instant of closing.

[0004] In substations, circuit breaker closing resistors, as key components, are responsible for limiting inrush current and overvoltage during closing operations. However, these resistors can also present problems and malfunctions during operation. For example, during operation, the closing resistors can rub against the supporting insulating rods, creating foreign matter and causing insulation degradation. This can lead to penetrating discharges through the supporting rods, causing current to flow through them and short-circuit the closing resistors, rendering them ineffective in suppressing inrush current.

[0005] The methods for detecting closing resistance in the prior art include:

[0006] Electrical test: Regular high voltage test and resistance test are carried out on the closing resistor to ensure that it can work normally under normal operating voltage and current;

[0007] Thermal imaging: Use an infrared thermal imager to detect temperature changes in the closing resistor during operation. If the temperature of the resistor rises abnormally during the closing process, this may indicate that the resistor is carrying an excessive current load.

[0008] Current monitoring: Install a current transformer in the circuit of the closing resistor to monitor the current waveform at the moment of closing in real time and analyze whether the current rise rate meets the design requirements;

[0009] Simulation analysis: Use power system simulation software to simulate the closing process and predict the working state of the closing resistor and the change of current;

[0010] Historical data analysis: Collect and analyze the historical operating data of the closing resistor, including closing times, current load records, etc., to evaluate its performance and lifespan;

[0011] Regular inspection and maintenance: Perform regular visual inspections on the closing resistor to check for obvious signs of burns, deformation, or looseness to ensure the reliability of its mechanical and electrical connections;

[0012] Operation ticket system: A strict operation ticket system is implemented in substation operations to ensure that each closing operation complies with operating procedures and avoid equipment damage caused by improper operation.

[0013] The above methods have many limitations. For example, electrical testing can only detect the performance of resistors under specific test conditions and cannot fully simulate the conditions in actual operation. Thermal imaging can only detect the temperature of the resistor surface and cannot directly measure the internal temperature and structural damage. Current monitoring requires continuous data analysis and professionals to interpret the test results. The accuracy and reliability of the simulation model in simulation analysis depends on the accuracy of the input data. Historical data analysis cannot reflect the current status of the resistor in real time. Regular inspection and maintenance require regular power outages for inspection, which affects industrial production efficiency. The operation ticket system cannot prevent operational errors caused by equipment failure.

[0014] Therefore, there is an urgent need for an electric power recording and analysis method that improves the limitations of current detection methods and reduces the complexity of detection results for professionals. Summary of the Invention

[0015] The problem to be solved by the present invention is to address the shortcomings of the existing technology and propose a fault wave analysis method for industrial substation closing resistor recording based on YOLOv8 neural network. This method is used to timely analyze the power recording data, and then accurately determine whether the closing resistor in the industrial substation is damaged due to the rapid increase of current at the moment of preventing closing.

[0016] The technical solution adopted by the present invention is a method for analyzing fault waves of industrial substation closing resistance recording based on the YOLOv8 neural network, comprising the following steps:

[0017] Step 1: The user uploads the industrial power transient recording comtrade dataset file;

[0018] Step 2: Based on Python 3 and using bytes data stream, parse the CFG configuration file and DAT file in the comtrade file to obtain the actual value of the current data to be tested, parse it, and save it as a CSV file to the specified location;

[0019] Step 3: Smooth the data in the saved CSV file, and fit the sine and cosine functions through Fourier transform to obtain the maximum value of the absolute value of the fitting curve.

[0020] Step 4: Compare the CSV file data extracted in step 2 with the maximum value. If any data point is greater than the maximum value, the file is recorded as a wave recording breakthrough error.

[0021] Step 5: Convert the portion of the image in step 4 that does not contain breakthrough errors into PNG format, and after manual review, input it into the YOLOv8 neural network for analysis.

[0022] Step 6: If there is no data point greater than the maximum value in step 4, record it as the part that has not been broken through. The starting point and ending point of the data in the CSV file that exceeds the State Grid Power Recording Fault Threshold α are recorded as t_1 and t_2 respectively;

[0023] Step 7: Subtract t_2 and t_1 obtained in step 6 to obtain the value t_2-t_1, and compare it with the standard threshold value β of the closing resistance input time of the national grid;

[0024] Step 8: If the value of t_2-t_1 in step 7 is greater than the threshold β, it is converted into PNG format, manually reviewed, and then input into the YOLOv8 neural network for analysis;

[0025] Step 9: If the value of t_2-t_1 in step 7 is less than the threshold β, it is recorded as a recording delay error and converted into PNG format and input into the YOLOv8 neural network for analysis;

[0026] In step 10, after reviewing the input PNG format file through the YOLOv8 neural network in steps 5, 8, and 9, the output current waveform error type is used to determine whether the closing resistor in the industrial substation is damaged due to the rapid increase in current at the moment of closing.

[0027] The present invention is also characterized in that:

[0028] 1. The use of discrete second-order derivative mathematical methods and computer vision neural network methods greatly improves the recognition rate, facilitating timely analysis of power wave recording data and accurately determining whether the closing resistor is damaged during the closing process;

[0029] 2. In the closing resistor detection of the State Grid's industrial substations, by parsing the comtrade dataset file, the actual value of the current data to be tested is obtained. After smoothing and Fourier transform fitting, the current waveform error type can be determined. This method has a high recognition rate and can effectively analyze whether the closing resistor in the industrial substation is damaged by the rapid increase in current at the moment of closing.

[0030] By improving the algorithm architecture and optimizing the model parameters, unique neural network parameters were obtained after YOLOv8 neural network training, thereby improving detection accuracy and inference speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 This is a system structure diagram of a method for analyzing fault waves of industrial substation closing resistance recording based on a YOLOv8 neural network according to the present invention;

[0032] Figure 2 The delayed error recording data in the image is fitted by Fourier transform of the data in the CSV file;

[0033] Figure 3 The breakthrough error recording data in the image is fitted by Fourier transform of the data in the CSV file;

[0034] Figure 4 The correct recorded data in the image is obtained by Fourier transform fitting of the data in the CSV file. DETAILED DESCRIPTION

[0035] The following is a further explanation of the fault wave analysis method of industrial substation closing resistance recording based on YOLOv8 neural network with reference to the accompanying drawings:

[0036] Step 1: Upload and obtain the comtrade file of industrial power transient recording data;

[0037] ,Among them, comtrade file is a common format for exchanging transient data of power system. ,Comtrade file contains industrial power transient recording data. ,Each comtrade file has a set of up to four related files.

[0038] The CFG configuration file and DAT data file are the files required in this method.

[0039] The CFG configuration file interprets the format of the data file and contains all the information, and the DAT data file contains the values ​​of all input channels for each sample in the record.

[0040] Step 2: Based on Python 3 and using bytes data stream, parse the CFG configuration file and DAT data file in the comtrade file to obtain the actual value of the current data to be tested, parse it, and save it as a CSV file to the specified location;

[0041] Use Python 3 version to facilitate file and character processing. It only parses comtrade's cfg configuration and dat data files, where the DAT data files are in binary format.

[0042] First, open the fault recording file and perform data slicing to accurately analyze the configuration information in the CFG configuration file variables to obtain the number of channels, read the detailed parameters and frequency of each analog channel, and calculate the sampling frequency and sampling length;

[0043] Use bytes data stream to read DAT data file, calculate the total length of dat, calculate the length of each packet at the corresponding sampling point, and use struct.unpack to parse the data from bytes to actual data.

[0044] Then save all the data corresponding to each sampling point as a CSV file to the specified location.

[0045] Step 3: Smooth the data in the saved CSV file, and fit the sine and cosine functions through Fourier transform to obtain the maximum value of the absolute value of the fitting curve.

[0046] Before processing the saved CSV file, block certain types of warnings and check and create a folder to save the PNG image files, obtain the length of the sequence, and calculate the time difference between two consecutive time points in the time series, that is, the time interval or sampling interval, to determine the frequency component;

[0047] Use the to_numpy() method to convert the Pandas Series object into a NumPy array as input to perform the actual Fourier transform;

[0048] After setting the numerical sequence to be smoothed, the filter window size, and the polynomial order, the Savitzky-Golay filter is used to smooth the data.

[0049] Let's identify and remove periodic segments from a set of time series data:

[0050] Use the scipy.signal.find_peaks function to find the peak value in the smoothed_values ​​array, then extract the value and time corresponding to the peak value from the original value sequence and the time series respectively, then print the ordinate value of the horizontal straight line part and delete the horizontal straight line part, and finally clear the NaN value or infinite value in the data;

[0051] Perform Fourier analysis on the original data to obtain frequency components and corresponding FFT values, and identify and filter frequency components through FFT determination;

[0052] Calculate the value that satisfies the cosine and sine function conditions, filter out the cosine and sine function parts, and find the maximum absolute value;

[0053] Finally, draw the graph. Before that, set the graph size, draw the filtered time series and data, and set the x-axis and y-axis labels to represent time and value, respectively. The graph size is 1920*1080 pixels, the graph file is a PNG file, and the resolution is 100dpi.

[0054] Through the above steps, specific frequency components are identified and filtered from the raw data to effectively remove specific periodic interference, thereby more clearly analyzing other features in the data.

[0055] Step 4: Compare the CSV file data extracted in step 2 with the maximum value. If any data point is greater than the maximum value, the file is recorded as a wave recording breakthrough error.

[0056] Analyze the CSV file data extracted from step 2 and compare each data point with the maximum absolute value. If the value of the data point is greater than the maximum absolute value, the file is recorded as a waveform breakthrough error.

[0057] Step 5: Convert the image without the recording breakthrough error in step 4 into PNG file format, and input it into the YOLOv8 neural network for analysis after manual review;

[0058] Step 6: If there are no data points greater than the maximum value in step 4, then record them as the part that has not been broken through. Perform a second derivative on all data points in the CSV file, and save the first-order derivative and second-order derivative to the CSV file. Record the second-order derivative value of the corresponding data point as d. Compare the second-order derivative value d of the data point with the State Grid Power Recording Fault Threshold α. Mark the starting point and end point of the data that exceeds the State Grid Power Recording Fault Threshold α:

[0059] d>α

[0060] The corresponding time is output as t1 and t2, which is convenient for analyzing the effective time of the closing resistor later.

[0061] Step 7: Subtract t2 and t1 obtained in step 6 to get the value of t2-t1, recorded as Δt:

[0062] Δt=t2-t1

[0063] The effective time of input is the comparison between the closing resistance and the standard threshold value β of the closing resistance input time of the national grid;

[0064] Step 8: If the value of t2-t1 in step 7 is greater than the threshold β:

[0065] Δt>β

[0066] The images are converted to PNG format and manually reviewed. The images are then analyzed and judged based on their characteristics before being fed into the YOLOv8 neural network for analysis.

[0067] Step 9: If the value of t2-t1 in step 7 is less than the threshold β:

[0068] Δt<β

[0069] It is recorded as a recording delay error and converted into PNG format, and input into the YOLOv8 neural network for analysis;

[0070] For the YOLOv8 neural network analysis part in step 5 and step 9:

[0071] The PNG image file of the portion without the recording breakthrough error in step 5 and the PNG file of the portion with the value of t2-t1 less than the threshold β in step 9 are input into the YOLOv8 neural network for feature extraction. A BP neural network fault detection model is established based on the extracted feature parameters and the BP neural network. The BP neural network is used to establish a mapping relationship between the feature parameters and the insulation fault information of the closing resistor of the circuit breaker used in the filter bank;

[0072] The actual characteristic parameters of the circuit breaker for the filter bank to be detected are obtained by using the existing characteristic parameter acquisition equipment; the actual characteristic parameters are used as input parameters of the BP neural network fault detection model, and the BP neural network fault detection model is used to determine whether the circuit breaker for the filter bank to be detected is faulty.

[0073] The established BP neural network fault detection model has 3 inputs and 1 output, wherein the 3 inputs and 2 outputs correspond to the characteristic parameters.

[0074] The image size, number of image channels, and image resolution are taken as input parameters of the model, and the waveform error type is taken as output parameter.

[0075] The actual feature parameters are used as the input parameters of the BP neural network model, and the BP neural network model is used to output the corresponding time-stamped moments and adjustment images as the detection results for reference by maintenance personnel.

[0076] Furthermore, the industrial substation closing resistance recording fault wave analysis method based on the YOLOv8 neural network also includes the following steps:

[0077] (1) In the data preparation stage:

[0078] a. Data collection and preprocessing:

[0079] After users upload the comtrade dataset file containing industrial power transient waveform recordings, we analyze the data using Python 3's byte stream processing and file parsing techniques, combined with CFG configuration files and DAT files. We use wavelet transforms to process the time-frequency information and automatically locate abnormal fault waveforms. The processed data is then Fourier transformed to locate the fault's frequency domain characteristics and saved as a CSV file. We then use a Savitzky-Golay filter to smooth the data in the CSV file, while also avoiding excessive distortion of the signal's high-frequency characteristics. The filtered waveform data is then converted into an image (PNG format) for neural network learning.

[0080] b. Data Annotation

[0081] Use the labelm tool manually or in combination with an automatic peak detection algorithm to create and label the bounding box of the recorded fault wave in the image, and define the start and end time points of the fault wave.

[0082] c. Training / validation / test set split

[0083] Based on the prepared dataset above, we divide it into training set, validation set and test set in a ratio of 0.8:0.1:0.1.

[0084] (2) During the network configuration phase

[0085] a. Model selection optimization

[0086] We chose the lightweight YOLOv8-n network structure to accommodate real-time inference requirements. We introduced discrete second-order derivatives and curve fitting algorithms into the YOLOv8 backbone layer to enhance the processing of waveform data characteristics. Furthermore, we used a two-stream network structure: one focused on processing long-term dependencies in time series data (Transformer) and the other on extracting local image features (CNN). We also restructured the original residual connection module into an embedded block module to enhance feature fusion capabilities.

[0087] b. Hyperparameter setting and training

[0088] Pre-training was performed using default hyperparameters and fine-tuned based on preliminary results. We appropriately set the batch_size based on computing resource constraints and adjusted model hyperparameters, such as the learning rate, after initial iterations. Data augmentation and gradient clipping strategies were applied to prevent overfitting. A focal loss function was introduced to emphasize the identification of a small number of fault waves. Training was optimized by gradually reducing the learning rate and employing a learning rate decay strategy.

[0089] c. Evaluation and Tuning

[0090] In resource-constrained environments, statistical metrics such as accuracy, recall, F1 score, and mAP are used to evaluate model performance and fine-tune network architecture and hyperparameters based on these metrics. Visualization tools such as TensorBoard can be used to monitor training progress and performance.

[0091] (3) In the implementation and deployment phase

[0092] After training, the model is quantized and pruned to reduce model size and improve inference speed in edge deployment environments. The quantized and optimized model is integrated into the fault detection system and tested and used in a real-world environment.

[0093] In step 10, after reviewing the input PNG format file through the YOLOv8 neural network in steps 5, 8, and 9, the output current waveform error type is used to determine whether the closing resistor in the industrial substation is damaged due to the rapid increase in current at the moment of closing.

[0094] The above content is only for explaining the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. Any changes made on the basis of the technical solution in accordance with the technical idea proposed by the present invention shall fall within the protection scope of the claims of the present invention.

Claims

1. A method for analyzing fault waves of industrial substation closing resistance recordings based on a YOLOv8 neural network, comprising the following steps: Step 1: The user uploads the industrial power transient recording comtrade dataset file; Step 2: Based on Python 3 and using bytes data stream, parse the CFG configuration file and DAT file in the comtrade file to obtain the actual value of the current data to be tested, and parse and save it as a CSV file to the specified location; Step 3: Smooth the data in the saved CSV file, and fit the sine function and cosine function parts through Fourier transform to obtain the maximum value of the absolute value part of the fitting curve; Step 4: Compare the CSV file data extracted in step 2 with the maximum value. If any data point is greater than the maximum value, the file is recorded as a waveform breakthrough error. Step 5: Convert the part of the image in step 4 that does not contain breakthrough errors into PNG format, compare it with the preset sample, and then input it into the YOLOv8 neural network for analysis; Step 6: If the data in step 4 does not contain a data point greater than the maximum value, record it as the part that has not been broken through. The starting point and the ending point of the data in the CSV file that exceeds the State Grid standard power recording fault threshold α are recorded as t_1 and t_2 respectively. Step 7: Subtract t_2 and t_1 obtained in step 6 to obtain the value t_2-t_1, and compare it with the national grid standard closing resistance input time threshold β; Step 8: If the value of t_2-t_1 in step 7 is greater than the threshold β, it is converted into PNG format, manually reviewed, and then input into the YOLOv8 neural network for analysis; Step 9: If the value of t_2-t_1 in step 7 is less than the threshold β, it is recorded as a recording delay error and converted into PNG format and input into the YOLOv8 neural network for analysis; Step 10: After reviewing the input PNG format file through the YOLOv8 neural network in steps 5, 8, and 9, the output current waveform error type is used to determine whether the closing resistor in the industrial substation is damaged due to the rapid increase in current at the moment of closing.

2. The method for analyzing power fault waves recorded based on the YOLOv8 neural network according to claim 1 is characterized in that: The specific steps in step 2 are as follows: parsing the CFG file in the comtrade file to obtain the number of channels and the name of each channel, the sampling frequency, and the sampling length; using the bytes data stream to read the DAT file in the comtrade file, and calculate the total length and the length of each packet, parsing the data in each packet, and saving it to a CSV file.

3. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 1, characterized in that: The specific steps in step 3 are: performing Fourier transform on the data extracted from the CSV file, smoothing the data, deleting NaN values ​​or infinite values ​​in the data, finding the frequency components, fitting the cosine and sine function parts, and finding the maximum value of the absolute value part of the fitting curve.

4. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 1, characterized in that: The step 4 further includes the following steps: fitting the part of the data that satisfies the cosine function and the sine function, selecting data points in the CSV file one by one and comparing them with the maximum value of the absolute value part of the fitting curve, and determining the recording breakthrough error.

5. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 1, characterized in that: The step 6 also includes the following steps: performing a second derivative on all data points in the CSV file, saving the first-order derivative and the second-order derivative to a CSV file, and comparing them with the State Grid standard power recording fault threshold α. The time corresponding to the starting point and the end point of the data exceeding the threshold α is recorded as t_1 and t_2 respectively.

6. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 5, characterized in that: The value t_2-t_1 obtained by subtracting the corresponding times t_1 and t_2 is the effective time of the closing resistor in the industrial substation.

7. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 1, characterized in that: Said step 8 also includes the following steps: manually reviewing and determining whether the closing resistor in the industrial substation is damaged due to the rapid increase in current at the moment of preventing closing.

8. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 6, characterized in that: The specific steps in steps 5 and 9 are: The waveform reading breakthrough portion after manual review in step 4 is input into the YOLOv8 neural network for analysis; In step 7, the PNG files whose t_2-t_1 values ​​are less than the threshold β are input into the YOLOv8 neural network for feature extraction. The image size, number of image channels, and image resolution are used as input parameters of the model, and whether it is a waveform error type is used as the output parameter.

9. The method for analyzing electric power fault waves recorded based on the YOLOv8 neural network according to claim 8, characterized in that: The specific steps in step 10 are: outputting the current waveform error CSV file obtained through manual review of the YOLOv8 neural network analysis, and judging whether the closing resistor in the industrial substation is damaged due to the rapid increase in current at the moment of preventing closing.

Images