Arcing data intelligent labeling method, system and equipment based on pre-labeling and post-processing and storage medium
By performing time-frequency transformation and model pre-annotation on arc data, combined with manual correction, the problems of low efficiency and high cost in existing arc data annotation technologies have been solved, realizing an efficient, low-cost and accurate intelligent annotation method.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-31
- Publication Date
- 2026-04-17
AI Technical Summary
In existing technologies, arc data annotation relies on purely manual operation, which is inefficient, costly, and lacks objectivity and repeatability. Furthermore, it is difficult for humans to accurately capture millisecond-level changes, resulting in poor model training performance.
An intelligent annotation method based on pre-annotation and post-processing is adopted. The original current time series data is processed by time-frequency transformation, the pre-trained arc detection model is used for feature recognition, and the optimized arc interval is generated by binarization, short-time elimination and neighbor merging. Combined with manual review and correction, the final structured annotation result is generated.
It achieves efficient and low-cost arc data annotation, improves annotation efficiency several times, ensures the rationality and consistency of annotation results, reduces labor costs, and improves annotation accuracy.
Smart Images
Figure CN121881091A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent annotation technology, and in particular to an intelligent annotation method, device and storage medium for arc data based on pre-annotation and post-processing. Background Technology
[0002] As a crucial component of clean energy, photovoltaic (PV) power generation has seen rapid and continuous growth in installed capacity. With increasingly complex application scenarios and rising system voltage and power levels, the electrical safety risks on the DC side of PV systems are becoming increasingly prominent. Among these risks, DC arcing faults, due to their high concealment, concentrated energy, and flammability of surrounding combustibles, have become a major cause of power plant fires. To effectively control this risk, artificial intelligence (AI)-based arcing detection technology is becoming a mainstream solution. The performance of this technology highly depends on the massive, high-quality, and highly consistent arcing current time-series labeled data used for model training.
[0003] Currently, the annotation of arcing data in the industry mainly relies on manual operation. The specific process is as follows: the annotator uses a general signal analysis tool (such as MATLAB, Python script, or a simple graphical interface program) on the local computer to load the raw current signal file (usually in .wav or .csv format), observes its time domain waveform or spectrum with the naked eye, subjectively judges whether arcing characteristics exist based on personal experience, manually marks its start and end time points, and finally exports the results into a structured file.
[0004] First, this purely manual annotation method often requires several minutes or even longer to carefully annotate the arcing data of each inverter. Faced with thousands of hours of raw data, the manual annotation cycle is extremely long, severely slowing down the algorithm iteration rhythm, resulting in low efficiency and high labor costs. Second, the annotation results rely entirely on the personal experience of the annotators. The judgment standards of different people, and even the same person at different times, may fluctuate, leading to a lack of objectivity and repeatability in the annotation results, directly affecting the model training effect, and exhibiting strong subjectivity and poor consistency. In addition, the start and end of arcing often manifest as a gradual process, making it difficult for humans to accurately capture the millisecond-level change points. The annotation boundaries generally have ambiguity and random errors, which restricts the model's ability to accurately characterize arcing features and results in limited positioning accuracy. Summary of the Invention
[0005] The purpose of this invention is to overcome the shortcomings of existing technologies and provide an intelligent annotation method for arc data based on pre-annotation and post-processing, comprising the following steps: S1: Obtain the original current time series data to be labeled, and perform time-frequency transformation processing on the original current time series data to obtain the corresponding time-frequency feature data; S2: Input the time-frequency feature data into the pre-trained arc detection model, and use the arc detection model to identify the arc features in the time-frequency feature data, and output the corresponding arc probability sequence; S3: Set a preset classification threshold, and perform binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to a candidate arc point and the second value corresponds to a non-arc point. S4: Traverse the initial binary mask sequence, identify all candidate arc intervals formed by consecutive first values, calculate the duration of each candidate arc interval, and remove candidate arc intervals with a duration less than a first preset threshold to obtain the intermediate binary mask sequence. S5: For the remaining candidate arcing intervals of the intermediate binary mask sequence, merge adjacent intervals with a time interval less than or equal to the second preset threshold to form an optimized arcing interval; S6: Visualize the optimized arc range as a pre-annotation result, receive and respond to the user's correction operation, integrate the correction operation with the pre-annotation result, and generate the final annotation result; S7: Output the final annotation result according to the preset structured format, which includes at least the file name, the start time of the arc interval, the end time of the arc interval, and the annotation source field.
[0006] Preferably, in step S1, the original current time-series data is subjected to time-frequency transformation processing to obtain corresponding time-frequency feature data, which further includes: S11: Cut the original current timing data stream into equal-length segments according to the preset frame length and frame shift to obtain a series of signal frames; S12: Multiply each of the signal frames with a preset window function to obtain a windowed signal frame; S13: Perform a fast Fourier transform algorithm on each windowed signal frame to convert the signal from the time domain to the frequency domain and calculate the amplitude spectrum of each windowed signal frame on different frequency components. S14: Arrange and combine all the amplitude spectra according to the time sequence of all signal frames to form the time-frequency feature data containing the time dimension and the frequency dimension.
[0007] Preferably, in step S2, the arc detection model is used to identify arc features in the time-frequency feature data and output the corresponding arc probability sequence, further including: S21: Standardize the time-frequency feature data to obtain standard time-frequency feature data; S22: The standard time-frequency feature data is input into the pre-trained arc detection model. The arc detection model learns and extracts deep abstract features related to arcing from the standard time-frequency feature data through multi-layer convolution operations. The deep abstract features are analyzed in the time dimension through recurrent layers or deep convolutional layers to capture the evolution rules, contextual dependencies and temporal correlations in the time dimension to obtain deep spatiotemporal features. The deep spatiotemporal features are activated by a classification layer to calculate the arcing probability sequence.
[0008] Preferably, in step S3, the arc probability sequence is binarized based on the preset classification threshold to generate a preliminary binary mask sequence, which further includes: S31: Traverse each time point in the arcing probability sequence, compare the current probability value of the current time point with the preset classification threshold. If the current probability value is greater than or equal to the preset classification threshold, then determine that the time point is a candidate arcing point and map the decision result to a first value. If the current probability value is less than the preset classification threshold, then determine that the time point is a non-arcing point and map the decision result to a second value. Wherein, the first value is 1 and the second value is 0. S32: Arrange the mapping results of all time points in sequence according to the original time order corresponding to the arcing probability sequence to form the preliminary binary mask sequence that corresponds one-to-one with the time axis of the original current time series data.
[0009] Preferably, in step S4, identifying candidate arc intervals formed by all consecutive first values, calculating the duration of each candidate arc interval, eliminating candidate arc intervals with durations less than a first preset threshold, and generating an intermediate binary mask sequence further includes: S41: Scan the preliminary binary mask sequence in chronological order, monitor the jump state of the values in the sequence, and when a value is detected to jump from the second value to the first value, mark the jump point as the starting point of a candidate arcing interval. Continue scanning along the time axis, and when a value is detected to jump from the first value to the second value, mark the jump point as the ending point of the corresponding candidate arcing interval. Repeat the above operation to identify all independent candidate arcing intervals composed of consecutive first values in the preliminary binary mask sequence, and record the coordinates of the starting point and ending point of each interval on the time axis of the original current timing data. S42: For each identified candidate arcing interval, calculate the actual duration of the interval based on the time coordinates of its start and end points, and compare the actual duration with a first preset threshold. If the actual duration is greater than or equal to the first preset threshold, the interval is determined to meet the minimum effective physical duration condition of the arcing event and is retained. If the actual duration is less than the first preset threshold, the interval is determined to be an invalid short-term interference interval and is eliminated. S43: Based on the judgment result, the preliminary binary mask sequence is modified by changing the first value in all short-term interference intervals that are judged to be invalid to the second value, while the first value in the remaining candidate arcing intervals remains unchanged. After the modification is completed, an intermediate binary mask sequence containing only valid candidate arcing intervals is generated.
[0010] Preferably, in step S5, for the remaining candidate arcing intervals of the intermediate binary mask sequence, adjacent intervals with a time interval less than or equal to a second preset threshold are merged to form an optimized arcing interval, further including: S51: Sort all valid candidate arcing intervals retained in the intermediate binary mask sequence in chronological order, extract the time coordinates of the termination point of each valid candidate arcing interval and the time coordinates of the starting point of the next adjacent valid candidate arcing interval, calculate the time difference between the termination point and the starting point, and obtain the adjacent time interval between two adjacent valid candidate arcing intervals. S52: Compare each of the adjacent time intervals with the second preset threshold one by one. If the adjacent time interval is greater than the second preset threshold, the two adjacent intervals are determined to be two independent events, and the intervals are kept independent and not merged. If the adjacent time interval is less than or equal to the second preset threshold, the two adjacent intervals are determined to be continuous stages of the same arcing event, and the merging condition is met, so the interval merging operation needs to be performed. S53: Perform merging processing on adjacent valid candidate arc interval pairs that meet the merging conditions. Modify all second values between the end point of the previous interval and the start point of the next interval to first values, so that the two intervals form a continuous sequence of first values. After traversing all adjacent interval pairs and completing the merging, output the optimized arc interval composed of continuous first values.
[0011] Preferably, in step S6, receiving and responding to the user's correction operation, integrating the correction operation with the pre-annotation result, and generating the final annotation result further includes: S61: Map the optimized arcing intervals one by one to the time axis of the original current time series data, determine the specific time range corresponding to each optimized arcing interval, construct a visualization display layer in the graphical user interface, draw the waveform corresponding to the original current time series data at the bottom layer, and overlay a visualization annotation layer on the top layer, using a preset highlight color band to annotate the time range of each optimized arcing interval. S62: Based on the superimposed display effect of the original current waveform and the highlighted color band in the graphical user interface, the user evaluates the rationality of the pre-labeling results and performs operations including boundary adjustment, interval deletion, adding labels and interval splitting / merging. S63: The graphical user interface responds to user correction operations in real time, and synchronously updates the superimposed display effect of the original current waveform and the highlight color band. After the user completes all review and correction operations and confirms them, the arcing interval corresponding to the highlight color band finally displayed in the current graphical interface is taken as the final annotation result. The final annotation result is associated with the user's correction operation type and operation parameters.
[0012] Based on the same concept, the present invention also provides an intelligent annotation system for arc data based on pre-annotation and post-processing, including: The time-frequency feature module acquires the original current time-series data to be labeled, performs time-frequency transformation processing on the original current time-series data, and obtains the corresponding time-frequency feature data. The model inference module inputs the time-frequency feature data into the pre-trained arc detection model, performs arc feature recognition on the time-frequency feature data through the arc detection model, and outputs the corresponding arc probability sequence. The threshold decision module sets a preset classification threshold and performs binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to a candidate arc point and the second value corresponds to a non-arc point. The short pulse filtering module traverses the initial binary mask sequence, identifies all candidate arc intervals formed by consecutive first values, calculates the duration of each candidate arc interval, removes candidate arc intervals with a duration less than a first preset threshold, and generates an intermediate binary mask sequence. The time continuity repair module performs adjacent interval correlation analysis on the intermediate binary mask sequence, calculates the time interval between two adjacent candidate arcing intervals, and merges adjacent intervals with a time interval less than or equal to a second preset threshold into an optimized arcing interval. The visualization and interactive correction module takes the original current time series data and the optimized arcing interval as pre-labeling results and displays them as a visual layer on the waveform of the original current time series data. It receives user review and correction operations on the pre-labeling results and generates the final labeling results based on the operation results. The annotation result output module outputs the final annotation result according to a preset structured format, which includes at least the file name, the start time of the arc interval, the end time of the arc interval, and the annotation source field.
[0013] Based on the same concept, the present invention also provides a computer device, including a memory and a processor, wherein the memory stores computer-readable instructions, and when executed by the processor, the computer-readable instructions cause the processor to perform the steps of the arc data intelligent annotation method based on pre-annotation and post-processing as described in any one of the embodiments.
[0014] Based on the same concept, the present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent annotation method for arc data based on pre-annotation and post-processing as described in any one of the embodiments.
[0015] Compared with the prior art, the beneficial effects of the present invention are: This invention obtains an arc probability sequence by performing time-frequency transformation on the original current time-series data to be labeled and inputting it into a pre-trained arc detection model. Based on this sequence, it generates an optimized arc interval as a pre-labeling result through binarization, short-time elimination, and neighbor merging post-processing. Finally, after manual review and correction, it outputs a structured labeling result. This upgrades the time-consuming and labor-intensive manual labeling mode to a collaborative intelligent labeling paradigm of machine intelligent pre-labeling, rule post-processing optimization, and efficient manual review. While ensuring labeling quality, it improves data labeling efficiency several times and significantly reduces labor costs.
[0016] This invention binarizes the probability sequence output by the model and then performs two post-processing steps: short-term interval elimination based on duration and neighboring interval merging based on interval time. This enables automatic repair and optimization of the two core defects in the original model pre-labeling results: short-term noise false alarms and breaks in valid events. This significantly enhances the rationality and completeness of the pre-labeling results, thereby shifting the focus of manual review from processing a large number of invalid segments to fine-tuning the boundaries, further improving overall work efficiency.
[0017] This invention overlays the optimized pre-annotation results onto the original waveform as a visual layer for manual review and allows users to directly and interactively correct them. This achieves a deep integration of the high efficiency and consistency of machine intelligence with the high precision and strong judgment of human experts. While utilizing machines to process massive amounts of data and unify annotation standards, it relies on human wisdom to solve problems of ambiguity, complexity, and boundary judgment, thereby achieving an optimal balance between efficiency and accuracy and ultimately producing high-quality, highly consistent annotation data. Attached Figure Description
[0018] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention.
[0019] Figure 1 This is a flowchart of an intelligent annotation method for arc data based on pre-annotation and post-processing according to the present invention; Figure 2 This is another flowchart of the intelligent annotation method for arc data based on pre-annotation and post-processing according to the present invention. Detailed Implementation
[0020] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention. Obviously, the described embodiments are only some, not all, of the embodiments described in this application. All other embodiments obtained by those skilled in the art based on the embodiments in this application without creative effort are within the scope of protection of this application.
[0021] Those skilled in the art will understand that, unless otherwise stated, the singular forms “a” and “an” used herein, and “the”, may also include the plural forms. It should be further understood that the term “comprising” as used in this specification means the presence of the stated features, integers, steps, operations, elements, and / or components, but does not exclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and / or groups thereof.
[0022] First Embodiment Please see Figure 1 and Figure 2 As shown in the figure, this embodiment provides an intelligent annotation method for arc data based on pre-annotation and post-processing, which includes the following steps: S1: Obtain the original current time series data to be labeled (such as a current signal in .wav format), perform time-frequency transformation processing on the original current time series data, and obtain the corresponding time-frequency feature data.
[0023] Preferably, in step S1, the original current time-series data is subjected to time-frequency transformation processing to obtain corresponding time-frequency characteristic data, which further includes: S11: The original current timing data stream is cut into equal lengths according to a preset frame length (e.g., 100 milliseconds) and frame shift to obtain a series of signal frames. Specifically, in this embodiment, in order to ensure the continuity between frames and avoid feature loss caused by boundary effects, a certain proportion of frame shift is set between adjacent frames (e.g., 50% overlap, that is, 50% of the time period data of adjacent frames overlap). S12: Multiply each signal frame with a preset window function (e.g., Hamming window) to obtain a windowed signal frame. Specifically, in this embodiment, the start and end points of each frame signal are smoothed to reduce the spectral leakage caused by direct truncation, so that the subsequent frequency domain analysis results are more accurate and more focused on the information of the center time period of the frame. S13: Perform a fast Fourier transform algorithm on each windowed signal frame to convert the signal from the time domain to the frequency domain and calculate the amplitude spectrum of each windowed signal frame at different frequency components (i.e., the distribution of signal energy at different frequencies). S14: Arrange and combine all amplitude spectra according to the time sequence of all signal frames to form time-frequency feature data (i.e., two-dimensional data structure) containing time and frequency dimensions. Specifically, in this embodiment, each row corresponds to a specific frequency point or frequency range, and each column corresponds to a time segment represented by an original signal frame. Each element value in the matrix represents the amplitude or energy intensity of the current signal at a specific time segment (column) and a specific frequency (row).
[0024] S2: Input the time-frequency feature data into the pre-trained arc detection model, and use the arc detection model to identify the arc features in the time-frequency feature data and output the corresponding arc probability sequence.
[0025] Preferably, in step S2, the arc detection model is used to identify arc features in the time-frequency feature data and output the corresponding arc probability sequence, further including: S21: Standardize the time-frequency feature data to obtain standard time-frequency feature data. Specifically, in this embodiment, the standardization process includes scaling the amplitude data in the matrix to a fixed numerical range (e.g., [0, 1] or [-1, 1]) and possibly adjusting the size of the matrix (e.g., height, width) to match the fixed dimensions of the model input layer. S22: Input the standard time-frequency feature data into the pre-trained arc detection model. The arc detection model learns and extracts deep abstract features related to arcing (such as specific frequency band energy patterns, time-frequency texture structures, etc.) from the standard time-frequency feature data through multi-layer convolutional operations. At the same time, it suppresses background noise features unrelated to arcing in this way. The deep abstract features are analyzed in the time dimension through recurrent layers or deep convolutional layers to capture the evolution rules, contextual dependencies and temporal correlations in the time dimension to obtain deep spatiotemporal features (to identify the typical dynamic patterns of arcing events, such as the onset, continuation and extinction). The deep spatiotemporal features are activated by the classification layer to calculate the arcing probability sequence.
[0026] S3: Set a preset classification threshold, and perform binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to the candidate arc point and the second value corresponds to the non-arc point.
[0027] Preferably, in step S3, the arc probability sequence is binarized based on a preset classification threshold to generate a preliminary binary mask sequence, further including: S31: Iterate through each time point in the arc probability sequence and obtain the current probability value for the current time point. Compared with the preset classification threshold Comparison ( =0.75), if the current probability value is greater than or equal to the preset classification threshold ( If the current probability value is less than the preset classification threshold, then the current time point is determined to be a candidate arc point, and the decision result is mapped to the first value. If the current probability value is less than the preset classification threshold, then the current time point is determined to be a non-arc point, and the decision result is mapped to the second value, where the first value is 1 and the second value is 0. S32: Arrange the mapping results of all time points in sequence according to the original time order corresponding to the arcing probability sequence to form a preliminary binary mask sequence that corresponds one-to-one with the time axis of the original current time series data.
[0028] S4: Traverse the initial binary mask sequence, identify all candidate arc intervals formed by consecutive first values, calculate the duration of each candidate arc interval, and discard those with shorter durations. Less than the first preset threshold The candidate arc intervals are used to obtain the intermediate binary mask sequence.
[0029] Preferably, in step S4, identifying candidate arc intervals formed by all consecutive first values, calculating the duration of each candidate arc interval, eliminating candidate arc intervals with durations less than a first preset threshold, and generating an intermediate binary mask sequence further includes: S41: Scan the preliminary binary mask sequence in chronological order, monitor the jump state of the values in the sequence, and when the value is detected to jump from the second value to the first value, mark the jump point as the starting point of a candidate arcing interval. Continue scanning along the time axis, and when the value is detected to jump from the first value to the second value, mark the jump point as the ending point of the corresponding candidate arcing interval. Repeat the above operation to identify all independent candidate arcing intervals composed of consecutive first values in the preliminary binary mask sequence, and record the coordinates of the starting point and ending point of each interval on the time axis of the original current time series data. S42: For each identified candidate arcing interval, calculate the actual duration of the interval based on the time coordinates of its start and end points, and compare the actual duration with a first preset threshold. If the actual duration is greater than or equal to the first preset threshold, the interval is determined to meet the minimum effective physical duration condition of the arcing event and is retained. If the actual duration is less than the first preset threshold, the interval is determined to be an invalid short-term interference interval and is removed. S43: Based on the judgment result, the preliminary binary mask sequence is corrected. The first value in all short-term interference intervals that are judged to be invalid is changed to the second value, while the first value in the remaining candidate arcing intervals remains unchanged. After the correction is completed, an intermediate binary mask sequence containing only valid candidate arcing intervals is generated.
[0030] S5: For the remaining candidate arcing intervals of the intermediate binary mask sequence, merge adjacent intervals with a time interval less than or equal to the second preset threshold to form an optimized arcing interval.
[0031] Preferably, in step S5, for the remaining candidate arcing intervals of the intermediate binary mask sequence, adjacent intervals with a time interval less than or equal to a second preset threshold are merged to form an optimized arcing interval, further including: S51: Sort all valid candidate arcing intervals retained in the intermediate binary mask sequence in chronological order, extract the time coordinates of the termination point of each valid candidate arcing interval (interval A) and the time coordinates of the starting point of the next adjacent valid candidate arcing interval (interval B), calculate the time difference between the termination point and the starting point, and obtain the adjacent time interval between two adjacent valid candidate arcing intervals. S52: Each adjacent time interval With the second preset threshold A comparison is performed one by one. If the time interval between adjacent intervals is greater than the second preset threshold, the two adjacent intervals are determined to be two independent events, and the intervals are kept independent and not merged. If the time interval between adjacent intervals is less than or equal to the second preset threshold, the intervals are not merged. If two adjacent intervals are determined to be continuous phases of the same arc-drawing event, and the merging condition is met, then the interval merging operation needs to be performed. S53: Perform merging processing on adjacent valid candidate arc interval pairs that meet the merging conditions. Modify all second values between the end point of the previous interval and the start point of the next interval to first values, so that the two intervals form a continuous sequence of first values. After traversing all adjacent interval pairs and completing the merging, output the optimized arc interval composed of continuous first values.
[0032] S6: Visualize the optimized arc range as a pre-annotation result, receive and respond to user correction operations, integrate the correction operations with the pre-annotation results, and generate the final annotation result. Specifically, it receives the optimized arc range and the original time series data waveform.
[0033] Preferably, in step S6, receiving and responding to the user's correction operation, integrating the correction operation with the pre-annotation results, and generating the final annotation result further includes: S61: Map the optimized arcing intervals one by one to the time axis of the original current time series data, determine the specific time range corresponding to each optimized arcing interval, build a visualization display layer in the graphical user interface, draw the waveform corresponding to the original current time series data at the bottom layer, and overlay a visualization annotation layer on the top layer, using a preset highlight color band to annotate the time range of each optimized arcing interval. S62: Based on the overlay display effect of the original current waveform and the highlighted color band in the graphical user interface, the user evaluates the rationality of the pre-annotation results and performs operations including boundary adjustment, interval deletion, adding annotations, and interval splitting / merging. Specifically, in this embodiment, boundary fine-tuning: by dragging the left or right boundary of the color band with the mouse, the start or end time point of the arcing event is directly adjusted to correct the machine's recognition deviation at the event boundary; interval deletion: clicking the delete button removes the entire highlighted color band. This operation is used to correct the machine's false alarms, that is, to delete those pre-annotated intervals that are not actually arcing; interval supplementation: in areas on the waveform that are not highlighted but the user believes there is arcing, a new highlighted color band is directly drawn by dragging the mouse. This operation is used to compensate for the machine's false negatives, that is, to supplement the annotation of arcing events that the machine has not recognized. S63: The graphical user interface responds to user correction operations in real time, and synchronously updates the superimposed display effect of the original current waveform and the highlight color band. After the user completes all review and correction operations and confirms them, the arcing interval corresponding to the highlight color band finally displayed in the current graphical interface is taken as the final annotation result. The final annotation result is associated with the user's correction operation type and operation parameters.
[0034] S7: Output the final annotation result according to the preset structured format. The structured format includes at least the file name, the start time of the arcing interval, the end time of the arcing interval, and the annotation source field. Specifically, in this embodiment, the file name corresponds to the unique identifier of the original current data file, which is used to establish the traceability relationship between the annotation result and the original data. The start time is the precise start timestamp of the arcing event on the original data timeline, and the end time is the precise end timestamp of the arcing event on the original data timeline. The annotation source field records the generation method of the event entry, including pre-annotation and manual correction. Pre-annotation indicates that the event is completely generated by the intelligent process of steps S1-S5 and has not been modified by the user in step S6. Manual correction indicates that the event originates from intelligent pre-annotation but has been modified by the user's boundary, or is completely created by the user manually in step S6. The extracted field information is logically organized and encapsulated according to the preset, general structured data format, usually using JSON or XML formats.
[0035] Second Embodiment Based on the same concept, the present invention also provides an intelligent annotation system for arc data based on pre-annotation and post-processing, including: The time-frequency feature module acquires the original current time-series data to be labeled, performs time-frequency transformation processing on the original current time-series data, and obtains the corresponding time-frequency feature data. The model inference module inputs the time-frequency feature data into the pre-trained arc detection model, performs arc feature recognition on the time-frequency feature data through the arc detection model, and outputs the corresponding arc probability sequence. The threshold decision module sets a preset classification threshold and performs binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to a candidate arc point and the second value corresponds to a non-arc point. The short pulse filtering module traverses the initial binary mask sequence, identifies all candidate arc intervals formed by consecutive first values, calculates the duration of each candidate arc interval, removes candidate arc intervals with a duration less than a first preset threshold, and generates an intermediate binary mask sequence. The time continuity repair module performs adjacent interval correlation analysis on the intermediate binary mask sequence, calculates the time interval between two adjacent candidate arcing intervals, and merges adjacent intervals with a time interval less than or equal to a second preset threshold into an optimized arcing interval. The visualization and interactive correction module takes the original current time series data and the optimized arcing interval as pre-labeling results and displays them as a visual layer on the waveform of the original current time series data. It receives user review and correction operations on the pre-labeling results and generates the final labeling results based on the operation results. The annotation result output module outputs the final annotation result according to a preset structured format, which includes at least the file name, the start time of the arc interval, the end time of the arc interval, and the annotation source field.
[0036] Third Embodiment In some embodiments of this application, a computer device is also provided, including a memory and a processor, wherein the memory stores computer-readable instructions, and when executed by the processor, the computer-readable instructions cause the processor to perform the steps of the arc data intelligent annotation method based on pre-annotation and post-processing as described in any one of the first embodiments.
[0037] The present invention also provides a storage medium storing computer-readable instructions, which, when executed by one or more processors, cause the one or more processors to perform the steps of the intelligent annotation method for arc data based on pre-annotation and post-processing as described in any one of the first embodiments.
[0038] It is understood that, for the aforementioned intelligent annotation methods for arc data based on pre-annotation and post-processing, if all are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer server or a network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes: USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, optical disks, and other media capable of storing program code.
[0039] Computer-readable storage media may include data signals propagated in baseband or as part of a carrier wave, carrying readable program code. Such propagated data signals may take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A readable storage medium may also be any readable medium other than a readable storage medium that can transmit, propagate, or transfer a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the readable storage medium may be transmitted using any suitable medium, including but not limited to wireless, wired, optical fiber, RF, etc., or any suitable combination thereof.
[0040] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.
Claims
1. A smart annotation method for arc data based on pre-annotation and post-processing, characterized in that, Includes the following steps: S1: Obtain the original current time series data to be labeled, and perform time-frequency transformation processing on the original current time series data to obtain the corresponding time-frequency feature data; S2: Input the time-frequency feature data into the pre-trained arc detection model, and use the arc detection model to identify the arc features in the time-frequency feature data, and output the corresponding arc probability sequence; S3: Set a preset classification threshold, and perform binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to a candidate arc point and the second value corresponds to a non-arc point. S4: Traverse the initial binary mask sequence, identify all candidate arc intervals formed by consecutive first values, calculate the duration of each candidate arc interval, and remove candidate arc intervals with a duration less than a first preset threshold to obtain the intermediate binary mask sequence. S5: For the remaining candidate arcing intervals of the intermediate binary mask sequence, merge adjacent intervals with a time interval less than or equal to the second preset threshold to form an optimized arcing interval; S6: Visualize the optimized arc range as a pre-annotation result, receive and respond to the user's correction operation, integrate the correction operation with the pre-annotation result, and generate the final annotation result; S7: Output the final annotation result according to the preset structured format, which includes at least the file name, the start time of the arc interval, the end time of the arc interval, and the annotation source field.
2. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S1, the original current time-series data is subjected to time-frequency transformation processing to obtain corresponding time-frequency characteristic data, which further includes: S11: Cut the original current timing data stream into equal-length segments according to the preset frame length and frame shift to obtain a series of signal frames; S12: Multiply each of the signal frames with a preset window function to obtain a windowed signal frame; S13: Perform a fast Fourier transform algorithm on each windowed signal frame to convert the signal from the time domain to the frequency domain and calculate the amplitude spectrum of each windowed signal frame on different frequency components. S14: Arrange and combine all the amplitude spectra according to the time sequence of all signal frames to form the time-frequency feature data containing the time dimension and the frequency dimension.
3. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S2, the arc detection model is used to identify arc features in the time-frequency feature data, and the corresponding arc probability sequence is output, which further includes: S21: Standardize the time-frequency feature data to obtain standard time-frequency feature data; S22: The standard time-frequency feature data is input into the pre-trained arc detection model. The arc detection model learns and extracts deep abstract features related to arcing from the standard time-frequency feature data through multi-layer convolution operations. The deep abstract features are analyzed in the time dimension through recurrent layers or deep convolutional layers to capture the evolution rules, contextual dependencies and temporal correlations in the time dimension to obtain deep spatiotemporal features. The deep spatiotemporal features are activated by a classification layer to calculate the arcing probability sequence.
4. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S3, the arc probability sequence is binarized based on the preset classification threshold to generate a preliminary binary mask sequence, further including: S31: Traverse each time point in the arcing probability sequence, compare the current probability value of the current time point with the preset classification threshold. If the current probability value is greater than or equal to the preset classification threshold, then determine that the time point is a candidate arcing point and map the decision result to a first value. If the current probability value is less than the preset classification threshold, then determine that the time point is a non-arcing point and map the decision result to a second value. Wherein, the first value is 1 and the second value is 0. S32: Arrange the mapping results of all time points in sequence according to the original time order corresponding to the arcing probability sequence to form the preliminary binary mask sequence that corresponds one-to-one with the time axis of the original current time series data.
5. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S4, candidate arc intervals consisting of all consecutive first values are identified, the duration of each candidate arc interval is calculated, candidate arc intervals with durations less than a first preset threshold are removed, and an intermediate binary mask sequence is generated. This further includes: S41: Scan the preliminary binary mask sequence in chronological order, monitor the jump state of the values in the sequence, and when a value is detected to jump from the second value to the first value, mark the jump point as the starting point of a candidate arcing interval. Continue scanning along the time axis, and when a value is detected to jump from the first value to the second value, mark the jump point as the ending point of the corresponding candidate arcing interval. Repeat the above operation to identify all independent candidate arcing intervals composed of consecutive first values in the preliminary binary mask sequence, and record the coordinates of the starting point and ending point of each interval on the time axis of the original current timing data. S42: For each identified candidate arcing interval, calculate the actual duration of the interval based on the time coordinates of its start and end points, and compare the actual duration with a first preset threshold. If the actual duration is greater than or equal to the first preset threshold, the interval is determined to meet the minimum effective physical duration condition of the arcing event and is retained. If the actual duration is less than the first preset threshold, the interval is determined to be an invalid short-term interference interval and is eliminated. S43: Based on the judgment result, the preliminary binary mask sequence is modified by changing the first value in all short-term interference intervals that are judged to be invalid to the second value, while the first value in the remaining candidate arcing intervals remains unchanged. After the modification is completed, an intermediate binary mask sequence containing only valid candidate arcing intervals is generated.
6. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S5, for the remaining candidate arcing intervals of the intermediate binary mask sequence, adjacent intervals with a time interval less than or equal to a second preset threshold are merged to form an optimized arcing interval, which further includes: S51: Sort all valid candidate arcing intervals retained in the intermediate binary mask sequence in chronological order, extract the time coordinates of the termination point of each valid candidate arcing interval and the time coordinates of the starting point of the next adjacent valid candidate arcing interval, calculate the time difference between the termination point and the starting point, and obtain the adjacent time interval between two adjacent valid candidate arcing intervals. S52: Compare each of the adjacent time intervals with the second preset threshold one by one. If the adjacent time interval is greater than the second preset threshold, the two adjacent intervals are determined to be two independent events, and the intervals are kept independent and not merged. If the adjacent time interval is less than or equal to the second preset threshold, the two adjacent intervals are determined to be continuous stages of the same arcing event, and the merging condition is met, so the interval merging operation needs to be performed. S53: Perform merging processing on adjacent valid candidate arc interval pairs that meet the merging conditions. Modify all second values between the end point of the previous interval and the start point of the next interval to first values, so that the two intervals form a continuous sequence of first values. After traversing all adjacent interval pairs and completing the merging, output the optimized arc interval composed of continuous first values.
7. The intelligent annotation method for arc data based on pre-annotation and post-processing according to claim 1, characterized in that, In step S6, the user's correction operation is received and responded to, the correction operation is integrated with the pre-annotation result, and the final annotation result is generated, further including: S61: Map the optimized arcing intervals one by one to the time axis of the original current time series data, determine the specific time range corresponding to each optimized arcing interval, construct a visualization display layer in the graphical user interface, draw the waveform corresponding to the original current time series data at the bottom layer, and overlay a visualization annotation layer on the top layer, using a preset highlight color band to annotate the time range of each optimized arcing interval. S62: Based on the superimposed display effect of the original current waveform and the highlighted color band in the graphical user interface, the user evaluates the rationality of the pre-labeling results and performs operations including boundary adjustment, interval deletion, adding labels and interval splitting / merging. S63: The graphical user interface responds to user correction operations in real time, and synchronously updates the superimposed display effect of the original current waveform and the highlight color band. After the user completes all review and correction operations and confirms them, the arcing interval corresponding to the highlight color band finally displayed in the current graphical interface is taken as the final annotation result. The final annotation result is associated with the user's correction operation type and operation parameters.
8. A smart annotation system for arc data based on pre-annotation and post-processing, characterized in that, include: The time-frequency feature module acquires the original current time-series data to be labeled, performs time-frequency transformation processing on the original current time-series data, and obtains the corresponding time-frequency feature data. The model inference module inputs the time-frequency feature data into the pre-trained arc detection model, performs arc feature recognition on the time-frequency feature data through the arc detection model, and outputs the corresponding arc probability sequence. The threshold decision module sets a preset classification threshold and performs binarization processing on the arc probability sequence based on the preset classification threshold to generate a preliminary binary mask sequence, wherein the first value in the preliminary binary mask sequence corresponds to a candidate arc point and the second value corresponds to a non-arc point. The short pulse filtering module traverses the initial binary mask sequence, identifies all candidate arc intervals formed by consecutive first values, calculates the duration of each candidate arc interval, removes candidate arc intervals with a duration less than a first preset threshold, and generates an intermediate binary mask sequence. The time continuity repair module performs adjacent interval correlation analysis on the intermediate binary mask sequence, calculates the time interval between two adjacent candidate arcing intervals, and merges adjacent intervals with a time interval less than or equal to a second preset threshold into an optimized arcing interval. The visualization and interactive correction module takes the original current time series data and the optimized arcing interval as pre-labeling results and displays them as a visual layer on the waveform of the original current time series data. It receives user review and correction operations on the pre-labeling results and generates the final labeling results based on the operation results. The annotation result output module outputs the final annotation result according to a preset structured format, which includes at least the file name, the start time of the arc interval, the end time of the arc interval, and the annotation source field.
9. A computer device, characterized in that, The system includes a memory and a processor, wherein the memory stores computer-readable instructions that, when executed by the processor, cause the processor to perform the steps of the intelligent annotation method for arc data based on pre-annotation and post-processing as described in any one of claims 1 to 8.
10. A storage medium storing computer-readable instructions, characterized in that, When the computer-readable instructions are executed by one or more processors, the one or more processors perform the steps of the intelligent annotation method for arc data based on pre-annotation and post-processing as described in any one of claims 1 to 8.