A method and system for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning

Abstract

This invention discloses a method and system for extracting feature points from the dynamic response waveform of asphalt pavement based on incremental learning. First, a configuration file is loaded, then edited to obtain an edited file. Next, the sensor's acquisition frequency and filtering parameters are input into the edited file to obtain a parameter-input file. This parameter-input file is then processed to obtain a processed file. Finally, it is determined whether the processed file contains machine learning data, and the results are obtained indicating whether machine learning data is present or absent. This invention achieves the function of using a depth-based extreme point detection method to sort local extremes by amplitude and prioritize the extraction of the two most significant valleys or peaks in the waveform. This not only effectively eliminates interference from local small fluctuations and boundary noise but also ensures that the extreme point detection results conform to the physical meaning of the mechanical response while possessing good stability and interpretability.

Description

Technical Field

[0001] This invention relates to the field of asphalt pavement performance monitoring technology, specifically to a method and system for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning. Background Technology

[0002] Long-term performance monitoring of asphalt pavement is an essential means of conducting long-term performance research on asphalt pavement. The most direct method is to analyze the response changes of stress-strain sensors pre-embedded inside the asphalt pavement structure layers. That is, when a vehicle load passes over the sensor, the sensor will produce corresponding changes. When estimating the pavement structure life, the life is estimated by the number of times a standard axle load is applied. Therefore, the true stress characteristics and patterns of the pavement can be obtained and analyzed by analyzing the response changes of the sensors under the action of a standard vehicle. Currently, the commonly used method is to use a standard vehicle to load under specific working conditions. During loading, multi-factor coupling analysis is required based on specific temperature, speed, load, and loading location. At the same time, different types of sensors are embedded in the bottom of different asphalt layers. Therefore, in multi-condition coupling analysis, each sensor corresponds to one acquisition signal, resulting in a huge amount of data and inconsistent data availability. If manual judgment is used, it will require a lot of effort.

[0003] Currently, in the performance monitoring of asphalt pavements, the loading position of vehicles is often not completely fixed, which easily leads to errors in the quality of the response waveform. Furthermore, due to the influence of the number of parallel tests, the number of sensors, and coupling factors, the number of response waveforms to be analyzed is enormous, making manual analysis extremely labor-intensive. In addition, the response waveforms of different types of sensors differ, resulting in differences in the required feature points. Therefore, it is necessary to design a method and system for extracting feature points of asphalt pavement dynamic response waveforms based on incremental learning. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and to better and more effectively solve the problems in the current performance monitoring of asphalt pavements. These problems stem from the fact that the loading position of vehicles is not always completely fixed, leading to errors in the quality of the response waveform. Furthermore, the large number of response waveforms to be analyzed due to the influence of parallel test numbers, sensor counts, and coupling factors makes manual analysis extremely labor-intensive. Additionally, the differences in response waveforms between different types of sensors result in differences in the required feature points. This invention provides a method and system for extracting feature points from the dynamic response waveform of asphalt pavements based on incremental learning. It achieves the function of sorting local extrema by amplitude using a depth-based extreme point detection method and prioritizing the extraction of the two most significant valleys or peaks in the waveform. This not only effectively eliminates the interference of local small fluctuations and boundary noise but also ensures that the extreme point detection results conform to the physical meaning of the mechanical response while possessing good stability and interpretability.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A method for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning includes the following steps: Step A: Load the configuration file, then edit the configuration file and obtain the edited file; Step B involves inputting the sensor's acquisition frequency and filtering parameters into the edited file and obtaining the parameter input file, then processing the parameter input file and obtaining the processed file. Step C: Determine whether the processed file contains machine learning data and obtain the judgment results including those containing machine learning data and those not containing machine learning data. Step D: If the judgment result is that there is no machine learning data, then first predict the waveform feature point rules and then generate the visualization interaction window. Step E: If the judgment result indicates the existence of machine learning data, then a visual interactive window is directly generated. Step F: Adjust the machine learning offset of the visual interactive window and obtain the machine learning offset amount; Step G: Calculate the final prediction time based on the machine learning offset and search for the closest waveform point near the final prediction time; Step H involves discarding the closest waveform points that meet the discard criteria and using the retained closest waveform points as the result of extracting the dynamic response waveform feature points of the asphalt pavement.

[0006] The aforementioned method for extracting feature points from the dynamic response waveform of asphalt pavement based on incremental learning includes step A, loading a configuration file, editing the configuration file, and obtaining the edited file. The specific steps are as follows. Step A1: Load the configuration file. Specifically, this involves checking if the file exists. If the file exists, the file list in the file is read. If the file does not exist, the file is created. Step A2 involves editing the configuration file and obtaining the edited file. Specifically, it asks whether to edit the 6-point pattern file list. The specific steps are as follows: Step A21: After entering Y, manually edit by entering the corresponding number as needed. Number 1 indicates using the current list directly, number 2 indicates adding a file, number 3 indicates deleting a file, number 4 indicates clearing and re-entering, and number 5 indicates opening the configuration file for manual editing. Step A22: After entering N, the current list will be used directly.

[0007] The aforementioned method for extracting feature points from the dynamic response waveform of asphalt pavement based on incremental learning, in step B, involves inputting the sensor's acquisition frequency and filtering parameters into an edited file to obtain the parameter input file, and then processing the parameter input file to obtain the processed file. The specific steps are as follows. Step B1: Input the sensor's acquisition frequency and filtering parameters into the edited file and obtain the file after parameter input. Specifically, the sensor acquisition frequency is 1000Hz, the sensor filtering method is low-pass filtering, and the cutoff frequency of the sensor filtering parameters is 40Hz and the filtering order is 4. Step B2 involves processing the input file and obtaining the processed file. The specific steps are as follows: Step B21: Process folders containing the same filename together; specifically, process the next file with the same name after processing the previous one. Step B22: Determine whether the file is in 4-point or 6-point mode. If it is in 4-point mode, multiply the absolute value of the original data of the selected file by -1. If it is in 4-point mode, the order of the points is flat point A, valley 1, valley 2 and flat point B. If it is in 6-point mode, the order of the points is flat point A, peak 1, peak 2, flat point B, valley 1 and valley 2.

[0008] In the aforementioned incremental learning-based method for extracting feature points of dynamic response waveforms of asphalt pavement, step C involves determining whether the processed file contains machine learning data and obtaining a judgment result that includes both non-existent and existing machine learning data. If the file does not exist, learning is restarted and a judgment result indicating the absence of machine learning data is obtained. If the file exists, historical learning records are loaded and a judgment result indicating the presence of machine learning data is obtained.

[0009] In the aforementioned incremental learning-based method for extracting waveform feature points from the dynamic response of asphalt pavement, step D involves determining that no machine learning data exists. If the result indicates that waveform feature point rules are not found, then the visualization interaction window is generated after predicting the waveform feature point rules. The specific steps for predicting the waveform feature point rules are as follows. Step D1, if it is a 4-point mode, the valley prediction process in the 4-point mode is as follows: first, exclude the boundary region between the first 0.5s and the last 0.5s; then calculate all local minima and sort them in ascending order of signal value; finally, take the two deepest valleys a. 4j And based on the chronological order, they were identified as valley point 1 and valley point 2 in the 4-point model; Step D2, if it is a 6-point mode, the peak prediction process in the 6-point mode is as follows: first, exclude the boundary region of the first 0.5s and the last 0.5s; then calculate all local maxima and sort them in descending order of signal value; finally, take the two highest peak values ​​b. 6j The peak points 1 and 2 in the 6-point mode are determined in chronological order. The valley prediction process in the 6-point mode is as follows: find the deepest valley value within 0.15 seconds before peak point 1, which is valley value 1 in the 6-point mode; and find the deepest valley value within 0.15 seconds before peak point 2, which is valley value 2 in the 6-point mode. The valley value time in the 6-point mode must be earlier than the corresponding peak point. Step D3: Predict the flat point A and flat point B. The average value within the range of 1.0 to 1.5 seconds before the first valley point is the flat point A, and the average value within the range of 1.0 to 1.5 seconds after the last valley point is the flat point B, as shown in formula (1). ; (1) in, The value of the flat point A, when i=4 The value of the smooth point A in the 4-point pattern, when i=6. This is the value of the smooth point A in the 6-point pattern; The value of the flat point B, when i=4 This refers to the value of the smooth point B in the 4-point pattern, when i=6. The value of the flat point B in the 6-point mode; S is the value within the range of 1 to 1.5 seconds before the valley value 1; f is the sampling frequency. The value is the value within the range of 1 to 1.5 seconds after the valley value 2.

[0010] In the aforementioned incremental learning-based method for extracting feature points of dynamic response waveforms of asphalt pavement, step E involves directly generating a visualization window if the judgment result indicates the existence of machine learning data. In this visualization window, the gray line represents the original signal, the blue line represents the filtered signal line, the cyan star points represent the basic prediction points, and the magenta star points represent the prediction points adjusted by machine learning. Left-clicking in the visualization window allows recording and marking correct points, thereby outputting the time and voltage values ​​corresponding to the correct points. Right-clicking skips the current file and does not output any values.

[0011] In the aforementioned incremental learning-based method for extracting feature points from the dynamic response waveform of asphalt pavement, step F involves adjusting the visualization window using machine learning offset and obtaining the machine learning offset. Specifically, when the mouse is left-clicked, the corresponding time point is recorded, and the click time point is compared with the predicted point T. aij Calculate the corresponding offset T, as shown in formula (2). T=t1-T aij (2) Where t1 is the time the user clicked.

[0012] The aforementioned method for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning, in step G, calculates the final prediction time based on machine learning offsets and searches for the closest waveform point near the final prediction time. Specifically, it reads the learning offsets of the most recent 30 times from the file and limits the offsets to within a range of ±0.2 seconds before calculating the final prediction time, as shown in formula (3). (3) in, For the final predicted time, For preliminary time prediction, It is any learning offset from the most recent 30 historical learning offsets.

[0013] The aforementioned method for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning, in step H, discards the closest waveform points that meet the discard conditions and uses the retained closest waveform points as the result of extracting feature points of dynamic response waveforms of asphalt pavement. The specific discard conditions include: peak 1 is greater than peak 2 in the 6-point mode, or valley 1 is greater than valley 2 in the 4-point mode and the 6-point mode; the interval between two valleys in the 4-point mode or between two peaks and valleys in the 6-point mode is less than 0.2s or greater than 1.3s; the difference between valley 1 and smooth point A in the 4-point mode is less than 10% of the difference between valley 2 and smooth point A; the difference between peak 1 and smooth point A in the 6-point mode is less than 10% of the difference between peak 2 and smooth point A; and the time difference between peak 1 and valley 1 or peak 2 and valley 2 in the 6-point mode is greater than 0.15s.

[0014] A system for extracting waveform feature points of dynamic response of asphalt pavement based on incremental learning includes a file loading module, a parameter input module, a machine learning data judgment module, a waveform feature point prediction module, a visualization interactive window generation module, a machine learning offset adjustment module, a waveform point extraction module, and a waveform feature point output module. The file loading module loads a configuration file, edits it, and obtains an edited file. The parameter input module inputs the sensor's acquisition frequency and filtering parameters into the edited file, obtains a file with input parameters, processes the file, and obtains a processed file. The machine learning data judgment module determines whether the processed file contains machine learning data and obtains files containing either non-existent or present machine learning data. The data judgment results; the waveform feature point prediction module shown is used to predict waveform feature point rules and generate a visualization interaction window when the judgment result is that there is no machine learning data; the visualization interaction window generation module shown is used to directly generate a visualization interaction window when the judgment result is that there is machine learning data; the machine learning offset adjustment module shown is used to adjust the machine learning offset of the visualization interaction window and obtain the machine learning offset amount; the waveform point extraction module shown is used to calculate the final prediction time based on the machine learning offset amount and search for the closest waveform point near the final prediction time; the waveform feature point output module shown is used to discard the closest waveform point that meets the discard condition and use the retained closest waveform point as the extraction result of the asphalt pavement dynamic response waveform feature point.

[0015] The beneficial effects of this invention are: (1) The present invention first loads the configuration file, then edits the configuration file and obtains the edited file, then inputs the sensor's acquisition frequency and filtering parameters into the edited file and obtains the parameter input file, then processes the parameter input file and obtains the processed file, then determines whether the processed file contains machine learning data and obtains the judgment result containing machine learning data. If the judgment result is that machine learning data does not exist, then the waveform feature point rules are predicted first and then a visualization interaction window is generated. If the judgment result is that machine learning data exists, then the visualization interaction window is directly generated, and then the machine learning offset is adjusted on the visualization interaction window and the machine learning data is obtained. The offset is used to calculate the final predicted time based on the machine learning offset, and the closest waveform point is searched near the final predicted time. Then, the closest waveform point that meets the discard criteria is discarded, and the retained closest waveform point is used as the result of extracting the waveform feature points of the dynamic response of asphalt pavement. This method and system effectively realizes the function of using a depth-based extreme point detection method to sort local extreme values ​​by amplitude and prioritize the extraction of the two most significant valleys or peaks in the waveform. This not only effectively eliminates the interference of local small fluctuations and boundary noise, but also makes the extreme point detection results not only consistent with the physical meaning of mechanical response, but also have good stability and interpretability.

[0016] (2) This invention supports dual-mode switching between 4-point mode and 6-point mode. By adopting an interactive error handling mechanism of left-click confirmation and right-click skip, combined with the persistent storage of learning data, the system can continuously improve itself in human-machine collaboration, which significantly improves the efficiency and accuracy of waveform feature point extraction.

[0017] (3) The present invention can record the deviation between the user-annotated position and the rule-predicted position by introducing a user feedback offset learning mechanism. Furthermore, by using a sliding window to average the historical offset trend, the learned offset can be applied to subsequent predictions. At the same time, the maximum adjustment range is limited to prevent over-correction. This hybrid strategy of rule-based learning and learning-assisted learning maintains the interpretability of deterministic algorithms and continuously optimizes the prediction accuracy through continuous user feedback. Attached Figure Description

[0018] Figure 1 This is an overall flowchart of a method for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning, according to the present invention. Figure 2 This is a schematic diagram illustrating the operating principle of an incremental learning-based asphalt pavement dynamic response waveform feature point extraction system according to the present invention. Figure 3 This is a schematic diagram of the correct prediction of the 6-point pattern of sensor 3-2-2 in an embodiment of the present invention; Figure 4 This is a schematic diagram of the correct prediction of the 6-point pattern of sensor 1-2-2 in an embodiment of the present invention; Figure 5 This is a schematic diagram of the correct prediction of the 4-point pattern of sensor 1-4-4 in an embodiment of the present invention; Figure 6 This is a schematic diagram of the correct prediction of the 4-point pattern of sensor 3-5-2 in an embodiment of the present invention; Figure 7 This is a schematic diagram of the 4-point mode error waveform of sensor 3-4-3 in an embodiment of the present invention; Figure 8 This is a schematic diagram of the 4-point mode error waveform of sensor 1-3-2 in an embodiment of the present invention; Figure 9 This is a schematic diagram of the 6-point mode error waveform of sensor 3-2-2 in an embodiment of the present invention. Detailed Implementation

[0019] The present invention will now be further described with reference to the accompanying drawings.

[0020] like Figure 1 and Figure 2 As shown, the present invention provides a method for extracting feature points of asphalt pavement dynamic response waveforms based on incremental learning, comprising the following steps: Step A: Load the configuration file, then edit the configuration file and obtain the edited file. The specific steps are as follows: Step A1: Load the configuration file. Specifically, this involves checking if the file exists. If the file exists, the file list in the file is read. If the file does not exist, the file is created. Step A2 involves editing the configuration file and obtaining the edited file. Specifically, it asks whether to edit the 6-point pattern file list. The specific steps are as follows: Step A21: After entering Y, manually edit by entering the corresponding number as needed. Number 1 indicates using the current list directly, number 2 indicates adding a file, number 3 indicates deleting a file, number 4 indicates clearing and re-entering, and number 5 indicates opening the configuration file for manual editing. Step A22: After entering N, the current list will be used directly.

[0021] Step B involves inputting the sensor's acquisition frequency and filtering parameters into the edited file to obtain the parameter input file. The parameter input file is then processed to obtain the processed file. The specific steps are as follows: Step B1: Input the sensor's acquisition frequency and filtering parameters into the edited file and obtain the file after parameter input. Specifically, the sensor acquisition frequency is 1000Hz, the sensor filtering method is low-pass filtering, and the cutoff frequency of the sensor filtering parameters is 40Hz and the filtering order is 4. Step B2 involves processing the input file and obtaining the processed file. The specific steps are as follows: Step B21: Process folders containing the same filename together; specifically, process the next file with the same name after processing the previous one. Step B22: Determine whether the file is in 4-point or 6-point mode. If it is in 4-point mode, multiply the absolute value of the original data of the selected file by -1. If it is in 4-point mode, the order of the points is flat point A, valley 1, valley 2 and flat point B. If it is in 6-point mode, the order of the points is flat point A, peak 1, peak 2, flat point B, valley 1 and valley 2.

[0022] Step C: Determine whether the processed file contains machine learning data and obtain a result that includes whether machine learning data exists or not. If the file does not exist, start learning again and obtain a result indicating that machine learning data does not exist. If the file exists, load the historical learning records and obtain a result indicating that machine learning data exists.

[0023] like Figure 3-6 As shown in step D, if the judgment result is that there is no machine learning data, then the waveform feature point rules are predicted first, and then a visualization interaction window is generated. The specific steps for predicting the waveform feature point rules are as follows. Step D1, if it is a 4-point mode, the valley prediction process in the 4-point mode is as follows: first, exclude the boundary region between the first 0.5s and the last 0.5s; then calculate all local minima and sort them in ascending order of signal value; finally, take the two deepest valleys a. 4j And based on the chronological order, they were identified as valley point 1 and valley point 2 in the 4-point model; Step D2, if it is a 6-point mode, the peak prediction process in the 6-point mode is as follows: first, exclude the boundary region of the first 0.5s and the last 0.5s; then calculate all local maxima and sort them in descending order of signal value; finally, take the two highest peak values ​​b. 6j The peak points 1 and 2 in the 6-point mode are determined in chronological order. The valley prediction process in the 6-point mode is as follows: find the deepest valley value within 0.15 seconds before peak point 1, which is valley value 1 in the 6-point mode; and find the deepest valley value within 0.15 seconds before peak point 2, which is valley value 2 in the 6-point mode. The valley value time in the 6-point mode must be earlier than the corresponding peak point. Step D3: Predict the flat point A and flat point B. The average value within the range of 1.0 to 1.5 seconds before the first valley point is the flat point A, and the average value within the range of 1.0 to 1.5 seconds after the last valley point is the flat point B, as shown in formula (1). ; (1) in, The value of the flat point A, when i=4 The value of the smooth point A in the 4-point pattern, when i=6. This is the value of the smooth point A in the 6-point pattern; The value of the flat point B, when i=4 This refers to the value of the smooth point B in the 4-point pattern, when i=6. The value of the flat point B in the 6-point mode; S is the value within the range of 1 to 1.5 seconds before the valley value 1; f is the sampling frequency. The value is the value within the range of 1 to 1.5 seconds after the valley value 2.

[0024] Step E: If the judgment result indicates the existence of machine learning data, a visualization interaction window is directly generated. In the visualization interaction window, the gray line represents the original signal, the blue line represents the filtered signal line, the cyan star points represent the basic prediction points, and the magenta star points represent the prediction points adjusted by machine learning. Left-clicking in the visualization interaction window allows you to record and mark the correct points, thereby outputting the time and voltage values ​​corresponding to the correct points. Right-clicking skips the current file and does not output any values.

[0025] Step F involves adjusting the machine learning offset in the visualization window and obtaining the machine learning offset. Specifically, when the mouse is left-clicked, the corresponding time point is recorded, and the click time point is compared with the prediction point T. aij Calculate the corresponding offset T, as shown in formula (2). T=t1-T aij (2) Where t1 is the time the user clicked.

[0026] Step G involves calculating the final prediction time based on the machine learning offset and searching for the closest waveform point near the final prediction time. Specifically, this involves reading the learning offsets from the last 30 times in the file, limiting the offsets to within a range of ±0.2 seconds, and then calculating the final prediction time, as shown in formula (3). (3) in, For the final predicted time, For preliminary time prediction, It is any learning offset from the most recent 30 historical learning offsets.

[0027] like Figure 7-9As shown, in step H, the closest waveform point that meets the discard criteria is discarded, and the retained closest waveform point is used as the result of extracting the waveform feature points of the dynamic response of the asphalt pavement. The specific discard criteria include: peak value 1 is greater than peak value 2 in the 6-point mode, or valley value 1 is greater than valley value 2 in the 4-point mode and the 6-point mode; the interval between two valley values ​​in the 4-point mode or between two peak values ​​and valley values ​​in the 6-point mode is less than 0.2s or greater than 1.3s; the difference between valley value 1 and smooth point A in the 4-point mode is less than 10% of the difference between valley value 2 and smooth point A; the difference between peak value 1 and smooth point A in the 6-point mode is less than 10% of the difference between peak value 2 and smooth point A; and the time difference between peak value 1 and valley value 1 or peak value 2 and valley value 2 in the 6-point mode is greater than 0.15s.

[0028] A system for extracting waveform feature points of dynamic response of asphalt pavement based on incremental learning includes a file loading module, a parameter input module, a machine learning data judgment module, a waveform feature point prediction module, a visualization interactive window generation module, a machine learning offset adjustment module, a waveform point extraction module, and a waveform feature point output module. The file loading module loads a configuration file, edits it, and obtains an edited file. The parameter input module inputs the sensor's acquisition frequency and filtering parameters into the edited file, obtains a file with input parameters, processes the file, and obtains a processed file. The machine learning data judgment module determines whether the processed file contains machine learning data and obtains files containing either non-existent or present machine learning data. The data judgment results; the waveform feature point prediction module shown is used to predict waveform feature point rules and generate a visualization interaction window when the judgment result is that there is no machine learning data; the visualization interaction window generation module shown is used to directly generate a visualization interaction window when the judgment result is that there is machine learning data; the machine learning offset adjustment module shown is used to adjust the machine learning offset of the visualization interaction window and obtain the machine learning offset amount; the waveform point extraction module shown is used to calculate the final prediction time based on the machine learning offset amount and search for the closest waveform point near the final prediction time; the waveform feature point output module shown is used to discard the closest waveform point that meets the discard condition and use the retained closest waveform point as the extraction result of the asphalt pavement dynamic response waveform feature point.

[0029] In summary, the present invention provides a method and system for extracting waveform feature points of asphalt pavement dynamic response based on incremental learning. First, a configuration file is loaded, then edited to obtain an edited file. Next, the sensor's acquisition frequency and filtering parameters are input into the edited file to obtain a parameter-input file. This file is then processed to obtain a processed file. Subsequently, it is determined whether machine learning data exists in the processed file, yielding a result indicating whether machine learning data is present or absent. If the result indicates the absence of machine learning data, waveform feature point rules are predicted first, and then a visual interactive window is generated. If the result indicates the presence of machine learning data, a visual interactive window is directly generated. Then, a machine learning offset adjustment is applied to the visual interactive window to obtain the machine learning offset. Based on the machine learning offset, the final prediction time is calculated, and the closest waveform point is searched near the final prediction time. The closest waveform points that meet the discard criteria are then discarded, and the retained closest waveform points are used as the extracted waveform feature points of the asphalt pavement dynamic response. This effectively realizes the extraction of waveform feature points of asphalt pavement dynamic response. The method and system for extracting feature points from force response waveforms employ a depth-based extreme point detection approach, sorting local extremes by amplitude and prioritizing the extraction of the two most significant valleys or peaks in the waveform. This effectively eliminates interference from local small fluctuations and boundary noise, ensuring that the extreme point detection results conform to the physical meaning of the mechanical response while exhibiting good stability and interpretability. The invention supports dual-mode switching between 4-point and 6-point modes. An interactive error handling mechanism using left-click confirmation and right-click skip, combined with persistent storage of learning data, allows the system to continuously improve itself through human-computer collaboration, significantly enhancing the efficiency and accuracy of waveform feature point extraction. Furthermore, the invention introduces a user-feedback offset learning mechanism to record the deviation between the user-annotated position and the rule-predicted position. The use of a sliding window averaging method to calculate historical offset trends allows the learned offsets to be applied to subsequent predictions, while limiting the maximum adjustment range to prevent over-correction. This hybrid strategy of rule-based primary and learning-secondary approaches maintains the interpretability of deterministic algorithms while continuously optimizing prediction accuracy through ongoing user feedback.

[0030] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning, characterized in that: Includes the following steps, Step A: Load the configuration file, then edit the configuration file and obtain the edited file; Step B involves inputting the sensor's acquisition frequency and filtering parameters into the edited file and obtaining the parameter input file, then processing the parameter input file and obtaining the processed file. Step C: Determine whether the processed file contains machine learning data and obtain the judgment results including those containing machine learning data and those not containing machine learning data. Step D: If the judgment result is that there is no machine learning data, then first predict the waveform feature point rules and then generate the visualization interaction window. Step E: If the judgment result indicates the existence of machine learning data, then a visual interactive window is directly generated. Step F: Adjust the machine learning offset of the visual interactive window and obtain the machine learning offset amount; Step G: Calculate the final prediction time based on the machine learning offset and search for the closest waveform point near the final prediction time; Step H involves discarding the closest waveform points that meet the discard criteria and using the retained closest waveform points as the result of extracting the dynamic response waveform feature points of the asphalt pavement.

2. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 1, characterized in that: Step A: Load the configuration file, then edit the configuration file and obtain the edited file. The specific steps are as follows: Step A1: Load the configuration file. Specifically, this involves checking if the file exists. If the file exists, the file list in the file is read. If the file does not exist, the file is created. Step A2 involves editing the configuration file and obtaining the edited file. Specifically, it asks whether to edit the 6-point pattern file list. The specific steps are as follows: Step A21: After entering Y, manually edit by entering the corresponding number as needed. Number 1 indicates using the current list directly, number 2 indicates adding a file, number 3 indicates deleting a file, number 4 indicates clearing and re-entering, and number 5 indicates opening the configuration file for manual editing. Step A22: After entering N, the current list will be used directly.

3. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 2, characterized in that: Step B involves inputting the sensor's acquisition frequency and filtering parameters into the edited file to obtain the parameter input file. The parameter input file is then processed to obtain the processed file. The specific steps are as follows: Step B1: Input the sensor's acquisition frequency and filtering parameters into the edited file and obtain the file after parameter input. Specifically, the sensor acquisition frequency is 1000Hz, the sensor filtering method is low-pass filtering, and the cutoff frequency of the sensor filtering parameters is 40Hz and the filtering order is 4. Step B2 involves processing the input file and obtaining the processed file. The specific steps are as follows: Step B21: Process folders containing the same filename together; specifically, process the next file with the same name after processing the previous one. Step B22: Determine whether the file is in 4-point or 6-point mode. If it is in 4-point mode, multiply the absolute value of the original data of the selected file by -1. If it is in 4-point mode, the order of the points is flat point A, valley 1, valley 2 and flat point B. If it is in 6-point mode, the order of the points is flat point A, peak 1, peak 2, flat point B, valley 1 and valley 2.

4. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 3, characterized in that: Step C: Determine whether the processed file contains machine learning data and obtain a result that includes whether machine learning data exists or not. If the file does not exist, start learning again and obtain a result indicating that machine learning data does not exist. If the file exists, load the historical learning records and obtain a result indicating that machine learning data exists.

5. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 4, characterized in that: Step D: If the judgment result indicates that there is no machine learning data, then first predict the waveform feature point rules and then generate the visualization interaction window. The specific steps for predicting the waveform feature point rules are as follows. Step D1, if it is a 4-point mode, the valley prediction process in the 4-point mode is as follows: first, exclude the boundary region between the first 0.5s and the last 0.5s; then calculate all local minima and sort them in ascending order of signal value; finally, take the two deepest valleys a. 4j And based on the chronological order, they were identified as valley point 1 and valley point 2 in the 4-point model; Step D2, if it is a 6-point mode, the peak prediction process in the 6-point mode is as follows: first, exclude the boundary region of the first 0.5s and the last 0.5s; then calculate all local maxima and sort them in descending order of signal value; finally, take the two highest peak values ​​b. 6j The peak points 1 and 2 in the 6-point mode are determined in chronological order. The valley prediction process in the 6-point mode is as follows: find the deepest valley value within 0.15 seconds before peak point 1, which is valley value 1 in the 6-point mode; and find the deepest valley value within 0.15 seconds before peak point 2, which is valley value 2 in the 6-point mode. The valley value time in the 6-point mode must be earlier than the corresponding peak point. Step D3: Predict the flat point A and flat point B. The average value within the range of 1.0 to 1.5 seconds before the first valley point is the flat point A, and the average value within the range of 1.0 to 1.5 seconds after the last valley point is the flat point B, as shown in formula (1). ； （1） in, The value of the flat point A, when i=4 The value of the smooth point A in the 4-point pattern, when i=6. This is the value of the smooth point A in the 6-point pattern; The value of the flat point B, when i=4 This refers to the value of the smooth point B in the 4-point pattern, when i=6. The value of the flat point B in the 6-point mode; S is the value within the range of 1 to 1.5 seconds before the valley value 1; f is the sampling frequency. The value is the value within the range of 1 to 1.5 seconds after the valley value 2.

6. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 5, characterized in that: Step E: If the judgment result indicates the existence of machine learning data, a visualization interaction window is directly generated. In the visualization interaction window, the gray line represents the original signal, the blue line represents the filtered signal line, the cyan star points represent the basic prediction points, and the magenta star points represent the prediction points adjusted by machine learning. Left-clicking in the visualization interaction window allows you to record and mark the correct points, thereby outputting the time and voltage values ​​corresponding to the correct points. Right-clicking skips the current file and does not output any values.

7. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 6, characterized in that: Step F involves adjusting the machine learning offset in the visualization window and obtaining the machine learning offset. Specifically, when the mouse is left-clicked, the corresponding time point is recorded, and the click time point is compared with the prediction point T. aij Calculate the corresponding offset T, as shown in formula (2). T=t1-T aij (2) Where t1 is the time the user clicked.

8. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 7, characterized in that: Step G involves calculating the final prediction time based on the machine learning offset and searching for the closest waveform point near the final prediction time. Specifically, this involves reading the learning offsets from the last 30 times in the file, limiting the offsets to within a range of ±0.2 seconds, and then calculating the final prediction time, as shown in formula (3). （3） in, For the final predicted time, For preliminary time prediction, It is any learning offset from the most recent 30 historical learning offsets.

9. The method for extracting feature points of asphalt pavement dynamic response waveform based on incremental learning according to claim 8, characterized in that: Step H involves discarding the closest waveform points that meet the discard criteria and using the retained closest waveform points as the results of extracting the dynamic response waveform feature points of the asphalt pavement. The specific discard criteria include: peak value 1 being greater than peak value 2 in the 6-point mode, or valley value 1 being greater than valley value 2 in the 4-point and 6-point modes; the interval between two valley values ​​in the 4-point mode or between two peak values ​​and valley values ​​in the 6-point mode being less than 0.2s or greater than 1.3s; the difference between valley value 1 and smooth point A in the 4-point mode being less than 10% of the difference between valley value 2 and smooth point A; the difference between peak value 1 and smooth point A in the 6-point mode being less than 10% of the difference between peak value 2 and smooth point A; and the time difference between peak value 1 and valley value 1 or peak value 2 and valley value 2 in the 6-point mode being greater than 0.15s.

10. A system for extracting feature points of dynamic response waveforms of asphalt pavement based on incremental learning, wherein the specific extraction process of the dynamic response waveform feature point extraction system is based on the dynamic response waveform feature point extraction method according to any one of claims 1-9, characterized in that: It includes a file loading module, a parameter input module, a machine learning data judgment module, a waveform feature point prediction module, a visualization interactive window generation module, a machine learning offset adjustment module, a waveform point extraction module, and a waveform feature point output module. The file loading module shown is used to load the configuration file, then edit the configuration file and obtain the edited file. The parameter input module shown is used to input the sensor's acquisition frequency and filtering parameters into the edited file and obtain the parameter input file. Then, the parameter input file is processed to obtain the processed file. The machine learning data judgment module shown is used to determine whether machine learning data exists in the processed file and to obtain judgment results including those containing machine learning data that does not exist and those containing machine learning data. The waveform feature point prediction module shown is used to predict waveform feature point rules and then generate a visual interactive window when the judgment result is that there is no machine learning data. The visualization interaction window generation module shown is used to directly generate a visualization interaction window when the judgment result indicates the existence of machine learning data. The machine learning offset adjustment module shown is used to adjust the machine learning offset of the visual interactive window and obtain the machine learning offset amount; The waveform point extraction module shown is used to calculate the final prediction time based on machine learning offset and search for the closest waveform point near the final prediction time. The waveform feature point output module shown is used to discard the closest waveform point that meets the discard criteria and use the retained closest waveform point as the result of extracting the waveform feature points of the dynamic response of asphalt pavement.

Images