Pavement texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation

Abstract

The invention provides a road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation. According to the method, twelve kinds of input feature data and one kind of output feature data are matched to form an original pavement texture data set of multi-section small samples. The method comprises the following steps: taking road sections as a grouping basis, and adopting a Holt exponential smoothing model to carry out transverse missing value supplementation on an original road surface texture data set through a horizontal updating and trend updating method; the method comprises the following steps: carrying out longitudinal interpolation expansion on a data set by adopting a PchipInterpolator interpolation method, and generating more dense interpolation points for all features at the same moment; compared with the prior art, the problems of missing and discontinuity of pavement texture data are solved, the time and economic cost of multiple data acquisition can be reduced, and a pavement texture data set is formed.

Description

Technical Field

[0001] The invention relates to the field of road texture data set expansion. Background Art

[0002] The slippery state of asphalt pavement is closely related to the safety of people's travel. Machine learning and deep learning methods are expected to achieve accurate prediction of pavement anti-skid performance. However, the prediction accuracy of such artificial intelligence methods is highly related to the data set. The better the data volume and accuracy of the data set, the higher the prediction accuracy.

[0003] The pavement texture dataset is mainly used for the prediction of the anti-skid performance of the road surface, including pavement conditions, climate conditions, traffic load, aggregate type and structural depth data. The structural depth data is used to evaluate the anti-skid performance of the road surface. It generally contains twelve input features and one output feature. The input features include: in terms of pavement conditions, the representative layer thickness is selected as the input feature index, which is helpful for evaluating the pavement performance; in terms of climate conditions, the influence of seven major factors such as precipitation, wind speed, average temperature, maximum temperature, freezing index, surface shortwave radiation, and humidity are comprehensively considered; in terms of traffic load, since heavy traffic has a greater impact on the anti-skid performance of the pavement, the annual cumulative traffic volume of the 9th type of large vehicles is selected. This data represents the 9th type of large vehicles (usually heavy trucks or large vehicles) carried by the lane in a specific year; in addition, considering the superposition of various types of heavy-duty trucks, the annual cumulative traffic volume of heavy trucks (vehicle categories 4, 6-13) is selected as the characteristic parameter; in terms of aggregate type, the percentage of passing through 4.75 mm and 0.075 mm sieve holes is selected as the characteristic parameter, which can better characterize the particle gradation of the aggregate. The output feature is the pavement texture depth, which is used to characterize skid resistance.

[0004] The main data of the pavement texture dataset are collected manually or by machine, with a collection period of monthly or annually, and are small sample data of multiple road sections. Limited by the collection technology, there are often problems such as mismatch of collected features and missing data, which greatly affects the prediction of pavement anti-skid performance and is not conducive to the development of artificial intelligence technology in the field of pavement anti-skid performance prediction.

[0005] For example, in the prediction of pavement skid resistance based on deep learning, the collected data must first be cleaned and preprocessed. However, after data cleaning, it is found that due to the limitation of the collection year, there is often a problem of insufficient data. Since the texture characteristics of the road surface change slowly, in most cases in units of years, it is extremely difficult to collect supplementary data. Even if multiple sections are collected at the same time, this multi-section small sample data will still affect the accuracy of deep learning predictions and prevent deep learning from capturing its dependence on time series. Therefore, in response to this problem, existing methods often reduce the complexity of the model and use linear interpolation methods to supplement data, but the effect is minimal. Summary of the invention

[0006] In view of the shortcomings of the prior art and to overcome the shortcomings of the prior art, the present invention proposes a road texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation. The method horizontally supplements the missing values ​​of the collected road texture data set based on Holt exponential smoothing to form a complete data set represented by 12 input features and 1 output feature; then, using the Python environment and the PchipInterpolator interpolation method, the data of each road section is vertically interpolated to expand the small sample data to meet the needs of deep learning models to predict the anti-skid performance of the road surface. The present invention can fully explore the temporal correlation between road texture data, realize the nonlinear expansion of the road texture data set, and effectively solve the problem of insufficient training data faced by the existing convolutional neural network using road texture data as the training set.

[0007] To achieve the above object, the technical solution of the present invention is as follows: The road texture dataset expansion method based on Holt exponential smoothing and PchipInterpolator interpolation includes the following steps: Step 1: From the long-term pavement performance database, collect matching data based on thickness m1, precipitation m2, wind speed m3, average temperature m4, maximum temperature m5, freezing index m6, surface shortwave radiation m7, humidity m8, Class 9 vehicle traffic volume m9, heavy truck traffic volume m 10 , 4.75 mm sieve hole passing rate m 11 , 0.075 mm sieve hole passing rate m 12 The input feature data represented by m i , and output feature data n represented by constructed depth data, forming an original road surface texture data set A1; Step 2: Based on the original road surface texture dataset, the data features are classified and sorted, and the road surface texture dataset is updated to the road surface texture dataset A2; Step 3: Construct the Holt exponential smoothing model, fill in the lateral missing values, and update the road texture dataset to the road texture dataset A3; Step 4: Construct the PchipInterpolator interpolator to fill in the longitudinal missing values, and update the road texture dataset to the road texture dataset A4; Step 5: Dynamically adjust steps 3 and 4 to form a multi-level data filling and expansion method, automatically adjust the smoothing parameters of the Holt model and the interpolation control points of PchipInterpolator according to the characteristic changes of the data, and form the final road surface texture data set A5; Step 6: Compare the data before and after interpolation to ensure that the trend of the interpolated data is consistent with the original data.

[0008] Specifically, the step 2: classify based on road sections to ensure that each road section is independent, and then arrange the data of each feature based on the collection year and month as the time series feature to ensure that the independent variable during interpolation, i.e., time, is monotonically increasing, and the road surface texture dataset is updated to the road surface texture dataset A2.

[0009] Specifically, the step 3: first use the time series clustering technology to perform cluster analysis on time series with similar features, then smooth the original data to capture the trend and horizontal changes of each feature, fit the data, and generate a smoothed sequence, use the smoothing result of the model to horizontally fill in the missing values ​​of each feature in the pavement texture dataset, and update the pavement texture dataset to pavement texture dataset A3.

[0010] Specifically, step 4: using road sections as the basis for classification, generating denser interpolation points in each adjacent year and month, and expanding the data vertically; after the data of all road sections are interpolated, merging the results back into the road surface texture dataset A3 to form a denser road surface texture dataset A4.

[0011] Preferably, the collection frequency of the 13 types of characteristic data is in years, but the years may not be consecutive, and each type contains at least 3 years of collection records on the same road section.

[0012] Preferably, the units used for the various characteristic data must be international units. If not, they need to be converted into international units.

[0013] Preferably, Holt exponential smoothing and PchipInterpolator interpolation are dynamic mechanisms that can be automatically adjusted according to changes in data characteristics.

[0014] Preferably, the heavy truck traffic volume is the annual cumulative traffic volume of heavy trucks (vehicle categories 4, 6-13).

[0015] The present invention has the following advantages and beneficial effects: 1. The present invention matches 12 input features and 1 output feature, constructs a Holt exponential smoothing model, first performs K-means time series clustering, performs cluster analysis on time series with similar features, and then performs lateral missing value supplementation to form a complete texture data set. Then, the PchipInterpolator interpolator is constructed to generate denser longitudinal interpolation points to form the final road surface texture data set. The data expansion has high accuracy, controllable expansion amount, and can achieve nonlinear expansion of the data set.

[0016] 2. The present invention adopts computer programming technology to supplement the data set, which is different from the previous manual data collection and supplement. On the basis of the existing data, it fully explores the time correlation within each feature data, which can effectively avoid the high time and economic costs caused by multiple data collection on the road section.

[0017] 3. All processes and algorithms of the present invention can be quickly implemented by computer editing cycle programs. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 It is a complete flow chart of the present invention.

[0019] Figure 2 The data distribution selected for the long-term pavement performance database.

[0020] Figure 3 Supplementary diagram for missing values ​​of Holt exponential smoothing model.

[0021] Figure 4 This is a schematic diagram of PchipInterpolator interpolation.

[0022] Figure 5 Interpolate the fitting curve for PchipInterpolator.

[0023] Figure 6 The precision after the PchipInterpolator interpolation is expanded. DETAILED DESCRIPTION

[0024] The technical solution provided by the present application will be further described below in conjunction with specific embodiments and accompanying drawings. The advantages and features of the present application will become more apparent with the following description.

[0025] The present invention proposes a road texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation, which mainly includes data matching, arrangement of characteristic time series data, data missing value supplementation PchipInterpolator interpolation and other processing steps. All steps can be completed in computer memory through program programming.

[0026] Specifically, Figure 1 As shown in FIG. 1 , the road texture dataset expansion method based on Holt exponential smoothing and PchipInterpolator interpolation includes the following steps: Step 1: Match characteristic variables to form the original data set: from the long-term pavement performance database, match thickness (m1), precipitation (m2), wind speed (m3), average temperature (m4), maximum temperature (m5), freezing index (m6), surface shortwave radiation (m7), humidity (m8), traffic volume of Class 9 vehicles (m9), traffic volume of heavy trucks (m 10 ), 4.75 mm sieve hole pass rate (m 11 ), 0.075 mm sieve hole pass rate (m 12 ) is represented by the input feature data m i , and the output feature data n represented by the constructed depth data, form the original road surface texture data set A1.

[0027] like Figure 2 As shown in the figure, the data comes from asphalt pavement data from 42 states / cities in the United States and 7 states / cities in Canada, totaling 301 road sections and 2,353 sets of data. Among these states / cities, Arizona has the most test sections, with 23 sections selected, followed by Oklahoma and Texas, with 21 and 20 sections selected respectively.

[0028] Step 2, based on the original road surface texture dataset A1, classify and sort the data features: classify by road section, and then arrange the data of each feature by the collection year and month as the time series feature. The data is arranged from low to high according to the collection year and month to ensure that the independent variable (time) during interpolation is monotonically increasing. The road surface texture dataset is updated to the road surface texture dataset A2.

[0029] Step 3: Construct the Holt exponential smoothing model to fill in the missing values ​​horizontally: First, use K-means time series clustering to perform cluster analysis on time series with similar characteristics, such as Figure 3 As shown in the figure, the time series structure and missing value of each feature data are clarified, and the Holt exponential smoothing model is constructed for each feature. In the case of missing feature data, Holt exponential smoothing mainly involves three aspects: horizontal update, trend update and missing value filling. After the horizontal missing value filling, the road surface texture dataset is updated to the road surface texture dataset A3.

[0030] 1. Horizontal update: in, It's time t The smoothing level value of It's time t The actual observed value of is the horizontal smoothing parameter (0≤ ≤1), It's time t− A smoothing level value of 1, It's time t− A smoothed trend value of 1.

[0031] 2. Trend Update: in, It's time t The smoothed trend value of is the trend smoothing parameter (0≤ ≤1), Indicates the difference between the current level value and the previous level value (i.e. the trend change amount).

[0032] 3. Missing value filling, using the smoothing level and trend of the previous moment to estimate the missing value: in, It's about time t The predicted value of .

[0033] Through the above steps, the Holt exponential smoothing model can capture the time correlation in the data while smoothing the time series data, and effectively fill the missing values ​​in the 13 feature sequences. In this way, the missing values ​​will be estimated as reasonable values ​​under the smoothing trend, retaining the overall trend characteristics of the data. It should be noted that Holt exponential smoothing depends on two smoothing parameters: the horizontal smoothing parameter and trend smoothing parameter Therefore, Python can be used to select the most suitable parameters for filling missing values ​​for each feature through cross-validation or grid search.

[0034] Step 4: Construct the PchipInterpolator interpolator to fill in the missing values ​​in the longitudinal direction: After the Holt exponential smoothing model supplements the missing values, a relatively complete data set has been formed, that is, 1 output feature corresponds to 12 input features without missing values. However, at this time, the data is still limited by the collection time, and the data is a small sample data of multiple sections. Therefore, if Figure 4 As shown in the figure, construct the PchipInterpolator interpolator, and for each feature, use the road segment as the classification basis to generate multiple sub-datasets, each of which corresponds to a road segment. In each sub-dataset, sort the data by year and month to ensure that the interpolator can correctly handle time series.

[0035] For each road segment, extract the time series , … and the corresponding feature m 1， m 2， ……m 12 and n.

[0036] in, represents the time of each observation (e.g. year and month), i= 1,2, ……,q ; m i and n represents a certain characteristic value at that time point, i =1,……,12.

[0037] Using the PchipInterpolator interpolator, a piecewise cubic interpolator is constructed: PchipInterpolator ( t , m i / n ) in t It's the time point. m i / n is the corresponding eigenvalue, and “ / ” means or.

[0038] In the original time series , … By adding multiple new time points between the original time points , … t ' w , to generate a denser time series, where w < q The specific number of new time points generated can be adjusted according to actual needs.

[0039] Use the PchipInterpolator interpolator to calculate the eigenvalue m1 corresponding to the newly generated time point ′ , m2 ′ ……m 12 ′ and n′: Among them, here Represents a new time point, and the interpolator generates the corresponding eigenvalue , “ / ” means or.

[0040] The new interpolation points ( , ) is added to the road texture dataset A3 to form a more dense road texture dataset A4, achieving complete data expansion.

[0041] Step 5, dynamic adjustment: Compare the data before and after interpolation to ensure that the trend of the interpolated data is consistent with the original data. Each feature is classified based on the road section to check the degree of fit of the curve after data expansion. In order to further clarify whether the data after expansion conforms to the trend of the original data, you can select the same number of interpolation points as the actual data points for comparison and calculate the determination coefficient. R 2 。 If you are not satisfied with the accuracy after data expansion, the Holt exponential smoothing model (horizontal smoothing parameter and trend smoothing parameter ) and the number of expansion points of the PchipInterpolator interpolator until the accuracy is satisfactory, forming the final road surface texture dataset A5.

[0042] Step 6, final proofreading: Compare the data before and after interpolation to ensure that the trend of the interpolated data is consistent with the original data.

[0043] Example 12 input features and 1 output feature were matched from the long-term pavement performance database to obtain 2353 sets of data. The Holt exponential smoothing model was used to supplement the missing values ​​of these data to obtain 2353 complete sets of data. The data of each test section was expanded using PchipInterpolator interpolation. Taking the output feature of depth as an example, the original data of each test section was expanded into 20 evenly distributed interpolation points. The final database data volume after expansion was 7339 sets.

[0044] Take the typical test sections: Arizona, 4, AA01, Oklahoma, 40, AA02, Texas, 40, AA03 as examples to expand the data. Figure 5 The figure shows the curve fitted after PchipInterpolator interpolation. It can be found that the fit is very good and the curve is very smooth. In order to further clarify the accuracy of PchipInterpolator interpolation, the same number of interpolation points as the actual data points are selected for comparison and the determination coefficient is calculated. R 2 ,like Figure 6 As shown in the figure, it can be found that after PchipInterpolator interpolation, the R 2The accuracy of Oklahoma, 40, AA02 is 1.00, the accuracy of Texas, 40, AA03 is 0.97, and the accuracy of Oklahoma, 40, AA02 is 0.96, and the accuracy of Texas, 40, AA03 is 0.97, all of which are above 0.90. The traditional linear interpolation method has an expansion accuracy of only about 0.7. In comparison, the method of the present invention can improve the accuracy by about 28.57% (significant effect), indicating that the method of the present invention can ensure that the trend of the interpolated data is consistent with the original data, and realize the effective expansion of road surface texture data.

[0045] The above description is only a description of the preferred embodiments of the present application, and is not intended to limit the scope of the present application. Any changes or modifications made by any person skilled in the art based on the above disclosed technical contents shall be deemed as equivalent effective embodiments and shall fall within the scope of protection of the technical solution of the present application.

Claims

1. A road texture dataset expansion method based on Holt exponential smoothing and PchipInterpolator interpolation, characterized in that: The following steps are involved: Step 1: From the long-term pavement performance database, collect matching data based on thickness m1, precipitation m2, wind speed m3, average temperature m4, maximum temperature m5, freezing index m6, surface shortwave radiation m7, humidity m8, Class 9 vehicle traffic volume m9, heavy truck traffic volume m 10 , 4.75 mm sieve hole passing rate m 11 , 0.075 mm sieve hole passing rate m 12 The input feature data represented by m i , and output feature data n represented by constructed depth data, forming an original road surface texture data set A1; Step 2: Based on the original road surface texture dataset, the data features are classified and sorted, and the road surface texture dataset is updated to the road surface texture dataset A2; Step 3: Construct the Holt exponential smoothing model, fill in the lateral missing values, and update the road texture dataset to the road texture dataset A3; Step 4: Construct the PchipInterpolator interpolator to fill in the longitudinal missing values, and update the road texture dataset to the road texture dataset A4; Step 5: Dynamically adjust steps 3 and 4 to form a multi-level data filling and expansion method, automatically adjust the smoothing parameters of the Holt model and the interpolation control points of PchipInterpolator according to the characteristic changes of the data, and form the final road surface texture data set A5; Step 6: Compare the data before and after interpolation to ensure that the trend of the interpolated data is consistent with the original data.

2. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 1, characterized in that: Step 2: Classification is performed based on the road section as the grouping basis to ensure that each road section is independent, and then the data of each feature is arranged based on the collection year and month as the time series feature to ensure that the independent variable during interpolation, i.e. time, is monotonically increasing. The road surface texture dataset is updated to the road surface texture dataset A2.

3. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 1, characterized in that: Step 3: First, the time series clustering technology is used to perform cluster analysis on time series with similar characteristics. Then, the original data is smoothed to capture the trend and horizontal change of each feature. The data is fitted and a smoothed sequence is generated. The smoothing results of the model are used to horizontally fill in the missing values ​​of each feature in the pavement texture dataset. The pavement texture dataset is updated to pavement texture dataset A3.

4. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 3, characterized in that: First, K-means time series clustering is used to perform cluster analysis on time series with similar features, clarify the time series structure and missing value of each feature data, and construct a Holt exponential smoothing model for each feature. In the case of missing feature data, Holt exponential smoothing involves three aspects: horizontal update, trend update, and missing value filling. After the horizontal missing value filling, the road surface texture dataset is updated to the road surface texture dataset A3. Horizontal Update: in, It's time t The smoothing level value of It's time t The actual observed value of is the horizontal smoothing parameter (0≤ ≤1), It's time t− A smoothing level value of 1, It's time t− A smoothed trend value of 1; Trend Update: in, It's time t The smoothed trend value of is the trend smoothing parameter (0≤ ≤1), It indicates the difference between the current level value and the level value at the previous moment, that is, the trend change amount; Missing value filling, using the smoothing level and trend of the previous moment to estimate the missing value: in, It's about time t The predicted value of .

5. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 1, characterized in that: Step 4: Taking the road sections as the classification basis, more dense interpolation points are generated in each adjacent year and month to expand the data vertically. After the data of all road sections are interpolated, these results are merged back into the road surface texture dataset A3 to form a more dense road surface texture dataset A4.

6. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 5, characterized in that: Construct the PchipInterpolator interpolator, and for each feature, use the road segment as the classification basis to generate multiple sub-datasets, each sub-dataset corresponds to a road segment; In each sub-dataset, sort the data by year and month to ensure that the interpolator can handle the time series correctly; For each road segment, extract the time series , … and the corresponding feature m 1， m 2， ……m 12 and n; in, represents the time of each observation, i= 1,2, ……,q ; m i and n represents a certain characteristic value at that time point, i =1,……,12; Using the PchipInterpolator interpolator, a piecewise cubic interpolator is constructed: PchipInterpolator ( t , m i / n ) in t It's the time point. m i / n is the corresponding eigenvalue, " / " means or; In the original time series , … By adding multiple new time points between the original time points , … t ' w , to generate a denser time series, where w < q ; Use the PchipInterpolator interpolator to calculate the eigenvalue m1 corresponding to the newly generated time point ′ , m2 ′ ……m 12 ′ and n′: Among them, here Represents a new time point, and the interpolator generates the corresponding eigenvalue , " / " means or; The new interpolation points ( , ) is added to the road texture dataset A3 to form a more dense road texture dataset A4, achieving complete data expansion.

7. The road surface texture data set expansion method based on Holt exponential smoothing and PchipInterpolator interpolation as claimed in claim 1, characterized in that: The collection frequency of the 13 types of characteristic data is in years, and each type contains at least 3 years of collection records on the same road section.

Images