Pavement maintenance and condition data abnormal value comprehensive identification method based on data driving

Abstract

The invention belongs to the technical field of pavement condition data monitoring, and particularly relates to a pavement maintenance and condition data abnormal value comprehensive identification method based on data driving. The method comprises the following steps: respectively training a maintenance measure probability prediction model based on an artificial neural network and a pavement condition prediction model based on a Bayesian neural network by using a normal data set; the method comprises the following steps: performing abnormal maintenance data identification and correction by using a trained maintenance measure probability prediction model based on an artificial neural network, and performing abnormal pavement condition index data identification and correction by using a pavement condition prediction model based on a Bayesian neural network. And dynamic adjustment of maintenance data abnormal threshold values and road condition index abnormal threshold values of different road sections is carried out in combination with a 2 sigma principle. The method can be used for improving the data quality of the pavement management system and helping managers to make scientific and reasonable maintenance decisions.

Description

Technical Field

[0001] The present invention belongs to the technical field of road condition data monitoring, and in particular relates to a data-driven comprehensive recognition method for abnormal values ​​of road condition data and maintenance. Background Art

[0002] With the continuous expansion of highway networks and the continuous growth of traffic volume, the contradiction between maintenance needs and limited financial investment has become increasingly prominent. In this context, the scientificity and effectiveness of pavement management systems are crucial to ensuring road performance, service life, and the rational allocation of maintenance funds. The core of pavement management systems lies in accurately assessing pavement conditions and formulating reasonable maintenance plans. However, existing pavement management systems still have many limitations in data processing and model building, which limit the reliability of the system and the scientificity of decision-making.

[0003] First, the pavement management system is highly dependent on the quality of pavement condition data and maintenance data. These data reflect the physical state, performance, maintenance history and results of the pavement, and are the basis for making maintenance decisions and rationally allocating maintenance funds. However, due to possible errors in data collection, transmission and storage, pavement condition data and maintenance data often contain outliers, which significantly reduce the reliability of the system.

[0004] Secondly, existing methods have obvious deficiencies in feature extraction, model generalization ability and uncertainty quantification. Traditional models are mostly based on static or single-time point data. For example, the performance changes of the same road section in the rainy season and the dry season are significantly different, but existing methods are difficult to capture such dynamic characteristics. In addition, the road surface state is affected by multiple random factors (such as sudden traffic loads and extreme weather). Existing deterministic models cannot quantify the uncertainty of the prediction results, which limits the reliability of decision-making.

[0005] Finally, existing research on pavement management data identification mostly focuses on a single type of data, that is, either only on pavement condition data or only on maintenance data, lacking a method that can comprehensively identify and correct outliers in pavement condition and maintenance data. For example, some studies identify outliers in pavement condition data through statistical methods or machine learning models, while other studies focus on anomaly detection in maintenance data. These methods fail to fully consider the inherent relationship between pavement condition data and maintenance data, resulting in the inability to fully identify anomalies in both types of data.

[0006] In response to the above problems, there is an urgent need for a comprehensive identification method that can integrate multi-source data, dynamically optimize model structure and quantify uncertainty, so as to improve the scientificity and adaptability of pavement maintenance decisions. Summary of the invention

[0007] The purpose of the present invention is to provide a data-driven comprehensive identification method for abnormal values ​​in pavement maintenance and condition data, so as to solve the problems raised in the background technology.

[0008] In order to achieve the above technical objectives, the technical solution adopted by the present invention is as follows: A data-driven comprehensive identification method for outliers in pavement maintenance and condition data includes the following steps: S1: Identify outliers on the pavement management data and divide the pavement management data into a normal data set and an abnormal data set; S2: Use the normal data set to train the maintenance measure probability prediction model based on artificial neural network and the road condition prediction model based on Bayesian neural network respectively; S3: First, use the trained maintenance measure probability prediction model based on artificial neural network (ANN) to identify and correct abnormal maintenance data, then use the road condition prediction model based on Bayesian neural network (BNN) to identify and correct abnormal road condition index data, and combine the 2σ principle to dynamically adjust the abnormal thresholds of maintenance data and road condition index of different road sections. Finally, the data set after comprehensive identification and correction is obtained.

[0009] Further preferably, the pavement management data includes pavement condition data, maintenance data, road attribute data, traffic data, climate and environment data.

[0010] Further preferably, the pavement condition data include: the pavement condition index (PCI) of the previous year and the pavement condition index of the current year; the maintenance data include: no maintenance, thin anti-skid surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-situ hot regeneration, milling and replacing one layer, milling and replacing two layers; the road attribute data include: surface layer thickness, surface layer material, base layer thickness, base layer material, road age; the traffic data include: annual average daily traffic volume, annual average daily truck traffic volume.

[0011] Further preferably, the maintenance measure probability prediction model based on artificial neural network (ANN) includes: 13 input variables, including: pavement condition index (PCI) of the previous year, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, and precipitation; Hidden layer: 2 layers, using ReLU activation function; Output layer: 9 output nodes, corresponding to the probabilities of 9 maintenance measures, including: no maintenance, thin anti-slip surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-situ hot regeneration, milling and replacing one layer, milling and replacing two layers; Use the Softmax function to output probability distribution.

[0012] Further preferably, the road condition prediction model based on Bayesian neural network (BNN) includes: Input layer: 14 input variables, including: pavement condition index (PCI) of the previous year, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, precipitation, and maintenance measures; Hidden layer: 2 layers, using ReLU activation function; Output layer: 1 output node, corresponding to the predicted value of the road condition index for the current year.

[0013] Further preferably, the loss value of the maintenance measure probability prediction model based on artificial neural network (ANN) is cross entropy loss.

[0014] Further preferably, the loss value of the road condition prediction model based on Bayesian neural network (BNN) is the mean square error (MSE), and the accuracy is the determination coefficient R 2 .

[0015] Further preferably, in step S3, the trained maintenance measure probability prediction model based on artificial neural network (ANN) is used to calculate the probability of each maintenance measure for each road section, and abnormal maintenance data is identified and corrected; according to the 2σ principle, if the probability of a maintenance measure is lower than 0.05, the maintenance measure is determined to be abnormal and replaced with the maintenance measure with the highest probability.

[0016] Further preferably, in step S3, the reasonable range of the pavement condition index of each road section in the current year is calculated according to the 2σ principle; if the pavement condition index of the current year exceeds the reasonable range, the pavement condition index is determined to be abnormal and replaced by the mean pavement condition index predicted by the pavement condition prediction model based on the Bayesian neural network (BNN).

[0017] The present invention analyzes the inherent laws between pavement condition data and maintenance data, and divides the data into normal data sets and abnormal data sets; constructs a maintenance measure probability prediction model based on an artificial neural network (ANN) and a pavement condition prediction model based on a Bayesian neural network (BNN), which are used to identify and correct abnormal maintenance data and pavement condition data respectively; through actual case verification, the proportion of abnormal data has dropped from 47.79% to 6.35%, proving the effectiveness of this method. The present invention achieves comprehensive identification of abnormal values ​​of pavement maintenance and condition data by integrating multi-source data, constructing a phased single-channel neural network model architecture, and combining a dynamic threshold optimization mechanism, significantly improving data quality, helping managers make scientific and reasonable maintenance decisions, and also promoting the rational allocation of pavement maintenance funds. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The present invention is further illustrated by means of the non-limiting examples given in the accompanying drawings.

[0019] Figure 1 is a flow chart of the method of the present invention; Figure 2 This is a graph showing the change in loss value and accuracy of the maintenance measures probability prediction model based on artificial neural network; Figure 3 This is a graph showing the change in loss value and accuracy of the road condition prediction model based on the Bayesian neural network (BNN); Figure 4 is the probability value of all maintenance measures in the maintenance measure probability prediction model based on artificial neural network; Figure 5 It is a comparison chart of the percentage of abnormal data before and after the implementation of the comprehensive identification and correction method of the present invention. DETAILED DESCRIPTION

[0020] In order to enable those skilled in the art to better understand the present invention, the technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0021] like Figure 1 As shown, a data-driven comprehensive identification method for abnormal values ​​of pavement maintenance and condition data of the present invention comprises the following steps: S1: Identify outliers on the pavement management data and divide the pavement management data into a normal data set and an abnormal data set; S2: Use the normal data set to train the maintenance measure probability prediction model based on artificial neural network and the road condition prediction model based on Bayesian neural network respectively; S3: First, use the trained maintenance measure probability prediction model based on artificial neural network (ANN) to identify and correct abnormal maintenance data, then use the road condition prediction model based on Bayesian neural network (BNN) to identify and correct abnormal road condition index data, and combine the 2σ principle to dynamically adjust the abnormal thresholds of maintenance data and road condition index of different road sections. Finally, the data set after comprehensive identification and correction is obtained.

[0022] In step S1 of this embodiment, the pavement management data includes pavement condition data, maintenance data, road attribute data, traffic data, climate and environmental data; the pavement condition data includes: pavement condition index (PCI) of the previous year and pavement condition index of the current year; the maintenance data includes: no maintenance, thin anti-skid surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-situ hot regeneration, milling and replacing one layer, milling and replacing two layers; the road attribute data includes: surface layer thickness, surface layer material, base layer thickness, base layer material, road age; the traffic data includes: annual average daily traffic volume, annual average daily truck traffic volume; the climate and environmental data include: high temperature days, consecutive high temperature days, low temperature days, consecutive low temperature days, precipitation.

[0023] In this embodiment, daily average temperatures > 25°C and < 0°C are defined as high temperature and low temperature, respectively.

[0024] (1) Total high temperature days: the total number of days in a year with an average daily temperature > 25°C; (2) Consecutive high temperature days: the number of consecutive days with an average daily temperature > 25°C for three or more days in a year; (3) Total low temperature days: the total number of days in a year with an average daily temperature < 0°C; (4) Consecutive low temperature days: the number of consecutive days with an average daily temperature of < 0°C for 3 or more days in a year; (5) Precipitation: The total amount of precipitation in the form of rain, snow, sleet or hail in a specific area in a year.

[0025] In step S1 of this embodiment, data standardization is also performed before outlier identification. Input variables (such as PCI, road attributes, traffic data, climate data, etc.) are standardized to ensure that the data are in the same dimension, which is convenient for subsequent model training.

[0026] Based on the inherent laws of pavement maintenance measures and the pavement condition index (PCI): when no maintenance is implemented, the pavement condition will gradually deteriorate over time; when maintenance measures are taken, the pavement condition index of the year will be improved to a certain extent. The pavement management data is identified for outliers, including three abnormal situations, as shown in Table 1: Table 1 Judgment of data anomalies

[0027] The maintenance measures probability prediction model based on artificial neural network (ANN) includes: 13 input variables, including: pavement condition index (PCI) of the previous year, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, precipitation. Hidden layer: 2 layers, using ReLU activation function. Output layer: 9 output nodes, corresponding to the probabilities of 9 maintenance measures, including: no maintenance, thin anti-skid surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-place hot regeneration, milling and replacing one layer, milling and replacing two layers. Use Softmax function to output probability distribution.

[0028] Based on experience and multiple attempts, the hyperparameters of the maintenance measure probability prediction model based on artificial neural network (ANN) were determined: including loss function (cross entropy loss), learning rate (0.001), batch size (500), number of hidden layer neurons (36, 36), optimizer (Adam), activation function (ReLU), and number of training rounds (1000).

[0029] The road condition prediction model based on Bayesian neural network (BNN) includes: Input layer: 14 input variables, including: pavement condition index (PCI) of the previous year, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, precipitation, and maintenance measures. Hidden layer: 2 layers, using ReLU activation function. Output layer: 1 output node, corresponding to the pavement condition index prediction value of the current year. Monte Carlo Dropout is applied: random dropout (probability 0.3) is introduced in the hidden layer, and the uncertainty of pavement condition index prediction is simulated through multiple forward propagations to increase the robustness of the model.

[0030] Based on experience and multiple attempts, the hyperparameters of the road condition prediction model based on Bayesian neural network (BNN) were determined: including loss function (mean square error MSE), learning rate (0.001), batch size (500), number of hidden layer neurons (36, 36), optimizer (Adam), activation function (ReLU), number of training rounds (1000), number of simulations (20), and discard probability (0.3).

[0031] The loss value of the maintenance measure probability prediction model based on artificial neural network (ANN) is the cross entropy loss: ; The prediction accuracy formula is as follows: ; in, is the loss value of the maintenance measure probability prediction model based on artificial neural network (ANN), i is the category, Indicates the true label, only the position corresponding to the target category is 1, and other positions are 0; It is the predicted probability output by the maintenance measure probability prediction model based on artificial neural network (ANN), that is, the output value after being processed by the softmax function; is the number of correct predictions made by the model; is the total quantity, is the prediction accuracy of the maintenance measure probability prediction model based on artificial neural network.

[0032] The loss value of the road condition prediction model based on Bayesian neural network (BNN) is the mean square error (MSE), and the formula is: ; The accuracy is the coefficient of determination R 2 , the formula is: ; in, is the loss value of the road condition prediction model based on Bayesian neural network (BNN), is the total number of predicted or true values; Indicates the number of simulations; Indicates True value; Indicates In the first run predicted values; Yes all The average value of is the accuracy of the road condition prediction model based on Bayesian neural network (BNN).

[0033] According to the 2σ principle, whether the current value is abnormal is judged, that is, 95% of the data will fall within the range of the mean plus or minus two times the standard deviation, otherwise it will be regarded as an abnormal value. The normal interval formula is as follows: ; in, is the mean, is the standard deviation.

[0034] In step S3 of this embodiment, the trained maintenance measure probability prediction model based on artificial neural network (ANN) is used to calculate the probability of each maintenance measure for each road section, and abnormal maintenance data is identified and corrected. According to the 2σ principle, if the probability of a maintenance measure is lower than 0.05, the maintenance measure is judged to be abnormal and replaced with the maintenance measure with the highest probability. This process realizes the dynamic adjustment of the abnormal threshold of maintenance data for different road sections.

[0035] With the help of the trained road condition prediction model based on Bayesian neural network (BNN), abnormal road condition index data is identified and corrected, and the reasonable range of road condition index of each road section in the current year is calculated according to the 2σ principle. If the road condition index of the current year exceeds the reasonable range, the road condition index is judged to be abnormal and replaced with the mean of the road condition index predicted by the road condition prediction model based on Bayesian neural network (BNN). This operation realizes the dynamic adjustment of the abnormal threshold of road condition index of different road sections.

[0036] For the first two abnormal situations in Table 1, if the maintenance data is normal, the pavement condition index is abnormal; if the maintenance data is abnormal, it is necessary to continue to determine whether the pavement condition index is abnormal. The abnormal maintenance measures are replaced by the maintenance measures with the highest probability value predicted by the maintenance measures probability prediction model based on artificial neural network (ANN).

[0037] For the third abnormal situation in Table 1, it is considered that the pavement condition index of the current year must be abnormal. Then it is determined whether the maintenance measure is abnormal. If it is abnormal, it is replaced with the maintenance measure with the highest probability value.

[0038] This embodiment takes a highway in a certain place as an example. According to the judgment rules in Table 1, the abnormal data set and the normal data set are distinguished. There are 12632 (47.79%) abnormal data and 13802 (52.21%) normal data. Subsequently, the maintenance measure probability prediction model based on artificial neural network (ANN) and the road condition prediction model based on Bayesian neural network (BNN) are constructed using the normal data set.

[0039] The values ​​of the 13 input features are linearly transformed to obey the standard normal distribution (i.e. Gaussian distribution) with a mean of 0 and a standard deviation of 1. The 9 maintenance measures are processed using one-hot encoding, and each maintenance measure is represented as a vector of length 9, where the corresponding position is 1 and the rest are 0.

[0040] The maintenance measure probability prediction model based on artificial neural network (ANN) was trained using a normal data set, and the prediction accuracy eventually converged to 93.6%, and the loss value converged to 0.156. Figure 2As shown in the figure, the road condition prediction model based on Bayesian neural network (BNN) is trained using a normal data set, and the prediction accuracy eventually converges to 0.733 and the loss value converges to 0.266. Figure 3 It shows that the established model has good accuracy.

[0041] In order to understand how the present invention uses the maintenance measure probability prediction model based on artificial neural network (ANN) and the road condition prediction model based on Bayesian neural network (BNN) for comprehensive identification, four road sections are selected as case studies. Table 2 lists the detailed data of these four road sections.

[0042] Table 2 Detailed data of road sections

[0043] First, the type of abnormal road section data needs to be determined. For road section 1, maintenance measures were applied but the road surface condition deteriorated. Road section 2 was not maintained, and road section 3 took maintenance measures, but its road surface condition did not change. Road section 4 did not take maintenance measures, but the road surface condition coefficient improved.

[0044] Subsequently, the values ​​of the first 13 variables of the four road sections in Table 2 were input into the maintenance measure probability prediction model based on artificial neural network (ANN) to obtain the probability values ​​of all maintenance measures, such as Figure 4 shown.

[0045] For section 1, the maintenance measure is super-sticky wear layer. Figure 4 , the probability value of the maintenance measure is less than the threshold value of 0.05, indicating that the maintenance measure is abnormal. Then, this abnormal maintenance measure is replaced with the maintenance measure with the highest probability value. Although the maintenance measure is corrected, it is still necessary to determine whether the pavement condition index (PCI) of the current year is abnormal.

[0046] For sections 2 and 3, their PCI data for the current year must be abnormal, and the maintenance measures are no maintenance and super-adhesive wearing layer, respectively. Figure 4 In the figure, the probability value of no maintenance for section 2 is greater than 0.05, indicating that no action is normal. The probability value of super-sticky wear layer for section 3 is less than 0.05, indicating that the super-sticky wear layer is abnormal. Then, this abnormal maintenance measure (super-sticky wear layer) is replaced with the maintenance measure with the highest probability value (no maintenance).

[0047] For section 4, the maintenance measure is no maintenance. Figure 4 In the data, the probability value of no maintenance is greater than 0.05, indicating that the maintenance measure is normal and does not need to be replaced. This also further indicates that the PCI data of the current year must be abnormal.

[0048] Finally, the pavement condition prediction model based on Bayesian neural network (BNN) and the 2σ principle are used to identify abnormal PCI data. The first 13 variables in Table 1 and the corrected maintenance data are input into the pavement condition prediction model based on Bayesian neural network (BNN). Each set of input data is repeatedly input into the pavement condition prediction model based on Bayesian neural network (BNN) 1000 times to generate 1000 PCI prediction values. Then, the mean (μ) and standard deviation (σ) of these prediction values ​​are calculated to obtain the reasonable range of PCI values. , as shown in Table 3.

[0049] Table 3 Reasonable PCI value (pavement condition index) range of road sections

[0050] Referring to Table 3, for Section 1, its PCI value for the current year is 89.25, which is between (56.94, 97.65), indicating that the PCI value for the current year is reasonable; the PCI values ​​for Sections 2 and 3 in the current year are abnormal, and are corrected to the corresponding means in Table 3, which are 94.55 and 89.67, respectively; for Section 4, since the maintenance measures are normal, the PCI value for the current year must be abnormal, and is corrected to 72.12.

[0051] According to the above comprehensive identification process, other road section data were also identified and corrected. With the help of the abnormality judgment criteria in Table 1, it was found that there were still 1,678 abnormal road section data, accounting for 6.35% of the total data. This may be because the normal data used to train the model is not suitable or the classification method for normal data and abnormal data is relatively simple, which may inevitably contain some hidden abnormal situations. Figure 5 As shown, the proportion of abnormal data is 47.79% before correction and 6.35% after correction, which proves the effectiveness of the comprehensive identification method proposed in the present invention.

[0052] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects: the present invention follows the inherent characteristics and changing laws of asphalt pavement maintenance and condition data, and constructs a comprehensive identification method for pavement maintenance and condition data anomalies based on multi-source data fusion and staged neural network model. In addition, the artificial neural network model is coupled with the Bayesian neural network model, and through reasonable model architecture design and dynamic threshold adjustment mechanism, efficient and accurate identification and correction of anomalies in pavement maintenance and condition data are achieved. The present invention provides a more accurate and reliable decision support tool for the field of asphalt pavement management technology, which is of great significance for improving the scientificity and rationality of pavement maintenance management and ensuring the operating quality of the highway network. It is a strong support for promoting the development of pavement maintenance work towards intelligence and refinement.

[0053] The above embodiments are merely illustrative of the principles and effects of the present invention, and are not intended to limit the present invention. Anyone familiar with the technology may modify or change the above embodiments without violating the spirit and scope of the present invention. Therefore, all equivalent modifications or changes made by those with ordinary knowledge in the technical field without departing from the spirit and technical ideas disclosed by the present invention should still be covered by the claims of the present invention.

Claims

1. A data-driven comprehensive identification method for outliers in pavement maintenance and condition data, characterized in that: The following steps are involved: S1: Identify outliers on the pavement management data and divide the pavement management data into a normal data set and an abnormal data set; S2: Use the normal data set to train the maintenance measure probability prediction model based on artificial neural network and the road condition prediction model based on Bayesian neural network respectively; S3: First, use the trained artificial neural network-based maintenance measure probability prediction model to identify and correct abnormal maintenance data, then use the Bayesian neural network-based pavement condition prediction model to identify and correct abnormal pavement condition index data, and combine the 2σ principle to dynamically adjust the abnormal thresholds of maintenance data and pavement condition index in different sections. Finally, the data set is comprehensively identified and corrected.

2. The data-driven comprehensive identification method for outliers in pavement maintenance and condition data according to claim 1 is characterized in that: The road surface management data includes road surface condition data, maintenance data, road attribute data, traffic data, climate and environment data.

3. The data-driven comprehensive identification method for abnormal values ​​of pavement maintenance and condition data according to claim 2 is characterized in that: The pavement condition data include: the pavement condition index of the previous year and the pavement condition index of the current year; the maintenance data include: no maintenance, thin anti-skid surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-situ hot regeneration, milling and replacing one layer, milling and replacing two layers; the road attribute data include: surface layer thickness, surface layer material, base layer thickness, base layer material, road age; the traffic data include: annual average daily traffic volume, annual average daily truck traffic volume.

4. The data-driven comprehensive identification method for outliers in pavement maintenance and condition data according to claim 1 is characterized in that: The maintenance measure probability prediction model based on artificial neural network includes: 13 input variables, including: pavement condition index, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, and precipitation in the previous year; Hidden layer: 2 layers, using ReLU activation function; Output layer: 9 output nodes, corresponding to the probabilities of 9 maintenance measures, including: no maintenance, thin anti-slip surface, micro-surfacing, fine surface treatment, super-sticky wear layer, ultra-thin wear layer, in-situ hot regeneration, milling and replacing one layer, milling and replacing two layers; Use the Softmax function to output probability distribution.

5. The data-driven comprehensive identification method for abnormal values ​​of pavement maintenance and condition data according to claim 1 is characterized in that: The road condition prediction model based on Bayesian neural network includes: Input layer: 14 input variables, including: pavement condition index, surface thickness, surface material, base thickness, base material, road age, annual average daily traffic volume, annual average daily truck traffic volume, number of hot days, number of consecutive hot days, number of cold days, number of consecutive cold days, precipitation, and maintenance measures in the previous year; Hidden layer: 2 layers, using ReLU activation function; Output layer: 1 output node, corresponding to the predicted value of the road condition index for the current year.

6. The data-driven comprehensive identification method for outliers in pavement maintenance and condition data according to claim 1 is characterized in that: The loss value of the maintenance measure probability prediction model based on artificial neural network is the cross entropy loss.

7. The data-driven comprehensive identification method for abnormal values ​​of pavement maintenance and condition data according to claim 1 is characterized in that: The loss value of the road condition prediction model based on Bayesian neural network is the mean square error, and the accuracy is the determination coefficient R 2 .

8. The data-driven comprehensive identification method for abnormal values ​​of pavement maintenance and condition data according to claim 1 is characterized in that: In step S3, the trained maintenance measure probability prediction model based on artificial neural network is used to calculate the probability of each maintenance measure for each road section, and abnormal maintenance data is identified and corrected; according to the 2σ principle, if the probability of a maintenance measure is lower than 0.05, the maintenance measure is judged to be abnormal and replaced by the maintenance measure with the highest probability.

9. The data-driven comprehensive identification method for outliers in pavement maintenance and condition data according to claim 1 is characterized in that: In step S3, the reasonable range of the pavement condition index of each road section in the current year is calculated according to the 2σ principle; if the pavement condition index of the current year exceeds the reasonable range, the pavement condition index is judged to be abnormal and replaced by the mean pavement condition index predicted by the pavement condition prediction model based on the Bayesian neural network.

Images