A Method for Predicting the Service Performance of Highway Asphalt Pavement Based on Prior Recurrent Neural Networks

Abstract

This invention discloses a method for predicting the service performance of highway asphalt pavement based on a priori recurrent neural networks, comprising: constructing a highway pavement maintenance information matrix; analyzing the characteristics and influencing factors of performance degradation of highway asphalt pavement and selecting relevant features; constructing a classical recurrent neural network model RNN-EAPPP; constructing a long short-term neural network model LSTM-EAPPP; constructing a gated recurrent unit neural network model GRU-EAPPP; constructing a priori classical recurrent neural network model PK-RNN-EAPPP; constructing a priori long short-term neural network model PK-LSTM-EAPPP; constructing a priori gated recurrent unit neural network model PK-GRU-EAPPP; conducting comparative experiments and selecting the model PK-GRU-EAPPP with the best performance. This invention proposes a method for constructing a highway pavement maintenance information matrix and constructs a model for predicting the service performance of highway asphalt pavement, which is of great significance for the digital and intelligent development of road maintenance.

Description

Technical Field

[0001] This invention belongs to the research fields of intelligent transportation and smart highways, specifically involving a method for predicting the service performance of highway asphalt pavement based on a priori recurrent neural networks. Background Technology

[0002] Data mining and intelligent decision-making based on multi-source heterogeneous highway pavement maintenance data is a key technology in intelligent transportation systems, and it is of great significance to the digital and intelligent development of road maintenance. Traditional highway pavement maintenance schemes suffer from problems such as high labor intensity, low decision-making efficiency, and inability to make timely and scientific maintenance decisions. However, efficient and scientific pavement maintenance decisions can extend the service life of highway pavements.

[0003] In research on the performance prediction of highway pavements, techniques such as feature encoding and autoencoders provide technical support for the mining and analysis of character-based data in highway pavement maintenance data. The development of artificial intelligence technologies such as information fusion, data mining, and deep learning offers new solutions and approaches for intelligent decision-making in highway maintenance. Therefore, this invention proposes a method for predicting the performance of highway asphalt pavements based on prior recurrent neural networks. Summary of the Invention

[0004] Purpose of the invention: In order to overcome the shortcomings of the existing technology, this invention provides a method for predicting the service performance of highway asphalt pavement based on a priori recurrent neural network. It establishes a highway maintenance information matrix and uses a priori recurrent neural network to effectively predict the service performance of highway asphalt pavement, which can provide technical support for highway maintenance decision-making.

[0005] Technical Solution: To achieve the above objectives, this invention provides a method for predicting the service performance of highway asphalt pavement based on a priori recurrent neural networks, comprising the following steps:

[0006] S1: Construct a highway pavement maintenance information matrix.

[0007] S2: Analyze the performance degradation characteristics and influencing factors of asphalt pavement on highways and select relevant features.

[0008] S3: Construct a classic recurrent neural network model RNN-EAPPP for predicting the service performance of asphalt pavements on highways.

[0009] S4: Construct a long short-time neural network model LSTM-EAPPP for predicting the service performance of asphalt pavements on highways.

[0010] S5: Construct a gated recurrent unit neural network model GRU-EAPPP for predicting the service performance of asphalt pavements on highways.

[0011] S6: Construct a prior classical recurrent neural network model PK-RNN-EAPPP for predicting the service performance of asphalt pavements on highways.

[0012] S7: Construct a prior long short-time neural network model PK-LSTM-EAPPP for predicting the service performance of asphalt pavement on highways.

[0013] S8: Construct a priori gated recurrent unit neural network model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways.

[0014] S9: Conduct comparative experiments and select the best-performing model PK-GRU-EAPPP for predicting the performance of asphalt pavement on highways.

[0015] Furthermore, the specific method for constructing the highway pavement maintenance information matrix in step S1 is as follows:

[0016] S1-1: Collect highway pavement maintenance data, including descriptions of basic highway characteristics such as technical grade, pavement structure, maintenance unit, start and end station numbers, lanes, year of construction, traffic volume, and meteorological data; descriptions of highway pavement performance such as Pavement Damage Index (PCI), Road Quality Index (RQI), Road Rutting Index (RDI), Road Skid Resistance Index (SRI), Pavement Wear Index (PWI), Road Bump Index (PBI), and Structural Strength Index (PSSI); and descriptions of historical highway maintenance conditions such as the start and end station numbers of defects, defect names, and defect lengths.

[0017] S1-2: Extract, transform, and load highway pavement maintenance data from different sources and in different formats to construct an information matrix.

[0018] S1-3: Preprocess highway pavement maintenance data.

[0019] S1-4: Before training each model, randomly select image data to divide it into training and test sets.

[0020] Furthermore, the specific steps for constructing the information matrix in steps S1-2 are as follows:

[0021] ① Determine the index data and construct the information table. This invention selects 1 kilometer as the length of the road segment unit, that is, the length of the highway segment of the unit sample. The index data of the data table is set as "starting point chainage - ending point chainage - up / down direction - lane". At the same time, in order to facilitate indexing, the "K...+..." of the starting and ending point chainages are converted into integer values, for example, "K80+300" is converted into "80300".

[0022] ② Manual data entry. For highway pavement maintenance data with small data volume and simple sources, highway pavement maintenance data from different sources can be collected into a single data table by manual entry. This includes data such as climate data, construction time, road age, and pavement structure. Such data usually does not need to be updated or is updated infrequently.

[0023] ③ Batch Loading. For large volumes of highway pavement maintenance data, the data is extracted from various sources and processed through a series of transformations and integrations before being loaded into the database. These sources include periodic inspection reports, construction drawing design tables, detailed damage tables, and traffic volume survey tables. For structured data, data reading and format conversion are performed, and the data is matched against the index data before being loaded into the data table. For unstructured data, character segmentation is used to clarify the meaning of each part. For example, for a construction drawing design table, the location information of a road segment might be in the form of "downward K47+000-K50+000". The string can be segmented into five parts based on "K", "+", and "-", representing the up and down directions, and the start and end point locations, thus finding the corresponding index data. Simultaneously, for descriptions of different damage types and maintenance schemes, such as "transverse crack", "transverse joint", and "lateral fissure", character segmentation and fuzzy matching are used to convert them into a unified description. For example, "transverse and longitudinal joints" are converted into ['transverse crack', 'longitudinal crack'], and then matched against the index data before being loaded into the data table.

[0024] ④ Constructing the information matrix. For each road segment unit, the data for each year is used as a feature vector. The road segment unit data are stacked based on spatial coordinates to form a two-dimensional information matrix. Then, considering the time dimension, a time-series-based three-dimensional information matrix is ​​constructed according to the chronological order of the highways.

[0025] Furthermore, the specific method for preprocessing highway pavement maintenance data in steps S1-3 is as follows:

[0026] ①One-hot encoding. Use one-hot encoding for features with few states, such as up / down and lane features. "Up / down, lane one" is converted to [1,0,1,0,0].

[0027] ② Label encoding. First, by segmenting, combining, and removing duplicates from the data on the pavement structure and disease types of Shaanxi Expressways, the following vocabulary is generated: {'AC':0,'medium':5,'lime-fly ash':6,'fine':19,'ATB':1,'coarse':17,'SMA':3,'micro-surfacing peeling':11,'cement':14,'insufficient pavement skid resistance':22,'crack':21,'pothole':9,'SBS':2,'longitudinal crack':18,'rut':23,'transverse crack':13,'repair':7,'pavement heave':12,'pumping':8,'dense transverse cracks':10,'subsidence':15,'mesh crack':20,'burnout':16,'sinking':4}. Then, map the text data to numerical data according to the vocabulary. If the number of disease types in different sections of the expressway is different and the same-length feature vectors cannot be obtained after conversion, padding operations are performed to fill in 0s at the end to make the data align. The converted data form is as follows: [1,19,3,17,6,13,18,23,0,0,0,0].

[0028] ③ Missing value handling. In the pavement maintenance data of expressways, the pavement performance data, traffic volume data, climate data, etc. are all continuous variables in time series data. If there are missing values, interpolation methods can be used to fill in the missing values; for categorical data such as maintenance measures, the missing values can be regarded as a new category; for other null values, placeholders can be used for filling.

[0029] ④ Outlier detection. The present invention uses the interquartile range (IQR) detection method to process outliers in the pavement maintenance data of expressways. First, calculate the third quartile Q3, the first quartile Q1, and the median of the data; second, calculate the box height IQR = Q3 - Q1; third, determine the outlier range, the upper limit is Q3 + 1.5 × IQR, and the lower limit is Q1 - 1.5 × IQR; finally, regard the data points outside this range as outliers and remove them.

[0030] ⑤ Data normalization. To prevent the model from assigning different weights to features of different magnitudes and relatively weakening the role of features with smaller values, maximum-minimum normalization is used for scaling:

[0031]

[0032] ⑥ Data standardization. To adjust the data distribution to a standard normal distribution for the subsequent use of machine learning models, the z-Score method is used for data standardization, where μ is the mean of all samples and σ is the standard deviation of all samples:

[0033]

[0034] Furthermore, the specific method for analyzing the performance degradation characteristics and influencing factors of highway asphalt pavement and selecting relevant features in step S2 is as follows:

[0035] S2-1: Newly built or heavily repaired highways exhibit relatively stable pavement degradation under the influence of loads, environmental factors, etc. Therefore, the performance of highway asphalt pavements should refer to the standard degradation equation:

[0036] PPI = PPI0{1 - exp[-(a / t)} b ]}

[0037] The performance variation curves of four typical highway pavements were fitted by different values ​​of a and b: concave, convex, inverse S-shaped, and rapid decay followed by stabilization.

[0038] S2-2: Select the following features for the prediction model: mixed equivalent, passenger-to-freight ratio, road age, up and down traffic, lane characteristics, annual average precipitation, annual average temperature, maximum temperature, minimum temperature, pavement structure, maintenance plan, historical PCI data, PQI, SRI, RQI, RDI.

[0039] S2-3: Input highway maintenance data into the decision tree, predict the highway pavement PCI for the following year based on the data of the current year, and calculate the impact of each feature on the future performance of the highway pavement according to the feature importance scoring function of the decision tree, and rank the importance of each feature.

[0040] Furthermore, the specific method for constructing the classic recurrent neural network model RNN-EAPPP for predicting the service performance of highway asphalt pavement in step S3 is as follows:

[0041] S3-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0042] S3-2: Set up an RNN-EAPPP model with 128 recurrent units. Each recurrent unit passes the hidden state sequentially between time steps. The specific method of passing the hidden state is as follows:

[0043] h t =f(W i x t +W h h t-1 +b)

[0044] Where f is the activation function, typically tanh, W i W is the weight matrix between the input node and the hidden state. h This is the weight matrix between hidden states.

[0045] S3-3: RNN-EAPPP outputs upwards at the last time step, and the output method is as follows:

[0046] y t =g(W o h t )

[0047] S3-4: After the RNN-EAPPP outputs upwards, it connects to three fully connected layers with dimensions of 64, 64, and 1 respectively, outputting the PCI prediction result for the following year. t+1 .

[0048] Furthermore, the specific method for constructing the LSTM-EAPPP long-short-time neural network model for predicting the service performance of highway asphalt pavements in step S4 is as follows:

[0049] S4-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0050] S4-2: Set up the LSTM-EAPPP model with 128 recurrent units, and each recurrent unit passes the hidden state sequentially between time steps.

[0051] S4-3: LSTM-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1, outputting the PCI prediction result for the following year. t+1 .

[0052] Furthermore, the calculation method for the hidden state passed by each loop unit in step S4-2 is as follows:

[0053] ① Set a forget gate f in the loop unit to view h t-1 and x t Output a vector between [0, 1] to represent the cell state c. t-1 The retention and discarding of information in the calculation method, where is the sigmoid activation function:

[0054] f t =σ(W f ·[h t-1 x t ]+b f )

[0055] ② In the loop unit, set input gate i to control information updates, h t-1 and x t After entering the input gate, you will receive the information that determines what needs to be updated. t and the state of the current loop unit. The calculation method is as follows, where tanh is the activation function:

[0056] i t =σ(W i ·[h t-1 x t ]+b i )

[0057]

[0058] ③ Update kernel state c t The calculation method is as follows:

[0059]

[0060] ④ According to h t-1 and x t The output of the loop unit at the current time is obtained, and the update rule is as follows:

[0061] o t =σ(W o [h t-1 x t ]+b o )

[0062] ⑤ Based on the current kernel state c t Update the current hidden state h t The calculation method is as follows:

[0063] h t =o t ·tanh(c t )

[0064] Furthermore, the specific method for constructing the gated recurrent neural network model GRU-EAPPP for predicting the service performance of highway asphalt pavement in step S5 is as follows:

[0065] S5-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0066] S5-2: The GRU-EAPPP model is configured with 128 recurrent units, each of which sequentially passes the hidden state between time steps. Based on LSTM-EAPPP, the GRU-EAPPP model integrates the input gate and forget gate into a single update gate, using a reset gate r and an update gate u to control the hidden state h. t The update and transfer are calculated as follows, where h′ is the intermediate state of the memory unit.

[0067] r t =σ(W r ·[h t-1 x t ]+br )

[0068] u t =σ(W u ·[h t-1 x t ]+b u )

[0069] h′ t-1 =h t-1 ·r t

[0070] h′ t =tanh(W h ·[h t-1 ·r t x t ]+b h )

[0071] h t =(1-u)·h t -1+u·h′ t

[0072] S5-3: GRU-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1, outputting the PCI prediction result for the following year. t+1 .

[0073] Furthermore, the specific steps in step S6 for constructing the prior classical recurrent neural network model PK-RNN-EAPPP for predicting the service performance of highway asphalt pavements are as follows:

[0074] S6-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0075] S6-2: The model input dimension is (T,X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0076] S6-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S3-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data with the PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model.

[0077] S6-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0078] Furthermore, the specific method for constructing the prior long short-time neural network model PK-LSTM-EAPPP for predicting the service performance of highway asphalt pavement in step S7 is as follows:

[0079] S7-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0080] S7-2: The model input dimension is (T,X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0081] S7-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S4-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data to the PCI index. The total feature dimension remains unchanged. This is equivalent to adding a prior knowledge to the model.

[0082] S7-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0083] Furthermore, the specific method for constructing the prior gated recurrent neural network model PK-GRU-EAPPP for predicting the service performance of highway asphalt pavement in step S8 is as follows:

[0084] S8-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0085] S8-2: The model input dimension is (T,X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0086] S8-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S5-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data with the PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model.

[0087] S8-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0088] Furthermore, in step S9, a comparative experiment is conducted to select the best-performing model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways. The specific method is as follows: construct the selection decision regression tree DT-EAPPP, gradient boosting decision tree GBDT-EAPPP, and random forest model RF-EAPPP for predicting the service performance of asphalt pavement on highways, and conduct experiments to compare the above nine models, and select the best-performing prior gated recurrent unit neural network model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways.

[0089] Beneficial effects: Compared with the prior art, this invention uses deep learning technology to predict the performance of asphalt pavement on highways, improving prediction accuracy, model stability and generalization ability, and can provide technical support for highway maintenance decisions. Attached Figure Description

[0090] Figure 1 A schematic diagram of a three-dimensional information matrix for highway pavement maintenance data.

[0091] Figure 2 This is a schematic diagram of the PK-GRU-EAPPP prior gated recurrent unit neural network model used for predicting the service performance of asphalt pavement on highways. Detailed Implementation

[0092] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments.

[0093] This invention provides a method for predicting the service performance of highway asphalt pavement based on a priori recurrent neural networks, comprising the following steps:

[0094] S1: Construct a highway pavement maintenance information matrix, specifically as follows:

[0095] S1-1: Collect highway pavement maintenance data, including descriptions of basic highway characteristics such as technical grade, pavement structure, maintenance unit, start and end station numbers, lanes, year of construction, traffic volume, and meteorological data; descriptions of highway pavement performance such as Pavement Damage Index (PCI), Road Quality Index (RQI), Road Rutting Index (RDI), Road Skid Resistance Index (SRI), Pavement Wear Index (PWI), Road Bump Index (PBI), and Structural Strength Index (PSSI); and descriptions of historical highway maintenance conditions such as the start and end station numbers of defects, defect names, and defect lengths.

[0096] S1-2: Extract, transform, and load highway pavement maintenance data from different sources and in different formats to construct an information matrix. After construction, there are 486 data samples per year, covering 7 years from 2014 to 2020. The data is organized chronologically with dimensions (7, 486, X), and after expansion, there are 3402 data records.

[0097] ① Determine the index data and construct the information table. This invention selects 1 kilometer as the length of the road segment unit, that is, the length of the highway segment of the unit sample. The index data of the data table is set as "starting point chainage - ending point chainage - up / down direction - lane". At the same time, in order to facilitate indexing, the "K...+..." of the starting and ending point chainages are converted into integer values, for example, "K80+300" is converted into "80300".

[0098] ② Manual data entry. For highway pavement maintenance data with small data volume and simple sources, highway pavement maintenance data from different sources can be collected into a single data table by manual entry. This includes data such as climate data, construction time, road age, and pavement structure. Such data usually does not need to be updated or is updated infrequently.

[0099] ③ Batch Loading. For large volumes of highway pavement maintenance data, the data is extracted from various sources and processed through a series of transformations and integrations before being loaded into the database. These sources include periodic inspection reports, construction drawing design tables, detailed damage tables, and traffic volume survey tables. For structured data, data reading and format conversion are performed, and the data is matched against the index data before being loaded into the data table. For unstructured data, character segmentation is used to clarify the meaning of each part. For example, for a construction drawing design table, the location information of a road segment might be in the form of "downward K47+000-K50+000". The string can be segmented into five parts based on "K", "+", and "-", representing the up and down directions, and the start and end point locations, thus finding the corresponding index data. Simultaneously, for descriptions of different damage types and maintenance schemes, such as "transverse crack", "transverse joint", and "lateral fissure", character segmentation and fuzzy matching are used to convert them into a unified description. For example, "transverse and longitudinal joints" are converted into ['transverse crack', 'longitudinal crack'], and then matched against the index data before being loaded into the data table.

[0100] ④ Constructing the information matrix. For each road segment unit, the data for each year is used as a feature vector. The road segment unit data are stacked based on spatial coordinates to form a two-dimensional information matrix. Then, considering the time dimension, a time-series-based three-dimensional information matrix is ​​constructed according to the chronological order of the highways.

[0101] S1-3: Preprocessing of highway pavement maintenance data:

[0102] ①One-hot encoding. Use one-hot encoding for features with few states, such as up / down and lane features. "Up / down, lane one" is converted to [1,0,1,0,0].

[0103] ② Label encoding. First, by segmenting, combining, and removing duplicates from the data on the pavement structure and disease types of Shaanxi Expressways, the following vocabulary is generated: {'AC':0,'medium':5,'lime-fly ash':6,'fine':19,'ATB':1,'coarse':17,'SMA':3,'micro-surfacing peeling':11,'cement':14,'insufficient pavement skid resistance':22,'crack':21,'pothole':9,'SBS':2,'longitudinal crack':18,'rut':23,'transverse crack':13,'repair':7,'pavement heave':12,'pumping':8,'dense transverse cracks':10,'subsidence':15,'net crack':20,'burnout':16,'sinking':4}. Then, map the text data to numerical data according to the vocabulary. If the number of disease types in different sections of the expressway is different and the same-length feature vectors cannot be obtained after conversion, padding operations are performed to append 0s at the end to align the data. The converted data form is as follows: [1, 19, 3, 17, 6, 13, 18, 23, 0, 0, 0, 0].

[0104] ③ Missing value handling. In the expressway pavement maintenance data, pavement performance data, traffic volume data, climate data, etc. are all continuous variables in time series data. If there are missing values, interpolation methods can be used to fill in the missing values; for categorical data such as maintenance measures, the missing values can be regarded as a new category; for other null values, placeholders can be used for filling.

[0105] ④ Outlier detection. This invention uses the interquartile range (IQR) detection method to process outliers in expressway pavement maintenance data. First, calculate the third quartile Q3, the first quartile Q1, and the median of the data; second, calculate the box height IQR = Q3 - Q1; third, determine the outlier range, with the upper limit being Q3 + 1.5×IQR and the lower limit being Q1 - 1.5×IQR; finally, consider the data points outside this range as outliers and remove them.

[0106] ⑤ Data normalization. To prevent the model from assigning different weights to features of different magnitudes and relatively weakening the role of features with smaller values, maximum-minimum normalization is used for scaling:

[0107]

[0108] ⑥ Data standardization. To adjust the data distribution to a standard normal distribution for the subsequent use of machine learning models, the z-Score method is used for data standardization, where μ is the mean of all samples and σ is the standard deviation of all samples:

[0109]

[0110] S1-4: Before training each model, image data is randomly sampled in a 3:1 ratio to divide the training set and the test set.

[0111] S2: Analyze the performance degradation characteristics and influencing factors of asphalt pavement on highways and select relevant features:

[0112] S2-1: Newly built or heavily repaired highways exhibit relatively stable pavement degradation under the influence of loads, environmental factors, etc. Therefore, the performance of highway asphalt pavements should refer to the standard degradation equation:

[0113] PPI = PPI0{1 - exp[-(a / t)} b ]}

[0114] The performance variation curves of four typical highway pavements were fitted by different values ​​of a and b: concave, convex, inverse S-shaped, and rapid decay followed by stabilization.

[0115] S2-2: Select the following features for the prediction model: mixed equivalent, passenger-to-freight ratio, road age, up and down traffic, lane characteristics, annual average precipitation, annual average temperature, maximum temperature, minimum temperature, pavement structure, maintenance plan, historical PCI data, PQI, SRI, RQI, RDI.

[0116] S2-3: Input highway maintenance data into the decision tree, predict the highway pavement PCI for the following year based on the data of the current year, and calculate the impact of each feature on the future performance of the highway pavement according to the feature importance scoring function of the decision tree, and rank the importance of each feature.

[0117] S3: Constructing a classic recurrent neural network model RNN-EAPPP for predicting the service performance of asphalt pavements on highways:

[0118] S3-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0119] S3-2: Set up an RNN-EAPPP model with 128 recurrent units. Each recurrent unit passes the hidden state sequentially between time steps. The specific method of passing the hidden state is as follows:

[0120] h t =f(W i x t +W h h t-1 +b)

[0121] Where f is the activation function, typically tanh, W i W is the weight matrix between the input node and the hidden state. h This is the weight matrix between hidden states.

[0122] S3-3: RNN-EAPPP outputs upwards at the last time step, and the output method is as follows:

[0123] y t =g(W o h t )

[0124] S3-4: After the RNN-EAPPP outputs upwards, it connects to three fully connected layers with dimensions of 64, 64, and 1 respectively, outputting the PCI prediction result for the following year. t+1 .

[0125] S4: LSTM-EAPPP, a long-term and short-term neural network model for predicting the service performance of asphalt pavements on highways.

[0126] S4-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0127] S4-2: Set up the LSTM-EAPPP model with 128 recurrent units. Each recurrent unit passes the hidden state sequentially between time steps. The specific calculation method is as follows:

[0128] ① Set a forget gate f in the loop unit to view h t-1 and x t Output a vector between [0, 1] to represent the cell state c. t-1 The retention and discarding of information in the calculation method, where is the sigmoid activation function:

[0129] f t =σ(W f ·[h t-1 x t ]+b f )

[0130] ② In the loop unit, set input gate i to control information updates, h t-1 and x t After entering the input gate, you will receive the information that determines what needs to be updated. t and the state of the current loop unit. The calculation method is as follows, where tanh is the activation function:

[0131] i t =σ(IW i ·[h t-1 x t ]+b t )

[0132]

[0133] ③ Update kernel state c t The calculation method is as follows:

[0134]

[0135] ④ According to h t-1 and x t The output of the loop unit at the current time is obtained, and the update rule is as follows:

[0136] o t =σ(W o [h t-1 x t ]+b o )

[0137] ⑤ Based on the current kernel state c t Update the current hidden state h t The calculation method is as follows:

[0138] h t =o t ·tanh(c t )

[0139] S4-3: LSTM-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1, outputting the PCI prediction result for the following year. t+1 .

[0140] S5: Constructing a gated recurrent neural network model GRU-EAPPP for predicting the service performance of asphalt pavements on highways:

[0141] S5-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample.

[0142] S5-2: The GRU-EAPPP model is configured with 128 recurrent units, each of which sequentially passes the hidden state between time steps. Based on LSTM-EAPPP, the GRU-EAPPP model integrates the input gate and forget gate into a single update gate, using a reset gate r and an update gate u to control the hidden state h. t The update and transfer are calculated as follows, where h′ is the intermediate state of the memory unit.

[0143] r t =σ(W r ·[h t-1 x t ]+b r )

[0144] u t =σ(Wu ·[h t-1 x t ]+b u )

[0145] h′ t-1 =h t-1 ·r t

[0146] h′ t =tanh(W h ·[h t-1 ·r t x t ]+b h )

[0147] h t =(1-u)·h t-1 +u·h′ t

[0148] S5-3: GRU-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1, outputting the PCI prediction result for the following year. t+1 .

[0149] S6: Constructing a prior classical recurrent neural network model PK-RNN-EAPPP for predicting the service performance of asphalt pavements on highways:

[0150] S6-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0151] S6-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0152] S6-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S3-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data with the PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model.

[0153] S6-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0154] S7: Constructing a priori long-short-time neural network model PK-LSTM-EAPPP for predicting the service performance of asphalt pavements on highways:

[0155] S7-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0156] S7-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0157] S7-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S4-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data to the PCI index. The total feature dimension remains unchanged. This is equivalent to adding a prior knowledge to the model.

[0158] S7-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0159] S8: Constructing a priori gated recurrent neural network model PK-GRU-EAPPP for predicting the service performance of asphalt pavements on highways:

[0160] S8-1: Based on the result of feature importance ranking in step S2-3, select the highway pavement PCI value as the important feature x. s In order to reduce the impact of data normalization, the PCI value input in the shallow layer is the unnormalized value, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year.

[0161] S8-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps.

[0162] S8-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as in step S5-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. Then, perform Add feature fusion with the PCI index input in the last time step. That is, add each dimension of data with the PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model.

[0163] S8-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1, and output the PCI prediction result for the following year. t+1 .

[0164] S9: A comparative experiment was conducted. Based on the established highway maintenance data matrix, ReLU was selected as the activation function and Adam as the optimizer. Three loss functions—MAE, Huber, and RSME—were selected for the experiment. The calculation methods for each loss function are as follows:

[0165] MAE is the mean absolute error, where e t It represents the difference between the true value and the predicted value. Its advantage is that it is less affected by outliers and can intuitively reflect the difference between the predicted value and the true value.

[0166]

[0167] RSME, or Root Mean Square Error, is the square root of the mean square error. It makes the loss value have the same dimensions as the sample and, like MSE, is prone to bias towards outliers.

[0168]

[0169] Huber loss is less sensitive to outliers in the data than MSE. When the error is large, Huber loss tends to be MAE, while when the error is small, Huber loss becomes MSE. This means that Huber loss combines the advantages of MAE and MSE and is more robust to outliers.

[0170]

[0171] When evaluating the model, two indicators are selected: the mean absolute error (MAE) and the root mean square error (RMSE) between the predicted and actual values. The smaller the values, the closer the predicted and actual values ​​are. The correlation coefficient R is also used. 2 R represents the goodness of fit of the data. 2 The closer a value is to 1, the better the model interprets the data.

[0172] The experimental results under different loss functions are as follows: Untrained test sets were imported into each model.

[0173] Table 1. Experimental results of recurrent neural network models on the test set under different loss functions.

[0174]

[0175]

[0176] MAE: Mean Absolute Error; RSME: Root Mean Square Error; R 2 Correlation coefficient

[0177] As shown in the table, when using a model without prior information, the GRU-EAPPP model outperforms both the RNN-EAPPP and LSTM-EAPPP models. The R² of the GRU-EAPPP model is significantly higher. 2 Around 0.73; after adding prior information, the R-values ​​of the three prior recurrent neural network prediction models, PK-RNN-EAPPP, PK-LSTM-EAPPP, and PK-GRU-EAPPP, are... 2 Between 0.82 and 0.84, the model's accuracy shows a significant improvement. The PK-GRU-EAPPP model's accuracy and stability are slightly better than PK-RNN-EAPPP and PK-LSTM-EAPPP. A horizontal comparison of loss function performance shows that Huber outperforms MAE and RMSE. The prediction model using the Huber loss function achieves the best MAE (2.12) and best RMSE (2.70) on the test set. 2 It reached a maximum of 0.839.

[0178] To compare and analyze the advantages of the highway asphalt pavement service performance prediction model constructed in this paper with other data-driven machine learning models in predicting the service performance of highway asphalt pavements, this invention further constructed the selection decision regression tree DT-EAPPP, gradient boosting decision tree GBDT-EAPPP, and random forest model RF-EAPPP for highway asphalt pavement service performance prediction, and conducted comparative experiments. The performance of each model on the test set is shown in Table 2. It can be seen that the PK-GRU-EAPPP model constructed in this invention outperforms the three prediction models DT-EAPPP, GBDT-EAPPP, and RF-EAPPP, with its MAE, RMSE, and R... 2 The values ​​are 2.12, 2.70, and 0.839, respectively.

[0179] Table 2 shows the performance of each model on the test set.

[0180]

[0181] The technical means disclosed in this invention are not limited to those disclosed in the above embodiments, but also include technical solutions composed of any combination of the above technical features. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of this invention, and these improvements and modifications are also considered within the scope of protection of this invention.

Claims

1. A method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks, characterized in that: Includes the following steps: S1: Construct a highway pavement maintenance information matrix; S2: Analyze the performance degradation characteristics and influencing factors of asphalt pavement on highways and select relevant features; S3: Construct a classic recurrent neural network model RNN-EAPPP for predicting the service performance of asphalt pavements on highways; S4: Construct a long short-time neural network model LSTM-EAPPP for predicting the service performance of asphalt pavement on highways; S5: Construct a gated recurrent unit neural network model GRU-EAPPP for predicting the service performance of asphalt pavement on highways; S6: Construct a prior classical recurrent neural network model PK-RNN-EAPPP for predicting the service performance of asphalt pavement on highways; S7: Construct a prior long short-time neural network model PK-LSTM-EAPPP for predicting the service performance of asphalt pavement on highways; S8: Construct a priori gated recurrent unit neural network model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways; S9: Conduct comparative experiments and select the best-performing model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways; The specific content of constructing the prior long-short-time neural network model PK-LSTM-EAPPP for predicting the service performance of highway asphalt pavement in step S7 is as follows: S7-1: Based on the result of feature importance ranking in step S2-3, the Highway Pavement Damage Index (PCI) value is selected as an important feature. In order to reduce the impact of data normalization, the road damage index (PCI) value input in the shallow layer is before normalization, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year. S7-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps; S7-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as step S4-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. And perform Add feature fusion with the road damage index PCI index input in the last time step. That is, add each dimension of data with the road damage index PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model. S7-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1 to output the PCI (Pavement Damage Index) prediction result for the following year. ; The specific content of constructing the prior gated recurrent neural network model PK-GRU-EAPPP for predicting the service performance of asphalt pavement on highways in step S8 is as follows: S8-1: Based on the result of feature importance ranking in step S2-3, the Highway Pavement Damage Index (PCI) value is selected as an important feature. In order to reduce the impact of data normalization, the road damage index (PCI) value input in the shallow layer is before normalization, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year. S8-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps; S8-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as step S5-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. And perform Add feature fusion with the road damage index PCI index input in the last time step. That is, add each dimension of data with the road damage index PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model. S8-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1 to output the PCI (Pavement Damage Index) prediction result for the following year. .

2. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific method for constructing the highway pavement maintenance information matrix in step S1 is as follows: S1-1: Collect highway pavement maintenance data, including descriptions of highway basic characteristics such as technical grade, pavement structure, maintenance unit, starting and ending station numbers, lanes, year of construction, traffic volume, and meteorological data; and pavement damage index describing highway pavement performance. The following are included: Roadway Quality Index (RQI), Roadway Rutting Index (RDI), Roadway Skid Resistance Index (SRI), Roadway Wear Index (PWI), Roadway Bump Index (PBI), and Structural Strength Index (PSSI); and descriptions of the starting and ending points, names, and lengths of defects in the historical maintenance of highways. S1-2: Extract, transform, and load highway pavement maintenance data from different sources and in different formats to construct an information matrix; S1-3: Preprocessing highway pavement maintenance data; S1-4: Before training each model, randomly sample image data to divide it into training and test sets; The specific content of constructing the information matrix in step S1-2 is as follows: ① Determine the index data and construct the information table; select 1 kilometer as the length of the road segment unit, that is, the length of the highway segment of the unit sample; set the index data of the data table as: starting station number - ending station number - up and down lanes - lane; at the same time, in order to facilitate indexing, convert the K...+... of the starting and ending station numbers to integer values; ② Manual data input; For highway pavement maintenance data with small data volume and simple sources, highway pavement maintenance data from different sources can be collected into a data table by manual input, including climate data, construction time, road age, and pavement structure. This type of data does not need to be updated or has a low update frequency. ③ Batch loading; For a large amount of highway pavement maintenance data, the data is extracted from different sources and loaded into the database after a series of transformations and integrations. These sources include periodic inspection reports, construction drawing design tables, defect details tables, and traffic volume survey tables. For structured data, data is read and formatted, and then matched with the index data to load it into the data table. For unstructured data, character segmentation is used to clarify the meaning of each part, representing the upstream and downstream, and start and end point information, so as to find the corresponding index data. At the same time, for descriptions of different disease types and maintenance schemes, character segmentation and fuzzy matching are used to convert them into a unified description, and then matched with the index data to load them into the data table. ④ Construct an information matrix; For each road segment unit, the data for each year is used as a feature vector. The road segment unit data are stacked based on spatial coordinates to form a two-dimensional information matrix; Then, considering the time dimension, a time-series-based three-dimensional information matrix is ​​constructed according to the time sequence of the expressway. The specific content of highway pavement maintenance data preprocessing in steps S1-3 is as follows: ①One-hot encoding; use one-hot encoding for features with few states; ② Tag Encoding; First, by segmenting, combining, and deduplicating data on pavement structure and defect types of Shaanxi expressways, the following vocabulary was generated: {'AC': 0, 'Medium': 5, 'Fei': 6, 'Fine': 19, 'ATB': 1, 'Coarse': 17, 'SMA': 3, 'Micro-surfacing detachment': 11, 'Cement': 14, 'Insufficient pavement skid resistance': 22, 'Crack': 21, 'Pothole': 9, 'SBS': 2, 'Longitudinal crack': 18, 'Ruts': 23, 'Transverse crack': 13, 'Repair': 7, 'Bursting': 12, 'Pumping': 8, 'Dense transverse cracks': 10, 'Subsidence': 15, 'Network crack': 20, 'Burn':} 16, 'sink':4}, and then map the text data into numerical data according to the vocabulary; if the number of disease types in different sections of the highway is different, resulting in the inability to obtain feature vectors of the same length after conversion, then padding is performed to add 0s to the end to make the data aligned; ③ Handling missing values; In highway pavement maintenance data, pavement performance data, traffic volume data, and climate data are all continuous variables in time series data. If there are missing values, interpolation is used to fill in the missing values; For maintenance measure category data, missing values ​​are treated as a new category; For other empty values, placeholders are used to fill in the missing values. ④ Outlier detection; the interquartile range (IQR) method is used to handle outliers in highway pavement maintenance data; first, the third quartile of the data is calculated. First quartile And the median; then calculate the height of the box. ; Re-determine the range of outliers, with the upper limit being... The lower limit is Finally, data points outside this range are considered outliers and removed. ⑤ Data normalization; To avoid assigning different weights to features of different magnitudes in the model, which would relatively weaken the effect of features with smaller values, min-max normalization is used for scaling: ⑥ Data Standardization; To adjust the data distribution to a standard normal distribution for easier use in subsequent machine learning models, the z-Score method is used for data standardization, where μ is the mean of all samples and σ is the standard deviation of all samples: 。 3. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific content of analyzing the performance degradation characteristics and influencing factors of highway asphalt pavement in step S2 and selecting relevant features is as follows: S2-1: Newly built or heavily repaired highways exhibit relatively stable pavement degradation under load and environmental factors. Therefore, the performance of highway asphalt pavements is referenced to the standard degradation equation. Among them, four typical highway pavement performance change curves were fitted by different values ​​of a and b: concave, convex, inverse S-shaped, and rapid decay followed by stability. S2-2: The following features are selected for the prediction model: mixed equivalent, passenger-to-freight ratio, road age, up and down traffic, lane characteristics, annual average precipitation, annual average temperature, maximum temperature, minimum temperature, pavement structure, maintenance plan, historical data of pavement damage index PCI, pavement technical condition index PQI, pavement skid resistance index SRI, pavement ride quality index RQI, and pavement rutting index RDI. S2-3: Input highway maintenance data into the decision tree, predict the highway pavement damage index (PCI) for the following year based on the data of the current year, and calculate the impact of each feature on the future performance of the highway pavement according to the feature importance scoring function of the decision tree, and rank the importance of each feature.

4. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific content of constructing the classic recurrent neural network model RNN-EAPPP for predicting the service performance of highway asphalt pavement in step S3 is as follows: S3-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample; S3-2: Set up an RNN-EAPPP model with 128 recurrent units. Each recurrent unit passes hidden states sequentially between time steps. The specific method of passing hidden states is as follows: in, For the activation function, tanh is chosen. This is the weight matrix between the input nodes and the hidden states. This is the weight matrix between hidden states; S3-3: RNN-EAPPP outputs upwards at the last time step, and the output method is as follows: S3-4: After the RNN-EAPPP outputs upwards, it connects to three fully connected layers with dimensions of 64, 64, and 1 respectively, outputting the predicted pavement damage index (PCI) for the following year. .

5. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific content of constructing the LSTM-EAPPP long-short-time neural network model for predicting the service performance of highway asphalt pavement in step S4 is as follows: S4-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample; S4-2: Set up the LSTM-EAPPP model with 128 recurrent units, and each recurrent unit passes the hidden state sequentially between time steps; S4-3: LSTM-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1 to output the predicted pavement damage index (PCI) for the following year. ; The calculation method for the hidden state passed to each loop unit in step S4-2 is as follows: ① Set a forget gate f in the loop unit to check. and Output a A vector between them, used to represent cell states. The retention and discarding of information in the calculation method, where is the sigmoid activation function: ② Set input gate i in the loop unit to control information updates. and After entering the input gate, you will receive the information that determines which update to make. and the state of the current loop unit. The calculation method is as follows, where tanh is the activation function: ③ Update kernel status The calculation method is as follows: ④According to and The output of the loop unit at the current time is obtained, and the update rule is as follows: ⑤ Based on the current nuclear state Update the current hidden state The calculation method is as follows: 。 6. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific content of constructing the gated recurrent neural network model GRU-EAPPP for predicting the service performance of highway asphalt pavement in step S5 is as follows: S5-1: Input highway maintenance data with dimensions (T, X), where T is the time step and X is the feature dimension of each highway segment sample; S5-2: The GRU-EAPPP model is configured with 128 recurrent units, each of which sequentially passes the hidden state between time steps; based on LSTM-EAPPP, the GRU-EAPPP model integrates the input gate and forget gate into a single update gate, using a reset gate. and the update gate Controlling hidden states The update and transmission are calculated as follows, where... This represents the intermediate state of a memory unit; = = S5-3: GRU-EAPPP outputs upwards at the last time step and connects three fully connected layers with dimensions of 64, 64, and 1, outputting the PCI prediction result for the following year. .

7. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: The specific content of constructing the prior classical recurrent neural network model PK-RNN-EAPPP for predicting the service performance of asphalt pavement on highways in step S6 is as follows: S6-1: Based on the result of feature importance ranking in step S2-3, the Highway Pavement Damage Index (PCI) value is selected as an important feature. In order to reduce the impact of data normalization, the road damage index (PCI) value input in the shallow layer is before normalization, so that the network structure before adding prior knowledge is similar to learning the difference between the PCI of the predicted year and the current year. S6-2: The model input dimension is (T, X), where X is the feature dimension of each road segment sample, T is the time step, and the hidden state is passed sequentially between time steps; S6-3: Set the dimension of the loop operation unit to 128. The operation method of the loop unit is the same as step S3-2. Only output upwards in the last time step. Use a 64-dimensional fully connected layer for further feature extraction. And perform Add feature fusion with the road damage index PCI index input in the last time step. That is, add each dimension of data with the road damage index PCI index. The total feature dimension remains unchanged, which is equivalent to adding a prior knowledge to the model. S6-4: After feature fusion, connect two fully connected layers with dimensions of 64 and 1 to output the PCI (Pavement Damage Index) prediction result for the following year. .

8. The method for predicting the service performance of highway asphalt pavement based on prior recurrent neural networks according to claim 1, characterized in that: We constructed the selection decision regression tree DT-EAPPP, gradient boosting decision tree GBDT-EAPPP, and random forest model RF-EAPPP for predicting the service performance of asphalt pavement on highways. We conducted experiments to compare the above nine models and selected the prior gated recurrent unit neural network model PK-GRU-EAPPP, which has the best performance for predicting the service performance of asphalt pavement on highways.

Images