A method for analyzing the strain of road base based on temporal convolutional network

Abstract

The present invention discloses a method for analyzing the strain of road base layers based on a temporal convolutional network. First, the present invention preprocesses the original data of the long-term monitoring data obtained by the automated road monitoring system. On this basis, Pearson correlation analysis is used to achieve the purpose of reducing the dimension of the input data, thereby improving the prediction accuracy of the model. Finally, the data in the appropriate time period is selected and input into the temporal convolutional network model for training to learn the characteristics of the monitoring data. To address the problem that the data between each time period is not continuous, a training method with weight self-transfer is adopted, making the historical information features relatively more obvious in the next time period and easier to be learned by the model, which helps to improve the memory ability of the model and reduce overfitting. In addition, the present invention can not only save labor and time costs, but also lay a foundation for subsequent road maintenance work.

Description

Technical Field

[0001] The present invention belongs to the field of time series data analysis, and relates to a road base strain analysis method based on a time series convolutional network. The present invention is applied to perform road base strain analysis on pavement structure monitoring data obtained by a road monitoring system using multi-sensor fusion. Background Art

[0002] In the past two decades, China's highway network has been gradually built and formed, and the future situation of highway infrastructure construction has gradually changed from "construction" to "maintenance". Diseases such as cracks and potholes affect the safety and comfort of driving to a certain extent. Therefore, accurate road base strain analysis results can largely avoid serious structural damage to the road in the future and provide reliable and effective technical support for subsequent road maintenance management work.

[0003] Due to the lack of sufficient basic theory and technical support, currently traditional road base strain prediction and analysis methods generally perform regression analysis of mathematical statistics based on the stress and strain data themselves obtained from long-term monitoring. The calculation process is cumbersome, time-consuming, and the accuracy is often difficult to guarantee. Benefiting from the development of artificial intelligence technology, machine learning methods have been gradually applied to the field of monitoring big data analysis, and can relatively quickly and efficiently predict and analyze the obtained data, saving time and reducing costs.

[0004] Currently, road base strain prediction and analysis based on deep learning often face problems such as large input data dimensions and discontinuous monitoring data time. Performing correlation analysis on the data to reduce the data dimension at the input end, and adopting "weight self-transfer" during the training process can bring considerable performance improvement, improve accuracy and stability, and reduce overfitting.

[0005] Therefore, the present invention proposes a road base strain analysis method based on a time series convolutional network. The present invention first preprocesses the sensor data collected by the road monitoring system to obtain a data set that can be input into the network for training, then obtains the Pearson correlation coefficient through feature correlation analysis of the data set, so as to remove the feature data with a smaller correlation coefficient to achieve the purpose of data dimension reduction. Finally, the final data set is trained with a time series convolutional neural network to predict and analyze the road base strain data, so as to pre-understand the change trend of the road base strain and lay a foundation for subsequent road maintenance work. Summary of the Invention

[0006] The object of the present invention is to analyze the pavement monitoring data obtained by sensors from 2012 to 2020 for eight years through a road base strain prediction and analysis method based on a temporal convolutional network. The long-term monitoring data includes asphalt strain, embedded triaxial strain, soil layer strain, soil pressure, temperature, osmotic pressure, and soil moisture, so as to predict the base strain according to appropriate influencing factors and obtain the change trend of the base strain in the future for a period of time.

[0007] I. Temporal Convolutional Network

[0008] The temporal convolutional network (TCN) adopted in the present invention consists of dilated and causal 1D convolutional layers with the same input and output lengths. The traditional causal convolutional layer is replaced by a dilated causal convolutional layer to better capture longer time-dependent information. The temporal convolutional network consists of two parts: a dilated causal convolutional layer and a residual link, as Figure 1 shown. The dilated causal convolutional layer obtains a larger receptive field according to different dilation factors for the convolutional kernel, obtains longer dependencies at the output end, and reduces the number of network layers; the residual module enables the network to transmit information in a cross-layer manner, which solves the problem of gradient disappearance in deep networks.

[0009] The technical solution adopted in the present invention is a road base strain prediction and analysis method based on a temporal convolutional network, which includes four major parts: preprocessing of original monitoring data, feature correlation analysis, dataset production, and temporal convolutional network-based road base strain prediction, as Figure 2 shown. The specific steps are as follows:

[0010] Step 1: Preprocessing of original monitoring data;

[0011] First, integrate the road monitoring data obtained by various sensors.

[0012] Secondly, according to the time span of the data obtained by the road monitoring system, select the data in the appropriate time period, then delete the duplicate values and jump values of the data, and complete the filling of the missing values to prevent poor training effects of the monitoring data.

[0013] Finally, according to the characteristics of the time series data, unify the time interval between all data points to 1 hour.

[0014] Step 2: Feature correlation analysis;

[0015] First step, perform correlation analysis on the first three data points obtained by each sensor after processing, and measure the correlation between features according to the Pearson correlation coefficient. Second step, remove the soil layer strain due to its low correlation.

[0016] Step 3: Dataset production;

[0017] For all the monitored data points of asphalt strain, osmotic pressure, soil moisture, and subbase triaxial strain, take the average value. Then, according to the burial positions of the temperature sensors, the temperature will be divided into three parts: surface layer, subbase, and soil subgrade, and the average value will be taken for each part. Finally, divide the monitored data set into a training set and a test set according to a ratio of approximately 8:2.

[0018] Step 4: Prediction of subbase strain by the temporal convolutional network;

[0019] The temporal convolutional network consists of 13 one-dimensional convolutional layers. The convolutional kernel size of each one-dimensional convolutional layer is 2, and the numbers are 64 respectively. The dilation factors are 1, 2, 4, 8, 16, and 32. Each layer uses the ReLU activation function and is followed by a SpatialDropout1D layer with a decay rate set to 0.05. Use the preprocessed and dimension-reduced data set as the input of the temporal convolutional network, and set the time step to 8.

[0020] The present invention can utilize the pavement sensor monitoring data obtained by the automated road monitoring system for a long time, analyze the long-term and multi-dimensional road monitoring data through deep learning methods, so as to predict and analyze the subbase strain of the road. At the same time, transfer the learning weights after training in each time period to the training in the next time period, improve the accuracy and stability of the prediction model, and reduce overfitting. In addition, using the present invention can effectively reduce labor costs and time losses, and the processed data can lay a foundation for subsequent monitoring data analysis, mining, and training of the prediction model. Brief Description of the Drawings

[0021] Figure 1 It is a schematic diagram of the temporal convolutional network structure

[0022] Figure 2 It is a diagram of the method implementation steps

[0023] Figure 3 It is a heat map of correlation analysis

[0024] Figure 4 It is a diagram of the input data set description

[0025] Figure 5 It is a diagram of the prediction effect of the temporal convolutional network Detailed Embodiment

[0026] The original pavement sensor monitoring data set adopted by the present invention is the data obtained by the automated road monitoring system. The specific implementation steps are as follows:

[0027] (1) Preprocessing of the original monitoring data

[0028] Due to systematic error problems such as data loss during monitoring data transmission and temporary failures of the acquisition devices, issues such as duplication, missing values, and jumps may occur in the original data, which have a great impact on the network prediction accuracy. Therefore, duplicate values and jump values in the original data obtained by the sensors are deleted, and missing values are supplemented. To ensure the time continuity of the monitoring data, we selected data from nine appropriate time periods and unified the time interval to one hour, which will be more in line with the characteristics of time series data and facilitate model training.

[0029] (2) Feature correlation analysis

[0030] The data monitored by the three-way strain gauge is the base strain value, which is the measured value of the predicted target data. All the data obtained by this monitoring system includes many features. It is extremely crucial to explore the relationships between these features and the correlation degree with the predicted target. The initial idea of this invention is to predict the base strain through factors such as asphalt layer strain, earth pressure, seepage pressure, water content, temperature, and soil base strain. If some feature quantities can be correctly reduced, it will help the model perform a large number of operations, such as improving accuracy and reducing the risk of overfitting. Therefore, it is necessary to conduct a correlation analysis on these data features. First, since there are many monitoring positions for each sensor, resulting in a large data dimension, only the data of the first three monitoring points obtained by each sensor are selected for correlation analysis; then, the Pearson correlation coefficient is used to measure the correlation between features, and a correlation analysis heat map is obtained, as Figure 3 shown. The closer the value of the Pearson correlation coefficient is to 1 or -1, the stronger the positive or negative correlation. Finally, according to the analysis results, the correlation between soil layer strain and other features is very small, and it is removed from the input end.

[0031] (3) Dataset production

[0032] To meet the needs of supervised learning, the road surface monitoring dataset is divided according to a ratio of approximately 8:2 for the training set and the test set. The description of the input dataset is as Figure 4 shown. During the training process, in order to monitor the training effect of the model, 20% is divided from the training set as the validation set.

[0033] (5) Time series convolutional network base strain prediction analysis

[0034] The temporal convolutional network adopted in the present invention utilizes its dilated causal convolution structure to expand the receptive field of the convolutional kernel, thereby capturing the long-term dependence relationship of historical data and learning the mapping relationship between features; and utilizes its residual structure to add an identity mapping of cross-layer connections to the deep network, solving the problem of gradient disappearance in the training of deep networks. The invention replaces the traditional causal convolution layer with a causal convolution layer with dilation properties, and connects the deep network with residual modules, thereby better retaining the long-term dependence relationship and historical feature information, improving the model training effect, and enhancing the robustness and accuracy of the prediction model.

[0035] The temporal convolutional network consists of dilated and causal 1D convolutional layers with the same input and output lengths, with a total of 13 layers. The size of the one-dimensional convolutional kernel is 2 for all, and the number is 64 respectively. Each layer uses the ReLU activation function, and a cross-layer connection is connected after two one-dimensional convolutional layers to implement a residual module. At this time, the convolutional kernel of each convolutional layer increases its receptive field according to different dilation factors to learn more distant input information and reduce the network complexity, and then the information is transmitted across layers in the residual module. Finally, the predicted information obtained in sequence is used as the output at the output end. The final prediction effect is as Figure 5 shown.

[0036] The ReLU function used in the present invention is expressed as follows:

[0037]

[0038] During the training process of the temporal convolutional network for road surface monitoring data, the input multi-dimensional data and the target data are jointly trained to continuously learn the data features and the physical relationship between the data, and the optimization objective is to minimize the error between the predicted value and the measured value. The Adadelta (Adaptive Learning Rate Method), an adaptive learning rate adjustment algorithm, is used as the optimization algorithm for gradient descent in the backpropagation process. The advantage of this algorithm is that it can adaptively adjust the learning rate during gradient descent without manual setting.

[0039] For the learning method, the Adam method is used to optimize the model parameters. The Adam method is a simple and computationally efficient stochastic objective function gradient optimization algorithm. This method has two advantages in dealing with sparse gradients and non-stationary targets. Since Adam can be well applied to a wide range of non-convex optimization problems, it is used in the present invention.

[0040] Adam maintains the exponentially decaying trend of the past average squared gradient vt. It also has an exponentially decaying trend of the average value of the past gradient mt and has a preference for a flat minimum on the error surface. Then, the past decaying average value and the past squared gradient m t and v tCorrespondingly as follows:

[0041] m t = β1m t-1 +(1 - β1)g t (2)

[0042]

[0043] where m t and v t are estimates of the first moment (mean) and the second moment (uncentered variance) of the gradient, respectively. This algorithm maintains a single learning rate for stochastic gradient descent of multi-dimensional data and updates all weights in the temporal convolutional network.

[0044] Since m t and v t are vectors initialized to 0, they tend to 0, and these tendencies can be calculated as:

[0045]

[0046]

[0047] Then these t are used and the parameters are updated as:

[0048]

[0049] The default value of β1 is 0.9, the default value of β2 is 0.999, and the default value of ∈ is 10 -8 . Each epoch is the whole process of a neural network training once through the entire dataset, including forward and backward. In the present invention, the learning rate is 0.002.

[0050] It has been observed that as the number of iterations increases, the loss value continuously decreases. After training for 100 generations, the predicted values of the temporal convolutional network used can be well fitted to the measured values.

Claims

1. A method for analyzing the strain of road base based on temporal convolutional network, characterized in that: Use an automated road monitoring system to obtain historical road surface monitoring data, and combine deep learning methods to predict key data. The specific steps are as follows: Step 1: Preprocess the original monitoring data; First, integrate the road monitoring data obtained by various sensors; Second, according to the time span of the data obtained by the road monitoring system, select the data for a time period, then delete the duplicate values and jump values of the data, and complete the filling of missing values to prevent poor training effects of the monitoring data; Finally, unify the time interval between data points of all time series data to 1 hour; Step 2: Feature correlation analysis; In the first step, perform correlation analysis on the first three data points obtained by each sensor after processing, and measure the correlation between features according to the Pearson correlation coefficient; in the second step, remove those with very little correlation between soil layer strain and other features; Step 3: Dataset production; Take the average value of all monitoring data points of asphalt strain, osmotic pressure, soil moisture, and base triaxial strain. Then, according to the burial position of the temperature sensor, the temperature is divided into three parts: surface layer, base layer, and subgrade for taking the average value; finally, divide the monitoring dataset according to the ratio of the training set to the test set of 8:2; Step 4: Prediction of base layer strain by temporal convolutional network; The temporal convolutional network consists of 13 one-dimensional convolutional layers; the convolutional kernel size of the one-dimensional convolutional layers is 2, the number of convolutional layers is 64 respectively, and the dilation factors are 1, 2, 4, 8, 16, and 32; each layer uses the ReLU activation function and is followed by a SpatialDropout1D layer with a decay rate set to 0.05; use the preprocessed and dimension-reduced dataset as the input of the temporal convolutional network, and set the time step to 8; The Adam method is used to optimize the model parameters. Adam maintains an exponentially decaying estimate of the second moment of past gradients without the centered variance, denoted as v t with an exponentially decaying trend; it also has an exponentially decaying trend for the first moment mean of the gradients, denoted as m t , and has a preference for a flat minimum on the error surface; then, the first moment mean of the gradients and the estimate of the second moment without the centered variance, m t and v t are calculated as follows: m t = β1m t-1 + (1 - β1)g t (2) where m t and v t are the estimated values of the first moment mean and the second moment non-centered variance of the gradient, respectively; the Adam method maintains a single learning rate for the stochastic gradient descent of multi-dimensional data and updates all weights in the temporal convolutional network; Since m t and v t are vectors initialized to 0, they are biased towards 0, and these biases are calculated as: Then use t and update the parameters as follows: β1 is 0.9, β2 is 0.999, and ∈ is 10 -8 ; Each epoch is the entire process of a neural network training once through the entire dataset, including forward and backward; As the number of iterations increases, the loss value is continuously decreasing. After training for 100 epochs, the predicted values of the temporal convolutional network used are fitted with the measured values.

2. The method for analyzing the strain of road base based on temporal convolutional network according to claim 1, characterized in that, During the training process of the temporal convolutional network for road surface monitoring data, jointly train the input multi-dimensional data and the target data, learn the data features and the physical relationships between the data, and the optimization goal is to minimize the error between the predicted value and the measured value; use the Adadelta algorithm with an adaptive learning rate as the optimization algorithm for gradient descent in the backpropagation process.

3. The method for analyzing the strain of road base based on temporal convolutional network according to claim 1, characterized in that, Make the following improvements to the original model using the temporal convolutional network: (1) Replace causal convolution with dilated causal convolution; (2) Add an operation of residual connection to prevent the disappearance of gradients during the learning and training process; (3) Follow a SpatialDropout1D layer after two one-dimensional convolutional layers; (4) Use the ReLU activation function for all layers except the last output layer which uses the Linear activation function.

Images