CN115374903A - Long-term pavement monitoring data enhancement method based on expressway sensor network layout - Google Patents
Long-term pavement monitoring data enhancement method based on expressway sensor network layout Download PDFInfo
- Publication number
- CN115374903A CN115374903A CN202210693861.7A CN202210693861A CN115374903A CN 115374903 A CN115374903 A CN 115374903A CN 202210693861 A CN202210693861 A CN 202210693861A CN 115374903 A CN115374903 A CN 115374903A
- Authority
- CN
- China
- Prior art keywords
- data
- network
- monitoring data
- sequence
- model
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000012544 monitoring process Methods 0.000 title claims abstract description 63
- 238000000034 method Methods 0.000 title claims abstract description 48
- 230000007774 longterm Effects 0.000 title claims abstract description 22
- 238000012549 training Methods 0.000 claims abstract description 35
- 238000009499 grossing Methods 0.000 claims abstract description 17
- 238000013136 deep learning model Methods 0.000 claims abstract description 10
- 238000004458 analytical method Methods 0.000 claims abstract description 5
- 230000006870 function Effects 0.000 claims description 32
- 239000013598 vector Substances 0.000 claims description 14
- 238000012360 testing method Methods 0.000 claims description 13
- 230000008569 process Effects 0.000 claims description 12
- 230000003068 static effect Effects 0.000 claims description 12
- 239000002689 soil Substances 0.000 claims description 11
- 238000005457 optimization Methods 0.000 claims description 10
- 238000009826 distribution Methods 0.000 claims description 8
- 238000001914 filtration Methods 0.000 claims description 8
- 238000013507 mapping Methods 0.000 claims description 8
- 238000011084 recovery Methods 0.000 claims description 8
- 238000004422 calculation algorithm Methods 0.000 claims description 7
- 230000002123 temporal effect Effects 0.000 claims description 7
- 238000010219 correlation analysis Methods 0.000 claims description 6
- 238000013459 approach Methods 0.000 claims description 4
- 238000012545 processing Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 230000006872 improvement Effects 0.000 claims description 3
- 230000001360 synchronised effect Effects 0.000 claims description 3
- 230000003042 antagnostic effect Effects 0.000 claims description 2
- 238000013135 deep learning Methods 0.000 claims description 2
- 230000007246 mechanism Effects 0.000 claims description 2
- 238000007781 pre-processing Methods 0.000 claims description 2
- 230000002708 enhancing effect Effects 0.000 claims 2
- 238000012163 sequencing technique Methods 0.000 claims 1
- 238000005516 engineering process Methods 0.000 abstract description 6
- 238000013528 artificial neural network Methods 0.000 abstract description 3
- 238000012423 maintenance Methods 0.000 abstract 1
- 238000004088 simulation Methods 0.000 abstract 1
- 230000000694 effects Effects 0.000 description 5
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 4
- 230000001364 causal effect Effects 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000008901 benefit Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000007637 random forest analysis Methods 0.000 description 3
- 230000003044 adaptive effect Effects 0.000 description 2
- 239000010426 asphalt Substances 0.000 description 2
- 230000005540 biological transmission Effects 0.000 description 2
- 238000004364 calculation method Methods 0.000 description 2
- 239000002131 composite material Substances 0.000 description 2
- 238000010276 construction Methods 0.000 description 2
- 230000004927 fusion Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 230000003204 osmotic effect Effects 0.000 description 2
- 206010063385 Intellectualisation Diseases 0.000 description 1
- 230000003321 amplification Effects 0.000 description 1
- 238000013473 artificial intelligence Methods 0.000 description 1
- 230000006399 behavior Effects 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000002457 bidirectional effect Effects 0.000 description 1
- 230000000295 complement effect Effects 0.000 description 1
- 230000001808 coupling effect Effects 0.000 description 1
- 238000005336 cracking Methods 0.000 description 1
- 238000007405 data analysis Methods 0.000 description 1
- 238000013480 data collection Methods 0.000 description 1
- 230000007423 decrease Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 230000036541 health Effects 0.000 description 1
- 238000012804 iterative process Methods 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000005259 measurement Methods 0.000 description 1
- 230000006855 networking Effects 0.000 description 1
- 238000003199 nucleic acid amplification method Methods 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000002441 reversible effect Effects 0.000 description 1
- 238000012546 transfer Methods 0.000 description 1
- 238000013526 transfer learning Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/08—Construction
Landscapes
- Engineering & Computer Science (AREA)
- Business, Economics & Management (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Human Resources & Organizations (AREA)
- Economics (AREA)
- Strategic Management (AREA)
- General Physics & Mathematics (AREA)
- General Health & Medical Sciences (AREA)
- Marketing (AREA)
- Tourism & Hospitality (AREA)
- General Business, Economics & Management (AREA)
- Health & Medical Sciences (AREA)
- Biomedical Technology (AREA)
- Computing Systems (AREA)
- Artificial Intelligence (AREA)
- Primary Health Care (AREA)
- Biophysics (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Molecular Biology (AREA)
- Life Sciences & Earth Sciences (AREA)
- General Engineering & Computer Science (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Development Economics (AREA)
- Game Theory and Decision Science (AREA)
- Entrepreneurship & Innovation (AREA)
- Operations Research (AREA)
- Quality & Reliability (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a long-term pavement monitoring data enhancement method based on the arrangement of a highway sensor network, which adopts a computer technology to perform data completion and data smoothing on original monitoring data acquired by the highway sensor network; deep data enhancement is carried out on the preprocessed monitoring data based on a time sequence countermeasure generation network, and key information in the monitoring data is learned, so that more high-quality simulation data are synthesized, the physical characteristics and time dynamics of time sequence data samples are relatively obvious, the prediction precision and generalization capability of the model are improved, and overfitting is reduced; and finally, training the synthesized monitoring large data set based on a deep learning model time sequence convolution neural network and a sequence-to-sequence model, and transferring the training weight to the model to obtain a pre-training model so as to realize the road base strain prediction based on the original data set. In addition, the method can carry out accurate strain analysis of the road base layer and lay a foundation for subsequent road maintenance work.
Description
Technical Field
The invention belongs to the field of time sequence data analysis, and relates to a long-term pavement monitoring data enhancement method based on expressway sensor network layout. The method is applied to data enhancement of the pavement structure monitoring data acquired by the long-term monitoring system integrated with the highway multisensor.
Background
As the construction of a large number of road infrastructures is gradually completed, the structures are susceptible to various road defects, such as rutting, fatigue cracking, and the like, which are greatly related to the elasticity and permanent deformation of the base layer and the underlayer under the coupling action of repeated traffic loads and the external environment. Thanks to the high-speed development of the computer technology and the sensor technology, the expressway gradually develops towards high-order digitalization, networking and intellectualization. The intelligent highway is a new form of highway in the big data era and is an extension and innovation of the traditional highway system. Therefore, a model of selecting the type of the multifunctional scene sensor and a model of a sensor network layout method under different intelligent high speeds are constructed based on design to guide the layout design and the construction of the sensor network, so that the advanced artificial intelligence technology is applied to analyze and predict the mechanical behavior of the road by utilizing massive monitoring big data, the method is vital to the road health monitoring and the structural design, and simultaneously, the method is also beneficial to developing a complete road analysis and management system in the future to prolong the service life of the expressway to the maximum extent.
Compared with the numerical analysis of the traditional mechanics theory, the advanced machine learning technology has the advantages of high efficiency, low calculation amount and the like. However, these intelligent calculation methods have high requirements on the structure and the number of long-term monitoring data. Through iterative computations, they are constantly learned from the experience of the data, so deep learning models may not learn more important features if there is not a sufficient amount of time series data. According to the existing long-term monitoring data obtained based on expressway sensing net equipment, due to the structural fracture problem of the time sequence data, the input sparse data can possibly cause the problems of poor prediction performance of a model, sensor layout optimization, total cost and the like, the data depth enhancement technology can reduce the workload of a monitoring system, can bring considerable performance improvement to a deep learning model, improves the accuracy and the stability, and reduces overfitting.
Therefore, the invention provides a long-term pavement monitoring data enhancement method based on the arrangement of the expressway sensor network. Firstly, preprocessing raw data of sensor data collected by a highway monitoring system, wherein the raw data comprises data completion and data smoothing; and finally, generating a network based on time sequence counterwork to perform data amplification so as to expand the number of samples and lay a foundation for the subsequent pavement base strain prediction work of the expressway.
Disclosure of Invention
The invention aims to deeply enhance the pavement monitoring data which are acquired by a sensor and are as long as eight years from 2012 to 2020 by a long-term pavement monitoring data enhancement method based on the arrangement of a highway sensor network, thereby effectively improving the accuracy of predicting the base layer strain by a deep learning model. The long-term pavement monitoring data comprises asphalt strain, embedded three-dimensional strain, soil layer strain, soil pressure, temperature, osmotic pressure and soil moisture.
1. Sequential countermeasure generation network
The time sequence countermeasure generating network model adopted by the invention uses the gated recurrent neural network to replace the multi-layer perceptron structure of a discriminator and a generator in the traditional countermeasure generating network, and the following improvements are made on the basis of the original countermeasure generating network model: (1) The introduction of the embedding and restoring functions provides a mapping between features and underlying spaces, allowing the antagonistic network to characterize the underlying temporal dynamics of the learning data through a low dimension; (2) The countermeasure network operates in the potential space provided by the embedded network, with the potential temporal dynamics of the real and synthetic data synchronized by supervised losses; (3) The autoencoder and the countermeasure network are jointly trained based on three different loss functions. The Reconstruction Loss is used for the optimization of the self-encoder parameters, the Unsupervised Loss is used for the optimization of the countermeasure network parameters, and the Supervised Loss is directed to the learning of the "temporal dynamics" by the generator.
The countermeasure generation network adopted by the invention comprises four parts of structures, namely an embedding function, a recovery function, a sequence generator and a sequence discriminator. The four parts of components realize co-training, and through continuous iteration and updating, the time sequence countermeasure generation network simultaneously learns the coding characteristics, generates the expression and iterates along with time, so that the generator has good capability of generating monitoring data, and the discriminator has good capability of discriminating the true and false data, as shown in fig. 1.
The technical scheme adopted by the invention is a long-term pavement monitoring data enhancement method based on expressway sensor network layout, which comprises four major parts, namely completion and smoothing of monitoring data, characteristic correlation analysis based on a mutual information method, time sequence confrontation generation network data enhancement and road base layer strain prediction based on a deep learning model, and is shown in figure 2, and the method comprises the following specific steps:
the method comprises the following steps: completing and smoothing monitoring data acquired by a highway sensor network;
and (3) data completion:
firstly, dividing a data set subjected to preliminary processing and obtained by monitoring a highway sensor network into a training set and a testing set, wherein the training set does not contain missing values, and the testing set contains missing values;
secondly, setting the characteristics including the missing values in the training set as target values Y, and setting the other characteristics as X, and starting training;
and finally, performing missing value prediction on the test set by using the trained model, and filling the predicted value into the original data set, so that the predicted value can be ensured to be closer to the distribution of real data.
Data smoothing:
first, savitzky-Golay (SG) filtering is implemented using the scipy.
Secondly, after a plurality of times of operation tests, two important parameter polynomial orders of smoothing filtering and sliding windows are set to be 3 and 11.
And finally, smoothing all the monitoring data by a set filtering algorithm.
Step two: analyzing the characteristic correlation of the monitoring data of the highway sensor network based on a mutual information method;
in the first step, mutual information quantity between the multidimensional characteristic and the target prediction characteristic is calculated by using a mutual information method, so that all characteristics are sequenced. And secondly, removing the soil layer strain due to small mutual information amount of the soil layer strain.
Step three: the method comprises the steps of performing long-term pavement monitoring data enhancement on the basis of a time sequence countermeasure generation network;
the time sequence confrontation generation network is composed of four parts, namely an embedding function, a recovery function, a sequence generator and a sequence discriminator. The embedding function is used as a loop structure to respectively convert two dynamic and static characteristics of characteristic spaces S and X into potential characteristic codes, and the recovery function is used for returning the potential codes to characteristic representations of the potential codes. The generator randomly takes 100-dimensional static and dynamic random vectors as input, and first synthesizes the random vectors into potential codes. Dynamic and static encoding of the input arbiter then achieves a two-classification distinction of composite or true. Three Loss functions are introduced in the four parts for carrying out combined training to standardize the learning process, the repetition Loss represents the mastery degree of an internal mode of the self-encoder on input data, the Unsupervised Loss represents the game situation of a sequence generator and a sequence discriminator, and the time sequence data generated by the super Loss generator can approach the data of the real time sequence data after the self-encoding coding.
Step four: predicting the strain of a road base layer based on a deep learning model;
the time-series convolutional network is composed of 13 layers of one-dimensional convolutional layers. The convolution kernel sizes of the one-dimensional convolution layers are all 2, the number of the convolution kernels is 64, and the expansion factors are 1, 2, 4, 8, 16 and 32. Each layer was followed by a SpatialDropout1D layer using the ReLU activation function, with the decay rate set to 0.05. And taking the preprocessed and dimensionality-reduced data set as an input of the time sequence convolution network, and setting the time step length to be 7.
The encoder structure of the sequence-to-sequence model is a bidirectional LSTM with 64 elements of the hidden layer, where an attention-driven layer is introduced, while the decoder structure is a LSTM with 32 elements of the hidden layer. A 64-unit bi-directional LSTM layer is connected after the encoder-decoder structure, and finally the prediction value is output through the fully-connected layer.
The invention can utilize the sensing network laid on the highway to obtain the monitoring data of the pavement sensor for a long time, and synthesize high-quality multidimensional monitoring data which is highly similar to the real data distribution by inputting the monitoring data into the time sequence confrontation generation network, so that the time dynamic characteristics of the monitoring data are more obvious, the number of samples is expanded, the accuracy and the stability of a time sequence prediction model are improved, and overfitting is reduced. In addition, the invention can combine the effective fusion of the enhanced data and the real data to train a more stable deep learning prediction model so as to effectively overcome the problem of poor prediction performance caused by the sparse monitoring data of the expressway sensor network, reduce labor cost and time loss, and lay a foundation for the effective training of the subsequent road base layer strain prediction model by the processed data.
Drawings
FIG. 1 is a diagram of the steps of a method.
Fig. 2 is a schematic diagram of partial monitoring data completion and smoothing effect.
Fig. 3 is a mutual information quantity histogram.
FIG. 4 is a diagram of a time-ordered pair generation network framework.
Fig. 5 is a data set description diagram.
FIG. 6 is a comparison graph of the base layer strain prediction results of the deep learning model before and after enhancement. a is a sequence-to-sequence model and b is a time-series convolution model.
Detailed Description
The original monitoring data set adopted by the invention is data obtained by a long-term monitoring system based on a highway sensing network. The specific implementation steps are as follows:
(1) Completing and smoothing monitoring data acquired by highway sensor network
Problems with sensors, transmission equipment and networks in the overall monitoring system, which are difficult to avoid in the overall system, may result in data loss during data collection. Inputting a data set containing missing values into a prediction model may cause a large prediction error, resulting in a low prediction accuracy. In addition, the raw sensor data collected from the deployed sensor network transmission to the system has significant non-smoothness, i.e., the resulting data curve is noisy, resulting in the risk of over-fitting during the prediction process, and less accurate predictions on the test set. Therefore, it is necessary to complement and smooth the raw monitoring data. For the completion of the monitoring data, because the missing value of the original monitoring data is of a continuous missing type, a random forest model is adopted for completion. Firstly, taking the data part which is not lost as a training set, and taking the lost data part as a test set; secondly, setting labels for the random forest model to perform supervised learning, setting the characteristics including missing values in the training set as target values Y, setting the other characteristics as X, and inputting the characteristics into the random forest model to perform training so that the model can rapidly learn the mapping relation between the characteristics and the prediction target; and finally, predicting missing values of the test set containing the missing values by using the well-trained model, and filling the missing values in corresponding positions, so that the missing values can be effectively filled by prediction based on a machine learning method and are closer to the distribution of original data. For the monitor data smoothing, we expect to denoise the original signal and preserve the original signal shape well, so the invention uses SG filtering to perform smoothing operation on the time series data obtained in each time segment. Firstly, calling a signal module in a scipy library by using a python language to realize SG filtering; secondly, setting the size of a parameter sliding window and the polynomial order of the filter function, and setting the size and the polynomial order to be 11 and 3 through multiple tests; and finally, smoothing all monitoring data of each time period by using a set filter, wherein the data smoothing effect is as shown in the figure.
(2) Monitoring data characteristic correlation analysis of highway sensor network
The method comprises the steps of acquiring long-term pavement monitoring data based on expressway sensing net equipment, wherein the road base layer strain is an actual measurement value of predicted target data. All data acquired by the sensors include many features, and it is crucial to explore the links between these features and the degree of correlation with the predicted target. The invention primarily contemplates predicting future base strain through historical data of several factors, namely, asphalt layer strain, base strain, soil pressure, osmotic pressure, water content, temperature, and soil layer strain, and if the number of features can be reduced correctly, it will help the model to perform a large number of operations, such as improving accuracy, reducing risk of overfitting, etc. Therefore, it is necessary to perform correlation analysis on these characteristic variables. Firstly, carrying out correlation analysis on an influence variable and a predictive variable in monitoring data; then, the correlation between the features is measured by using the mutual information quantity to obtain a correlation analysis histogram, as shown in fig. 3. The larger the mutual information value is, the more correlation between the influence characteristic and the predicted characteristic is. And finally, according to the analysis result, displaying that the correlation between the soil layer strain and the target predicted value is small, and removing the soil layer strain from the input end.
(3) Long-term pavement monitoring data enhancement based on time sequence countermeasure generation network
The time sequence countermeasure generation network model adopted by the invention is composed of four parts, namely an embedding function, a recovery function, a sequence generator and a sequence discriminator, the embedding function and the recovery function are introduced to provide mapping between features and potential space, the countermeasure network is allowed to represent potential time dynamics of learning data through low dimension, the countermeasure network operates in the potential space provided by the embedding network, and the potential time dynamics of real data and synthetic data are synchronized through supervision loss. The autoencoder and the countermeasure network are jointly trained based on three different loss functions. The Reconstruction Loss is used for optimizing self-encoder parameters, the Unsupervised Loss is used for resisting optimization of network parameters, and the Supervised Loss is used for enabling the 'temporal dynamics' of synthetic data to approach real data as much as possible.
The timing countermeasure generation network framework of the present invention is schematically illustrated in fig. 4. The embedding function is used as a loop structure to convert two dynamic and static characteristics of the characteristic space S and X into potential characteristic codes respectively, and the recovery function is used for returning the potential codes to the characteristic representation of the potential codes. The generator randomly takes 100-dimensional static and dynamic random vectors as input, and first synthesizes the random vectors into potential codes. Dynamic and static encoding of the input arbiter then achieves a two-classification distinction of composite or true. An embedded network is introduced to provide a reversible mapping between features (features) and latent representations (latent representations), which allows the countermeasure module to train the dimensions of the parameter space that needs to be learned, better capturing the underlying dynamic features of the data in this low-dimensional mode. Three Loss functions are introduced in the joint training of the four parts to standardize the learning process, the repetition Loss represents the mastery degree of the self-encoder on the internal mode of the input data, the Unsupervised Loss represents the game condition of the sequence generator and the sequence discriminator, and the time sequence data generated by the Supervised Loss generator can approach the data of the real time sequence data after self-encoding to the extent. Compared with ordinary static data, the monitoring data not only has characteristic distribution on each time stamp, but also contains potential complex relation among variables at different time points. The timing countermeasure generation network can therefore better exploit the powerful capability of generating timing data. The sequence generator receives random noise in Gaussian distribution, the purpose of the random noise is to enable the distribution to be better similar to the distribution of real data, and because the cycle structure of the generator can better capture the dynamic characteristic information in a potential space, better generated static and dynamic codes are continuously output to the discriminator. In the whole game training process, the capability of distinguishing real data and false data of the discriminator is continuously improved, and the capability is output by a probability value. Finally, the generator can produce more realistic synthetic data until the discriminator is hard to distinguish.
The three loss functions in the time sequence countermeasure generation network applied by the invention are expressed as follows:
the Sigmoid and Tanh functions in the GRU unit used in the present invention are expressed as follows:
for the learning method, an Adam method is used for optimizing the model parameters, and the Adam method is a simple and high-calculation-efficiency random objective function gradient optimization algorithm. The method has two advantages in the aspects of processing sparse gradients and processing non-stationary targets. Adam is used in the present invention because it can be well adapted to a wide range of non-convex optimization problems.
Adam keeps the past trend of the mean squared gradient vt decaying exponentially. It also has an average of past gradients mt with an exponential decay trend and a flat minimum preference in the error plane. Then, the past attenuation mean and the past square gradient m are calculated t And v t The corresponding is as follows:
m t ＝β 1 m t-1 +(1-β 1 )g t (6)
wherein m is t And v t Are estimates of the first moment (mean) and the second moment (no central variance) of the gradient, respectively. The algorithm keeps the random gradient decline of time series data to keep a single learning rate, and updates all weights in the countermeasure generation network.
Due to m t And v t Vectors initialized to 0, which are biased toward 0, can be calculated as:
these t are then used and the parameters are updated as:
β 1 default value is 0.9, beta 2 Default value of (2) is 0.999, and the default value of epsilon is 10 -8 . Each epoch is the entire process of neural network training through the entire data set, including forward and backward. The learning rate in the present invention was 0.0002.
Through observation, the characteristics of the monitoring data are gradually clear along with the increase of the iteration times, and after 15000 generations of training, the used time sequence confrontation generation network can generate more vivid monitoring data.
(4) Monitoring data set manufacturing method for expressway sensor network
To meet the requirements of supervised learning, the proportion of the data set to the training set and the test set is about 8: and 2, dividing. The data set description is shown in fig. 5. In order to monitor the training effect of the model during the training process, 20% of the training set is divided into verification sets.
(5) Road base layer strain prediction based on deep learning model
The time sequence convolution network adopted by the invention utilizes the expansion causal convolution structure to expand the receptive field of the convolution kernel, thereby capturing the long-term dependence of historical data and learning the mapping relation between characteristics; the method adds the deep network into the cross-layer connected identity mapping by utilizing the residual structure of the deep network, and solves the problem that the gradient disappears in deep network training. According to the method, the traditional causal convolutional layer is replaced by the causal convolutional layer with the expansion property, and the deep networks are connected by using the residual error module, so that the long-term dependence relationship and the historical characteristic information are well reserved, the model training effect is improved, and the robustness and the precision of the prediction model are improved. Another employed sequence-to-sequence model encoder uses bi-directional LSTM layers and converts the input vector into its corresponding hidden vector at the last time step. The context vector contains the compressed memory information of the entire input sequence, i.e., the key feature information extracted from the input sequence. The decoder will receive the concealment vector from the encoder and combine its own concealment state with the output vector at the previous time to generate the next concealment vector. The invention uses the idea of 'transfer learning' to realize the effective fusion of real data and enhanced data, namely two deep learning models continuously learn the key time information of a synthetic data set in the training iterative process, and then transfer the last updated training weight to the training process of small data sets in the training process. This ensures that the training of the prediction model on the small data set has prior knowledge, promotes the learning process, and improves the generalization and accuracy of the model. The enhanced prediction effect pair is shown in fig. 6.
The time-sequential convolutional network consists of extended, causal 1D convolutional layers with the same input and output lengths, for a total of 13 layers. The sizes of the one-dimensional convolution kernels are all 2, the number of the one-dimensional convolution kernels is 64, each layer uses a ReLU activation function, and two one-dimensional convolution layers are connected with each other in a cross-layer mode to achieve a residual error module. At the moment, the convolution kernel of each convolution layer increases the receptive field thereof according to different expansion factors to learn more distant input information, so as to reduce the complexity of the network, and then the information is transferred in a residual error module in a cross-layer mode. And finally, the prediction information obtained in sequence is taken as output at the output end.
The ReLU function used in the present invention is represented as follows:
in the training process of the time sequence convolution network on the road surface monitoring data, input multidimensional data and target data are trained together, the physical relation between data characteristics and data is continuously learned, and the optimization target is to minimize the error between a predicted value and an actually measured value. An Adaptive Learning Rate adjusting algorithm Adadelta (Adaptive Learning Rate Method) is taken as an optimization algorithm for gradient descent in the back propagation process, and the algorithm has the advantage of being capable of adaptively adjusting the Learning Rate during gradient descent without manual setting.
The sequence-to-sequence model comprises an encoder and a decoder. Wherein the encoder is made up of 64-element bi-directional LSTM of the hidden layer and the decoder is made up of 32-element LSTM of the hidden layer. At this time, after the input data passes through the encoder, key information in the data and mapping relation between input and output are extracted by the network, and then the input data is output to the encoder for decoding by the context vector, and an attention mechanism is introduced so that the model can selectively focus on a useful part of the input sequence. A unit number 64 bi-directional LSTM layer is accessed after the decoder to extract the sequence key information after decoding. And finally, acquiring the prediction information at the full connection layer, and taking the prediction information sequentially acquired as output at the output end.
Claims (3)
1. A long-term pavement monitoring data enhancement method based on expressway sensor network layout is characterized by comprising the following steps: the method comprises the following steps of utilizing road surface structure monitoring data obtained by an automatic long-term monitoring system based on the arrangement of a highway sensing network, and combining a deep learning method to carry out data enhancement and predictive analysis;
the method comprises the following specific steps:
the method comprises the following steps: completing and smoothing monitoring data acquired by the highway sensor network;
and (3) data completion:
firstly, dividing a data set obtained by monitoring a highway sensor network and subjected to preliminary processing into a training set and a test set, wherein the training set does not contain missing values, and the test set contains missing values;
secondly, setting the characteristics including the missing values in the training set as target values Y, and setting the other characteristics as X, and starting training;
finally, performing missing value prediction on the test set by using the trained model, and filling the predicted value into the original data set, thereby ensuring that the predicted value is closer to the distribution of real data;
data smoothing:
firstly, a signal module in python is used for realizing Savitzky-Golay filtering;
secondly, through multiple operation tests, two important parameter polynomial orders of smooth filtering and sliding windows are set to be 3 and 11;
finally, smoothing all the monitoring data of the highway sensor network by a set filtering algorithm;
step two: analyzing the characteristic correlation of the monitoring data of the highway sensor network based on a mutual information method;
step one, calculating mutual information quantity between the multidimensional characteristic and the target prediction characteristic by using a mutual information method, and sequencing all the characteristics; secondly, removing the soil layer strain due to small mutual information amount of the soil layer strain;
step three: carrying out long-term pavement monitoring data enhancement on the basis of a time sequence countermeasure generation network;
the time sequence confrontation generation network consists of four parts, namely an embedding function, a recovery function, a sequence generator and a sequence discriminator; the embedded function is used as a loop structure to respectively convert two dynamic and static characteristics of characteristic spaces S and X into potential characteristic codes, and the recovery function returns the potential codes to the characteristic representation of the potential codes; a generator randomly obtains 100-dimensional static and dynamic random vectors as input, and firstly, the random vectors are synthesized into potential codes; then the dynamic and static codes input into the discriminator will realize a two-classification discrimination whether the two-classification discrimination is synthetic or real; three Loss functions are introduced in the joint training of the four parts to standardize the learning process, the repetition Loss represents the mastery degree of the self-encoder on the internal mode of the input data, the Unsupervised Loss represents the game condition of a sequence generator and a sequence discriminator, and the time sequence data generated by the Supervised Loss generator can approach the data of the real time sequence data after the self-encoding;
step four: predicting the strain of a road base layer based on a deep learning model;
the time sequence convolution network is composed of 13 layers of one-dimensional convolution layers; the convolution kernel sizes of the one-dimensional convolution layers are all 2, the number of the convolution kernels is 64, and the expansion factors are 1, 2, 4, 8, 16 and 32; each layer uses a ReLU activation function, and is followed by a spatialDropout1D layer, and the attenuation rate is set to be 0.05; taking the preprocessed and dimensionality-reduced data set as the input of a time sequence convolution network, and setting the time step length to be 7;
the encoder structure of the sequence-to-sequence model is a 64-element bi-directional LSTM of the concealment layer, in which an attention mechanism layer is introduced, while the decoder structure is a 32-element LSTM of the concealment layer; a 64-unit bi-directional LSTM layer is connected after the encoder-decoder structure, and finally the prediction value is output through the fully-connected layer.
2. The method for enhancing the long-term pavement monitoring data based on the expressway sensor network layout as recited in claim 1, wherein on the basis of data preprocessing of an original data set, data completion and smoothing are included, firstly, data with low correlation are removed through correlation analysis to reduce dimensions, then the data set of each time period is input into a time sequence confrontation generation network, important static and dynamic features of the monitoring data can be better learned, and high-quality monitoring data can be generated.
3. The method for enhancing the long-term pavement monitoring data based on the expressway sensor network layout as recited in claim 1, wherein the time-series countermeasure generation network model is adopted to perform the following improvements on the original model: (1) The introduction of the embedding and restoring functions provides a mapping between features and underlying spaces, allowing the antagonistic network to characterize the underlying temporal dynamics of the learning data through a low dimension; (2) The countermeasure network operates in the potential space provided by the embedded network, with the potential temporal dynamics of the real and synthetic data synchronized by supervised losses; (3) Performing joint training on the self-encoder and the countermeasure network based on three different loss functions; the Reconstruction Loss is used for the optimization of the self-encoder parameters, the Unsupervised Loss is used for the optimization of the countermeasure network parameters, and the Supervised Loss is directed to the learning of the "temporal dynamics" by the generator.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210693861.7A CN115374903A (en) | 2022-06-19 | 2022-06-19 | Long-term pavement monitoring data enhancement method based on expressway sensor network layout |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210693861.7A CN115374903A (en) | 2022-06-19 | 2022-06-19 | Long-term pavement monitoring data enhancement method based on expressway sensor network layout |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115374903A true CN115374903A (en) | 2022-11-22 |
Family
ID=84062544
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210693861.7A Pending CN115374903A (en) | 2022-06-19 | 2022-06-19 | Long-term pavement monitoring data enhancement method based on expressway sensor network layout |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115374903A (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115984146A (en) * | 2023-03-16 | 2023-04-18 | 中国海洋大学 | Global consistency-based marine chlorophyll concentration image completion method and network |
| CN116362714A (en) * | 2023-02-21 | 2023-06-30 | 中国公路工程咨询集团有限公司 | Prediction method and device for pavement maintenance period |
| CN116756493A (en) * | 2023-08-15 | 2023-09-15 | 湖南湘江智慧科技股份有限公司 | Data management method for security and fire control finger collecting platform |
-
2022
- 2022-06-19 CN CN202210693861.7A patent/CN115374903A/en active Pending
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN116362714A (en) * | 2023-02-21 | 2023-06-30 | 中国公路工程咨询集团有限公司 | Prediction method and device for pavement maintenance period |
| CN116362714B (en) * | 2023-02-21 | 2023-11-03 | 中国公路工程咨询集团有限公司 | Prediction method and device for pavement maintenance period |
| CN115984146A (en) * | 2023-03-16 | 2023-04-18 | 中国海洋大学 | Global consistency-based marine chlorophyll concentration image completion method and network |
| CN116756493A (en) * | 2023-08-15 | 2023-09-15 | 湖南湘江智慧科技股份有限公司 | Data management method for security and fire control finger collecting platform |
| CN116756493B (en) * | 2023-08-15 | 2023-10-27 | 湖南湘江智慧科技股份有限公司 | Data management method for security and fire control finger collecting platform |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN115374903A (en) | Long-term pavement monitoring data enhancement method based on expressway sensor network layout | |
| Zheng et al. | Unsupervised representation learning with long-term dynamics for skeleton based action recognition | |
| CN109492822B (en) | Air pollutant concentration time-space domain correlation prediction method | |
| CN110135319A (en) | A kind of anomaly detection method and its system | |
| Sun et al. | Prediction of short-time rainfall based on deep learning | |
| CN112541572A (en) | Residual oil distribution prediction method based on convolutional encoder-decoder network | |
| CN111242377A (en) | Short-term wind speed prediction method integrating deep learning and data denoising | |
| CN113470316B (en) | Debris flow monitoring and early warning method based on self-coding single classification model | |
| CN109614896A (en) | A method of the video content semantic understanding based on recursive convolution neural network | |
| CN115185937A (en) | SA-GAN architecture-based time sequence anomaly detection method | |
| CN114898121A (en) | Concrete dam defect image description automatic generation method based on graph attention network | |
| CN116933046A (en) | Deep learning-based multi-mode health management scheme generation method and system | |
| CN116844041A (en) | Cultivated land extraction method based on bidirectional convolution time self-attention mechanism | |
| CN116596150A (en) | Event prediction method of transform Hoxwell process model based on multi-branch self-attention | |
| CN115032602A (en) | Radar target identification method based on multi-scale convolution capsule network | |
| CN112347531B (en) | Brittle marble Dan Sanwei crack propagation path prediction method and system | |
| CN116306780B (en) | Dynamic graph link generation method | |
| CN115660221B (en) | Oil and gas reservoir economic recoverable reserve assessment method and system based on hybrid neural network | |
| CN116364203A (en) | Water quality prediction method, system and device based on deep learning | |
| CN116148864A (en) | Radar echo extrapolation method based on DyConvGRU and Unet prediction refinement structure | |
| CN109934352A (en) | The automatic evolvement method of model of mind | |
| CN115510748A (en) | Landslide displacement prediction method based on variational modal decomposition and CNN-GRU | |
| CN111860158B (en) | Time sequence InSAR high coherence point extraction method fusing 1D-CNN and BiLSTM neural network | |
| CN113269702A (en) | Low-exposure vein image enhancement method based on cross-scale feature fusion | |
| CN117454762B (en) | Coal-penetrating tunnel face gas concentration prediction method of Markov-neural network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |