CN111724074A - Pavement lesion detection early warning method and system based on deep learning - Google Patents

Pavement lesion detection early warning method and system based on deep learning Download PDF

Info

Publication number
CN111724074A
CN111724074A CN202010582258.2A CN202010582258A CN111724074A CN 111724074 A CN111724074 A CN 111724074A CN 202010582258 A CN202010582258 A CN 202010582258A CN 111724074 A CN111724074 A CN 111724074A
Authority
CN
China
Prior art keywords
road surface
network
parameters
early warning
segmented
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.)
Granted
Application number
CN202010582258.2A
Other languages
Chinese (zh)
Other versions
CN111724074B (en
Inventor
凌贺飞
黄昌喜
张鹏锋
李青松
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Huazhong University of Science and Technology
Original Assignee
Huazhong University of Science and Technology
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Application filed by Huazhong University of Science and Technology filed Critical Huazhong University of Science and Technology
Priority to CN202010582258.2A priority Critical patent/CN111724074B/en
Publication of CN111724074A publication Critical patent/CN111724074A/en
Application granted granted Critical
Publication of CN111724074B publication Critical patent/CN111724074B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • 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
    • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
    • G06Q10/063—Operations research, analysis or management
    • G06Q10/0639—Performance analysis of employees; Performance analysis of enterprise or organisation operations
    • G06Q10/06393—Score-carding, benchmarking or key performance indicator [KPI] analysis
    • 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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
    • Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
    • Y02T—CLIMATE CHANGE MITIGATION TECHNOLOGIES RELATED TO TRANSPORTATION
    • Y02T10/00—Road transport of goods or passengers
    • Y02T10/10—Internal combustion engine [ICE] based vehicles
    • Y02T10/40—Engine management systems

Landscapes

  • Engineering & Computer Science (AREA)
  • Business, Economics & Management (AREA)
  • Human Resources & Organizations (AREA)
  • Theoretical Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • General Physics & Mathematics (AREA)
  • Software Systems (AREA)
  • Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Biophysics (AREA)
  • Mathematical Physics (AREA)
  • Biomedical Technology (AREA)
  • Artificial Intelligence (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Computational Linguistics (AREA)
  • Development Economics (AREA)
  • Educational Administration (AREA)
  • Economics (AREA)
  • Entrepreneurship & Innovation (AREA)
  • Strategic Management (AREA)
  • Game Theory and Decision Science (AREA)
  • Marketing (AREA)
  • Operations Research (AREA)
  • Quality & Reliability (AREA)
  • Tourism & Hospitality (AREA)
  • General Business, Economics & Management (AREA)
  • Image Analysis (AREA)
  • Image Processing (AREA)

Abstract

The invention provides a pavement lesion detection and early warning method based on deep learning, which comprises the following steps: determining parameters of different time periods of different sections of the road surface; respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter; determining an alarm threshold value based on the POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning. The method calculates the sample energy of the input pavement sample data based on the two-way deep neural network, and the higher the sample energy value is, the higher the possibility of early lesion and deterioration of the pavement section is; and in the screening alarm stage, a POT model and a delay method are used for two-step screening, so that the recall rate is improved as much as possible, and the precision is not greatly influenced.

Description

Pavement lesion detection early warning method and system based on deep learning
Technical Field
The invention belongs to the technical field of pavement lesion detection, and particularly relates to a pavement lesion detection early warning method and system based on deep learning.
Background
With the rapid development and transportation of roads in China, the traffic flow is increased year by year, the asphalt pavement often has the defects of rutting, cracking, loosening, slurry turning, water damage or settlement and the like, the driving safety and the public safety are seriously influenced, and the asphalt pavement needs to be detected in order to ensure the effective development of pavement management work. At the present stage, the road disease detection method in China mainly adopts point sampling detection and multifunctional road detection vehicle detection. Although the point sampling detection method represented by core drilling sampling is simple and easy, the defects of damage to the integrity of the pavement structure, poor representativeness of the detection result, overlong traffic sealing time and the like exist. The multifunctional road detection vehicle cannot detect a large number of hidden diseases in the road surface structure layer, such as reflection cracks, interlayer hollowing and uneven settlement of the road surface structure. The diseases have high concealment, and cannot be identified, positioned and measured by a conventional detection means.
The technology for detecting the pavement structure diseases at home and abroad mainly comprises core drilling sampling, an optical fiber technology, a sound wave and ultrasonic flaw detection technology, a CT tomography technology and a ground penetrating radar detection technology. From the accuracy of detection, the accuracy of the CT scanning technology is optimal, but the CT scanning equipment can only carry out small-range scanning, the detection comprehensiveness is not good, the human body is injured, the equipment cost is high, and the popularization of the technology is not facilitated. Compared with CT scanning technology, the ultrasonic detection coverage is wide, and the equipment is light and cheap. However, the accuracy of ultrasonic detection is poor, it is difficult to accurately reflect the morphological characteristics of the road surface diseases, and the detection result is greatly influenced by environmental conditions. The ground penetrating radar technology has the characteristics of real time, multiple dimensions and high precision in crack disease detection, and the crack positioning and crack depth calculation can meet the requirements of pavement detection and maintenance. However, no matter whether crack detection is hidden in the structural layer or road surface crack detection is performed by adopting a ground penetrating radar of a road surface coupling type, the radar is required to be completely attached to the ground in the detection process, and the detection efficiency is low.
Some conventional machine learning algorithms can be used for pavement damage recognition and early warning, including K-nearest neighbor (kNN) algorithms, Local anomaly Factor (LOF) algorithms, K-means algorithms based on clustering, Gaussian Mixture Models (GMMs), and the like, but with more and more pavement detection parameters, new challenges are brought to degradation early warning. The dimension span is large, from several dimensions to hundreds of dimensions are possible, and the relevance among the parameters is complex. The traditional anomaly detection method and the degradation early warning technology cannot well process high-dimensional data, and the problem of dimension disaster exists. The existing pavement disease detection methods mainly utilize a convolutional neural network to identify ground penetrating radar images and visible light images shot by a camera, so that the method is effective in identifying obvious diseases, the identification accuracy is improved greatly compared with that of the traditional machine learning method, and the early deterioration and trend early warning method is not enough. In fact, the early-stage deterioration early warning of asphalt pavement collapse is very important, major traffic accidents can be avoided, and public safety hazards can be reduced.
In recent years, deep learning has shown a strong ability in various fields, and application of deep learning to the field of abnormality detection has been studied. To overcome the problem of dimension disaster, the mainstream method is to use a self-coding network to compress the features and reduce the dimension into a low-dimensional space. There are two treatment methods after that: firstly, abnormity is directly screened according to reconstruction errors of a self-encoding network, and the self-encoding network is trained by using normal data, so that the reconstruction errors of the normal data are small, and the reconstruction errors of the abnormal data are often large; secondly, other abnormal detection methods such as cluster analysis and the like are used for the compressed features. These two types of methods also have their own disadvantages. The former only identifies abnormal data according to reconstruction errors, and the reconstruction errors of the abnormal data can be very small under certain conditions, so that a large number of abnormal false positives are easily caused; the latter problem is that the self-encoder does not know the cluster analysis or other anomaly detection tasks to be performed next when performing feature compression, and may miss some key information useful for anomaly detection.
Therefore, many recent studies, such as the DAGMM model, attempt to train the self-coding network and the clustering task at the same time, and achieve good results. However, the methods still have room for improvement, for example, some models do not consider the time sequence dependence of data, most models are general models, and research and optimization on the acquisition parameters of the asphalt pavement are not specially performed. In addition, since the probability of occurrence of an abnormality is often low, studies have been made to detect an abnormality by using a peak-Over-Threshold (POT) model on the basis of an extremum theory. However, the POT model is directed to one-dimensional data and cannot process high-dimensional data.
Disclosure of Invention
Aiming at the defects of the prior art, the invention aims to provide a pavement lesion detection early warning method and a pavement lesion detection early warning system based on deep learning, so as to solve the following technical problems: traditional machine learning methods, such as K-nearest neighbor algorithm, Gaussian mixture model, etc., are difficult to deal with the problem of dimensionality disaster, and often have poor performance when facing high-dimensional data; the existing deep learning method has respective disadvantages, such as only considering reconstruction errors, no cooperative work of a self-coding network and a clustering analysis task, no consideration of time sequence characteristics of data, and no suitable method for determining an alarm threshold.
In order to achieve the above object, in a first aspect, the present invention provides a road surface lesion detection and early warning method based on deep learning, including the following steps:
determining parameters of different time periods of different sections of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter;
determining an alarm threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
It will be appreciated that the parameters are normalized using the Z-Score method before being input into the two-way deep neural network model.
Specifically, the road surface structure parameters include: the road surface damage condition index, the road surface running quality index, the road surface rutting depth index, the road surface jumping index, the road surface abrasion index, the road surface anti-skid performance index, the road surface structural strength index, the road surface material parameter and the road grade parameter;
the technical parameters of the pavement comprise: the method comprises the following steps of (1) obtaining a highway technical condition index, a pavement service performance index, a roadbed technical condition index, a bridge and tunnel structure technical condition index and a facility technical index along a line;
the road surface environmental parameters include: road surface climate parameters and road surface humidity parameters.
Optionally, the two-way deep neural network model comprises a self-coding network and an estimation network;
the self-encoding network comprises an encoding network and a decoding network:
the coding network is constructed by using a multilayer recurrent neural network, and the output end of the multilayer recurrent neural network is connected with a layer of full-connection layer network; full connection layer output characteristic z of the coding networkc(ii) a Each layer of neural network in the multilayer cyclic neural network receives an input parameter of a segmented road surface at one moment;
the decoding network is constructed by a full-connection layer network, the number of layers is consistent with that of the coding network, and the number of hidden layer nodes corresponding to the full-connection layer of the decoding network is consistent with that of the coding network; the decoding network obtains corresponding reconstruction data based on the original data of the road surface parameters, and calculates Euclidean distance and cosine similarity between the reconstruction data and the original data; the Euclidean distance and cosine similarity are combined into a vector zr；
The characteristic zcSum vector zrIs combined into a stationThe output z of the self-encoding network;
the estimation network is constructed by a full connection layer, and a dropout layer and a softmax function are used; the estimation network is used for estimating the classification probability of the input features z under the Gaussian mixture model
Figure BDA0002553591780000041
And based on the distribution probability
Figure BDA0002553591780000042
And updating parameters of the Gaussian mixture model.
In particular, the classification probability is obtained
Figure BDA0002553591780000043
Then, the parameters of the gaussian mixture model need to be updated, and the formula is as follows:
Figure BDA0002553591780000044
Figure BDA0002553591780000045
Figure BDA0002553591780000051
using K to represent the number of Gaussian distributions, then K is the serial number of Gaussian distributions, N is the number of training samples, αkThe mixing coefficient of the kth Gaussian component is referred, and the sum of the mixing coefficients of all components is 1;
Figure BDA0002553591780000052
and
Figure BDA0002553591780000053
respectively the mean and covariance matrices of the samples in the kth gaussian component,
Figure BDA0002553591780000054
a summary indicating that the ith sample belongs to the kth Gaussian componentAnd (4) rate.
Optionally, the loss function J of the two-way deep neural network model is:
Figure BDA0002553591780000055
wherein, the left side of the equation is thetac，θd，θmParameters of an encoding network, a decoding network and an estimation network are respectively set; the first term on the right side of the equation is the reconstruction error of the self-coding network, the second term is the classification loss of the Gaussian mixture model, and the third term is a linear correlation penalty term of the characteristics; lambda [ alpha ]1,λ2The weight coefficients are respectively two loss terms; n represents the number of training samples, L () represents the square of the L2 norm, xiI sample, x, representing the inputi' denotes the reconstructed data of the i-th sample from the coding network, ZiRepresenting the incoming compressed network's feature data for the ith sample, E () representing the energy of the incoming feature,
Figure BDA0002553591780000056
represents the average pearson correlation coefficient between the dimensions of the feature,
Figure BDA0002553591780000057
representing the covariance matrix in a gaussian mixture model.
Alternatively,
Figure BDA0002553591780000058
optionally, e (z) in the classification loss is called sample energy, and the calculation formula is as follows:
Figure BDA0002553591780000059
in the above formula, z is the incoming feature data, K represents the number of gaussian distributions, and then K is the serial number of gaussian distributions;
Figure BDA00025535917800000510
and
Figure BDA00025535917800000511
respectively are a mean value and a covariance matrix corresponding to the kth Gaussian component, and exp () is an exponential function with a natural constant e as a base; the superscript T represents the matrix arrangement, π is the circumferential ratio, |, is the determinant operation.
Alternatively,
Figure BDA0002553591780000061
the specific formula is as follows:
Figure BDA0002553591780000062
in the above formula, K is the number of gaussian components, and fd is the dimension of the low-dimensional compression feature; the value in the absolute value function abs is the pearson correlation coefficient for the ith and jth attributes in the kth gaussian component.
Optionally, the alarm threshold zqThe calculation formula of (a) is as follows:
Figure BDA0002553591780000063
wherein ,
Figure BDA0002553591780000064
in order to be a parameter of the shape,
Figure BDA0002553591780000065
the two parameters are scale parameters which are generally obtained by calculation by using maximum likelihood estimation according to a set Y of excess quantities; n is a radical oftThe parameter q is the expectation of the probability of occurrence of anomalous data, for the number of peaks.
Optionally, the method further comprises the following steps:
performing first-pass alarm point screening by using the alarm threshold obtained by the POT model; when the alarm threshold value is calculated, parameters of the POT model need to be properly adjusted, so that higher recall rate is obtained;
performing second-time alarm point screening by using a delay method; and setting proper delay method parameters, and screening the alarm points obtained in the first time again to obtain the finally confirmed alarm points.
In a second aspect, the present invention provides a road surface lesion detection and early warning system based on deep learning, including:
the parameter determining unit is used for determining parameters of different time periods of different sections of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
the energy detection unit is used for respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter;
the threshold value alarming unit is used for determining an alarming threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
Optionally, the two-way deep neural network model comprises a self-coding network and an estimation network;
the self-encoding network comprises an encoding network and a decoding network:
the coding network is constructed by using a multilayer recurrent neural network, and the output end of the multilayer recurrent neural network is connected with a layer of full-connection layer network; full connection layer output characteristic z of the coding networkc(ii) a Each layer of neural network in the multilayer cyclic neural network receives an input parameter of a segmented road surface at one moment;
the decoding network is constructed by a full-connection layer network, the number of layers is consistent with that of the coding network, and the number of hidden layer nodes corresponding to the full-connection layer of the decoding network is consistent with that of the coding network; the decoding network obtains corresponding reconstruction data based on the original data of the road surface parameters, and calculates Euclidean distance and cosine similarity between the reconstruction data and the original data; the Euclidean distance and cosine similarity are combined into a vector zr；
The characteristic zcSum vector zrCombining into the output z of the self-encoding network;
the estimation network is constructed by a full connection layer, and a dropout layer and a softmax function are used; the estimation network is used for estimating the classification probability of the input features z under the Gaussian mixture model
Figure BDA0002553591780000071
And based on the distribution probability
Figure BDA0002553591780000072
And updating parameters of the Gaussian mixture model.
Generally, compared with the prior art, the above technical solution conceived by the present invention has the following beneficial effects:
the invention provides a pavement lesion detection and early warning method and system based on deep learning, which can extract characteristics of high-dimensional data consisting of a large number of pavement parameters based on a model of a self-coding network so as to avoid the problem of dimension disaster which is difficult to process by a traditional machine learning method.
The invention provides a method for training a self-coding network and cluster analysis in a combined manner, which is matched with an improved loss function, so that extracted features are more meaningful, the cluster analysis is more facilitated, and the defects that only a self-coding network is used for reconstructing errors, and the task of feature extraction and cluster analysis does not work cooperatively are overcome.
The invention uses the recurrent neural network in the coding part of the self-coding network and uses the full connection layer in the decoding part, thereby solving the problem that some algorithms ignore the data time sequence characteristics, and simultaneously reducing the time cost of network training and use under the condition of not influencing the decoding capability.
The invention utilizes the POT model and the delay method to carry out two-step alarm information screening, determines a proper alarm threshold value, effectively improves the recall rate and the precision of the alarm information and solves the problem that some algorithms do not have proper alarm strategies.
Drawings
FIG. 1 is a flow chart of a road surface lesion detection and early warning method based on deep learning according to the present invention;
FIG. 2 is a schematic diagram of a pavement lesion and deterioration warning process according to the present invention;
FIG. 3 is a diagram of a deep neural network model architecture in accordance with the present invention;
fig. 4 is a structural diagram of a road surface lesion detection and early warning system based on deep learning provided by the invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the present invention more apparent, the present invention is described in further detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.
The invention relates to a bituminous pavement lesion detection and deterioration early warning method based on deep learning, which is characterized by comprising the following steps: the two-way deep neural network consists of a heterogeneous self-coding network and an estimation network; the self-coding network is used for feature extraction, the coding network is composed of a multilayer recurrent neural network, and the decoding network is composed of a full-connection network; the estimation network predicts the classification probability of the compression characteristics based on the frame of a Gaussian mixture model; the loss function consists of three parts, namely reconstruction loss of the self-coding network, prediction loss of the estimation network and linear correlation loss of compression characteristics; the trained model can calculate the sample energy of input sample data, and the larger the sample energy value is, the higher the possibility of early lesion and deterioration of the section of road surface is; and in the screening alarm stage, a POT model and a delay method are used for two-step screening, so that the recall rate is improved as much as possible, and the precision is not greatly influenced.
Fig. 1 is a flowchart of a road surface lesion detection and early warning method based on deep learning, as shown in fig. 1, including the following steps:
s110, determining parameters of different time periods of different sections of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
s120, respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter;
s130, determining an alarm threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
In order to achieve the above purpose, the technical scheme provided by the invention is as shown in fig. 2, and the specific steps are as follows:
1. and (6) data acquisition. Parameters of different time periods of different sections of the road surface are collected through various means such as a multifunctional road detection vehicle and a ground penetrating radar, and proper normal state section data are screened out from the parameters to serve as training data.
Specifically, the parameters of the road surface include: pavement materials, road grade, climate, humidity, etc., such as: road technical status Index (MQI); road surface usability index (PQI); subgrade Condition Index (SCI); bridge and Tunnel construction state of the art Index (BCI); the Traffic-facility Condition Index (TCI); road surface damage Condition Index (PCI); road Quality Index (RQI); road Rut Depth Index (RDI); road skip Index (PBI); road surface wear Index (PWI); road surface Skid Resistance Index (SRI); pavement Structure Strength Index (PSSI).
In one example, the road surface parameters are as follows:
TABLE 1 road surface parameter table at a certain time for different road sections (pile numbers)
Figure BDA0002553591780000101
2. And (4) preprocessing data. The obtained original multidimensional data x is normalized by using a Z-Score method, so that all dimensional attributes are changed into standard normal distribution with the mean value of 0 and the variance of 1, and the formula is as follows:
Figure BDA0002553591780000102
in the above formula, x' is normalized data,
Figure BDA0002553591780000103
is the mean, σ, of the raw dataxIs the standard deviation of the raw data. The Z-Score standardization method can remove the dimension of each dimension attribute, so that the dimension has comparability and the convergence speed is accelerated.
3. And constructing a deep neural network model. The overall network structure is shown in fig. 3, and is composed of a self-coding network and an estimation network, which will be described in detail below.
Let x1,x2…xtThe road surface acquisition parameters at the time 1 and the time 2 … … t are respectively. Specifically, XtThe method refers to a set of values of each parameter at the time t, namely an array of 1 × M, wherein M represents the specific number of the parameters of the road surface. DbIndicating the value at the time point t of the b-th road surface parameter. (1. ltoreq. b. ltoreq.M):
Figure BDA0002553591780000104
(31) an encoding network portion of a self-encoding network. The partial Network is constructed by using 1 to multi-layer Recurrent Neural Network (RNN), and a full connection layer is connected after the partial Network is output, and the obtained output is the extracted characteristic zc. When data is input, a sliding window mode can be adopted, namely at the time T, the latest T +1 continuous data are selected to be input into the RNN; a form without a sliding window may also be used.
(32) A decoding network portion of the self-encoding network. The partial network uses a network structure different from that of the coding network and is constructed by full connection layers, but the number of layers is consistent with that of the coding network, and the layers are hiddenThe number of nodes corresponds to the coding network. The obtained reconstructed data is x't. Calculating Euclidean distance and cosine similarity between the reconstructed data and the original data, and combining the Euclidean distance and the cosine similarity into a vector zrThe calculation formula is as follows:
Figure BDA0002553591780000111
will be characteristic zcSum vector zrCombined into the output z of the self-encoding network.
(33) The network part is evaluated. The estimation network is used for estimating the classification probability of the input characteristic z under the framework of Gaussian mixed model
Figure BDA0002553591780000112
The network is constructed by using a full connection layer, a dropout layer can be used in the middle, and a softmax function is finally used because probability values need to be output.
After the classification probability is obtained, the parameters of the gaussian mixture model need to be updated, and the formula is as follows:
Figure BDA0002553591780000113
Figure BDA0002553591780000114
Figure BDA0002553591780000115
k is used for representing the number of the Gaussian distribution, and the K is the serial number of the Gaussian distribution; n is the number of training samples, xiFor the (i) th sample,
Figure BDA0002553591780000116
α as the probability that the ith sample belongs to the kth Gaussian componentkThe mixing coefficient of the kth Gaussian component is referred, and the sum of the mixing coefficients of all components is 1;
Figure BDA0002553591780000117
and
Figure BDA0002553591780000118
respectively, the mean and covariance matrices of the samples in the kth gaussian component.
(34) A loss function. The loss function is formulated as follows:
Figure BDA0002553591780000119
in the left side of the equation, θc，θd，θmParameters of an encoding network, a decoding network and an estimation network are respectively set; the first term on the right side of the equation is the reconstruction error of the self-coding network, the second term is the classification loss of the Gaussian mixture model, and the third term is a linear correlation penalty term of the characteristics; lambda [ alpha ]1,λ2The weighting coefficients of the two loss terms are respectively.
The first term in the loss function is the reconstruction error, using the L2 norm, i.e.:
Figure BDA0002553591780000121
the second term is the classification loss, wherein E (z) is also called sample energy, and the calculation formula is as follows:
Figure BDA0002553591780000122
in the above formula, z is the incoming feature data, K represents the number of gaussian distributions, and then K is the serial number of gaussian distributions;
Figure BDA0002553591780000123
and
Figure BDA0002553591780000124
respectively are a mean value and a covariance matrix corresponding to the kth Gaussian component, and exp () is an exponential function with a natural constant e as a base; pi is the circumferential ratio, |, is the determinant operation.
The third term is the extracted feature, and the average value of the pearson correlation coefficients among the attributes is as follows:
Figure BDA0002553591780000125
in order to simplify the calculation amount, the covariance matrix of the compression characteristic is not directly calculated, but the covariance matrix obtained in the estimation network is used (since the covariance matrix is a symmetric matrix, only the lower triangle or the upper triangle part of the covariance matrix needs to be used in the calculation). In the above formula, K is the number of gaussian components, and fd is the dimension of the low-dimensional compression feature; the value in the absolute value function (abs) is the pearson correlation coefficient for the ith and jth attributes in the kth gaussian component. When the compression feature dimension is 1, the correlation coefficient cannot be calculated, and the loss value is set to 0. By utilizing the loss term, the self-coding network can be forced to learn the features with smaller linear correlation and even linearly independent, and the redundancy caused by the linear correlation is reduced, so that the network can extract more meaningful features.
4. And (5) off-line training. And (3) using the data prepared in the step 2, training the deep neural network model established in the step 3 by using a back propagation algorithm, and accelerating model convergence by using methods such as an Adam optimizer and the like. The batch size should be set to be appropriately large during training so that the network can better estimate the parameters of the Gaussian mixture model during training. After the training is finished, all the training data are run together to update the parameters of the Gaussian mixture model, so that a better effect is obtained.
5. Calculating an alarm threshold z by using a POT modelq. The input of the POT model is the energy of all training data output by the deep neural network model.
The energy of N training data samples is sorted from small to large, the N × th level value is taken as the peak value threshold value t, if the sample energy X isiGreater than the peak threshold t, this is called a super-threshold or peak, Yi＝Xi-t is the excess. Finally, the alarm threshold zqThe calculation formula of (a) is as follows:
Figure BDA0002553591780000131
wherein ,
Figure BDA0002553591780000132
in order to be a parameter of the shape,
Figure BDA0002553591780000133
the two parameters are scale parameters, and are generally obtained by calculation by using maximum likelihood estimation according to an excess set Y; n is a radical oftThe quantity of peaks parameter q is the expectation of the probability of occurrence of anomalous data. If the sample energy is larger than the alarm threshold value zqThen the link is considered to be diseased or deteriorated.
6. The model is run online. And (3) acquiring the current acquisition parameters of the road surface by using the method same as the steps 1 and 2, preprocessing, then sending the parameters into the model trained in the step 4, operating the model and calculating to obtain the sample energy corresponding to each data sample.
7. And (5) screening alarm information. And (5) screening alarm information on the sample energy obtained in the step 6 by using a POT model and a delay method.
The specific steps for screening abnormal data are as follows:
(71) alarm threshold z obtained using POT modelqAnd performing the first-pass alarm point screening. When calculating the alarm threshold in step 5, the parameters of the POT model need to be properly adjusted, so that higher recall rate is obtained.
(72) The second screening pass was performed using the delayed method. And setting proper delay method parameters, and screening the alarm points obtained in the previous step again to obtain finally confirmed alarm information.
In a specific embodiment, the historical data collected from a certain asphalt pavement is 40-dimensional raw data, which is 20000 bars in total, and the sampling interval is 1 minute. The Z-Score method is used for normalizing the data so that all dimensional attributes of the data are in a standard normal distribution with the mean value of 0 and the variance of 1.
According to the method provided by the invention, a deep neural network model is established, and proper extracted feature dimension (namely the number of hidden layer nodes positioned in the middle of a self-coding network) and clustering number (namely the number of output layer nodes of an estimation network) are set, wherein the extracted feature dimension and the clustering number are respectively set to be 20 and 4. The number of nodes of the self-encoding network may be expressed as: 40-100-60-20-60-100-40, and the output of the self-coding network is 42-dimensional by adding two characteristics of Euclidean distance and cosine similarity; estimating the number of nodes of the network may be expressed as: 42-20-dropout-4, where "dropout" represents a dropout layer. The appropriate sliding window size, here set to 100, is set according to the periodic characteristics of the raw data.
Then off-line training is performed according to the present invention. The batch size was set to 1000, the initial learning rate was set to 0.0002, and the adom optimizer was used with the epoch set to 5000. When the dimensionality of the raw data is higher, convergence may be slower, a larger epoch may be set, or a larger initial learning rate may be set, and then smaller based on the loss.
After the model training is finished, the model can be used for on-line operation. And during online operation, the acquired real-time operation parameters need to be normalized by Z-Score, and then are input into the model, and the sample energy of each real-time data is obtained through calculation.
Finally, the appropriate POT model and delay method parameters are set, which need to be determined according to experience or test results, so that the recall rate is properly improved, but the accuracy cannot be too low, because the adjustment capability of the delay method is limited. Here, the parameter level is set to 0.95 and the parameter q is set to 0.001. Only the energy value of the sample data exceeds the anomaly threshold zqAn alarm signal is sent out.
Fig. 4 is a structural diagram of a road surface lesion detection and early warning system based on deep learning, as shown in fig. 4, including:
a parameter determining unit 410, configured to determine parameters of different time periods of different segments of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
the energy detection unit 420 is configured to input parameters of each segmented road surface in different time periods to a two-way deep neural network model obtained through pre-training, so as to detect sample energy corresponding to each segmented road surface parameter;
the threshold value alarming unit 430 is used for determining an alarming threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
Specifically, the functions of each unit in fig. 4 can be referred to the description in the foregoing method embodiment, and are not described herein again.
It will be understood by those skilled in the art that the foregoing is only a preferred embodiment of the present invention, and is not intended to limit the invention, and that any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the scope of the present invention.

Claims (10)

1. A pavement lesion detection early warning method based on deep learning is characterized by comprising the following steps:
determining parameters of different time periods of different sections of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter;
determining an alarm threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
2. The deep learning-based road surface lesion detection and early warning method as claimed in claim 1, wherein the two-way deep neural network model comprises a self-coding network and an estimation network;
the self-encoding network comprises an encoding network and a decoding network:
the coding network is constructed by using a multilayer recurrent neural network, and the output end of the multilayer recurrent neural network is connected with a layer of full-connection layer network; full connection layer output characteristic z of the coding networkc(ii) a Each layer of neural network in the multilayer cyclic neural network receives an input parameter of a segmented road surface at one moment;
the decoding network is constructed by a full-connection layer network, the number of layers is consistent with that of the coding network, and the number of hidden layer nodes corresponding to the full-connection layer of the decoding network is consistent with that of the coding network; the decoding network obtains corresponding reconstruction data based on the original data of the road surface parameters, and calculates Euclidean distance and cosine similarity between the reconstruction data and the original data; the Euclidean distance and cosine similarity are combined into a vector zr；
The characteristic zcSum vector zrCombining into the output z of the self-encoding network;
the estimation network is constructed by a full connection layer, and a dropout layer and a softmax function are used; the estimation network is used for estimating the classification probability of the input features z under the Gaussian mixture model
Figure FDA0002553591770000011
And based on the distribution probability
Figure FDA0002553591770000021
And updating parameters of the Gaussian mixture model.
3. The road surface lesion detection and early warning method based on deep learning of claim 1, wherein a loss function J of the two-way deep neural network model is as follows:
Figure FDA0002553591770000022
wherein, the left side of the equation is thetac，θd，θmRespectively encoding network, decodingA network, estimating parameters of the network; the first term on the right side of the equation is the reconstruction error of the self-coding network, the second term is the classification loss of the Gaussian mixture model, and the third term is a linear correlation penalty term of the characteristics; lambda [ alpha ]1,λ2The weight coefficients are respectively two loss terms; n represents the number of training samples, L () represents the square of the L2 norm, xiI sample, x, representing the inputi' denotes the reconstructed data of the i-th sample from the coding network, ZiRepresenting the incoming compressed network's feature data for the ith sample, E () representing the energy of the incoming feature,
Figure FDA0002553591770000023
represents the average pearson correlation coefficient between the dimensions of the feature,
Figure FDA0002553591770000024
representing the covariance matrix in a gaussian mixture model.
4. The road surface lesion detection and early warning method based on deep learning of claim 3,
Figure FDA0002553591770000025
5. the deep learning-based pavement lesion detection and early warning method according to claim 3, wherein E (z) in the classification loss is called sample energy, and the calculation formula is as follows:
Figure FDA0002553591770000026
in the above formula, z is the incoming feature data, K represents the number of gaussian distributions, and then K is the serial number of gaussian distributions;
Figure FDA0002553591770000027
and
Figure FDA0002553591770000028
respectively are a mean value and a covariance matrix corresponding to the kth Gaussian component, and exp () is an exponential function with a natural constant e as a base; the superscript T represents the matrix arrangement, π is the circumferential ratio, |, is the determinant operation.
6. The road surface lesion detection and early warning method based on deep learning of claim 3,
Figure FDA0002553591770000031
the specific formula is as follows:
Figure FDA0002553591770000032
in the above formula, K is the number of gaussian components, and fd is the dimension of the low-dimensional compression feature; the value in the absolute value function abs is the pearson correlation coefficient for the ith and jth attributes in the kth gaussian component.
7. The deep learning-based road surface lesion detection and early warning method as claimed in claim 3, wherein the warning threshold value z isqThe calculation formula of (a) is as follows:
Figure FDA0002553591770000033
wherein ,
Figure FDA0002553591770000034
in order to be a parameter of the shape,
Figure FDA0002553591770000035
the two parameters are scale parameters which are generally obtained by calculation by using maximum likelihood estimation according to a set Y of excess quantities; n is a radical oftThe parameter q is the expectation of the probability of occurrence of anomalous data, for the number of peaks.
8. The road surface lesion detection and early warning method based on deep learning of any one of claims 1 to 7, further comprising the following steps:
performing first-pass alarm point screening by using the alarm threshold obtained by the POT model; when the alarm threshold value is calculated, parameters of the POT model need to be properly adjusted, so that higher recall rate is obtained;
performing second-time alarm point screening by using a delay method; and setting proper delay method parameters, and screening the alarm points obtained in the first time again to obtain the finally confirmed alarm points.
9. The utility model provides a road surface pathological change detects early warning system based on degree of depth learning which characterized in that includes:
the parameter determining unit is used for determining parameters of different time periods of different sections of the road surface; the parameters comprise road surface structure parameters, road surface environment parameters and road surface technical parameters;
the energy detection unit is used for respectively inputting parameters of each segmented road surface in different time periods into a two-way deep neural network model obtained through pre-training so as to detect sample energy corresponding to each segmented road surface parameter;
the threshold value alarming unit is used for determining an alarming threshold value based on the threshold peak POT model and the sample energy corresponding to each segmented road surface parameter; and when the sample energy of a certain segmented road surface exceeds the alarm threshold value, the segmented road surface is considered to be pathological change, and the segmented road surface is set as an alarm point to carry out road surface pathological change early warning.
10. The deep learning based pavement pathology detection and early warning system of claim 9, wherein the two-way deep neural network model comprises a self-coding network and an estimation network;
the self-encoding network comprises an encoding network and a decoding network:
the coding network is constructed by using a multilayer recurrent neural network, and the output end of the multilayer recurrent neural network is connected with a layer of full-connection layer network; full connection layer transmission of the coding networkOut of feature zc(ii) a Each layer of neural network in the multilayer cyclic neural network receives an input parameter of a segmented road surface at one moment;
the decoding network is constructed by a full-connection layer network, the number of layers is consistent with that of the coding network, and the number of hidden layer nodes corresponding to the full-connection layer of the decoding network is consistent with that of the coding network; the decoding network obtains corresponding reconstruction data based on the original data of the road surface parameters, and calculates Euclidean distance and cosine similarity between the reconstruction data and the original data; the Euclidean distance and cosine similarity are combined into a vector zr；
The characteristic zcSum vector zrCombining into the output z of the self-encoding network;
the estimation network is constructed by a full connection layer, and a dropout layer and a softmax function are used; the estimation network is used for estimating the classification probability of the input features z under the Gaussian mixture model
Figure FDA0002553591770000041
And based on the distribution probability
Figure FDA0002553591770000042
And updating parameters of the Gaussian mixture model.
CN202010582258.2A 2020-06-23 2020-06-23 Pavement lesion detection early warning method and system based on deep learning Active CN111724074B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202010582258.2A CN111724074B (en) 2020-06-23 2020-06-23 Pavement lesion detection early warning method and system based on deep learning

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202010582258.2A CN111724074B (en) 2020-06-23 2020-06-23 Pavement lesion detection early warning method and system based on deep learning

Publications (2)

Publication Number Publication Date
CN111724074A true CN111724074A (en) 2020-09-29
CN111724074B CN111724074B (en) 2023-10-27

Family

ID=72568522

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202010582258.2A Active CN111724074B (en) 2020-06-23 2020-06-23 Pavement lesion detection early warning method and system based on deep learning

Country Status (1)

Country Link
CN (1) CN111724074B (en)

Cited By (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112329865A (en) * 2020-11-09 2021-02-05 平安科技（深圳）有限公司 Data anomaly identification method and device based on self-encoder and computer equipment
CN113591606A (en) * 2021-07-08 2021-11-02 武汉理工大学 Method and device for identifying hidden diseases of asphalt pavement, electronic equipment and storage medium
CN114169402A (en) * 2021-11-16 2022-03-11 哈尔滨工业大学 Ground penetrating radar underground cavity target automatic identification method based on gate control cycle unit
CN117351367A (en) * 2023-12-06 2024-01-05 鄄城县公路事业发展中心 Highway maintenance inspection method, medium and equipment

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109146161A (en) * 2018-08-07 2019-01-04 河海大学 Merge PM2.5 concentration prediction method of the stack from coding and support vector regression
CN110796497A (en) * 2019-10-31 2020-02-14 支付宝(杭州)信息技术有限公司 Method and device for detecting abnormal operation behaviors

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109146161A (en) * 2018-08-07 2019-01-04 河海大学 Merge PM2.5 concentration prediction method of the stack from coding and support vector regression
CN110796497A (en) * 2019-10-31 2020-02-14 支付宝(杭州)信息技术有限公司 Method and device for detecting abnormal operation behaviors

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
BO ZENG 等: "deep autoencoding Gaussian mixture model for unsupervised anomaly detdction", 《ICLR 2018 CONFERENCE》 *
谭敏生 等: "基于弹性网的深度去噪自编码器异常检测方法", 《计算机工程与设计》, pages 1516 - 1520 *

Cited By (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN112329865A (en) * 2020-11-09 2021-02-05 平安科技（深圳）有限公司 Data anomaly identification method and device based on self-encoder and computer equipment
WO2022095434A1 (en) * 2020-11-09 2022-05-12 平安科技（深圳）有限公司 Auto-encoder-based data anomaly identification method and apparatus and computer device
CN112329865B (en) * 2020-11-09 2023-09-08 平安科技（深圳）有限公司 Data anomaly identification method and device based on self-encoder and computer equipment
CN113591606A (en) * 2021-07-08 2021-11-02 武汉理工大学 Method and device for identifying hidden diseases of asphalt pavement, electronic equipment and storage medium
CN113591606B (en) * 2021-07-08 2024-06-04 武汉理工大学 Asphalt pavement hidden disease identification method and device, electronic equipment and storage medium
CN114169402A (en) * 2021-11-16 2022-03-11 哈尔滨工业大学 Ground penetrating radar underground cavity target automatic identification method based on gate control cycle unit
CN114169402B (en) * 2021-11-16 2024-10-18 哈尔滨工业大学 Ground penetrating radar underground cavity target automatic identification method based on gating circulating unit
CN117351367A (en) * 2023-12-06 2024-01-05 鄄城县公路事业发展中心 Highway maintenance inspection method, medium and equipment
CN117351367B (en) * 2023-12-06 2024-02-09 鄄城县公路事业发展中心 Highway maintenance inspection method, medium and equipment

Also Published As

Publication number Publication date
CN111724074B (en) 2023-10-27

Similar Documents

Publication Publication Date Title
CN111724074B (en) Pavement lesion detection early warning method and system based on deep learning
Liu et al. Novel YOLOv3 model with structure and hyperparameter optimization for detection of pavement concealed cracks in GPR images
CN110071913B (en) Unsupervised learning-based time series anomaly detection method
Cano-Ortiz et al. Machine learning algorithms for monitoring pavement performance
CN113009447B (en) Road underground cavity detection and early warning method based on deep learning and ground penetrating radar
Zhu et al. Weakly supervised network based intelligent identification of cracks in asphalt concrete bridge deck
CN117168331B (en) Real-time detection method for thickness of ice layer of natural ice rink based on optical fiber sensor
CN113239720B (en) Subway vehicle running gear fault diagnosis method based on deep migration learning
CN110222593A (en) A kind of vehicle real-time detection method based on small-scale neural network
CN117371337B (en) Water conservancy model construction method and system based on digital twin
CN116168356B (en) Vehicle damage judging method based on computer vision
CN110674858A (en) Traffic jam detection method based on space-time association and big data mining
CN117269954B (en) Real-time identification method for multiple hidden diseases of ground penetrating radar road based on YOLO
CN115059455B (en) Underground fault diagnosis method for sucker rod pumping unit based on machine learning
Shafabakhsh et al. Determining the relative importance of parameters affecting concrete pavement thickness
CN114486571B (en) Pavement comprehensive performance evaluation method based on pavement service performance and structural performance
Guo Threshold Segmentation Bridge Crack Detection Algorithm Based on Deep Learning
CN115620041B (en) Pavement disease sensing and service state evaluation method
Han et al. Pavement roughness level classification based on logistic and decision tree machine learnings
Chavan et al. An Overview of Machine Learning Techniques for Evaluation of Pavement Condition
Kavitha Lung ancer detection using machine learning
XU et al. Deep Learning-Based Corroded Crack Identification in Reinforced Concrete Using SCNet Model
CN117074291B (en) Non-contact texture friction prediction method
CN117197095B (en) Surface defect detection method and system for generating countermeasure segmentation model based on semi-supervision
CN116319262B (en) Communication fault data detection model of electric energy metering device

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
GR01 Patent grant
GR01 Patent grant