CN108304674A - A kind of railway prediction of soft roadbed settlement method based on BP neural network - Google Patents
A kind of railway prediction of soft roadbed settlement method based on BP neural network Download PDFInfo
- Publication number
- CN108304674A CN108304674A CN201810164191.3A CN201810164191A CN108304674A CN 108304674 A CN108304674 A CN 108304674A CN 201810164191 A CN201810164191 A CN 201810164191A CN 108304674 A CN108304674 A CN 108304674A
- Authority
- CN
- China
- Prior art keywords
- neural network
- prediction
- training
- roadbed
- settlement
- 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
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F30/00—Computer-aided design [CAD]
- G06F30/20—Design optimisation, verification or simulation
-
- 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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/061—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using biological neurons, e.g. biological neurons connected to an integrated circuit
Abstract
The railway prediction of soft roadbed settlement method based on BP neural network that the invention discloses a kind of, includes the following steps:The processing mode of roadbed, the engineering characteristic to banket, subgrade construction duration are obtained into training sample set as the input variable of neural network using the measured data of the settlement observation section corresponding to input variable as output vector;Neuron node number in hidden layer is determined by trial and error procedure;After all data of training sample set are normalized, the training of sample data is carried out using trainlm functions as training function, builds BP neural network;The processing mode of roadbed for collecting certain railroad bed settlement observation section, the engineering characteristic to banket, subgrade construction duration data, are normalized, and input the BP neural network of gained to get railway prediction of soft roadbed settlement result.Present invention introduces the methods of optimization to solve the problems, such as that BP neural network is encountered when predicting subgrade settlement, and precision of prediction significantly improves after introducing optimization.
Description
Technical field
The present invention relates to the realm of building construction, and in particular to a kind of railway soft soil roadbed settlement based on BP neural network is pre-
Survey method.
Background technology
Nerual network technique blank research starts from the 1940s, still just being applied until latter stage the last century 80's
In civil engineering subject.China is that Beijing Jiaotong University Zhang Qingjiao is awarded and was applied to mechanical behaviors of rocks at 1992 earliest
Forecasting research on.By in recent years, nerual network technique has obtained increasingly wider in civil engineering subject, geotechnical engineering subject
General application, many uncertain problems in effective solution civil engineering, especially geotechnical engineering subject.Such as:Nerve
Network technology is in the inverting and prediction of geotechnical engineering physical and mechanical parameter index, the prediction of ultimate bearing capacity of single pile, roadbed and stake
The settlement prediction of base, the inverting of deep foundation pit support parameter, the calculating of soil dynamics parameter, side slope deformation monitoring etc. have
Extensive use.
BP neural network is long-term to be widely used, have benefited from it have the advantages that it is following:BP neural network has first
Stronger Nonlinear Processing ability, therefore it can preferably solve the problems, such as non-linear birds of the same feather flock together.It was verified that for any one
A three layers or three layers or more of BP neural network can be achieved with as long as the neuron number of the network hidden layer is enough to arbitrary
One the unlimited of nonlinear function approaches.Secondly as BP neural network uses the information processing technology of distributed parallel, therefore
There is the ability of stronger association and memory to input information.Again, BP networks also have optimization computing capability.It can to
Under fixed constraints, one group of suitable parameter combination is found, determining object function error is made to reach minimum.
Although although BP neural network has many good qualities, it is widely used in practice in engineering, for a long time
Use during find BP neural network there is also some disadvantages, mainly have the following aspects.
(1) use BP networks in training, gradient descent method convergence rate is very slow.Learning rate is improved using momentum method,
Convergence rate slightly improves, but speed is still inadequate in actual treatment complex nonlinear problem, and both methods is typically only capable to
For incrementally training.
(2) multilayer neural network is applied to may be implemented to approach the simulation of arbitrary function in lineary system theory.But
Might not there can be solution in actual use.
(3) Multi-layer BP Neural Network is unable to get optimal solution sometimes, is primarily due to non-linear in Multi-layer BP Neural Network
Transmission function may have multiple locally optimal solutions.And the selection for seeking optimum point and initial point has much relations, initial point is such as
Fruit closer to be local best points rather than globe optimum when, global optimal solution will be hardly resulted in.
(4) selection of the structure of BP neural network is still not clear, generally can only be selected by experience.And the structure of network is good
The bad quality for directly influencing network performance.The determination of the number of neural network hidden neuron is also a difficult point.Neuron
If number is likely to result in the discomfort of network very little, and may cause the mistake adaptive of network too much.
Invention content
To solve the above problems, the present invention provides a kind of railway prediction of soft roadbed settlement side based on BP neural network
Method.
To achieve the above object, the technical solution that the present invention takes is:
A kind of railway prediction of soft roadbed settlement method based on BP neural network, includes the following steps:
S1, the processing mode of roadbed, the engineering characteristic to banket, subgrade construction duration are become as the input of neural network
Amount, using the measured data of the settlement observation section corresponding to the input variable as output vector, obtains training sample set;
S2, neuron node number in hidden layer is determined by trial and error procedure;
S3, all data of training sample set are normalized, between so that total data is transformed into 0~1 after, with
Trainlm functions carry out the training of sample data as training function, build BP neural network;
S4, the processing mode of roadbed for collecting certain railroad bed settlement observation section, the engineering characteristic to banket, subgrade construction work
Issue evidence, is normalized, between so that total data is transformed into 0~1 after, input gained BP neural network to get right
The railway prediction of soft roadbed settlement result answered.
Wherein, the step S2 specifically comprises the following steps:
S21, the range for determining neuron node number with theoretical method first
Determine that number of nodes can use n=13 using " 2N+1 " method, then marked centered on n=13 one it is relatively large
Range, optional access purpose ranging from 5~21;
S22, according to the 5~21 of selection this range, programmed using MATLAB and carry out tentative calculation, acquire number of nodes respectively
Amount to items BP neural network forecast error amounts in 17 cyclic processes from 5 to 21;
That minimum number of S23, corresponding BP neural network forecasts error amount is most suitable neuron node number.
Present invention introduces the methods of optimization to solve the problems, such as that BP neural network is encountered when predicting subgrade settlement, introduces optimization
Precision of prediction significantly improves afterwards.
Description of the drawings
Fig. 1 is the corresponding error of different the number of hidden nodes.
Fig. 2 is traingdx functions training figure.
Fig. 3 is that traingdx trains function error figure.
Fig. 4 is traingdm functions training figure.
Fig. 5 is that traingdm trains function error figure.
Fig. 6 is trainlm functions training figure.
Fig. 7 is that trainlm trains function error figure.
Fig. 8 is trainbfg functions training figure.
Fig. 9 is that trainbfg trains function error figure.
Specific implementation mode
In order to make objects and advantages of the present invention be more clearly understood, the present invention is carried out with reference to embodiments further
It is described in detail.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, it is not used to limit this hair
It is bright.
Embodiment
Collect certain railroad bed settlement observation section DK0+243, DK0+280, DK0+342, DK0+370, DK0+342, DK0+
445, the measured data of 48 settlement observation sections such as DK0+450, DK0+500, DK0+535, DK0+573, as shown in table 1.Choosing
Take wherein 36 groups of settlement observation data such as DK0+243, DK0+280, DK0+342, DK0+370 as training sample, remaining 37-
48 groups of settlement observation data are as test samples, to test to trained BP neural network, judge the effect of its training
Fruit.Since selection parameter is more, the parameters dimension of the input of network is inconsistent, such as the amount of thickness of soft soil, depth of fill
Guiding principle is m, and the construction period is day, and foundation treatment mode is precompressed processing, CFG processing and does not handle that these data are in magnitude
It is inconsistent, therefore first original data sequence should be normalized, so that total data is transformed between 0~1.Wait for network
After emulation, then reduction treatment is carried out to output valve.Since foundation treatment mode herein is not quantitative variable, herein
It is handled in the following ways:Ground, which is not handled, is assigned a value of 1, and precompressed processing is assigned a value of 2, CFG processing and is assigned a value of 3.
1 test samples tables of data of table
Continued 4.1
Hidden layer design
Hidden layer design it is most important be exactly the number of hidden nodes determination.If very little, the network of the number of hidden nodes choosing
It may be searched in Local Minimum, cause network that may not train not come out, it is difficult to identify the sample of " being unfamiliar with ", fault-tolerance
Difference, it is difficult to obtain reliable result.If but hidden layer neuron number of nodes is excessive, it will cause the training time to increase, be easy to make
Network training is excessive, and error is also not necessarily best.So it is also BP neural network that the selection of neuron node number, which is a difficult point,
The key of model optimization.About the selection of number of nodes, many scholars do a lot of work, and research has shown that three layers of BP nerve nets
Network assumes that it has m node with input layer, and hidden layer has 2m+1 node, output layer to have n node, then it can be accurate
Ground express any one continuous functionAs shown in formula (1).
It can be obtained by formula (1), if a BP network input layer has m node, number of nodes n=2m+1.In addition,
The number of hidden nodes is asked to also have formula (2)-(4) that can refer to:
N=log2m (3)
In formula:N- the number of hidden nodes;M- input number of nodes;n1Output node number;n2Number of training.
Determine that number of nodes also has other many methods, such as direct typing method, pruning method, growth method etc..But these are managed
By being all still not perfect, the case where many theories often will appear failure in practical applications.In view of the not perfect of theory, intend herein
Neuron node number is determined by trial and error procedure.Detailed process is as follows:
(1) range of neuron node number is determined with theoretical method first.Determine that number of nodes can use n using " 2N+1 " method
=13, a relatively large range, optional access purpose ranging from 5~21 are then marked centered on n=13;
(2) it according to the 5~21 of selection this range, is programmed using MATLAB and carries out tentative calculation, acquire number of nodes respectively
Amount to items BP neural network forecast error amounts in 17 cyclic processes from 5 to 21;
(3) that minimum number of corresponding BP neural network forecasts error amount is most suitable neuron node number.
Different neuron node numbers and corresponding BP neural network forecasts error amount are as shown in Figure 1.As can be seen from Figure 1 in nerve
During neural network forecast, Prediction sum squares constantly fluctuate fluctuation with the variation of number of nodes.When number of nodes is 13, BP god
Prediction sum squares through network are minimum.Thus may determine that prediction result most sorrow when the number of hidden nodes is 13.
The selection of training function
Training function in MATLAB Neural Network Toolbox is mainly the following:
(1) traingdx functions:With adjusting learning rate and additional momentum BP algorithm function;
(2) traingdm functions:Gradient declines Momentum BP Algorithm function;
(3) trainlm functions:Become the Levenberg-Marquardt algorithmic functions of gradient backpropagation;
(4) trainbfg functions:Reverse transmittance nerve network is trained, Bayesian normalization method functions.
To choose the BP neural network of most sorrow, sample data is trained with above-mentioned four kinds of functions respectively, obtains difference
The prediction effect of training function is as shown in Fig. 2~9:It can be seen from Fig. 2~9 traingdx functions, traingdm functions,
The four kinds of network convergence rates of trained function when being trained such as trainlm functions and trainbfg functions and prediction error feelings
Condition.For the prediction effect of clearer comparison function, the network convergence rate of each function and prediction error are now listed in table 2.
Each trained function convergence speed of table 2 and prediction error statistics table
It can be seen from Fig. 2-9 and table 2 traingdm functions carry out network training when network training number up to 20000 times
Still not converged, maximum relative error 13.47%, minimum relative error is -0.32%, is that four kinds of trained function errors are maximum
's;Traingdx Function Network frequency of training is 717 times, and for maximum relative error error up to 13.01%, training effect is bad;
Trainbfgt Function Network frequency of training is 88 times, and for maximum relative error error up to 12.70%, training effect is not good enough;It is using
When trainlm functions are trained, network convergence rate is most fast and relative error is minimum, maximum relative error 9.83%,
Minimum relative error is only -0.01%.Comprehensive analysis is than choosing it is found that using trainlm functions as training function best results.
The above is only a preferred embodiment of the present invention, it is noted that for the ordinary skill people of the art
For member, without departing from the principle of the present invention, it can also make several improvements and retouch, these improvements and modifications are also answered
It is considered as protection scope of the present invention.
Claims (2)
1. a kind of railway prediction of soft roadbed settlement method based on BP neural network, which is characterized in that include the following steps:
S1, using the processing mode of roadbed, the engineering characteristic to banket, subgrade construction duration as the input variable of neural network, will
The measured data of settlement observation section corresponding to the input variable obtains training sample set as output vector;
S2, neuron node number in hidden layer is determined by trial and error procedure;
S3, all data of training sample set are normalized, between so that total data is transformed into 0~1 after, with
Trainlm functions carry out the training of sample data as training function, build BP neural network;
S4, the processing mode of roadbed for collecting certain railroad bed settlement observation section, the engineering characteristic to banket, subgrade construction duration number
According to, be normalized, between so that total data is transformed into 0~1 after, input gained BP neural network to get corresponding
Railway prediction of soft roadbed settlement result.
2. a kind of railway prediction of soft roadbed settlement method based on BP neural network as described in claim 1, feature exist
In the step S2 specifically comprises the following steps:
S21, the range for determining neuron node number with theoretical method first
It determines that number of nodes can use n=13 using " 2N+1 " method, a relatively large model is then marked centered on n=13
It encloses, optional access purpose ranging from 5~21;
S22, according to the 5~21 of selection this range, programmed using MATLAB and carry out tentative calculation, acquire number of nodes respectively from 5
Amount to items BP neural network forecast error amounts in 17 cyclic processes to 21;
That minimum number of S23, corresponding BP neural network forecasts error amount is most suitable neuron node number.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810164191.3A CN108304674A (en) | 2018-02-09 | 2018-02-09 | A kind of railway prediction of soft roadbed settlement method based on BP neural network |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201810164191.3A CN108304674A (en) | 2018-02-09 | 2018-02-09 | A kind of railway prediction of soft roadbed settlement method based on BP neural network |
Publications (1)
Publication Number | Publication Date |
---|---|
CN108304674A true CN108304674A (en) | 2018-07-20 |
Family
ID=62848983
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201810164191.3A Pending CN108304674A (en) | 2018-02-09 | 2018-02-09 | A kind of railway prediction of soft roadbed settlement method based on BP neural network |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN108304674A (en) |
Cited By (6)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110593018A (en) * | 2019-06-28 | 2019-12-20 | 吉林大学 | Method for predicting settlement of high-speed railway subgrade |
CN111428297A (en) * | 2020-03-23 | 2020-07-17 | 交通运输部公路科学研究所 | BP neural network-based pile foundation P-S curve determination method |
CN111737808A (en) * | 2020-08-06 | 2020-10-02 | 北京大成国测科技有限公司 | Railway roadbed early warning system and method based on artificial neural network |
CN112163669A (en) * | 2020-10-09 | 2021-01-01 | 上海应用技术大学 | Pavement subsidence prediction method based on BP neural network |
CN113821863A (en) * | 2021-11-22 | 2021-12-21 | 中南大学 | Method for predicting vertical ultimate bearing capacity of pile foundation |
CN114626008A (en) * | 2022-03-15 | 2022-06-14 | 中铁二院工程集团有限责任公司 | Railway subgrade settlement prediction method and device based on power-related random process |
Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2003268756A (en) * | 2002-03-18 | 2003-09-25 | Fuji Kiso Consultant Kk | Foundation construction method selecting method and foundation construction method selecting program and computer readable recording medium recorded with foundation construction method selecting program |
CN102425148B (en) * | 2011-09-02 | 2014-01-08 | 铁道第三勘察设计院集团有限公司 | Rapid sub-grade settlement predicting method based on static sounding and BP (Back Propagation) neural network |
CN106845641A (en) * | 2017-03-03 | 2017-06-13 | 东南大学 | Subway settlement prediction method based on empirical mode decomposition and BP neural network |
-
2018
- 2018-02-09 CN CN201810164191.3A patent/CN108304674A/en active Pending
Patent Citations (3)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
JP2003268756A (en) * | 2002-03-18 | 2003-09-25 | Fuji Kiso Consultant Kk | Foundation construction method selecting method and foundation construction method selecting program and computer readable recording medium recorded with foundation construction method selecting program |
CN102425148B (en) * | 2011-09-02 | 2014-01-08 | 铁道第三勘察设计院集团有限公司 | Rapid sub-grade settlement predicting method based on static sounding and BP (Back Propagation) neural network |
CN106845641A (en) * | 2017-03-03 | 2017-06-13 | 东南大学 | Subway settlement prediction method based on empirical mode decomposition and BP neural network |
Non-Patent Citations (2)
Title |
---|
郭亚宇 等: "改进的BP神经网络在路基沉降预测中的应用", 《港工技术》 * |
郭亚宇 等: "遗传优化的BP神经网络预测路基沉降", 《低温建筑技术》 * |
Cited By (8)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN110593018A (en) * | 2019-06-28 | 2019-12-20 | 吉林大学 | Method for predicting settlement of high-speed railway subgrade |
CN111428297A (en) * | 2020-03-23 | 2020-07-17 | 交通运输部公路科学研究所 | BP neural network-based pile foundation P-S curve determination method |
CN111737808A (en) * | 2020-08-06 | 2020-10-02 | 北京大成国测科技有限公司 | Railway roadbed early warning system and method based on artificial neural network |
CN112163669A (en) * | 2020-10-09 | 2021-01-01 | 上海应用技术大学 | Pavement subsidence prediction method based on BP neural network |
CN113821863A (en) * | 2021-11-22 | 2021-12-21 | 中南大学 | Method for predicting vertical ultimate bearing capacity of pile foundation |
CN113821863B (en) * | 2021-11-22 | 2022-03-01 | 中南大学 | Method for predicting vertical ultimate bearing capacity of pile foundation |
CN114626008A (en) * | 2022-03-15 | 2022-06-14 | 中铁二院工程集团有限责任公司 | Railway subgrade settlement prediction method and device based on power-related random process |
CN114626008B (en) * | 2022-03-15 | 2023-03-21 | 中铁二院工程集团有限责任公司 | Railway subgrade settlement prediction method and device based on power-related random process |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN108304674A (en) | A kind of railway prediction of soft roadbed settlement method based on BP neural network | |
CN103853786B (en) | The optimization method and system of database parameter | |
CN104463359A (en) | Dredging operation yield prediction model analysis method based on BP neural network | |
CN103105246A (en) | Greenhouse environment forecasting feedback method of back propagation (BP) neural network based on improvement of genetic algorithm | |
CN103577888A (en) | Improved entropy weight AHP and application thereof | |
CN105931116A (en) | Automated credit scoring system and method based on depth learning mechanism | |
CN103942461A (en) | Water quality parameter prediction method based on online sequential extreme learning machine | |
CN110309609A (en) | A kind of architecture indoor air quality evaluation method based on rough set and wavelet neural network | |
CN108334943A (en) | The semi-supervised soft-measuring modeling method of industrial process based on Active Learning neural network model | |
CN111397901A (en) | Rolling bearing fault diagnosis method based on wavelet and improved PSO-RBF neural network | |
CN104915515A (en) | BP neural network based GFET modeling method | |
CN109508498A (en) | Rubber shock absorber formula designing system and method based on BP artificial neural network | |
CN112183935A (en) | River water quality comprehensive evaluation method and system | |
EP4080429A1 (en) | Technology readiness level determination method and system based on science and technology big data | |
CN110212975A (en) | A kind of OTDR fault signature judgment method based on differential evolution neural network | |
CN108764523A (en) | Predictive Methods of Road Accidents based on unbiased nonhomogeneous gray model and geneva model | |
Luo et al. | A novel nonlinear combination model based on support vector machine for stock market prediction | |
Liu et al. | Fuzzy optimization BP neural network model for pavement performance assessment | |
CN110969312A (en) | Short-term runoff prediction coupling method based on variational modal decomposition and extreme learning machine | |
Tu et al. | Evaluation of seawater quality in hangzhou bay based on TS fuzzy neural network | |
CN114091883A (en) | Method, device, medium and equipment for predicting well leakage risk horizon before drilling | |
Yang et al. | Concrete strength evaluation based on fuzzy neural networks | |
CN101425157A (en) | Overall evaluation method for railway emergency scheme | |
Wenxi et al. | Expressway management risk evaluation based on fuzzy neural networks | |
Liu et al. | Short-term traffic flow prediction based on Lagrange support vector regression |
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 | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20180720 |
|
RJ01 | Rejection of invention patent application after publication |