CN114810100B - Shield tunneling attitude prediction method based on deep neural network - Google Patents

Shield tunneling attitude prediction method based on deep neural network Download PDF

Info

Publication number
CN114810100B
CN114810100B CN202210739896.XA CN202210739896A CN114810100B CN 114810100 B CN114810100 B CN 114810100B CN 202210739896 A CN202210739896 A CN 202210739896A CN 114810100 B CN114810100 B CN 114810100B
Authority
CN
China
Prior art keywords
shield
data
attitude
neural network
shield tunneling
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.)
Active
Application number
CN202210739896.XA
Other languages
Chinese (zh)
Other versions
CN114810100A (en
Inventor
章龙管
徐进
刘绥美
李才洪
杨冰
林良宇
朱菁
陈鑫
梁博
郑军
白杰
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.)
Southwest Jiaotong University
China Railway Engineering Service Co Ltd
Original Assignee
Southwest Jiaotong University
China Railway Engineering Service Co Ltd
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 Southwest Jiaotong University, China Railway Engineering Service Co Ltd filed Critical Southwest Jiaotong University
Priority to CN202210739896.XA priority Critical patent/CN114810100B/en
Publication of CN114810100A publication Critical patent/CN114810100A/en
Application granted granted Critical
Publication of CN114810100B publication Critical patent/CN114810100B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • E—FIXED CONSTRUCTIONS
    • E21—EARTH DRILLING; MINING
    • E21D—SHAFTS; TUNNELS; GALLERIES; LARGE UNDERGROUND CHAMBERS
    • E21D9/00—Tunnels or galleries, with or without linings; Methods or apparatus for making thereof; Layout of tunnels or galleries
    • E21D9/003—Arrangement of measuring or indicating devices for use during driving of tunnels, e.g. for guiding machines
    • E—FIXED CONSTRUCTIONS
    • E21—EARTH DRILLING; MINING
    • E21D—SHAFTS; TUNNELS; GALLERIES; LARGE UNDERGROUND CHAMBERS
    • E21D9/00—Tunnels or galleries, with or without linings; Methods or apparatus for making thereof; Layout of tunnels or galleries
    • E21D9/06—Making by using a driving shield, i.e. advanced by pushing means bearing against the already placed lining
    • E—FIXED CONSTRUCTIONS
    • E21—EARTH DRILLING; MINING
    • E21D—SHAFTS; TUNNELS; GALLERIES; LARGE UNDERGROUND CHAMBERS
    • E21D9/00—Tunnels or galleries, with or without linings; Methods or apparatus for making thereof; Layout of tunnels or galleries
    • E21D9/06—Making by using a driving shield, i.e. advanced by pushing means bearing against the already placed lining
    • E21D9/0642—Making by using a driving shield, i.e. advanced by pushing means bearing against the already placed lining the shield having means for additional processing at the front end
    • E21D9/0678—Adding additives, e.g. chemical compositions, to the slurry or the cuttings
    • E—FIXED CONSTRUCTIONS
    • E21—EARTH DRILLING; MINING
    • E21D—SHAFTS; TUNNELS; GALLERIES; LARGE UNDERGROUND CHAMBERS
    • E21D9/00—Tunnels or galleries, with or without linings; Methods or apparatus for making thereof; Layout of tunnels or galleries
    • E21D9/12—Devices for removing or hauling away excavated material or spoil; Working or loading platforms
    • E—FIXED CONSTRUCTIONS
    • E21—EARTH DRILLING; MINING
    • E21D—SHAFTS; TUNNELS; GALLERIES; LARGE UNDERGROUND CHAMBERS
    • E21D9/00—Tunnels or galleries, with or without linings; Methods or apparatus for making thereof; Layout of tunnels or galleries
    • E21D9/12—Devices for removing or hauling away excavated material or spoil; Working or loading platforms
    • E21D9/124—Helical conveying means therefor
    • 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
    • G06N3/044—Recurrent networks, e.g. Hopfield 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
    • G06N3/04—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
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/048—Activation functions
    • 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
    • G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
    • G06N3/0675—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means using electro-optical, acousto-optical or opto-electronic means

Landscapes

  • Engineering & Computer Science (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Mining & Mineral Resources (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Geology (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • Geochemistry & Mineralogy (AREA)
  • General Life Sciences & Earth Sciences (AREA)
  • Environmental & Geological Engineering (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Computing Systems (AREA)
  • Molecular Biology (AREA)
  • General Engineering & Computer Science (AREA)
  • Evolutionary Computation (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Neurology (AREA)
  • Chemical & Material Sciences (AREA)
  • Chemical Kinetics & Catalysis (AREA)
  • General Chemical & Material Sciences (AREA)
  • Excavating Of Shafts Or Tunnels (AREA)

Abstract

The invention belongs to the technical field of shield tunneling, and particularly relates to a shield tunneling attitude prediction method based on a deep neural network; aiming at the problems of difficult prediction of the shield tunneling attitude, low manual decision efficiency and the like, the method comprehensively uses the wavelet transformation denoising and deep learning method, determines the set construction of various related parameters of the shield attitude, and realizes the prediction of the shield attitude parameters at the future moment. The shield tunneling attitude prediction model established by the invention can effectively reduce the influence of noise in construction data, and can still keep good prediction effect when facing mass and high-dimensional data objects; the method is applied to the shield project, and can assist field operators to judge the shield tunneling attitude through predicting the shield attitude parameters, so as to adjust and operate in time to avoid the risk of attitude abnormality.

Description

Shield tunneling attitude prediction method based on deep neural network
Technical Field
The invention belongs to the technical field of shield tunneling, and particularly relates to a shield tunneling attitude prediction method based on a deep neural network.
Background
With the introduction of urbanization process, shield construction has become one of the most common construction methods in tunnel construction, and the shield construction refers to a mechanized construction process in which a shield machine excavates soil, transports muck and assembles segments on the ground according to a design axis of a tunnel. The movement track of the shield machine is generally determined by the posture and the position of the shield machine, and if the shield and the movement track deviate from a design axis, an excavation route is changed, so that the subsequent segment assembling operation and assembling quality are influenced, and engineering problems such as water seepage in a tunnel, overlarge ground settlement or uplift and the like can be caused; therefore, a prediction method capable of predicting the attitude of the shield tunneling machine at the next moment is needed; the existing prediction method found by searching is as follows:
1. the invention has the name: the patent document discloses an invention patent with the publication number of CN112100841A, and the patent document predicts the shield attitude by using a GRU neural network, and inputs a historical pitch angle, a historical roll angle and a historical yaw angle into a model to predict related shield attitude data at the time of arrival.
2. The invention has the name: a dynamic prediction method, a system and equipment for shield attitude are disclosed in the invention patent with the publication number CN112879024A, in the patent document, the shield attitude prediction is carried out by using a bidirectional LSTM neural network and an attention mechanism, and bidirectional learning is carried out by shield attitude data at the past moment and the future moment, so that a good shield attitude prediction effect is obtained.
3. The invention has the name: the patent document discloses an invention patent with the number of CN113344256A, and predicts the shield attitude through an ANN network module, optimizes a model through massive historical construction data and a control performance evaluation module, and improves the shield attitude prediction effect.
The above method does not deal with noise generated in the original construction data. The methods 1 and 2 only use attitude parameter data as input, do not use construction data, and do not predict horizontal and vertical deviations of the shield head and the shield tail; the relevance of the input parameters and the output parameters of the method 3 is small, so that the selection of the construction parameters is not reasonable enough, and the ANN can not learn the time sequence characteristics of the shield data.
Disclosure of Invention
The invention discloses a shield tunneling attitude prediction method based on a deep neural network, aiming at solving the technical problem of noise in original shield construction data.
In order to solve the technical problems, the technical scheme adopted by the invention is as follows:
a shield tunneling attitude prediction method based on a deep neural network comprises the following steps:
step 1: collecting related historical shield construction data based on a sensor on a shield machine;
and 2, step: preprocessing historical shield construction data: the method comprises the steps of abnormal value processing, wavelet transformation denoising and data standardization; converting shield construction data into data suitable for a shield tunneling attitude prediction model after preprocessing;
and 3, step 3: determining the step length s and the prediction time period t of the shield tunneling attitude prediction model, and continuously reading the preprocessed data to generate time sequence data;
and 4, step 4: establishing a shield tunneling attitude prediction model based on an LSTM neural network based on the preprocessed data and time series data, dividing historical shield construction data serving as a historical data set into a training set and a testing set, training and testing the shield tunneling attitude prediction model, judging the prediction effect of the model through root mean square error, finally keeping the model parameters with the best prediction effect, and taking the model parameters with the best prediction effect as the parameters of the shield tunneling attitude prediction model to obtain the final shield tunneling attitude prediction model;
and 5: and (4) inputting the actually detected shield construction data into the shield tunneling attitude prediction model finally obtained in the step (4) to predict the attitude of the shield machine, so as to obtain the attitude of the shield machine at the next moment.
According to the method, the noise in the construction data is removed, the shield tunneling attitude at the future moment is accurately predicted according to the historical construction data, the efficient decision of a construction site manager and an operator is assisted, the abnormal tunneling attitude is avoided in advance, the problems that the manual decision of the shield construction is low in efficiency and unstable, the data is idle and the like are solved, and the method has important values for ensuring the stable tunneling of the shield and reducing the risk accidents of shield projects. The invention constructs a shield tunneling attitude prediction model based on an LSTM neural network, and the LSTM model is a special circulating neural network and can solve the problem of gradient explosion or gradient disappearance of a common RNN so as to obtain long-term data characteristics.
Preferably, the historical shield construction data includes: propulsion system parameters, slag tapping system parameters and grouting system parameters;
the propulsion system parameters include: the propelling speed of the shield machine, the rotating speed of the screw conveyer and the displacement parameters of 6 groups of guide propelling oil cylinders in the shield laser guide system;
the slag tapping system parameters include: pressure detection information in 10 soil pressure sensors in the slag tapping system;
the grouting system parameters include pressure detection information in 4 pressure sensors in the grouting system.
According to the method, 22 construction parameters in the shield system are selected as the input of the shield tunneling attitude prediction model, so that the relevance between the input parameters and the output parameters is ensured, and the linear and nonlinear relations between the parameters are better learned through a deep neural network, so that a better prediction effect is achieved.
Preferably, the method for processing the abnormal value comprises the following steps: will be distributed in [ mu-3 sigma, mu +3 sigma based on the 3 sigma criterion]The data outside is removed, whereinσWhich represents the standard deviation of the data,μrepresenting the mean of the data.
The outliers, also called outliers, are unreasonable values in the data; if an abnormal value exists in shield construction data, the final prediction result is inaccurate, so that the method eliminates data distributed outside [ mu-3 sigma, mu +3 sigma ] based on the 3 sigma criterion, realizes the elimination of the abnormal value, and ensures the prediction precision of the shield tunneling attitude prediction model.
Preferably, the wavelet transform denoising step includes:
decomposing the time sequence of the historical shield construction data processed by the abnormal values for n times through wavelet transformation;
and (4) carrying out wavelet reconstruction on the data subjected to n-time decomposition to obtain a new de-noising data sequence.
In the shield construction process, shield construction data are acquired by sensors distributed all over the body of the shield machine and transmitted to a cloud platform; due to the problems of sensor failure, data communication abnormity and the like possibly existing in the shield construction process, the acquired data has noise, and therefore, the time sequence of the shield machine parameters is decomposed by adopting wavelet transformation; and then, a new de-noising data sequence is generated through wavelet transformation reconstruction, so that the shield attitude prediction precision is improved.
The wavelet basis functions used by the wavelet transform are various, the common wavelet basis functions include DB (Daubechies) wavelet, haar wavelet, meyer wavelet and the like, and different wavelet basis selections can generate different results. The DB wavelet has good regularity, can make the data after noise reduction smoother, and is suitable for processing the data set.
Therefore, the invention selects DB wavelet to perform wavelet transformation denoising, the DB wavelet is usually written as dbN (N belongs to 1,10), N represents vanishing moment of the wavelet basis function, the higher N is, the smoother the processed data is, the stronger the dividing effect and localization capability of different frequency bands are, but the calculated amount is increased, and the timeliness is weakened; the invention sets the wavelet transform decomposition layer number as the maximum, and determines the wavelet basis dbN through experiments.
Preferably, the data standardization is that the historical shield construction data subjected to wavelet transformation noise reduction is subjected to Z-Score standardization processing, and the formula is as follows:
Figure 453691DEST_PATH_IMAGE001
；
in the formula:xfor historical shield construction data after wavelet transformation and noise reduction,x * is andxcorresponding to the standardized historical shield construction data,μthe mean value of the data is represented,
Figure 938025DEST_PATH_IMAGE002
data standard deviation is indicated.
In order to eliminate the influence of construction data dimension and provide the prediction accuracy of the shield tunneling attitude prediction model, the invention carries out Z-Score standardization processing on construction parameter data before inputting a data set into a deep neural network so as to balance comparability among data indexes.
Preferably, the output parameters of the shield tunneling attitude prediction model include: the method comprises the following steps that a guide rolling angle of a shield laser guide system, a guide pitch angle of the shield laser guide system, a guide horizontal front of the shield laser guide system, a guide vertical front of the shield laser guide system, a guide horizontal rear of the shield laser guide system, a guide vertical rear of the shield laser guide system, a guide horizontal trend RP of the shield laser guide system and a guide vertical trend RP of the shield laser guide system are respectively arranged, wherein the guide horizontal front of the shield laser guide system and the guide vertical front of the shield laser guide system represent horizontal and vertical deviations of a shield head; and the table shield is horizontally deviated from the table shield after the guide of the shield laser guide system is horizontal and after the guide of the shield laser guide system is vertical.
Preferably, the LSTM neural network is composed of an LSTM layer, a fully-connected layer, and an Adam algorithm;
the LSTM layer consists of an input layer, an LSTM unit layer and a hidden layer;
the input layer is used for receiving the preprocessed data, transmitting the data with time information to the LSTM unit layer, learning the time sequence characteristics in the data through the LSTM unit layer, outputting the data to the hidden layer under the processing of the LSTM unit layer, learning the relation between the time sequence construction data and the attitude parameters through the hidden layer, and transmitting the relation to the full connection layer;
all the neurons in the full connection layer are connected with the LSTM layer in the previous layer for calculation, and the features learned by the LSTM layer are transmitted and output.
Preferably, the operation steps of the LSTM unit layer during the LSTM neural network training are as follows:
step A, according to the output processed by the LSTM neural network at the last momenth t−1 And construction data at the current timex t By sigmoid function
Figure 487955DEST_PATH_IMAGE003
Produce af t The value is in the range of [0,1]So as to determine the information of the last timeC t−1 The process is shown as the formula:
Figure 369323DEST_PATH_IMAGE004
；
in the formula:Wa matrix of the weights is represented by,brepresenting a bias matrix;W f andb f representsf t The weight matrix and the bias matrix of the generation process, and the values of the elements in the matrix can be changed along with the continuous training of the LSTM networkAnd (4) transforming.
B, output after neural network processing at last momenth t−1 And construction data of the current timex t By sigmoid function
Figure 866163DEST_PATH_IMAGE003
Judging the value to be updated and obtainingi t A value; at the same time, the user can select the desired position,h t−1 andx t then through tanhFunction generation candidate value
Figure 500276DEST_PATH_IMAGE005
(ii) a The concrete formula is as follows:
Figure 158790DEST_PATH_IMAGE006
；
Figure 527455DEST_PATH_IMAGE007
；
in the formula:W i andb i representi t A weight matrix and a bias matrix of a value generation process;W C andb C represents a candidate value
Figure 562407DEST_PATH_IMAGE005
Generating a weight matrix and a bias matrix of the process;
step C. Based on the production of step A, Bf t 、i t Value and candidate value
Figure 129655DEST_PATH_IMAGE005
Generating aC t Value of,C t the value represents the initial output of data for the present moment o t The screening and scaling operations to be performed:
Figure 709803DEST_PATH_IMAGE008
；
step D, inputting the data updated in the step C into an input layer and passing through a sigmoid function
Figure 565763DEST_PATH_IMAGE003
Obtain initial outputo t (ii) a At the same time, tanhLaminating the layer obtained in step CC t Scaling the values to [ -1,1]Interval and multiplying the initial output pair by pair to obtain the output of the modelh t The concrete formula is as follows:
Figure 138827DEST_PATH_IMAGE009
；
Figure 498264DEST_PATH_IMAGE010
；
in the formula:W o andb o representing the initial outputo t A weight matrix and a bias matrix for the process are generated.
Preferably, the step 5 further includes optimizing parameters of the shield tunneling attitude prediction model based on the attitude of the shield tunneling attitude prediction model at the next moment and the actual attitude of the shield tunneling machine at the next moment.
Preferably, the shield tunneling attitude prediction model is trained by adopting a training set, and the shield tunneling attitude prediction model is optimized by adopting an Adam algorithm.
In summary, due to the adoption of the technical scheme, the invention has the beneficial effects that:
1. aiming at the problems of difficult prediction of the shield tunneling attitude, low manual decision efficiency and the like, the invention uses the wavelet transformation denoising and deep learning method to determine the set construction of various related parameters of the shield attitude and realize the prediction of the shield attitude parameters at the future time. The shield tunneling attitude prediction model is established based on the LSTM, so that the influence of noise in construction data can be effectively reduced, and a good prediction effect can be still kept when a large number of high-dimensional data objects are faced; by applying the method to the shield project and predicting the shield attitude parameters, the method can assist field operators to judge the shield tunneling attitude and further adjust and operate in time to avoid the risk of abnormal attitude.
2. According to the method, the noise in the construction data is removed, the shield tunneling attitude at the future moment is accurately predicted according to the historical construction data, the efficient decision of a construction site manager and an operator is assisted, the abnormal tunneling attitude is avoided in advance, the problems that the manual decision of the shield construction is low in efficiency and unstable, the data is idle and the like are solved, and the method has important values for ensuring the stable tunneling of the shield and reducing the risk accidents of shield projects. The invention constructs a shield tunneling attitude prediction model based on an LSTM neural network, and the LSTM model is a special circulating neural network and can solve the problem of gradient explosion or gradient disappearance of a common RNN so as to obtain long-term data characteristics.
3. In the shield construction process, shield construction data are acquired by sensors distributed all over the body of the shield machine and transmitted to a cloud platform; due to the problems of sensor failure, data communication abnormity and the like possibly existing in the shield construction process, the acquired data has noise, and therefore, the time sequence of the shield machine parameters is decomposed by adopting wavelet transformation; and then, a new de-noising data sequence is generated through wavelet transformation reconstruction, so that the shield attitude prediction precision is improved.
4. In order to eliminate the influence of construction data dimension and provide the prediction accuracy of the shield tunneling attitude prediction model, the invention carries out Z-Score standardization processing on construction parameter data before inputting a data set into a deep neural network so as to balance comparability among data indexes.
Drawings
The invention will now be described, by way of example, with reference to the accompanying drawings, in which:
fig. 1 is a diagram of a model structure for predicting shield tunneling attitude according to the present invention.
FIG. 2 is a diagram of the denoising process in wavelet transform according to the present invention.
Fig. 3 is a diagram of the LSTM unit structure of the present invention.
Fig. 4 is a flow chart of the shield tunneling attitude prediction of the present invention.
Detailed Description
In order to make the objects, technical solutions and advantages of the embodiments of the present application clearer, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all the embodiments. The components of the embodiments of the present application, as generally described and illustrated in the figures herein, could be arranged and designed in a wide variety of different configurations. Thus, the following detailed description of the embodiments of the present application, presented in the accompanying drawings, is not intended to limit the scope of the claimed application, but is merely representative of selected embodiments of the application. All other embodiments, which can be derived by a person skilled in the art from the embodiments of the present application without making any creative effort, shall fall within the protection scope of the present application.
Referring to fig. 1 to 4, a method for predicting a shield tunneling attitude based on a deep neural network includes the following steps:
step 1: collecting related historical shield construction data based on a sensor on a shield machine; the historical shield construction data comprises: propulsion system parameters, slag tapping system parameters and grouting system parameters;
the propulsion system parameters include: the propelling speed of the shield machine, the rotating speed of the screw conveyer and the displacement parameters of 6 groups of guide propelling oil cylinders in the shield laser guide system;
the slag tapping system parameters include: pressure detection information in 10 soil pressure sensors in the slag tapping system;
the grouting system parameters comprise pressure detection information in 4 pressure sensors in the grouting system.
For specific parameters, see table 1 below:
Figure 13428DEST_PATH_IMAGE011
according to the method, 22 construction parameters in the shield system are selected as the input of the shield tunneling attitude prediction model, so that the relevance between the input parameters and the output parameters is ensured, and the linear and nonlinear relations between the parameters are better learned through a deep neural network, so that a better prediction effect is achieved.
Step 2: preprocessing historical shield construction data: the method comprises the steps of abnormal value processing, wavelet transformation denoising and data standardization; converting shield construction data into data suitable for a shield tunneling attitude prediction model after pretreatment;
the abnormal value processing method comprises the following steps: will be distributed in [ mu-3 sigma, mu +3 sigma based on the 3 sigma criterion]The data outside is removed, whereinσWhich represents the standard deviation of the data,μrepresents the mean of the data. According to the basic ideas of probability theory and hypothesis testing, the probability generated by the numerical value is less than 5%, the numerical value is called as a small probability event, the probability of occurrence of the small probability event is small, and the model prediction is adversely affected; decomposing shield machine data by adopting a wavelet transform data denoising technology, determining a wavelet basis with the best denoising effect, generating new denoising data through wavelet reconstruction, and improving prediction precision; and finally, data standardization is used, the data are scaled, the influence of dimension among construction data is eliminated, and the model prediction effect is improved.
The outliers, also called outliers, are unreasonable values in the data; if an abnormal value exists in shield construction data, the final prediction result is inaccurate, so that the method eliminates data distributed outside [ mu-3 sigma, mu +3 sigma ] based on the 3 sigma criterion, realizes the elimination of the abnormal value, and ensures the prediction precision of the shield tunneling attitude prediction model.
The wavelet transform denoising step comprises:
decomposing the time sequence of the historical shield construction data processed by the abnormal values for n times through wavelet transformation;
and (4) carrying out wavelet reconstruction on the data subjected to n-time decomposition to obtain a new de-noising data sequence.
In the shield construction process, shield construction data are acquired by sensors distributed all over the body of the shield machine and transmitted to a cloud platform; due to the problems of sensor failure, data communication abnormity and the like possibly existing in the shield construction process, the acquired data has noise, and therefore, the time sequence of the shield machine parameters is decomposed by adopting wavelet transformation; and then, a new de-noising data sequence is generated through wavelet transformation reconstruction, so that the shield attitude prediction precision is improved.
The wavelet transformation denoising process is shown in FIG. 2, where X (t) is the original construction sequence data, D 1 (t) noise after the first wavelet transform decomposition, D 2 (t) noise after the second wavelet transform decomposition, D N (t) is the noise after the nth wavelet transform decomposition, S 1 (t) de-noising data after the first wavelet transform decomposition, S 2 (t) denoised data after a second wavelet transform decomposition, S N And (t) the denoised data decomposed by the nth wavelet transform is subjected to multiple wavelet transform decompositions and then restored by a wavelet data reconstruction method to obtain the denoised construction sequence data.
The wavelet basis functions used by the wavelet transform are various, the common wavelet basis functions include DB (Daubechies) wavelet, haar wavelet, meyer wavelet and the like, and different wavelet basis selections can generate different results. The DB wavelet has good regularity, can make the data after noise reduction smoother, and is suitable for processing the data set.
Therefore, the invention selects DB wavelet to perform wavelet transformation denoising, the DB wavelet is usually written as dbN (N belongs to 1,10), N represents vanishing moment of the wavelet basis function, the higher N is, the smoother the processed data is, the stronger the dividing effect and localization capability of different frequency bands are, but the calculated amount is increased, and the timeliness is weakened; the invention sets the wavelet transform decomposition layer number as the maximum, and determines the wavelet basis dbN through experiments.
The data standardization is that the Z-Score standardization processing is carried out on the historical shield construction data subjected to wavelet transformation noise reduction, and the formula is as follows:
Figure 91106DEST_PATH_IMAGE012
；
in the formula:xfor historical shield construction data subjected to wavelet transformation and noise reduction,x * is andxcorresponding to the standardized historical shield construction data,μthe mean value of the data is represented,
Figure 733439DEST_PATH_IMAGE013
data standard deviation is indicated.
In order to eliminate the influence of the dimension of the construction data piece and provide the prediction accuracy of the shield tunneling attitude prediction model, the invention carries out Z-Score standardization processing on the construction parameter data before inputting the data set into the deep neural network so as to balance the comparability between data indexes.
And step 3: determining the step length s and the prediction time period t of the shield tunneling attitude prediction model, continuously reading the preprocessed data, and generating time sequence data; taking the current time T as an example, taking data in the time period [ T-s, T ] as input, and predicting shield attitude data in the time period [ T +1, T + T ]; the converted time series data is directly input into the LSTM neural network.
And 4, step 4: establishing a shield tunneling attitude prediction model based on an LSTM neural network based on the preprocessed data and time series data, dividing historical shield construction data serving as a historical data set into a training set and a testing set, training the shield tunneling attitude prediction model by using the training set, optimizing the shield tunneling attitude prediction model by using an Adam algorithm, judging the prediction effect of the model through a root mean square error, finally keeping the model parameter with the best prediction effect, and taking the model parameter with the best prediction effect as the parameter of the shield tunneling attitude prediction model to obtain the final shield tunneling attitude prediction model;
the output parameters of the shield tunneling attitude prediction model comprise: the method comprises the following steps that a guide rolling angle of a shield laser guide system, a guide pitch angle of the shield laser guide system, a guide horizontal front of the shield laser guide system, a guide vertical front of the shield laser guide system, a guide horizontal rear of the shield laser guide system, a guide vertical rear of the shield laser guide system, a guide horizontal trend RP of the shield laser guide system and a guide vertical trend RP of the shield laser guide system are respectively arranged, wherein the guide horizontal front of the shield laser guide system and the guide vertical front of the shield laser guide system represent horizontal and vertical deviations of a shield head; the table shield is horizontally deviated from the table shield after the guide of the shield laser guide system is horizontal and after the guide of the shield laser guide system is vertical.
The LSTM neural network consists of an LSTM layer, a full connection layer and an Adam algorithm;
the LSTM layer consists of an input layer, an LSTM unit layer and a hidden layer;
the input layer is used for receiving the preprocessed data, transmitting the data with time information to the LSTM unit layer, learning the time sequence characteristics in the data through the LSTM unit layer, outputting the data to the hidden layer under the processing of the LSTM unit layer, learning the relation between the time sequence construction data and the attitude parameters through the hidden layer, and transmitting the relation to the full connection layer;
all the neurons in the full connection layer are connected with the LSTM layer in the previous layer for calculation, and the features learned by the LSTM layer are transmitted and output.
The long-term and short-term memory neural network is a special cyclic neural network, and can solve the problem of gradient explosion or gradient disappearance of a common RNN so as to acquire long-term data characteristics.
The LSTM unit layer is a special structure of LSTM neural network, and comprises a forgetting gate, an input gate, an output gate and a unit state module, and the structure is shown in FIG. 3, whereinf t 、i t 、o t Respectively a forgetting gate, an input gate and an output gate, x represents a matrix product, + represents a matrix addition,C t representstStatus information of the time of day.
The calculation procedure for each LSTM cell layer is as follows (WRepresents a matrix of weights that is a function of,brepresenting a bias matrix):
step A, according to the output processed by the LSTM neural network at the last momenth t−1 And the current timeCarved construction datax t By sigmoid function
Figure 9700DEST_PATH_IMAGE013
Produce af t The value is in the range of [0,1]So as to determine the information of the previous timeC t−1 The process is shown as the formula:
Figure 446498DEST_PATH_IMAGE014
；
in the formula:Wa matrix of the weights is represented by,brepresenting a bias matrix;W f andb f representsf t The weight matrix and the bias matrix of the generation process, and the values of the elements in the matrix can change along with the continuous training of the LSTM network.
Step B, output after neural network processing at last momenth t−1 And construction data of the current timex t By sigmoid function
Figure 762204DEST_PATH_IMAGE013
Judging the value to be updated and obtainingi t A value; at the same time, the user can select the desired position,h t−1 andx t then passing through tanhFunction generation candidate value
Figure 677070DEST_PATH_IMAGE015
(ii) a The following formula is specified:
Figure 11099DEST_PATH_IMAGE016
；
Figure 415536DEST_PATH_IMAGE017
；
in the formula:W i andb i representsi t Weight matrix and bias for value generation processA matrix;W C andb C represents candidate value
Figure 185915DEST_PATH_IMAGE015
Generating a weight matrix and a bias matrix of the process;
step C. Based on the production of step A, Bf t 、i t Value and candidate value
Figure 638893DEST_PATH_IMAGE018
Generating aC t Value of,C t the value represents the initial output of data for the present moment o t The screening and scaling operations to be performed (see step D).
Figure 93008DEST_PATH_IMAGE019
；
Step D, inputting the data updated in the step C into an input layer and passing through a sigmoid function
Figure 668346DEST_PATH_IMAGE020
Obtain initial outputo t (ii) a At the same time, tanhLaminating the layer obtained in step CC t Scaling the values to [ -1,1]Interval and multiplying the initial output pair by pair to obtain the output of the modelh t The concrete formula is as follows:
Figure 207911DEST_PATH_IMAGE021
；
Figure 480892DEST_PATH_IMAGE022
；
in the formula:W o andb o representing the initial outputo t A weight matrix and a bias matrix for the process are generated.
The invention utilizes the training set to train the model for a plurality of times,the model is optimized using Adam's algorithm. The prediction accuracy of the model is evaluated by Root Mean Square Error (RMSE), and the lower the RMSE value of the model is, the lower the RMSE value is, the attitude parameter prediction result is represented
Figure 523934DEST_PATH_IMAGE023
And true value
Figure 473436DEST_PATH_IMAGE024
The lower the degree of deviation, i.e., the higher the prediction accuracy. The calculation is shown by the following formula:
Figure 234718DEST_PATH_IMAGE025
after the model is trained, testing the trained model by using a test set, judging the model prediction effect through RMSE, and storing the model structure and parameters when the prediction effect is optimal.
And 5: and (4) inputting the actually detected shield construction data into the shield tunneling attitude prediction model finally obtained in the step (4) to predict the attitude of the shield machine, so as to obtain the attitude of the shield machine at the next moment.
And 5, optimizing parameters of the shield tunneling attitude prediction model based on the attitude of the shield tunneling attitude prediction model at the next moment and the actual attitude of the shield tunneling machine at the next moment.
According to the method, noise in construction data is removed, shield tunneling postures at future moments are accurately predicted according to historical construction data, efficient decisions of construction site managers and operators are assisted, abnormal tunneling postures are avoided in advance, the problems that manual decisions are low in efficiency and unstable and data are idle in shield construction at present are solved, and the method has important values for guaranteeing stable tunneling of the shield and reducing shield project risk accidents. The invention constructs a shield tunneling attitude prediction model based on an LSTM neural network, and the LSTM model is a special circulating neural network and can solve the problem of gradient explosion or gradient disappearance of a common RNN so as to obtain long-term data characteristics.
The above embodiments only express specific embodiments of the present application, and the description is specific and detailed, but not construed as limiting the scope of the present application. It should be noted that, for those skilled in the art, without departing from the technical idea of the present application, several changes and modifications can be made, which are all within the protection scope of the present application.

Claims (9)

1. A shield tunneling attitude prediction method based on a deep neural network is characterized by comprising the following steps:
step 1: collecting related historical shield construction data based on a sensor on a shield machine; the historical shield construction data comprises: propulsion system parameters, slag tapping system parameters and grouting system parameters;
the propulsion system parameters include: the propelling speed of the shield machine, the rotating speed of the screw conveyer and the displacement parameters of 6 groups of guide propelling oil cylinders in the shield laser guide system;
the slag tapping system parameters include: pressure detection information in 10 soil pressure sensors in the slag tapping system;
the parameters of the grouting system comprise pressure detection information of 4 pressure sensors in the grouting system;
step 2: preprocessing historical shield construction data: the method comprises the steps of abnormal value processing, wavelet transformation denoising and data standardization; converting shield construction data into data suitable for a shield tunneling attitude prediction model after preprocessing;
and step 3: determining the step length s and the prediction time period t of the shield tunneling attitude prediction model, and continuously reading the preprocessed data to generate time sequence data;
and 4, step 4: establishing a shield tunneling attitude prediction model based on an LSTM neural network based on the preprocessed data and time series data, dividing historical shield construction data serving as a historical data set into a training set and a testing set, training and testing the shield tunneling attitude prediction model, judging the prediction effect of the model through root mean square error, finally keeping the model parameters with the best prediction effect, and taking the model parameters with the best prediction effect as the parameters of the shield tunneling attitude prediction model to obtain the final shield tunneling attitude prediction model;
and 5: and (4) inputting the actually detected shield construction data into the shield tunneling attitude prediction model finally obtained in the step (4) to predict the attitude of the shield machine, so as to obtain the attitude of the shield machine at the next moment.
2. The method for predicting the shield tunneling attitude based on the deep neural network according to claim 1, wherein the method for processing the abnormal value is as follows: and eliminating data distributed outside [ mu-3 sigma, mu +3 sigma ] based on a 3 sigma criterion, wherein sigma represents the standard deviation of the data, and mu represents the mean value of the data.
3. The method for predicting the shield tunneling attitude based on the deep neural network according to claim 1, wherein the step of reducing the noise through wavelet transformation comprises the following steps:
decomposing the time sequence of the historical shield construction data processed by the abnormal values for n times through wavelet transformation;
and (4) carrying out wavelet reconstruction on the data subjected to n-time decomposition to obtain a new de-noising data sequence.
4. The method for predicting the shield tunneling attitude based on the deep neural network as claimed in claim 1, wherein the data is normalized by performing Z-Score normalization on historical shield construction data subjected to wavelet transform denoising, and the formula is as follows:
Figure FDA0003814654730000021
in the formula: x is historical shield construction data subjected to wavelet transformation noise reduction, and x * The normalized historical shield construction data corresponding to x is shown, mu represents a data mean value, and sigma represents a data standard deviation.
5. The method for predicting the shield tunneling attitude based on the deep neural network according to claim 1, wherein the output parameters of the shield tunneling attitude prediction model comprise: the method comprises the following steps that a guide rolling angle of a shield laser guide system, a guide pitch angle of the shield laser guide system, a guide horizontal front of the shield laser guide system, a guide vertical front of the shield laser guide system, a guide horizontal rear of the shield laser guide system, a guide vertical rear of the shield laser guide system, a guide horizontal trend RP of the shield laser guide system and a guide vertical trend RP of the shield laser guide system are respectively arranged, wherein the guide horizontal front of the shield laser guide system and the guide vertical front of the shield laser guide system represent horizontal and vertical deviations of a shield head; the table shield is horizontally deviated from the table shield after the guide of the shield laser guide system is horizontal and after the guide of the shield laser guide system is vertical.
6. The shield tunneling attitude prediction method based on the deep neural network according to claim 1, characterized in that the LSTM neural network is composed of an LSTM layer, a full connection layer and an Adam algorithm;
the LSTM layer consists of an input layer, an LSTM unit layer and a hidden layer;
the input layer is used for receiving the preprocessed data, transmitting the data with time information to the LSTM unit layer, learning the time sequence characteristics in the data through the LSTM unit layer, outputting the data to the hidden layer under the processing of the LSTM unit layer, learning the relation between the time sequence construction data and the attitude parameters through the hidden layer, and transmitting the relation to the full connection layer;
all the neurons in the full connection layer are connected with the LSTM layer in the previous layer for calculation, and the features learned by the LSTM layer are transmitted and output.
7. The method for predicting the shield tunneling attitude based on the deep neural network as claimed in claim 6, wherein during the training of the LSTM neural network, the operation steps of the LSTM unit layers are as follows:
step A, passing through LS according to last momentOutput h after TM neural network processing t-1 And the construction data x at the current time t Generating an f by sigmoid function sigma t The value is in the range of [0,1]Thus, the information C of the previous time is determined t-1 The process is shown as the formula:
f t ＝σ(W f [h t-1 ,x t ]+b f )；
in the formula: w represents a weight matrix, b represents a bias matrix; w f And b f Represents f t The values of elements in the weight matrix and the bias matrix in the generation process can change along with the continuous training of the LSTM network;
step B, output h processed by the neural network at the last moment t-1 And the construction data x at the current time t Judging the value to be updated through sigmoid function sigma and obtaining i t A value; at the same time, h t-1 And x t Generating candidate values by the tanh function
Figure FDA0003814654730000031
The following formula is specified:
i t ＝σ(W i [h t-1 ,x t ]+b i )；
Figure FDA0003814654730000032
step C. F produced based on step A, B t 、i t Value and candidate value
Figure FDA0003814654730000033
Generation of C t Value, C t The value representing the initial output o of the data for the present moment t The screening and scaling operations to be performed:
Figure FDA0003814654730000034
in the formula: w i And b i Represents i t A weight matrix and a bias matrix of a value generation process; w C And b C Represents a candidate value
Figure FDA0003814654730000035
Generating a weight matrix and a bias matrix of the process;
d, inputting the data updated in the step C into an input layer, and obtaining an initial output o through a sigmoid function sigma t (ii) a Meanwhile, the tanh layer is the C obtained in the step C t Scaling the values to [ -1,1]Interval and multiplying the initial output pair by pair to obtain the output h of the model t The concrete formula is as follows:
o t ＝σ(W o [h t-1 ,x t ]+b o )；
h t ＝o t ×tanh(C t )；
in the formula: w o And b o Representing the initial output o t A weight matrix and a bias matrix for the process are generated.
8. The method for predicting the shield tunneling attitude based on the deep neural network as claimed in claim 1, wherein the step 5 further comprises optimizing parameters of the shield tunneling attitude prediction model based on the attitude of the shield tunneling attitude prediction model at the next moment and the actual attitude of the shield tunneling machine at the next moment.
9. The shield tunneling attitude prediction method based on the deep neural network according to claim 1, characterized in that a training set is used for training the shield tunneling attitude prediction model, and an Adam algorithm is used for optimizing the shield tunneling attitude prediction model.
CN202210739896.XA 2022-06-28 2022-06-28 Shield tunneling attitude prediction method based on deep neural network Active CN114810100B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202210739896.XA CN114810100B (en) 2022-06-28 2022-06-28 Shield tunneling attitude prediction method based on deep neural network

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202210739896.XA CN114810100B (en) 2022-06-28 2022-06-28 Shield tunneling attitude prediction method based on deep neural network

Publications (2)

Publication Number Publication Date
CN114810100A CN114810100A (en) 2022-07-29
CN114810100B true CN114810100B (en) 2022-12-02

Family

ID=82523223

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202210739896.XA Active CN114810100B (en) 2022-06-28 2022-06-28 Shield tunneling attitude prediction method based on deep neural network

Country Status (1)

Country Link
CN (1) CN114810100B (en)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN115929406B (en) * 2022-12-07 2023-08-18 河海大学 Surrounding rock stability prediction method, system and storage medium based on neural network
CN117235874B (en) * 2023-11-14 2024-02-20 中电建铁路建设投资集团有限公司 Track deformation prediction method and system based on shield tunneling
CN117493837B (en) * 2024-01-03 2024-03-19 中铁南方投资集团有限公司 Machine learning-based shield tunneling machine attitude item prediction method

Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05171890A (en) * 1991-12-25 1993-07-09 Nippon Telegr & Teleph Corp <Ntt> Directional correction amount model formation method of tunnel robot
CN108710940A (en) * 2017-12-31 2018-10-26 中交第公路工程局有限公司 Method based on shield machine running orbit parameter in Neural Network Optimization dust stratum
CN110195592A (en) * 2019-04-30 2019-09-03 华中科技大学 Shield driving pose intelligent Forecasting and system based on interacting depth study
CN110242310A (en) * 2019-06-14 2019-09-17 西安电子科技大学 Shield axis method for correcting error based on deep neural network in conjunction with association analysis
CN113282990A (en) * 2021-05-31 2021-08-20 中铁十六局集团北京轨道交通工程建设有限公司 Intelligent real-time determination method, terminal and medium for shield motion trail
CN113323676A (en) * 2021-06-03 2021-08-31 上海市隧道工程轨道交通设计研究院 Method for determining cutter head torque of shield tunneling machine by using principal component analysis-length memory model
CN113344256A (en) * 2021-05-21 2021-09-03 上海隧道工程有限公司 System and method for predicting movement characteristics and evaluating control performance of multiple degrees of freedom of shield attitude
CN114329810A (en) * 2021-11-16 2022-04-12 中国水利水电科学研究院 Real-time shield tunneling machine working attitude prediction method based on big data
CN114662699A (en) * 2022-02-21 2022-06-24 上海城建市政工程（集团）有限公司 Shield attitude prediction method based on machine learning

Family Cites Families (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN108022027B (en) * 2018-02-06 2021-06-25 苏交科集团股份有限公司 Prediction method for soil output when shield passes through surrounding rock contact zone
CN114117599A (en) * 2021-11-22 2022-03-01 中铁高新工业股份有限公司 Shield attitude position deviation prediction method

Patent Citations (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH05171890A (en) * 1991-12-25 1993-07-09 Nippon Telegr & Teleph Corp <Ntt> Directional correction amount model formation method of tunnel robot
CN108710940A (en) * 2017-12-31 2018-10-26 中交第公路工程局有限公司 Method based on shield machine running orbit parameter in Neural Network Optimization dust stratum
CN110195592A (en) * 2019-04-30 2019-09-03 华中科技大学 Shield driving pose intelligent Forecasting and system based on interacting depth study
CN110242310A (en) * 2019-06-14 2019-09-17 西安电子科技大学 Shield axis method for correcting error based on deep neural network in conjunction with association analysis
CN113344256A (en) * 2021-05-21 2021-09-03 上海隧道工程有限公司 System and method for predicting movement characteristics and evaluating control performance of multiple degrees of freedom of shield attitude
CN113282990A (en) * 2021-05-31 2021-08-20 中铁十六局集团北京轨道交通工程建设有限公司 Intelligent real-time determination method, terminal and medium for shield motion trail
CN113323676A (en) * 2021-06-03 2021-08-31 上海市隧道工程轨道交通设计研究院 Method for determining cutter head torque of shield tunneling machine by using principal component analysis-length memory model
CN114329810A (en) * 2021-11-16 2022-04-12 中国水利水电科学研究院 Real-time shield tunneling machine working attitude prediction method based on big data
CN114662699A (en) * 2022-02-21 2022-06-24 上海城建市政工程（集团）有限公司 Shield attitude prediction method based on machine learning

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
基于LSTM的盾构掘进轨迹预测研究;佟雨泉;《建筑安全》;20210805;第36卷(第8期);第2-3节 *
隧道掘进机大数据研究进展：数据挖掘助推隧道挖掘;石茂林等;《机械工程学报》;20211120;第57卷(第22期);全文 *

Also Published As

Publication number Publication date
CN114810100A (en) 2022-07-29

Similar Documents

Publication Publication Date Title
CN114810100B (en) Shield tunneling attitude prediction method based on deep neural network
CN112561148B (en) Ship track prediction method and system based on one-dimensional convolutional neural network and LSTM
Ma et al. A hybrid attention-based deep learning approach for wind power prediction
CN110163433B (en) Ship flow prediction method
CN111428789A (en) Network traffic anomaly detection method based on deep learning
CN111242377B (en) Short-term wind speed prediction method integrating deep learning and data denoising
CN110096827B (en) Shield tunneling machine parameter optimization method based on deep neural network
CN111461463B (en) Short-term load prediction method, system and equipment based on TCN-BP
CN112116162B (en) Power transmission line icing thickness prediction method based on CEEMDAN-QFAOA-LSTM
CN111447217A (en) Method and system for detecting flow data abnormity based on HTM under sparse coding
CN116010900A (en) Multi-scale feature fusion gearbox fault diagnosis method based on self-attention mechanism
CN114528756A (en) Ship rolling integrated prediction method and system based on deep learning
CN115099603A (en) Rocket offshore recovery platform dynamic positioning system fault risk assessment method, computer equipment and readable storage medium
US11948079B2 (en) Multi-agent coordination method and apparatus
CN114152442A (en) Rolling bearing cross-working condition fault detection method based on migration convolutional neural network
Wei et al. Deterministic ship roll forecasting model based on multi-objective data fusion and multi-layer error correction
Zhou et al. Short-term wind power prediction based on EMD-LSTM
RU2600099C1 (en) Method of neural network forecasting of change of values of function with its complementary wavelet processing and device for its implementation
CN116070768A (en) Short-term wind power prediction method based on data reconstruction and TCN-BiLSTM
CN112464578B (en) Amplification method for predicting degradation data of residual service life
CN116934603A (en) Logging curve missing segment completion method and device, storage medium and electronic equipment
CN114297795A (en) Mechanical equipment residual life prediction method based on PR-Trans
CN113221450A (en) Dead reckoning method and system for sparse and uneven time sequence data
US11783194B1 (en) Evolutionary deep learning with extended Kalman filter for modeling and data assimilation
Wang et al. Multivariate USV Motion Prediction Method Based on a Temporal Attention Weighted TCN-Bi-LSTM Model

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