CN115522945A - Dynamic prediction method for shield tunneling attitude - Google Patents
Dynamic prediction method for shield tunneling attitude Download PDFInfo
- Publication number
- CN115522945A CN115522945A CN202211250468.7A CN202211250468A CN115522945A CN 115522945 A CN115522945 A CN 115522945A CN 202211250468 A CN202211250468 A CN 202211250468A CN 115522945 A CN115522945 A CN 115522945A
- Authority
- CN
- China
- Prior art keywords
- data
- shield
- attitude
- model
- construction
- 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
- 238000000034 method Methods 0.000 title claims abstract description 80
- 230000005641 tunneling Effects 0.000 title claims abstract description 33
- 238000010276 construction Methods 0.000 claims abstract description 75
- 238000012549 training Methods 0.000 claims abstract description 15
- 238000012545 processing Methods 0.000 claims abstract description 10
- 230000007246 mechanism Effects 0.000 claims abstract description 9
- 238000013527 convolutional neural network Methods 0.000 claims abstract description 7
- 238000000354 decomposition reaction Methods 0.000 claims abstract description 4
- 238000012360 testing method Methods 0.000 claims description 14
- 238000012163 sequencing technique Methods 0.000 claims description 9
- 238000007781 pre-processing Methods 0.000 claims description 7
- 238000004422 calculation algorithm Methods 0.000 claims description 5
- 230000010354 integration Effects 0.000 claims description 4
- 238000012935 Averaging Methods 0.000 claims description 3
- 230000003044 adaptive effect Effects 0.000 claims description 3
- 230000002123 temporal effect Effects 0.000 claims description 3
- 238000013461 design Methods 0.000 abstract description 8
- 238000013528 artificial neural network Methods 0.000 description 12
- 230000008569 process Effects 0.000 description 12
- 108091006146 Channels Proteins 0.000 description 7
- 230000006870 function Effects 0.000 description 7
- 235000019580 granularity Nutrition 0.000 description 7
- 238000012216 screening Methods 0.000 description 6
- 238000004364 calculation method Methods 0.000 description 5
- 238000000605 extraction Methods 0.000 description 4
- 238000001914 filtration Methods 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000013136 deep learning model Methods 0.000 description 3
- 230000009467 reduction Effects 0.000 description 3
- 238000012706 support-vector machine Methods 0.000 description 3
- 101000935123 Homo sapiens Voltage-dependent N-type calcium channel subunit alpha-1B Proteins 0.000 description 2
- 102100025342 Voltage-dependent N-type calcium channel subunit alpha-1B Human genes 0.000 description 2
- 238000013473 artificial intelligence Methods 0.000 description 2
- 230000006399 behavior Effects 0.000 description 2
- 230000008901 benefit Effects 0.000 description 2
- 125000004122 cyclic group Chemical group 0.000 description 2
- 238000013135 deep learning Methods 0.000 description 2
- 230000007547 defect Effects 0.000 description 2
- 238000006073 displacement reaction Methods 0.000 description 2
- 238000011156 evaluation Methods 0.000 description 2
- 230000007774 longterm Effects 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 239000011159 matrix material Substances 0.000 description 2
- 238000012986 modification Methods 0.000 description 2
- 230000004048 modification Effects 0.000 description 2
- 210000002569 neuron Anatomy 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 230000000306 recurrent effect Effects 0.000 description 2
- ORILYTVJVMAKLC-UHFFFAOYSA-N Adamantane Natural products C1C(C2)CC3CC1CC2C3 ORILYTVJVMAKLC-UHFFFAOYSA-N 0.000 description 1
- 241000288105 Grus Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 230000004913 activation Effects 0.000 description 1
- 230000004931 aggregating effect Effects 0.000 description 1
- 238000012217 deletion Methods 0.000 description 1
- 230000037430 deletion Effects 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000004519 manufacturing process Methods 0.000 description 1
- 238000003062 neural network model Methods 0.000 description 1
- 230000035515 penetration Effects 0.000 description 1
- 230000003014 reinforcing effect Effects 0.000 description 1
- 238000011160 research Methods 0.000 description 1
- 238000007363 ring formation reaction Methods 0.000 description 1
- 238000007493 shaping process Methods 0.000 description 1
- 230000006641 stabilisation Effects 0.000 description 1
- 238000011105 stabilization Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Images
Classifications
-
- E—FIXED CONSTRUCTIONS
- E21—EARTH OR ROCK 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
- E21D9/093—Control of the driving shield, e.g. of the hydraulic advancing cylinders
-
- 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
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Mining & Mineral Resources (AREA)
- Life Sciences & Earth Sciences (AREA)
- Computational Linguistics (AREA)
- Evolutionary Computation (AREA)
- Geochemistry & Mineralogy (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- General Life Sciences & Earth Sciences (AREA)
- Data Mining & Analysis (AREA)
- Geology (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Environmental & Geological Engineering (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention relates to a dynamic prediction method of shield tunneling attitude, which comprises the following steps: s11, data processing: acquiring historical construction data of shield construction, processing the historical construction data and using the historical construction data as sample data; s12, model training: establishing a learning model, inputting sample data into the learning model, denoising the sample data by the learning model through set empirical mode decomposition to form a data set, and learning the data set through a convolutional neural network model with channel attention, a time sequence model and a time attention mechanism to form a prediction model; s13, automatic prediction: and inputting the operation parameters of the shield in construction into the prediction model, so that the prediction model outputs corresponding attitude data of the shield. The method effectively solves the problem of poor control precision of the shield attitude, and a shield driver can adjust the shield attitude in advance according to a prediction result, so that the influence of the shield on the surrounding strata and the segment attitude is reduced, and the shield attitude is matched with a design axis as far as possible.
Description
Technical Field
The invention relates to the field of shield construction, in particular to a dynamic prediction method of shield tunneling attitude.
Background
The movement track of the shield machine in the advancing process is usually around the design axis to advance, the shield of great degree drives the drift and can directly make the shaping quality of shield tunnel reduce, the production reason of shield drift is many-sided, in complicated operational environment, the shield drift can lead to lining ring section of jurisdiction assemble the error that appears, great assembly error can make the assembly process of section of jurisdiction complicated, can lead to tunnel cyclization quality to go wrong, if the section of jurisdiction damages and leaks etc., also can bring the potential safety hazard during the operation simultaneously, furthermore, in the process of driving, the accumulation of shield drift can make the shield increase to the influence of surrounding stratum environment, thereby produce the influence to the building around the tunnel.
At present, the shield control technology is mainly based on that an automatic navigation system measures real-time attitude and position information under a shield advancing state, deviation in shield machine advancing engineering is adjusted through driving experience of operators, and engineering adjustment based on real-time engineering feedback has certain hysteresis and the defect of untimely control.
With the penetration of Artificial Intelligence (AI) methods to various fields, some artificial intelligence methods are explored in the aspect of shield attitude control at present, and some scholars establish a behavior model of a shield driver by using a Support Vector Machine (SVM) method in combination with data collected during construction, and the behavior model is used for operating a shield; the support vector machine can not be continuously competent for the real-time driving and correcting work of the shield along with the change of engineering conditions and operators. The traditional common shallow model (such as a nonlinear supervised learning method and a shallow neural network) extremely depends on the feature extraction problem of prior knowledge and the requirement of a large amount of marked data, and when facing practical problems, the shallow Neural Network (NNs) often rapidly reaches the limit, so how to solve the problem of shield attitude control by using the neural network becomes the most important difficulty at present.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide a dynamic prediction method of the shield tunneling attitude, which solves the problem of poor control precision of the shield attitude, and dynamically predicts the real-time attitude and position of the shield through a dynamic prediction method based on a mixed deep learning model, so that a shield driver can adjust the shield attitude in advance according to the prediction result, reduce the influence of the shield on the attitude of surrounding strata and segments, and ensure that the shield attitude is matched with the design axis as much as possible.
The technical scheme for realizing the purpose is as follows:
the invention provides a dynamic prediction method of a shield tunneling attitude, which comprises the following steps:
s11, data processing: acquiring historical construction data of shield construction, processing the historical construction data and using the historical construction data as sample data;
s12, model training: establishing a learning model, inputting sample data into the learning model, denoising the sample data by the learning model through set empirical mode decomposition to form a data set, and learning the data set through a convolutional neural network model with channel attention, a time sequence model and a time attention mechanism to form a prediction model;
s13, automatic prediction: and inputting the operation parameters of the shield under construction into the prediction model, so that the prediction model outputs the corresponding attitude data of the shield.
The invention provides a dynamic prediction method of shield tunneling attitude, which comprises the steps of obtaining historical construction data, preprocessing the historical construction data to form sample data, inputting the sample data into a learning model to perform noise reduction, feature extraction and training learning to form a prediction model, inputting operation parameters into the prediction model in the actual construction process of a shield, outputting corresponding attitude data of the shield by the prediction model, constructing the shield according to the operation parameters if the attitude data accords with construction expectation, correspondingly modifying the operation parameters and inputting the operation parameters into the prediction model again to obtain corresponding attitude data if the attitude data does not accord with the construction expectation, solving the problem of poor control precision of the shield attitude, and dynamically predicting the real-time attitude and position of the shield by a dynamic prediction method based on a mixed deep learning model, so that a shield driver can adjust the shield attitude in advance according to a prediction result, reducing the influence of the shield on surrounding strata and attitude segments, and enabling the shield attitude to be consistent with a design axis as far as possible.
The dynamic prediction method for the shield tunneling attitude of the invention is further improved in that when the historical construction data is preprocessed, the method further comprises the following steps:
and deleting the data which is larger than the standard deviation of 3 times of the average value in the historical construction data, and deleting the historical construction data recorded when the shield stops moving in the shield construction.
The dynamic prediction method for the shield tunneling attitude is further improved in that when historical construction data of the shield during stopping movement is deleted, the method further comprises the following steps:
when the total thrust, the total torque or the thrust speed of the shield is zero, the shield is in a stop state at the time point, and the historical construction data corresponding to the time point is deleted.
The dynamic prediction method for the shield tunneling attitude is further improved in that historical construction data comprises operation data automatically recorded by a PLC assembly of the shield according to set frequency and observation data recorded by constructors according to granularity of each ring of pipe piece, and the recording frequency of the operation data is greater than that of the observation data;
averaging the observation data recorded in the set time interval to form a first data set, expanding the observation data through an interpolation method to form a second data set, enabling the number of data in the second data set to be matched with the number of data in the first data set, and further taking the data in the first data set and the second data set as sample data.
The dynamic prediction method for the shield tunneling attitude of the invention is further improved in that when the historical construction data is preprocessed, the method further comprises the following steps:
and sequencing the importance of each parameter in the historical construction data, and performing weighted integration on sequencing results to obtain sample data.
The dynamic prediction method for the shield tunneling attitude of the invention is further improved in that when the prediction model is formed by training, the method further comprises the following steps:
the shape of the data set is maintained in the channel attention model by using a 1-by-1 convolution kernel, and the input shape of the time sequence model is protected by a dimension replacement layer, so that the time expansion steps of the cycle kernel of the time sequence model are consistent with the time dimension of the data set.
The dynamic prediction method of the shield tunneling attitude of the invention is further improved in that the method also comprises the following steps:
providing a plurality of test data, inputting the test data into the prediction model and correspondingly obtaining attitude data, comparing the attitude data with actual attitude data corresponding to the test data to obtain a comparison result, and adjusting the hyper-parameters of the prediction model according to the comparison result.
The dynamic prediction method for the shield tunneling attitude is further improved in that the hyper-parameters of the prediction model are adjusted by using an adaptive moment estimation algorithm.
The dynamic prediction method of the shield tunneling attitude of the invention is further improved in that the method also comprises the following steps:
after the historical construction data are preprocessed, 80% of the historical construction data are randomly selected as sample data, and the rest of the historical construction data are used as test data.
The dynamic prediction method for the shield tunneling attitude is further improved in that the shield attitude data comprises the horizontal deviation of the shield head, the vertical deviation of the shield head, the horizontal deviation of the shield tail and the vertical deviation of the shield tail.
Drawings
Fig. 1 is a flowchart of a dynamic prediction method of shield tunneling attitude according to the present invention.
Fig. 2 is a schematic structural diagram of a hybrid neural network model in the dynamic prediction method for the shield tunneling attitude of the present invention.
Detailed Description
The invention is further described with reference to the following figures and specific examples.
The invention provides a dynamic prediction method of shield tunneling attitude, which comprises the steps of obtaining historical construction data, preprocessing the historical construction data to form sample data, inputting the sample data into a learning model to perform noise reduction, feature extraction and training learning to form a prediction model, inputting operation parameters into the prediction model in the actual construction process of a shield, outputting corresponding attitude data of the shield by the prediction model, constructing the shield according to the operation parameters if the attitude data accords with construction expectation, correspondingly modifying the operation parameters and inputting the operation parameters into the prediction model again to obtain corresponding attitude data if the attitude data does not accord with the construction expectation, solving the problem of poor control precision of the shield attitude, and dynamically predicting the real-time attitude and position of the shield by a dynamic prediction method based on a mixed deep learning model, so that a shield driver can adjust the shield attitude in advance according to a prediction result, reducing the influence of the shield on surrounding strata and attitude segments, and enabling the shield attitude to be consistent with a design axis as far as possible. The dynamic prediction method of the shield tunneling attitude of the invention is explained below with reference to the accompanying drawings.
Fig. 1 is a flow chart of the dynamic prediction method of the shield tunneling attitude of the present invention. The dynamic prediction method of the shield tunneling attitude of the present invention is described below with reference to fig. 1.
As shown in fig. 1, the invention provides a dynamic prediction method of shield tunneling attitude, comprising the following steps:
s11, data processing: acquiring historical construction data of shield construction, processing the historical construction data and using the historical construction data as sample data;
s12, model training: establishing a learning model, inputting sample data into the learning model, denoising the sample data by the learning model through set empirical mode decomposition to form a data set, and learning the data set through a convolutional neural network model with channel attention, a time sequence model and a time attention mechanism to form a prediction model;
s13, automatic prediction: and inputting the operation parameters of the shield in construction into the prediction model, so that the prediction model outputs corresponding attitude data of the shield.
Specifically, the shield attitude data includes horizontal deviation of the shield head, vertical deviation of the shield head, horizontal deviation of the shield tail, vertical deviation of the shield tail, gradient of the shield, and rotation angle of the shield.
As a preferred embodiment of the present invention, when the historical construction data is preprocessed, the method further includes:
and deleting data which is larger than the standard deviation of 3 times of the average value in the historical construction data, and deleting the historical construction data recorded when the shield stops moving in the shield construction.
Specifically, when historical construction data when the shield stops moving is deleted, the method further comprises the following steps:
when the total thrust, the total torque or the thrust speed of the shield is zero, the shield is in a stop state at the time point, and the historical construction data corresponding to the time point is deleted.
Specifically, the historical construction data comprises operation data automatically recorded by a PLC assembly of the shield according to a set frequency and observation data recorded by a constructor according to the granularity of each ring of pipe piece, and the recording frequency of the operation data is greater than that of the observation data;
the method comprises the steps of averaging observation data recorded in a set time interval to form a first data set, expanding the observation data through an interpolation method to form a second data set, enabling the number of data in the second data set to be matched with the number of data in the first data set, and further enabling the data in the first data set and the second data set to serve as sample data.
Specifically, when the historical construction data is preprocessed, the method further comprises the following steps:
and sequencing the importance of each parameter in the historical construction data, and performing weighted integration on sequencing results to obtain sample data.
Further, when training and forming the prediction model, the method further comprises:
the shape of the data set is maintained in the channel attention model by using a 1-by-1 convolution kernel, and the input shape of the time sequence model is protected by a dimension replacement layer, so that the time expansion steps of the cycle kernel of the time sequence model are consistent with the time dimension of the data set.
Further, the method also comprises the following steps:
providing a plurality of test data, inputting the test data into the prediction model and correspondingly obtaining attitude data, comparing the attitude data with actual attitude data corresponding to the test data to obtain a comparison result, and adjusting the hyper-parameters of the prediction model according to the comparison result.
Preferably, the hyper-parameters of the predictive model are adjusted using an adaptive moment estimation algorithm.
Still preferably, it further comprises:
after the historical construction data are preprocessed, 80% of the historical construction data are randomly selected as sample data, and the rest of the historical construction data are used as test data.
The specific embodiment of the invention is as follows:
acquiring historical construction data of shield construction, deleting data which is larger than 3 times of standard deviation of the average value in the historical construction data, judging that the shield at the time point is in a shutdown state when the total thrust, the total torque or the thrust speed of the shield is zero, and deleting all the historical construction data at the time point;
and the historical construction data comprises operation data automatically recorded by a PLC assembly of the shield according to a set frequency and observation data recorded by a constructor according to the granularity of each ring of pipe pieces, the recording frequency of the operation data is greater than that of the observation data, the observation data recorded in each 60 seconds can be averaged to form a first data set, the observation data is expanded by an interpolation method to form a second data set, the number of data in the second data set is matched with that of the data in the first data set, the data in the first data set and the data in the second data set are used as sample data, and the calculation formula of the interpolation method is as follows:
wherein X (t) is the variable condition of a t ring, m is the number of samples of the ring with the granularity of 60 seconds, and i is the time sequence number of the samples;
too many feature parameters increase the complexity and overfitting probability of the deep learning network model. Discarding some of the useless features may improve the training speed and accuracy of the deep learning network. Common feature screening methods can be divided into three categories: (a) filtration (b) wrapped and embedded (c). The filtering method is used for sequencing the feature importance by using a statistical mode and has the advantage of high calculation speed. The wrapping method is used for forming the sequence of the importance degree of the features by calculating the weight of the features and deleting the features with smaller weight absolute values. The embedded method integrates the feature selection and the machine learning model training process, namely, the feature importance degree sequencing is automatically carried out in the machine learning model training process. The three feature screening methods have respective superiority in the feature screening task, and the feature parameters with high correlation with the prediction parameters are better obtained by combining the three feature screening methods.
The shield will generate more than one thousand characteristic parameters during the operation stage. Parameters (such as voltage, current, power and the like) irrelevant to the operation attitude of the shield are removed before feature screening by combining expert experience and domain knowledge in research. Preprocessing is carried out on abnormal data (missing and errors) of the original data, and data with different granularities are fused. And two classical methods are selected for each of the three screening modes to rank the feature importance. And performing weighted integration (formula 1) on the sequencing results of all the parameters to obtain a final comprehensive characteristic evaluation result. The filtration method selects a Pearson correlation coefficient and a kendaltau correlation coefficient, and evaluates the influence relationship of the parameters on the shield attitude from linear and nonlinear angles respectively. Wherein the Pearson correlation coefficient may be defined by formula (2) and the kendallta correlation coefficient may be defined by formula (3). The wrapping method selects two models of rfe _ LR and rfe _ SVC (rfe-Support Vector Classification) to carry out recursive deletion on the features, and judges the importance sequence of the features to prediction by using the accuracy of the models. The embedded method selects a Catboost model and a Randomforest model, and utilizes the evaluation indexes of the tree model to screen the characteristics.
R total ＝ω 1 ·R Pearson +ω 2 ·R Kendall +ω 3 ·R RFE-LR +ω 4 ·R RFE-SVC +ω 5 ·R CatBoost +ω 6 ·R RandomForest (1)
In the formula, R is the order of the feature importance, and ω is the weight coefficient of the feature (in this embodiment, R is
)。
The calculated importance ranks as follows:
by adding white noise before data processing, a signal component having a physical meaning is obtained. EEMD inherits the adaptivity of EMD algorithm, and simultaneously EEMD also solves the modal aliasing problem of EMD algorithm. EEMD can be defined by equation (4) -equation (9):
wherein U is 1 Is a fitted curve (upper envelope curve) of the data maximum points, L 1 Is a fitted curve of data minimum points (lower envelope curve), A 1 Is the average of the upper and lower envelope curves.
h 1 ＝X(t)-a 1 (5)
Where X (t) is the raw data curve, h 1 Is the first difference between the original data and the average curve of the envelope.
h 1k ＝h 1(k-1) -a 1k (6)
Wherein a is 1k Definition of (a) and 1 is similarly defined as can be represented by h 1(k-1) The upper and lower envelope curves of (1) and the average curve thereof.
Repeating the above process until r remains n When it becomes a monotonic function, r at this time n Is r rest I.e. the final residual value.
r 1 -c 1 ＝r 2 ,…,r n-1 -c n ＝r n (8)
The operation data acquired by the shield automatic acquisition system generally contains a large amount of noise and cannot be directly used in shield prediction. In conjunction with the time-granular nature of the engineering data set used in the study (data extracted at a granularity of every minute), the frequency of noise caused by the machine during propulsion is much higher than the data used in the study. In the noise reduction and stabilization of shield parameter data, the decomposed high-frequency IMF (Intrinsic Mode Function) component is taken as disturbance noise to be removed, and the data is reconstructed, so that the influence of mechanical inherent noise on the data can be effectively reduced. The low-frequency IMF (Intrinsic Mode Function) component can smoothly show the change of the parameter data of the shield during movement. Compared with the high-frequency IMF, the amplitude of the low-frequency IMF is greatly changed, and the data set condition with minute as time granularity is met. And selecting the data of each ring of pipe pieces for processing.
The time attention mechanism effectively solves the above problem by aggregating all hidden states and reinforcing the important moment data with a weight matrix. In the study, a soft attention mechanism was used to determine a weight matrix for each time instant, whose weight expression can be defined as:
wherein h is t In order to be in the current target state,
the states at all times.
Gated cyclic units (GRUs) are one of the commonly used gated cyclic neural networks. Current time step X t Hidden state H with last time step t-1 As a reset gate R t And a refresh door Z t The output result in the current state can be obtained through the full connection layer (the activation function is sigmoid). By introducing the candidate hidden state to further assist the calculation of the subsequent hidden state, the formula (13) can see that the reset gate controls the inflow of the previous time step information to the current time step candidate hidden information, thereby controlling the loss and retention of the short-term history information. The update gate (equation (14)) can control the hidden state to beThe current candidate hidden state is updated, which is a long-term dependency.
R t ＝σ(X t H xr +H t-1 W hr +b r ) (11)
Z t ＝σ(X t H xz +H t-1 W hz +b z ) (12)
H t ＝tanh(X t W xh +(R t ⊙H t-1 )W hh +b h ) (13)
Wherein h is the number of hidden units, t is the number of time steps, W is a weight parameter, b is a bias parameter, l is an element multiplication, σ is a sigmoid function, and tanh is a tanh function.
Learning a data set through a channel attention model, a time sequence model and a time attention mechanism, maintaining the shape of the data set by using a 1 × 1 convolution kernel in the channel attention model, protecting the input shape of the time sequence model through a dimension displacement layer, and enabling the time expansion step number of the circulation kernel of the time sequence model to be consistent with the time dimension of the data set, wherein the neural network can be composed of an eemd filtering layer, a channel attention mechanism convolution layer changed from VGG7, a dimension displacement layer, a gating circulation layer with 16 hidden neurons, a time attention layer and a full connection layer with 16 hidden neurons, and is combined with the neural network shown in FIG. 2;
verifying the prediction model, inputting test data into the prediction model and correspondingly obtaining attitude data, comparing a plurality of attitude data with actual attitude data corresponding to the test data to obtain a comparison result, adjusting the hyper-parameters of the prediction model according to the comparison result, using Adam as an optimizer of the prediction model, using MSE as a loss function of the model in a training process, obtaining the optimal learning rate of the prediction model of 0.01, the batch size of the model of 20, training for 200epoch, and a verification result showing that the attitude prediction value of the shield is closer to the actual condition, the root mean square error of the two is within the range of 0.07-0.036, wherein the calculation formula is as follows:
wherein y is i Is to input x i Corresponding desired output value, f (x) i ) Is to input x i The corresponding model output value,/, denotes the number of data samples;
inputting the operation parameters of the shield in the construction process into a prediction model, so that the prediction model outputs corresponding shield attitude data, if the shield attitude data meets the design requirements, setting the shield according to the operation parameters, if the shield attitude data does not meet the design requirements, modifying the operation parameters by referring to the output shield attitude data, after modification, inputting the prediction model and outputting the corresponding shield attitude data, and when the shield attitude data meets the design requirements, setting the shield according to the operation parameters, so that the shield can be constructed according to the set route.
We have deployed the proposed predictive model (CNNCA + GRUTA) and three other models (including multi-layer perceptron, RNN, convLSTM) on the dataset of the present invention. The multilayer perceptron is the simplest neural network structure and has the advantages of high calculation speed and convenience in deployment. The invention uses a three-layer perceptron, belonging to a shallow neural network. RNNs are the most classical recurrent neural networks and have good results in processing data with sequence properties. ConvLSTM is a neural network newly proposed in recent years, and data do not need to be connected with a CNN and a time sequence model through a flat layer in the model training process, so that the data structure of the model can be better maintained. It is noted that all model uses are datasets that are eemd filtered. The predicted results of the model are shown in the following table:
as can be seen from the table, the accuracy of the model for predicting the shield position and attitude is reduced continuously with the increase of time. Due to the change of the construction condition of the shield in the tunneling process, the prediction capability of the model can be reduced by external unknown factors. Comparing MLP with other models shows that the long-term prediction ability of the shallow neural network on time series data is weaker than that of the model with the recurrent neural network. With increasing time instants, it is difficult for MLP to capture the more spaced dependencies in time series data.
The accuracy of the CACNN + GRUTA (we propose) model is less influenced with the increase of the prediction time, and the effectiveness of the TA mechanism in the model is further proved. The model with the CNN feature extraction can effectively extract and fuse important features in the aspect of prediction, and has great contribution to the prediction accuracy of the model. The traditional CNN feature extractor often loses the position information of the features. Loss of location information after data is fed into the predictor can cause the predictor to lose the temporal characteristics of the spatial dimension (relative location between different features). Compared with the prediction results of other models, the maximum error of the model can be reduced by 20-60%. In general, the accuracy and robustness of the CACNN + GRUTA hybrid depth model in predicting the time series length and prediction are far superior to other models.
While the present invention has been described in detail and with reference to the embodiments thereof as shown in the accompanying drawings, it will be apparent to one skilled in the art that various changes and modifications can be made therein. Therefore, certain details of the embodiments are not to be interpreted as limiting, and the scope of the invention is to be determined by the appended claims.
Claims (10)
1. A dynamic prediction method for shield tunneling attitude is characterized by comprising the following steps:
s11, data processing: acquiring historical construction data of shield construction, preprocessing the historical construction data and using the preprocessed historical construction data as sample data;
s12, model training: establishing a learning model, inputting the sample data into the learning model, denoising the sample data by the learning model through set empirical mode decomposition to form a data set, and learning the data set through a convolutional neural network model with channel attention, a time sequence model and a time attention mechanism to form a prediction model;
s13, attitude prediction: and inputting the operation parameters of the shield under construction into the prediction model, so that the prediction model outputs the corresponding attitude data of the shield.
2. The dynamic prediction method of the shield tunneling attitude of claim 1, wherein the preprocessing of the historical construction data further comprises:
and deleting the data which is larger than the standard deviation of 3 times of the average value in the historical construction data, and deleting the historical construction data recorded when the shield stops moving in the shield construction.
3. The dynamic prediction method of the shield tunneling attitude according to claim 2, wherein when historical construction data is deleted when the shield stops moving, the method further comprises:
when the total thrust, the total torque or the thrust speed of the shield is zero, the shield is in a stop state at the time point, and the historical construction data corresponding to the time point is deleted.
4. The dynamic prediction method of the shield tunneling attitude according to claim 1, wherein the historical construction data includes operation data automatically recorded by a PLC component of the shield according to a set frequency and observation data recorded by a constructor according to a granularity of each ring of segments, and a recording frequency of the operation data is greater than a recording frequency of the observation data;
averaging the observation data recorded within a set time interval to form a first data set, expanding the observation data by an interpolation method to form a second data set, so that the number of data in the second data set is matched with the number of data in the first data set, and further taking the data in the first data set and the second data set as the sample data.
5. The dynamic prediction method of the shield tunneling attitude of claim 1, wherein the preprocessing of the historical construction data further comprises:
and sequencing the importance of each parameter in the historical construction data, and performing weighted integration on sequencing results to obtain the sample data.
6. The dynamic prediction method of shield tunneling attitude of claim 1, wherein training to form the prediction model further comprises:
maintaining a shape of the data set using a 1 x 1 convolution kernel in the channel attention model, protecting an input shape of the temporal model by a dimension permutation layer such that a cycle kernel time expansion step number of the temporal model coincides with a time dimension of the data set.
7. The dynamic prediction method of shield tunneling attitude of claim 1, further comprising:
providing a plurality of test data, inputting the test data into the prediction model and correspondingly obtaining attitude data, comparing the attitude data with actual attitude data corresponding to the test data to obtain a comparison result, and adjusting the hyper-parameters of the prediction model according to the comparison result.
8. The dynamic prediction method of shield tunneling attitude of claim 7, wherein the hyper-parameters of the predictive model are adjusted using an adaptive moment estimation algorithm.
9. The dynamic prediction method of shield tunneling attitude of claim 7, further comprising:
and after preprocessing the historical construction data, randomly selecting 80% of the historical construction data as the sample data, and using the rest historical construction data as the test data.
10. The method of claim 1, wherein the shield attitude data includes horizontal deviation of the shield head, vertical deviation of the shield head, horizontal deviation of the shield tail, vertical deviation of the shield tail, gradient of the shield, and rotation angle of the shield.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211250468.7A CN115522945A (en) | 2022-10-12 | 2022-10-12 | Dynamic prediction method for shield tunneling attitude |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211250468.7A CN115522945A (en) | 2022-10-12 | 2022-10-12 | Dynamic prediction method for shield tunneling attitude |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115522945A true CN115522945A (en) | 2022-12-27 |
Family
ID=84701277
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211250468.7A Pending CN115522945A (en) | 2022-10-12 | 2022-10-12 | Dynamic prediction method for shield tunneling attitude |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115522945A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115982515A (en) * | 2023-01-05 | 2023-04-18 | 西南交通大学 | Method for obtaining optimal value of attitude control parameter of shield tunneling machine |
-
2022
- 2022-10-12 CN CN202211250468.7A patent/CN115522945A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN115982515A (en) * | 2023-01-05 | 2023-04-18 | 西南交通大学 | Method for obtaining optimal value of attitude control parameter of shield tunneling machine |
| CN115982515B (en) * | 2023-01-05 | 2023-09-29 | 西南交通大学 | Method for obtaining optimal value of attitude control parameter of shield tunneling machine |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110942101B (en) | Rolling bearing residual life prediction method based on depth generation type countermeasure network | |
| CN110619430B (en) | Space-time attention mechanism method for traffic prediction | |
| CN110941928A (en) | Rolling bearing residual life prediction method based on dropout-SAE and Bi-LSTM | |
| CN110163433B (en) | Ship flow prediction method | |
| CN112561148A (en) | Ship track prediction method and system based on one-dimensional convolutional neural network and LSTM | |
| CN112183993B (en) | TBM (tunnel boring machine) tunneling control parameter intelligent prediction and optimization decision method | |
| CN110728698B (en) | Multi-target tracking system based on composite cyclic neural network system | |
| CN115522945A (en) | Dynamic prediction method for shield tunneling attitude | |
| CN110083065B (en) | Self-adaptive soft measurement method based on flow type variational Bayesian supervised factor analysis | |
| CN112213771A (en) | Seismic wave impedance inversion method and device | |
| CN115828754A (en) | Cutter wear state monitoring method based on multi-scale space-time fusion network model | |
| CN111460737B (en) | Intelligent settlement prediction method and system for slurry air pressure balance shield | |
| CN114090718A (en) | Bi-LSTM prediction and fuzzy analysis based interrupted track correlation method | |
| CN115630742A (en) | Weather prediction method and system based on self-supervision pre-training | |
| Liu et al. | Realtime prediction of hard rock TBM advance rate using temporal convolutional network (TCN) with tunnel construction big data | |
| CN111241749B (en) | Permanent magnet synchronous motor chaos prediction method based on reserve pool calculation | |
| CN116451162A (en) | Industrial modeling system and method based on federal life learning under cloud edge cooperative scene | |
| CN111413931A (en) | Self-adaptive industrial process monitoring method based on sliding window recursion principal component analysis | |
| CN116595465A (en) | High-dimensional sparse data outlier detection method and system based on self-encoder and data enhancement | |
| CN114297795B (en) | PR-Trans-based mechanical equipment residual life prediction method | |
| CN110826184B (en) | Variational Bayesian identification method for NARX model structure and parameters under time-varying lag | |
| Li et al. | Policy gradient methods with gaussian process modelling acceleration | |
| CN115879569B (en) | Online learning method and system for IoT observation data | |
| Raisi et al. | Value Summation: A novel scoring function for MPC-based model-based reinforcement learning | |
| Shen et al. | Relative motion prediction of pontoon bridge module offshore connection based on deep learning |
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 |