CN115859826A - Integrated learning-based shield equipment fault fusion prediction method - Google Patents
Integrated learning-based shield equipment fault fusion prediction method Download PDFInfo
- Publication number
- CN115859826A CN115859826A CN202211614749.6A CN202211614749A CN115859826A CN 115859826 A CN115859826 A CN 115859826A CN 202211614749 A CN202211614749 A CN 202211614749A CN 115859826 A CN115859826 A CN 115859826A
- Authority
- CN
- China
- Prior art keywords
- data
- fault
- model
- shield
- 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
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention provides a shield equipment fault fusion prediction method based on ensemble learning, which comprises the following steps: s1, collecting data of a shield machine construction process of related projects; s2, processing a series of data on the original shield data; s3, constructing a Stacking shield fault prediction model; s4, training the integrated model to obtain a preliminary Stacking integrated model fault prediction result; and S5, training the single model by using a Bayesian optimization method, finding the optimal hyper-parameter combination, and outputting the optimal fault prediction result to form an optimal single model. The invention has the following beneficial effects: the equipment failure occurrence probability is reduced, the equipment maintenance cost is reduced, and the economic loss caused by equipment shutdown is avoided; the probability of the problems of progress delay, construction period delay and the like caused by the failure of the shield equipment is reduced, the smooth proceeding of the project construction process is further ensured, the working efficiency is improved, and the project construction quality is guaranteed at the same time.
Description
Technical Field
The invention belongs to the field of shield equipment fault prediction, and particularly relates to a shield equipment fault fusion prediction method based on ensemble learning.
Background
The method adopts a moving window mode to carry out fault diagnosis on the hydraulic drive cutter head system of the shield tunneling machine, and calculates the T of the prediction error of an adaptive AR model of new sample data on line 2 And monitoring the out-of-control point by using the statistical quantity so as to realize the online monitoring and fault diagnosis of the multivariable statistical process of the cutter head system.
The method is a shield machine fault prediction method (application number: CN 201910517455.3) based on LSTM, which predicts the fault of the shield machine by using the LSTM, and predicts the fault condition of the shield machine at the future time by using construction time sequence data aiming at the problem that a plurality of subsystems have faults simultaneously in shield construction.
The method is a failure prediction and diagnosis control method (application number: CN 201610972597.5) suitable for the shield machine, a failure fault tree model of the shield machine is established by analyzing a failure mechanism of the shield machine, a discrete-time Bayesian network of the shield machine failure corresponding to the failure tree model is obtained according to the established fault tree model, and the failure prediction of the shield machine is carried out by adopting a forward reasoning technology and a posterior probability of the Bayesian network.
The method 1 analyzes the cause of the failure of the shield machine in the construction process, but does not pre-judge the failure to be caused in the shield construction process in advance; when the 2 nd and 3 rd methods are used for predicting the shield machine fault, only a single method is used for constructing a shield fault prediction model, so that the precision is limited, and the complex change of the shield machine fault in the construction process is difficult to deal with.
Therefore, aiming at some defects in the method, the invention provides a shield equipment fault fusion prediction method based on ensemble learning.
Disclosure of Invention
The invention provides a shield equipment fault fusion prediction method based on ensemble learning, aiming at realizing the prediction of the fault condition of shield equipment during tunneling at the future time from shield equipment operation data, preventing and reducing the fault problem in the construction process in time and pertinently, and further having practical application value for reducing the equipment fault occurrence probability, reducing the equipment maintenance cost and guaranteeing the project construction progress, safety and quality.
In order to solve the problems, the invention provides a shield equipment fault fusion prediction method based on ensemble learning, which comprises the following steps:
s1: collecting construction data, construction fault data, geological data and risk source data of a shield machine construction process of related projects;
s2: performing a series of data processing including data preprocessing, data set balancing, integrated processing and data standardization on original shield construction data, construction fault data, geological data and risk source data to serve as data for training a model;
the failure type used as prediction can be comprehensively selected according to two indexes of failure severity obtained by failure occurrence frequency and expert grading;
s3: constructing a Stacking shield fault prediction model, respectively establishing single fault prediction models based on four algorithms of SVM, kNN, RF and XGboost, and training the four single models by utilizing an experimental data set divided by a K-fold cross verification method to obtain a fault prediction result of each model;
s4: taking the four single models trained in the step S3 as prediction models of the layer 1 in the Stacking, fusing prediction values output by the four models into a new data set to be used as an input of a prediction model of the layer 2, namely a logistic regression model LR, and training the new data set to obtain a preliminary failure prediction result of the Stacking integrated model;
s5: respectively adjusting and optimizing the hyper-parameters of the 4 single models established in the S3 by using a Bayesian optimization method, training the hyper-parameters, finding the optimal hyper-parameter combination, and outputting the optimal fault prediction result to form an optimal single model;
s6: and fusing the four optimal single models again by using the Stacking so as to form an optimal Stacking integrated model, and finally outputting an optimized integrated fault prediction result.
Preferably, the construction data, the construction fault data, the geological data and the risk source data in S1 are specifically as follows:
construction data: recording important information during the excavation operation of the shield machine, wherein each piece of data comprises values of a plurality of construction parameters including excavation time, excavation ring number, propulsion displacement, propulsion speed and cutter head rotating speed;
construction fault data: recording key information of abnormal states of the shield tunneling machine during tunneling operation, wherein the key information comprises fault content, fault occurrence time, recovery time and duration;
geological data: project geological information obtained through field surveying comprises geological types, a starting ring number and an ending ring number, wherein the geological types comprise compact pebble soil, sandy pebble, moderately weathered sandstone, sandy pebble and moderately weathered mudstone composite geology;
risk source data: the information of the shield tunneling machine passing through some important buildings is recorded, wherein the information comprises risk source types, starting ring numbers and ending ring numbers, and the risk source types comprise bridges, roads, pipelines, rivers, houses and shield launching.
Preferably, the step of S2 is specifically as follows:
s21 data preprocessing
S211 missing data: processing the missing data according to the actual situation, and deleting the data if the sampled data without any parameter is in a certain time point; if the data of partial parameters in a sampled certain time point is missing, filling up null values by using the statistical values of the corresponding parameters according to specific data conditions;
s212 repeating data: deleting the existing repeated data;
s213 abnormal data: taking a reservation for the abnormal data;
s22 data integration
On the basis of construction data, adding construction fault, geology and risk source information into the same data set, namely adding 3 rows of artificial labels in the construction data set according to the tunneling time and the tunneling ring number of the construction data: a fault status column, a geological type column, and a source of risk type column; the specific processing mode of the 3 columns of artificial labels is as follows: the fault state column is used for marking whether each piece of data in the construction data has a fault or not according to the fault occurrence time and the end time recorded in the construction fault data and by combining the tunneling time in the construction data; the geological type column is to use pinyin initial capital to represent the geological types recorded in the geological data, and mark the current geological condition for each piece of data according to the number of tunneling rings in the construction data; the risk source type column is used for representing the risk source type recorded in the risk source data in English, marking the current risk source condition for each piece of data according to the number of tunneling rings in the construction data, and representing the risk source type by using a null value if the risk source type does not pass through the risk source;
s23 Balancing the data set
Extracting an original data set, namely setting a time window to be n, setting a node of a shield machine which is converted from a normal state to a fault state to be i, extracting the first n pieces of data of the node i as data with a label of '1', and taking the first 2n to n pieces of data of the node i as data with a label of '0', so as to achieve the purpose of balancing the data set;
s24 data normalization
And (4) carrying out standardization treatment on the treated new data set, namely all the construction parameter values are in the same quantity level, so that the subsequent model training speed is not influenced.
Preferably, the specific steps of constructing the Stacking shield fault prediction model in S3 are as follows:
constructing a two-layer Stacking integrated model, selecting an algorithm with excellent performance and large difference by a layer 1 base learner, and selecting a simple and effective algorithm by a layer 2 meta-learner to prevent overfitting of the model; the Stacking integration model selects SVM, kNN, RF and XGboost as a base learner at the 1 st layer, and the meta-learner selected at the 2 nd layer is a simple linear model, namely a logistic regression model LR.
Preferably, the specific procedure of training the model in S3 is as follows:
s31, dividing the sample data set into 5 sub-data sets with equal size by a K-fold cross verification method, selecting 4 sub-data sets as training data to train a base learner, and using the remaining 1 sub-data set as verification data;
the S32 SVM, kNN, RF and XGboost 4 base learners respectively output a prediction result for each test data set.
Preferably, the specific process of S5 is as follows:
s51, adjusting and optimizing the hyper-parameters of each single model at the layer 1 in the Stacking by using a Bayesian optimization method to enable the hyper-parameters to reach a state with an optimal prediction effect, and outputting an optimal prediction result to generate a new data set;
s52, training the LR model of the layer 2 based on the new data set generated in S51 to obtain the final optimized integrated prediction result, so that the training of the shield fault prediction model based on Bayesian optimization Stacking integrated learning is completed.
Preferably, the process of using the bayesian optimization single model in S51 specifically includes the following steps:
s511, randomly generating initialization hyper-parameters in a search range set by hyper-parameters of the 4 single models, inputting the generated initialization hyper-parameters into a TPE algorithm, inputting an experimental data set into a single-classification fault prediction model to obtain a fault prediction result, and correcting the TPE algorithm to enable the output result of the model to be closer to a true value;
s512, the collection function EI is used for actively selecting the next most potential hyper-parameter combination point x from the corrected TPE algorithm (i) The point can maximize EI, so that the TPE algorithm is closer to the real distribution of the objective function relative to other hyperparameter combination points;
s513, if the preset target is met, stopping algorithm execution and exiting, and outputting a corresponding optimal hyper-parameter combination and an optimal value of a target function; and if the preset target is not met, inputting the output hyper-parameter combination and the objective function value into the TPE algorithm, correcting the TPE algorithm again, and executing S52 again until the preset target is met.
Preferably, the method further comprises the following step S7 of predicting the data:
and processing the new data, inputting a shield fault prediction model based on Bayesian optimization and Stacking, namely predicting the fault condition of the shield machine in the future time period, correspondingly evaluating the effect of the model, and improving and applying the model.
Aiming at the problem that the shield machine breaks down in the construction process, the invention designs a shield fault prediction model based on Bayesian optimization Stacking, and performs experiments by using shield machine operation data in a certain urban subway tunnel project, and the result shows that the model has better prediction effect in fault prediction compared with a single model. The positive effects brought are as follows:
(1) The method can help constructors to carry out fault prejudgment timely and effectively, avoid hidden trouble accidents in the construction process in advance, and ensure the safety and reliability of equipment in operation, thereby effectively reducing the construction safety risk and ensuring the construction safety of related workers;
(2) The equipment failure occurrence probability is reduced, the equipment maintenance cost is reduced, and the economic loss caused by equipment shutdown is avoided;
(3) The probability of the problems of progress delay, construction period delay and the like caused by the failure of the shield equipment is reduced, the smooth proceeding of the project construction process is further ensured, the working efficiency is improved, and the project construction quality is guaranteed at the same time.
Drawings
FIG. 1 is a flow chart of shield fault prediction;
FIG. 2 is a diagram of a Stacking ensemble learning method;
FIG. 3 is a fault multi-model fusion design diagram based on Stacking;
FIG. 4 is a pseudo code diagram of a training process of the Stacking integration model;
FIG. 5 is a Bayesian optimization single model flow chart;
FIG. 6 is a diagram of a shield fault prediction model based on Bayesian optimization Stacking.
Detailed Description
The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.
It should be noted that the embodiments and features of the embodiments may be combined with each other without conflict.
The invention hopes to predict whether the shield equipment fails in a later period of time through the proposed prediction model and the historical data of the shield equipment construction process. The constructed shield fault prediction process is shown in figure 1, wherein the specific process steps are as follows:
(1) And acquiring data of the shield equipment in the construction process.
(2) And (4) processing data.
2.1 Description of data
The data selected by the method comprises construction data, construction fault data, geological data and risk source data in the shield construction process.
(1) Construction data: important information during the excavation operation of the shield machine is recorded, wherein each piece of data comprises values of a plurality of construction parameters such as excavation time, excavation ring number, propulsion displacement, propulsion speed, cutter head rotating speed and the like.
(2) Construction fault data: the method records key information of abnormal states of the shield tunneling machine during tunneling operation, and the key information comprises fault content, fault occurrence time, recovery time and duration.
(3) Geological data: project geological information obtained by field surveying through relevant experts and technicians comprises geological types, starting ring numbers and ending ring numbers, wherein the geological types comprise compact pebble soil, sandy pebble, moderately weathered sandstone and sandy pebble, stroke-weathered mudstone composite geology and the like.
(4) Risk source data: information of passing through some important buildings (structures) when the shield machine drives is recorded, wherein the information comprises risk source types, starting ring numbers and ending ring numbers, wherein the risk source types comprise bridges, roads, pipelines, rivers, houses, shield starting and the like.
2.2 Data preprocessing
The construction data and the construction fault data of the shield machine are acquired by the sensor, and due to the factors of severe underground construction environment, poor network signals and the like, the sensor is unstable, so that the problems of missing, repetition, abnormality and the like exist in the construction data and the construction fault in the data acquisition and transmission process. In order to make the training time of the model and the generated result not influenced, the following methods are adopted for the two types of data before the model training is carried out:
(1) missing data: the missing data in the data used in the present invention is handled as it is. If no data of any parameter exists in a certain sampling time point, deleting the data; if the data of the partial parameters in a certain sampling time point is missing, filling the null value by using the statistical value of the corresponding parameters according to the specific data condition.
(2) Repeating data: when the network signal is not good, the sensor may intermittently acquire data, which may cause that a plurality of pieces of data are stored at the same time, that is, the acquired repeated data. And thus removes the duplicate data that exists.
(3) Abnormal data: and considering that the abnormal values in the data and the construction faults of the shield machine equipment may have certain connection, and therefore, the retention is taken for the abnormal data.
2.3 ) data integration
On the basis of construction data, adding construction fault, geology and risk source information into the same data set, namely adding 3 rows of artificial labels in the construction data set according to the tunneling time and the tunneling ring number of the construction data: a failure status (risk) column, a geology type (geologtype) column, and a risk source type (riskSourceType) column. See table 1 for details. The specific processing mode of the 3 columns of artificial labels is as follows: the risk column is used for marking whether each piece of data in the construction data has a fault or not according to the fault occurrence time and the end time recorded in the construction fault data and by combining the tunneling time in the construction data; the geologType column indicates the geological types recorded in the geological data by capitalization of pinyin initial letters, and marks the current geological condition for each piece of data according to the number of tunneling rings in the construction data; the riskSourceType column is used for representing the risk source type recorded in the risk source data in English, marking the current risk source condition of each piece of data according to the number of tunneling rings in the construction data, and if the risk source does not pass through, representing the risk source by using a null value (none).
TABLE 1 manual tag instruction sheet
2.4 ) balanced datasets
Construction fault data generated in the construction process of the shield machine is less than construction data, and faults occurring each time are not continuous. When the model is trained, the large difference of the sample quantity causes the classifier to be more inclined to learn large samples and difficult to identify small samples, so that balanced and accurate learning cannot be completed. In order to avoid this situation, the original data set needs to be extracted, that is, a time window is set to be n, a node where the shield machine is changed from the normal state to the fault state is i, the first n pieces of data of the node i are extracted as data with a label of "1", and the first 2n to n pieces of data of the node i are extracted as data with a label of "0", so as to achieve the purpose of balancing the data set.
2.5 Data normalization
A plurality of construction parameters are recorded in the construction data of the shield machine, and the measuring range of the construction parameters is greatly different, such as: the range of the propulsion pressure is 0-1000 bar, and the range of the cutter head abrasion pressure is 0-10 bar. In order to eliminate such a difference between the data, the processed new data set is subjected to standardization processing, that is, all construction parameter values are on the same quantity level, so that the subsequent model training speed is not influenced.
(3) And constructing a Stacking shield fault prediction model.
3.1 Introduction to Stacking Algorithm
The Stacking method is a typical ensemble learning method, and by fusing a plurality of machine learning algorithm models, advantages of the machine learning algorithm models are complementary, so that the overall prediction capability is improved, and the obtained prediction result is better than that of each single machine learning algorithm model. It typically has two layers: the learner at level 1 is referred to as the base learner; the learner at level 2 is referred to as a meta-learner and is primarily used to fuse the results of the learner at level 1. The basic idea of the Stacking algorithm is as follows: firstly, dividing an original data set into a plurality of subdata sets, and using the subdata sets to train each base learner at the layer 1, wherein each base learner can output a corresponding training result; then, regarding the output result of the 1 st layer base learner as the input of the 2 nd layer meta learner; and finally, outputting a final prediction result by the 2 nd-layer meta learner through training. The specific structure is shown in fig. 2.
During the training process of Stacking, if the training data set using the training base learner is also used to train the meta learner, a high overfitting condition may occur. Therefore, K-fold cross validation can be used to avoid this situation, and the specific training process of the Stacking algorithm is: for the original training data set S = { (x) i ,y i ) I =1,2, \8230;, n }, where x i Feature vector, y, representing the ith training sample i Representing a predicted value corresponding to the ith training sample; meanwhile, the number of each feature vector is defined as p, i.e., each feature vector is represented as (x) 1 ,x 2 ,…,x p ). Randomly dividing a training data set S into K mutually exclusive training subsets S with the same size 1 ,S 2 ,…,S K In which S is (-k) ＝S-S k Will S (-k) And S k Defined as the training set and test set of the K-fold in K-fold cross validation, respectively. The layer 1 comprises T learning algorithms, and the T learning algorithm is used in a training set S (-k) Base learning device obtained by training
T =1,2, \8230;, T. Test set S for the k-th fold k Each training sample x in (1) i Let Z ti Based learning device>
For x i ToAnd (6) measuring the result. After the complete cross-validation process, a new data set, S, is generated by the T base learners new ＝{(Z i1 ,Z i2 ,…,Z iT ,y i ) I =1,2, \8230;, n }. The new data set S generated new Using the data set S as the input data of the Stacking layer 2 new Training the learner at level 2 results in a meta-learner l'.
3.2 Stacking-based fault multi-model fusion design
The Stacking algorithm is used for analyzing data from multiple angles by fusing multiple machine learning algorithm models, and plays a role in complementing advantages so as to improve the overall prediction capability. In the modeling of the invention, a two-layer Stacking integrated model is constructed, the layer 1 base learner selects an algorithm with excellent performance and large difference, and the layer 2 meta learner selects a simple and effective algorithm to prevent the model from being over-fitted. Support Vector Machines (SVM) have particular advantages for solving non-linearity and high latitude problems. The k-nearest neighbor (kNN) algorithm is simple in concept, mature in theory and suitable for the classification problem with large sample capacity. Random Forest (RF) is an algorithm integrating a plurality of decision trees through Bagging integration idea, is not easy to overfit when applied to a large data set, and has the capacity of resisting noise. The XGboost is an ensemble learning method which is optimized in a Boosting mode and has good performance, regularization is added to effectively prevent overfitting, and the XGboost has the characteristics of high training speed and high accuracy.
In summary, the Stacking integration model of the invention selects SVM, kNN, RF, XGBoost as the base learner at layer 1, and the meta learner selected at layer 2 is a simple linear model LR (logistic regression model). The specific model structure is shown in fig. 3, and the pseudo code of the model training process is shown in fig. 4.
(4) And training the model.
4.1 Evaluation index of prediction model
The essence of the shield prediction problem of the invention is a two-classification problem, namely, two conditions of failure or no failure of the shield equipment are judged. Therefore, the evaluation indexes adopted by the invention aiming at the two-classification model are Accuracy (Accuracy), precision (Precision), recall (Recall) and F1 score (F1-score), which are respectively shown in (1) to (4). All 4 evaluation indexes are constructed based on a confusion matrix, as shown in table 2.
Defining the state that the prediction result is 1, and indicating that a fault occurs; the prediction result is in a "0" state, indicating that no failure has occurred.
TABLE 2 confusion matrix
(1) The accuracy is as follows: the proportion of correctly predicted samples to all samples is calculated according to the following formula:
(2) precision ratio: in this context, it is understood that the ratio of the number of samples in which a failure actually occurs to the total number of samples in which a failure is predicted is calculated as follows:
(3) and (3) recall ratio: in this context, it is understood that the ratio of the number of samples predicted to have a failure to the total number of samples actually having a failure is calculated as follows:
(4) f1 fraction: is a harmonic mean value of two indexes (2) and (3), and the calculation formula is as follows:
4.2 ) training procedure of model:
(1) dividing the sample data set into 5 sub-data sets with equal size by a 5-fold cross validation method, selecting 4 sub-data sets as training data to train a base learner, and using the remaining 1 sub-data set as validation data;
(2) the SVM, the kNN, the RF and the XGboost 4 base learners respectively output a prediction result for respective test data sets;
(3) combining the 4 output prediction results to serve as a new data set for training an LR model of a layer 2, wherein after the training of the Stacking shield fault prediction model is finished, the output result of the layer 2 is a preliminary integrated prediction result;
(3) based on the step (1), optimizing the hyper-parameters of each single model at the layer 1 in the Stacking by using a Bayesian optimization method to enable the hyper-parameters to reach a state with an optimal prediction effect, and generating a new data set by the output optimal prediction result;
(4) and (4) training the LR model of the layer 2 based on the new data set generated in the step (3) to obtain a final optimized integrated prediction result, so that the training of the shield fault prediction model based on Bayesian optimization Stacking integrated learning is completed.
(5) And (5) Bayesian optimization of a single model.
5.1 Bayesian optimization algorithm)
Bayesian Optimization (BO) is an effective global Optimization algorithm, and its basic idea is to first perform random sampling in a given parameter space to generate a preliminary objective function distribution, then combine with historical information to continuously search and iterate, and select the next sampling point according to the distribution until the distribution fitted by the selected sampling points is close to the actual objective function.
The probabilistic proxy model and the collection function are two core parts of Bayesian optimization. The Bayesian optimization algorithm constructed by the invention adopts a tree structure Parzen estimation method (TPE) as a probability agent model, and EI (Expected Improvement) as an acquisition function.
(1) A probability agent model: TPE (thermoplastic elastomer)
The TPE algorithm models p (x | y) using nonparametric density instead of building a prior probability distribution of hyper-parameters, using { x | y } (1) ,x (2) ,……x (k) Represents the hyperparameter and y represents the loss value. TPE uses two densities to define p (x | y), as shown in equation (5):
wherein l (x) represents the hyperparametric value { x (i) The density formed, and the corresponding loss y = f (x) (i) ) Less than y * (ii) a g (x) represents the density formed by the remaining over-parameter value; y is * The loss value is the optimal hyperparameter f (x). TPE Algorithm selection y * Some quantile γ as current over-parameter loss value y, such that p (y)<y * ) = γ, without requiring a specific model of p (y).
(2) Collecting a function: EI (El)
The definition of EI is shown in equation (6):
according to bayes' theorem, equation (6) can be rewritten as:
according to p (y)<y * ) = gamma and p (x) =: [ integral ] n R P (x | y) P (y) dy = γ l (x) + (1- γ) g (x), and formula (7) is rewritten to obtain:
equation (6) ends up with:
from the final expression (9), it can be seen that in order to maximize improvement, the probability of the hyperparameter x at l (x) is high and the probability at g (x) is low. According to
To evaluate each hyperparameter x, at each iteration the algorithm will return the hyperparameter x with the largest EI.
5.2 Bayesian optimization single model process
The detailed process of using the Bayesian optimization single model is shown in FIG. 5, and comprises the following steps:
(1) randomly generating initialization hyper-parameters in a search range set by each hyper-parameter of the 4 single models, inputting the generated initialization hyper-parameters into a TPE algorithm, inputting an experimental data set into a single-classification fault prediction model to obtain a fault prediction result, and correcting the TPE algorithm to enable the output result of the model to be closer to a true value;
(2) and actively selecting the next most potential hyper-parameter combination point x from the corrected TPE algorithm by utilizing the acquisition function EI (i) The point can maximize EI, so that the TPE algorithm is closer to the real distribution of the objective function relative to other hyperparameter combination points;
(3) if the preset target is met, stopping algorithm execution and exiting, and outputting a corresponding optimal hyper-parameter combination and a target function optimal value; and (3) if the preset target is not met, inputting the output hyper-parameter combination and the objective function value into a TPE algorithm, correcting the TPE algorithm again, and executing the step (2) again until the preset target is met.
(6) And constructing a shield fault prediction model based on Bayesian optimization and Stacking.
A training process for optimizing the Stacking shield fault prediction integrated model constructed by the counter based on the Bayesian optimization algorithm is shown in figure 6.
(7) And predicting the data.
And processing the new data, inputting a shield fault prediction model based on Bayesian optimization and Stacking, namely predicting the fault condition of the shield machine in the future time period, correspondingly evaluating the effect of the model, and improving and applying the model.
In order to make the objects, technical solutions and advantages of the present invention clearer and clearer, the present invention is further described with reference to the accompanying drawings, and the specific implementation steps are as follows:
s1: and collecting construction data, construction fault data, geological data and risk source data of the shield machine construction process of related projects.
S2: the original shield construction data, construction fault data, geological data and risk source data are subjected to a series of data processing including data preprocessing, data set balancing, integrated processing and data standardization and then serve as data for training a model.
The failure type used as the prediction can be comprehensively selected according to the failure occurrence frequency and the failure severity index obtained by expert grading.
S3: respectively establishing single fault prediction models based on four algorithms of SVM, kNN, RF and XGboost, and training the four single models by utilizing an experimental data set divided by a 5-fold cross-validation method to obtain a fault prediction result of each model.
S4: with reference to fig. 3 and 4, the four single models trained in S3 are used as the prediction models of the 1 st layer in the Stacking, the predicted values output by the four models are fused into a new data set to be used as the input of the 2 nd layer prediction model LR, and the new data set is trained to obtain a preliminary Stacking integrated model fault prediction result.
S5: and (3) with reference to fig. 5, respectively adjusting and optimizing the hyper-parameters of the 4 single models in the step (S3) by using a Bayesian optimization method, training the hyper-parameters, finding the optimal hyper-parameter combination, and outputting the optimal fault prediction result to form the optimal single model.
S6: and combining the graphs 6 and S4, fusing the four optimal single models again by using Stacking to form an optimal Stacking integrated model, and finally outputting an optimized integrated fault prediction result.
While the invention has been described with reference to a preferred embodiment, it will be understood by those skilled in the art that various changes in form and detail may be made therein without departing from the spirit and scope of the invention.
Claims (8)
1. A shield equipment fault fusion prediction method based on ensemble learning is characterized by comprising the following steps:
s1: collecting construction data, construction fault data, geological data and risk source data of a shield machine construction process of related projects;
s2: performing a series of data processing including data preprocessing, data set balancing, integrated processing and data standardization on original shield construction data, construction fault data, geological data and risk source data to serve as data for training a model;
the failure type used as the prediction can be comprehensively selected according to two indexes of failure severity obtained by failure occurrence frequency and expert grading;
s3: constructing a Stacking shield fault prediction model, respectively establishing single fault prediction models based on four algorithms of SVM, kNN, RF and XGboost, and training the four single models by utilizing an experimental data set divided by a K-fold cross verification method to obtain a fault prediction result of each model;
s4: taking the four single models trained in the step (S3) as a prediction model of a layer 1 in the Stacking, fusing predicted values output by the four models into a new data set, and taking the new data set as an input of a layer 2 prediction model, namely a logistic regression model LR, and training the new data set to obtain a preliminary Stacking integrated model fault prediction result;
s5: respectively adjusting and optimizing the hyper-parameters of the 4 single models established in the S3 by using a Bayesian optimization method, training the hyper-parameters, finding the optimal hyper-parameter combination, and outputting the optimal fault prediction result to form an optimal single model;
s6: and fusing the four optimal single models again by using the Stacking so as to form an optimal Stacking integrated model, and finally outputting an optimized integrated fault prediction result.
2. The method for predicting the fault fusion of the shield equipment based on the ensemble learning of claim 1, wherein the construction data, the construction fault data, the geological data and the risk source data in the step S1 are as follows:
construction data: recording important information during the excavation operation of the shield machine, wherein each piece of data comprises values of a plurality of construction parameters including excavation time, excavation ring number, propulsion displacement, propulsion speed and cutter head rotating speed;
construction fault data: recording key information of abnormal states of the shield tunneling machine during tunneling operation, wherein the key information comprises fault content, fault occurrence time, recovery time and duration;
geological data: project geological information obtained through field surveying comprises geological types, a starting ring number and an ending ring number, wherein the geological types comprise compact pebble soil, sandy pebble, moderately weathered sandstone, sandy pebble and moderately weathered mudstone composite geology;
risk source data: recording information of passing some important buildings when the shield machine drives, wherein the information comprises risk source types, starting ring numbers and ending ring numbers, and the risk source types comprise bridges, roads, pipelines, rivers, houses and shield starting.
3. The method for predicting shield equipment failure fusion based on ensemble learning according to claim 2, wherein the step S2 is specifically as follows:
s21 data preprocessing
S211 missing data: processing the missing data according to the actual situation, and deleting the data if the sampled data without any parameter is in a certain time point; if the data of partial parameters in a sampled certain time point is missing, filling up null values by using the statistical values of the corresponding parameters according to specific data conditions;
s212 repeating data: deleting the existing repeated data;
s213 abnormal data: taking reservation for abnormal data;
s22 data integration
On the basis of construction data, adding construction fault, geology and risk source information into the same data set, namely adding 3 rows of artificial labels in the construction data set according to the tunneling time and the tunneling ring number of the construction data: a fault status column, a geological type column, and a source of risk type column; the specific processing mode of the 3 columns of artificial labels is as follows: the fault state column is used for marking whether each piece of data in the construction data has a fault or not according to the fault occurrence time and the end time recorded in the construction fault data and by combining the tunneling time in the construction data; the geological type column is to use pinyin initial capital to represent the geological types recorded in the geological data, and mark the current geological condition for each piece of data according to the number of tunneling rings in the construction data; the risk source type column is used for representing the risk source type recorded in the risk source data in English, marking the current risk source condition for each piece of data according to the number of tunneling rings in the construction data, and representing the risk source type by using a null value if the risk source type does not pass through the risk source;
s23 Balancing the dataset
Extracting an original data set, namely setting a time window to be n, setting a node of a shield machine which is changed from a normal state to a fault state to be i, extracting the first n pieces of data of the node i as data with a label of '1', and taking the first 2n to n pieces of data of the node i as data with a label of '0', so as to achieve the purpose of balancing the data set;
s24 data normalization
And (4) carrying out standardization treatment on the treated new data set, namely all the construction parameter values are in the same quantity level, so that the subsequent model training speed is not influenced.
4. The method for predicting the fault fusion of the shield equipment based on the ensemble learning according to claim 3, wherein the step S3 of constructing the Stacking shield fault prediction model comprises the following steps:
constructing a two-layer Stacking integrated model, selecting an algorithm with excellent performance and large difference by a layer 1 base learner, and selecting a simple and effective algorithm by a layer 2 meta-learner to prevent overfitting of the model; the Stacking integration model selects SVM, kNN, RF and XGboost as a base learner at the 1 st layer, and the meta-learner selected at the 2 nd layer is a simple linear model, namely a logistic regression model LR.
5. The method for predicting the fault fusion of the shield equipment based on the ensemble learning according to claim 4, wherein the specific process of training the model in the step S3 is as follows:
s31, dividing the sample data set into 5 sub-data sets with equal size by a K-fold cross verification method, selecting 4 sub-data sets as training data to train a base learner, and using the remaining 1 sub-data set as verification data;
the S32 SVM, kNN, RF and XGboost 4 base learners respectively output a prediction result for each test data set.
6. The method for predicting the fault fusion of the shield equipment based on the ensemble learning according to claim 5, wherein the specific process in the step S5 is as follows:
s51, optimizing the hyper-parameters of each single model at the layer 1 in the Stacking by using a Bayesian optimization method to enable the hyper-parameters to reach the state with the optimal prediction effect, and merging the optimal results output by the 4 single models to generate a new data set;
s52, training the LR model of the layer 2 based on the new data set generated in S51 to obtain the final optimized integrated prediction result, so that the training of the shield fault prediction model based on Bayesian optimization Stacking integrated learning is completed.
7. The method for predicting the fault fusion of the shield equipment based on the ensemble learning according to claim 6, wherein the process of using the Bayesian optimization single model in the S5 is specifically as follows:
s511, initializing hyper-parameters are randomly generated in a search range set by each hyper-parameter of the 4 single models, the generated initializing parameters are input into a TPE algorithm, an experimental data set is input into a single-classification fault prediction model, a fault prediction result is obtained, and the TPE algorithm is corrected to enable the output result of the model to be closer to a true value;
s512, the acquisition function EI is utilized to actively select the next most potential hyper-parameter combination point x from the corrected TPE algorithm (i) The point can maximize EI, so that the TPE algorithm is closer to the real distribution of the objective function relative to other hyperparameter combination points;
s513, if the preset target is met, stopping algorithm execution and exiting, and outputting a corresponding optimal hyper-parameter combination and an optimal value of a target function; and if the preset target is not met, inputting the output hyper-parameter combination and the objective function value into the TPE algorithm, correcting the TPE algorithm again, and executing S52 again until the preset target is met.
8. The method for predicting the fault fusion of the shield equipment based on the ensemble learning is characterized by further comprising the following step S7 of predicting data:
and processing the new data, inputting the processed new data into a shield fault prediction model based on Bayesian optimization and Stacking, predicting the fault condition of the shield machine in the future time period, correspondingly evaluating the effect of the model, and improving and applying the model.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211614749.6A CN115859826A (en) | 2022-12-15 | 2022-12-15 | Integrated learning-based shield equipment fault fusion prediction method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202211614749.6A CN115859826A (en) | 2022-12-15 | 2022-12-15 | Integrated learning-based shield equipment fault fusion prediction method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN115859826A true CN115859826A (en) | 2023-03-28 |
Family
ID=85673240
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202211614749.6A Pending CN115859826A (en) | 2022-12-15 | 2022-12-15 | Integrated learning-based shield equipment fault fusion prediction method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN115859826A (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117725488A (en) * | 2024-02-06 | 2024-03-19 | 河北元泰建中项目管理有限公司 | Building engineering project safety performance prediction method and system based on machine learning |
-
2022
- 2022-12-15 CN CN202211614749.6A patent/CN115859826A/en active Pending
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN117725488A (en) * | 2024-02-06 | 2024-03-19 | 河北元泰建中项目管理有限公司 | Building engineering project safety performance prediction method and system based on machine learning |
| CN117725488B (en) * | 2024-02-06 | 2024-04-30 | 河北元泰建中项目管理有限公司 | Building engineering project safety performance prediction method and system based on machine learning |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111914873A (en) | Two-stage cloud server unsupervised anomaly prediction method | |
| CN110210169B (en) | LSTM-based shield tunneling machine fault prediction method | |
| CN103711523B (en) | Based on the gas density real-time predicting method of local decomposition-Evolutionary Neural Network | |
| CN104765825B (en) | Social networks link prediction method and device based on collaboration fusion principle | |
| CN107145634B (en) | Multi-state dynamic reliability assessment method for shield cutter head and driving system | |
| CN111597175B (en) | Filling method of sensor missing value fusing time-space information | |
| CN114662699A (en) | Shield attitude prediction method based on machine learning | |
| CN114201920A (en) | Laser cutting numerical control system fault diagnosis method based on digital twinning and deep transfer learning | |
| CN112560327B (en) | Bearing residual life prediction method based on depth gradient descent forest | |
| Zheng et al. | Real-time transient stability assessment based on deep recurrent neural network | |
| CN115859826A (en) | Integrated learning-based shield equipment fault fusion prediction method | |
| CN112879024A (en) | Dynamic prediction method, system and equipment for shield attitude | |
| CN116402352A (en) | Enterprise risk prediction method and device, electronic equipment and medium | |
| CN112132102A (en) | Intelligent fault diagnosis method combining deep neural network with artificial bee colony optimization | |
| CN113806889A (en) | Processing method, device and equipment of TBM cutter head torque real-time prediction model | |
| CN115081749A (en) | Bayesian optimization LSTM-based shield tunneling load advanced prediction method and system | |
| CN104732067A (en) | Industrial process modeling forecasting method oriented at flow object | |
| Zhang et al. | Cross-project prediction for rock mass using shuffled TBM big dataset and knowledge-based machine learning methods | |
| Liu et al. | Residual useful life prognosis of equipment based on modified hidden semi-Markov model with a co-evolutional optimization method | |
| CN117076993A (en) | Multi-agent game decision-making system and method based on cloud protogenesis | |
| CN111400964B (en) | Fault occurrence time prediction method and device | |
| CN114297795A (en) | Mechanical equipment residual life prediction method based on PR-Trans | |
| Ramachandran et al. | Case-Based Reasoning for System Anomaly Detection and Management | |
| CN110866607A (en) | Machine learning-based penetration behavior prediction algorithm | |
| Morgenroth et al. | Comparison of Bayesian Belief Networks and Artificial Neural Networks for prediction of tunnel ground class |
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 |