Real-time multi-step prediction method for tunneling key parameters of shield tunneling machine based on ST-GCN-LSTM

Abstract

The invention discloses a shield tunneling machine tunneling key parameter real-time multi-step prediction method based on ST-GCN-LSTM, and the method comprises the following steps: obtaining the operation data of a shield tunneling machine under a normal working condition, and carrying out the preprocessing of the operation data; analyzing the preprocessed operation data, and determining related variables of total thrust, cutterhead torque, cutterhead rotating speed, penetration and propelling speed average value tunneling key variables; respectively calculating the control limit of each tunneling key variable for different soil layers, and training a multi-step prediction model based on an ST-GCN and LSTM combined model; acquiring operation data of the shield tunneling machine in real time, inputting the operation data into the trained model for preprocessing, dynamically predicting tunneling key parameter values of multiple steps in the future, and judging whether predicted values continuously exceed a control limit or not; according to the method, real-time multi-step prediction of the tunneling key parameters can be carried out in combination with soil layer classification, and prediction inaccuracy caused by changes of external factors such as soil layers is reduced.

Description

Technical Field

[0001] The present invention relates to the technical field of shield machine key parameter prediction, and particularly relates to a real-time multi-step prediction method for key parameters of shield machine tunneling based on ST-GCN-LSTM. Background Technique

[0002] With the rapid development of the national economy, the construction of railway tunnels in China has developed rapidly. As an important means of modern tunnel construction, the shield tunneling technology involves various complex parameters during the construction process, such as propulsion speed, cutterhead torque, total thrust, etc. The real-time monitoring and prediction of these parameters are of great significance for ensuring construction safety and improving construction efficiency. However, due to the complex and variable working conditions involved in shield construction, traditional prediction methods are difficult to accurately capture the dynamic change laws of parameters. Therefore, using advanced machine learning methods for multi-step prediction has become a research hotspot.

[0003] Currently, there are few studies on classifying soil layers and training different models according to different soil layers. The shield machine is a large-scale underground operation equipment and will inevitably be affected by geological factors. Facing different soil layers, the tunneling parameters of the shield machine will change significantly. Due to the lack of consideration for soil layer classification, the prediction is inaccurate. Summary of the Invention

[0004] Object of the Invention: The object of the present invention is to provide a real-time multi-step prediction method for key parameters of shield machine tunneling based on ST-GCN-LSTM, which solves the problems of prediction lag, untimely response, weak early warning ability, etc. in traditional methods due to ignoring soil layer differences and only supporting single-step prediction. The present invention can predict the future trend of key tunneling parameters during the tunneling process of the shield machine in real time, so that the shield machine operator can adjust the set parameters of the shield machine in time according to whether the key tunneling parameters exceed the control limit.

[0005] Technical Solution: The real-time multi-step prediction method for key parameters of shield machine tunneling based on ST-GCN-LSTM described in the present invention includes the following steps: S1, based on the geological exploration report of a specific project, divide the soil layer types according to national standards; S2, obtain the operation data of the shield machine under normal working conditions, and preprocess the operation data, including removing shutdown data, segment assembly data and abnormal data; S3, analyze the preprocessed operation data to determine the relevant variables of the key variables of total thrust, cutterhead torque, cutterhead rotation speed, penetration degree, and average propulsion speed during tunneling; S4. Calculate the control limits of each tunneling key variable for different soil layers respectively, and train a multi-step prediction model based on the combined model of ST-GCN and LSTM. Among them, ST-GCN is used to extract the spatial features between the shield machine sensors; LSTM is used to capture the dynamic features of the time series. S5. Collect the operation data of the shield machine in real time, input it into the trained model for preprocessing, dynamically predict the tunneling key parameter values in the future for multiple steps, and judge whether the predicted values continuously exceed the control limits. The prediction accuracy is evaluated by the root mean square error RMSE, the mean absolute error MAE, and the accuracy ACC.

[0006] Further, step S1 specifically includes: mapping the shield tunneling ring number to the corresponding soil layer type according to the soil layer distribution area in the geological exploration report.

[0007] Further, the 3σ criterion is adopted for abnormal data rejection in step S2, which includes the following steps: S21. Calculate the mean μ and the standard deviation σ of the preprocessed data. S22. If the data point x satisfies x - μ > 3σ, it is determined as abnormal data and rejected.

[0008] Further, the standard deviation σ is calculated by the Bessel formula: ; where n represents the number of parameter data, represents the i-th parameter.

[0009] Further, the analysis of relevant variables in step S3 is realized through the XGBoost regression model, which includes the following steps: S31. Construct an XGBoost regression model for each tunneling key variable respectively. S32. Calculate the importance score FScore of each variable based on the feature split gain Gain. S33. Select the top 30 variables with the highest importance scores as relevant variables.

[0010] Further, the calculation formula of the feature split gain is: ; where, represents the total weight before splitting, and represent the total weights of the left child node and the right child node after splitting respectively.

[0011] S41. For the tunneling key variable data under each soil layer, use the Gaussian kernel function to estimate its probability density distribution. S42. Determine the control limit corresponding to the specified probability threshold p based on the cumulative distribution function. , the formula is as follows: ; Among them, represents the data of the tunneling key parameters under the j-th soil layer, is the cumulative distribution function CDF value, that is, the probability that the variable is less than or equal to , and p is the set probability threshold.

[0012] Furthermore, the construction of the ST-GCN-LSTM combined model includes: S43, set the time window length C and the output step length J, and construct a spatio-temporal graph structure; the nodes of the spatio-temporal graph are sensor variables, and the edge weights are the correlations between variables; S44, aggregate the node spatial features through multi-layer graph convolution, and the formula is: ; Among them, represents the hidden state of the nodes in the l-th layer, is the set of neighbor nodes of node v, is the weight of the graph convolution layer, is the activation function, is the normalization coefficient; S45, input the spatial features into the LSTM network, and capture the time-dependent relationship through the forget gate, input gate and output gate.

[0013] Furthermore, in step S5, the judgment criterion for continuous overrun is: an early warning is triggered when the predicted value exceeds the control limit in the next 3 consecutive time steps.

[0014] An electronic device according to the present invention includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of any one of the methods when executing the program.

[0015] Beneficial effects: Compared with the prior art, the present invention has the following remarkable advantages: the present invention can combine soil layer classification for real-time multi-step prediction of tunneling key parameters, reducing the inaccuracy of prediction caused by changes in external factors such as soil layers. ST-GCN can effectively capture the complex spatial dependence relationships between various components or sensors of the shield machine, especially when the mutual influence between multiple sensors or parameters is large. LSTM is good at processing time series data and can capture the long-term and short-term dependencies and change trends of parameters in the time dimension. Description of the Drawings

[0016] Figure 1 is the flowchart of the present invention; Figure 2 is the basic structure diagram of ST-GCN-LSTM of the present invention; Figure 3 It is the multi - prediction flow chart of the key parameters of shield tunneling in the present invention. Specific implementation manners

[0017] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0018] The embodiment of the present invention provides a cell segmentation method based on boundary uncertainty estimation, including the following steps: S1. According to the geological exploration report of a specific project, the soil layer types are divided according to the national standard. For example, according to the geological exploration report of the east line of Haizhu Bay, the soil layers are divided into four types according to the surrounding rock grades of the national standard.

[0019] S2. Obtain the operation data of the shield machine under normal working conditions of the project, and pre - process the operation data of the shield machine.

[0020] For the data collected by the shield machine, which is collected through the PLC system or sensors and contains shutdown and segment erection data, first, the shutdown and erection data and the data containing null values need to be removed, and the operation status data of each ring is retained.

[0021] In the shield construction data, the existing outliers will affect the prediction effect of the model. The present invention selects the 3σ criterion to remove outliers. For a set of data subject to normal distribution, the probability that its data values are distributed in (μ - 3σ, μ + 3σ) is 0.9974, where σ represents the standard deviation and μ represents the mean. The calculation formula is: ; If the residual error of a certain key parameter data of shield tunneling satisfies the absolute value , then this data is considered as abnormal data and is excluded.

[0022] S3. According to the pre - processed operation data, analyze the related variables of the key tunneling variables such as the total thrust, cutterhead torque, cutterhead rotation speed, penetration degree, and average propulsion speed.

[0023] When the shield machine works underground, it is equipped with a large number of sensors, and the real - time transmitted data reaches as many as thousands. If all variables are input into the model for training, it will cause a great deviation in the final prediction result, and the consumed resources and time are also very large. Therefore, for each key tunneling parameter, we need to separately perform variable screening on the pre - processed data. The present invention uses the XGBoost algorithm for regression prediction of variables to screen the related variables of the key shield tunneling variables, including the following steps: S31, the total thrust of the shield machine, cutter head torque, cutter head speed, penetration, and average thrust speed are five key parameters that are very important for the operation of the shield machine. X pre Select five variables, namely thrust, cutterhead torque, cutterhead speed, penetration, and average propulsion speed, and build an XGBoost regression tree model for each variable for regression prediction. ; n is the total number of sensors, vector x Represents the measurement values ​​of different sensors at different times over a period of time.

[0024] The regression prediction of variables using XGBoost is based on the tree model. When constructing the tree model, each level greedily selects a feature segmentation point as a leaf node to maximize the overall gain of the entire tree. Here, the feature is the variable collected by different sensors. When each feature is divided, the greater its average gain, the more important the feature is. In the segmentation process, the weight of each leaf node can be expressed as w ( g i , h i ).in , , training error Indicates the target value and predicted values The difference is g i Represents a leaf node i The sum of the first-order gradients of all samples on h i Represents a leaf node i The sum of the second-order gradients of all samples.

[0025] Traverse all features, for each feature, traverse all possible split points, calculate the gain of the split, and select the feature with the maximum gain from all splits of all features. The gain formula is: ; in, represents the total weight before splitting, and Respectively represent the total weights of the left child node and the right child node after the split.

[0026] The gain of each split point is expressed as the difference between the loss after the split and the loss before the split. For each feature, XGBoost evaluates its importance in the feature set according to the importance metric and sets the number of feature splits. FScore To measure feature importance: ; Among them, X is the set of leaf nodes to which the required features are classified, FScore represents the number of times each feature is selected as a splitting feature when constructing all trees, and is a simple counting index to measure the importance of features.

[0027] S32. By constructing a regression tree model through the above method, the 5 key tunneling parameters are respectively used as regression prediction targets to train 5 XGBoost regression tree models, and feature importance analysis is performed on each regression tree model to record the degree of association between each variable in the preprocessed shield machine operation data and each key tunneling variable. The degree of association is measured by FScore index, so as to obtain 5 sets of variable association degree tables. S33. Perform dimensionless (normalized) summation on the association degrees of each variable in the 5 sets of association degree tables to obtain the total association degree scores of the relevant variables of the key tunneling variables. Sort the association degree tables of the relevant variables of each key tunneling variable respectively, perform cumulative calculation on the sorted variables according to the importance ratio, screen out the top 30 variables with cumulative ratio, and select these 30 variables as the features used for the correlation of each subsequent key variable.

[0028] S4. Calculate the control limits of different key tunneling variables under different soil layers. And train multi-step prediction models for different key tunneling variables for different soil layers respectively; the specific steps include: S41: Use the kernel density estimation method to calculate the control limits of each key parameter under different soil layers. For example, the training set data of the total thrust in a certain soil layer is , where n represents the length of the total thrust in the training set, and then use the Gaussian kernel function to perform kernel density estimation. The kernel density estimation function is expressed as: ; Among them, h is the bandwidth parameter, is the i-th sample point in the total thrust data set. k is the kernel function, defined as: ; Then create an equally spaced point set from the minimum value of the total thrust to the maximum value , use the kernel density estimation function to calculate the density value of each point in the total thrust value range set , calculate the cumulative distribution function , which represents the cumulative probability below each point of the total thrust: ; Among them is the discrete point interval. Subsequently, the cumulative distribution function is normalized so that the maximum value is 1: ; Finally, through the cumulative distribution function find the control limit corresponding to the specified probability threshold p , defined as, , generally let the total thrust threshold be when, the corresponding control limit is . Output the calculated total thrust under the control limit of this soil layer for subsequent monitoring.

[0029] S42, standardize each tunneling key variable and related variables in the shield machine operation data, and the method used is maximum-minimum normalization: ; Among them is the minimum value of the data, is the maximum value of the data.

[0030] S43, train the model through ST-GCN -LSTM. For each tunneling key variable and related variables, the data under each soil layer is divided into a training set (which accounts for 80% of the total data) and a test set (which accounts for 20% of the total data). Put the training set into the set ST-GCN -LSTM model. In the ST-GCN model, the spatio-temporal graph is constructed for the graph structure, and the adjacency matrix A is used to encode the spatial relationship between nodes. The output dimension K of each layer of the graph convolution layer is set to 32, and the activation function used is ReLU. The number of LSTM hidden layers is set to 64 or 128, and is adjusted according to the data volume and complexity. The time step is initially set to 50, the learning rate is initially set to 0.001, the number of iterations is set to 100, the loss function uses the mean absolute error, and the optimization algorithm uses Adam. Use the cross-validation method Grid Search to find the best combination of hyperparameters, and save the trained models of each tunneling key variable under different soil layers.

[0031] S5, obtain the real-time operation data of the shield machine, preprocess the real-time operation data, input the online real-time data into the model, and dynamically update the predicted values of the tunneling key parameters in the future multi-steps. And judge whether there will be a situation where the predicted values are continuously over the limit. The prediction accuracy is measured by the root mean square error, mean absolute error, and accuracy. The specific steps are as follows: S51, for example, input the real-time data of the related variables of the total thrust into the trained model under the corresponding soil layer, and output the predicted values of the total thrust in the next ten moments in real time, is the predicted value of the total thrust, For the true value, three indices are selected in the present invention for evaluation, namely the root mean square error RMSE, the mean absolute error MAE, and the ACC accuracy rate. Among them, for RMSE and MAE, the closer the result is to 0, the better the performance, and the closer ACC is to 100%, the better. The formulas for the root mean square error, the mean absolute error, and the accuracy rate are as follows:

[0032] Among them, is the true value at time t, is the predicted value corresponding to the prediction at time t, and l is the total number of data. In the present invention, the total number l is set to 10.

[0033] S52. Predict the values of the total thrust at the next ten moments in real time. Using the 95% control limit of the total thrust in the corresponding soil layer calculated previously, if the data of individual predicted values exceed the control limit, but the vast majority of time points are within the range below the control limit. For this situation, there are various possible reasons. It may be caused by sensor errors or there may be deviations during data transmission. Generally, it is considered to be in a normal state. If the predicted values continuously exceed the control limit, the shield machine operator needs to adjust the relevant parameters.

Claims

1. A real-time multi-step prediction method for key parameters of shield tunneling based on ST-GCN-LSTM, characterized in that, It includes the following steps: S1 Divide the soil layer types based on the geological exploration report of a specific project according to national standards; S2 Obtain the operation data of the shield machine under normal working conditions and preprocess the operation data, including removing shutdown data, segment assembly data, and abnormal data; S3 Analyze the preprocessed operation data to determine the relevant variables of the key tunneling variables such as total thrust, cutterhead torque, cutterhead rotation speed, penetration, and average propulsion speed; S4 Calculate the control limits of each key tunneling variable for different soil layers respectively, and train a multi-step prediction model based on the combined model of ST-GCN and LSTM. Among them, ST-GCN is used to extract the spatial features between the sensors of the shield machine; LSTM is used to capture the dynamic features of the time series; S5 Collect the operation data of the shield machine in real time, input it into the trained model for preprocessing, dynamically predict the key tunneling parameter values in the future for multiple steps, and judge whether the predicted values continuously exceed the control limits; The prediction accuracy is evaluated by the root mean square error RMSE, mean absolute error MAE, and accuracy ACC.

2. The real-time multi-step prediction method for key parameters of shield tunneling based on ST-GCN-LSTM according to claim 1, characterized in that In step S2, the abnormal data is removed using the 3σ criterion, which includes the following steps: S21 Calculate the mean μ and standard deviation σ of the preprocessed data; S22 If the data point x satisfies x - μ > 3σ, it is determined as abnormal data and removed.

3. The real-time multi-step prediction method for key parameters of shield tunneling based on ST-GCN-LSTM according to claim 1, characterized in that, In step S3, the analysis of the relevant variables is realized through the XGBoost regression model, which includes the following steps: S31 Construct an XGBoost regression model for each key tunneling variable respectively; S32 Calculate the importance score FScore of each variable based on the feature split gain Gain; S33 Select the top 30 variables with the highest importance scores as the relevant variables.

4. The real-time multi-step prediction method for key tunneling parameters of a shield machine based on ST-GCN-LSTM according to claim 1, characterized in that, In step S4, the calculation of the control limits is realized through kernel density estimation KDE, which includes the following steps: S41 Use the Gaussian kernel function to estimate the probability density distribution of the key tunneling variable data under each soil layer; S42 Determine the control limit corresponding to the specified probability threshold p based on the cumulative distribution function , the formula is as follows: ； Among them, represents the data of the tunneling key parameters under the j-th soil layer, is the cumulative distribution function CDF value, that is, the probability that the variable is less than or equal to , and p is the set probability threshold.

5. The real-time multi-step prediction method for key parameters of shield tunneling based on ST-GCN-LSTM according to claim 1, characterized in that, The construction of the ST-GCN-LSTM combined model includes: S43 Set the time window length C and output step length J, and construct a spatio-temporal graph structure; The nodes of the spatio-temporal graph are sensor variables, and the edge weights are the correlations between variables; S44 Aggregate the node spatial features through multi-layer graph convolution, and the formula is: ； Among them, represents the hidden state of the nodes in the l-th layer, is the set of neighbor nodes of node v, is the weight of the graph convolutional layer, is the activation function, is the normalization coefficient; S45 Input the spatial features into the LSTM network, and capture the time-dependent relationship through the forget gate, input gate, and output gate.

6. The real-time multi-step prediction method for key parameters of shield tunneling based on ST-GCN-LSTM according to claim 1, wherein In step S5, the judgment criterion for continuous over-limit is: A warning is triggered when the predicted value exceeds the control limit in the next 3 consecutive time steps in the future.

7. An electronic device, characterized in that, It includes a memory and a processor. The memory stores a computer program, and when the processor executes the program, it realizes the steps of the method described in any one of claims 1-6.

Images