A method and device for predicting deformation of soft rock surrounding rock of a tunnel by fusing multiple sliding window models

Abstract

The present application relates to the technical field of tunnel deformation prediction, in particular to a soft rock tunnel surrounding rock deformation prediction method and equipment based on multi-sliding window model fusion, specifically comprising the following steps: S1, inputting soft rock tunnel surrounding rock deformation monitoring data; S2, preprocessing the soft rock tunnel surrounding rock deformation monitoring data according to time sequence to obtain a test data set; S3, taking BiGRU as a main prediction model, constructing a multi-sliding window prediction model through different sliding window scales, comprehensively evaluating the performance of the model, and fusing the model by using the error reciprocal method to construct a multi-sliding window BiGRU model; S4, inputting the test data set into the multi-sliding window BiGRU model and outputting soft rock tunnel surrounding rock deformation prediction results. The multi-sliding window BiGRU fusion model constructed by the present application can fully consider the prediction effects of multiple models, and the final output comprehensive prediction results are often more reliable than single models.

Description

Technical Field

[0001] This invention relates to the field of tunnel deformation prediction technology, and in particular to a method and equipment for predicting the deformation of surrounding rock in soft rock tunnels using a multi-sliding window model fusion approach. Background Technology

[0002] With the continuous expansion of highway construction in my country, more and more regions are undertaking large-scale mountain tunnel construction. Especially in southwestern regions such as Yunnan and Guangxi, the complex geological conditions at tunnel sites often present safety risks due to the deformation of surrounding rock in soft rock tunnels. Once the deformation rate and cumulative settlement exceed the design safety margin, it will inevitably endanger the safety of on-site tunnel construction, and in severe cases, may even cause serious property damage and adverse social impacts. Therefore, the safety control of surrounding rock deformation during soft rock tunnel construction is crucial, and reliable prediction of surrounding rock deformation during tunnel construction is an important guarantee for construction safety.

[0003] For tunnel deformation prediction, the earliest methods can be divided into two main categories: numerical simulation and mathematical statistics. The former involves collecting tunnel geological and design parameters and using numerical simulation software such as finite element method (FEM) and discrete element method (DIM) to model the tunnel and ultimately obtain the deformation of the surrounding rock in a specific area. This method suffers from drawbacks such as low accuracy of modeling parameters and high model simplification, resulting in tunnel deformation values ​​that only match the actual values ​​in magnitude, leading to poor accuracy. The latter method uses mathematical statistics to fit functions based on existing deformation monitoring data to predict the deformation of the surrounding rock. This method is generally suitable for tunnel projects with simple engineering geological conditions and stable surrounding rock deformation, but it is difficult to handle soft rock tunnels with variable settlement rates.

[0004] With the development of computer and artificial intelligence technologies, more and more scholars are introducing them into data analysis across various fields in hopes of obtaining relatively efficient and accurate solutions. In the field of tunnel surrounding rock deformation prediction, there are currently two main technical solutions:

[0005] One approach is to use conventional machine learning models as the basic prediction models, such as artificial neural network (ANN) models and support vector machine (SVM) models. By utilizing the nonlinear mapping capabilities and adaptive performance of these models, good application results have been achieved for the aforementioned improved conventional surrounding rock deformation prediction methods. However, these models are generally difficult to apply to univariate time series datasets of tunnel surrounding rock deformation in reality, and due to the limitations of the model principles, they cannot extract the high-dimensional features hidden in the time series data.

[0006] Secondly, from the perspective of time series prediction of surrounding rock deformation, time series data prediction models based on recurrent neural network models (RNN), such as long short-term memory recurrent neural network model (LSTM) and gated recurrent unit neural network model (GRU), are selected to carry out time series prediction of tunnel surrounding rock deformation.

[0007] In general, the second approach—selecting time-series data prediction models based on recurrent neural network (RNN) models, such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) models—for tunnel surrounding rock deformation time-series prediction is currently the mainstream research direction. This method can effectively consider the time-series characteristics of surrounding rock deformation, especially in univariate time-series datasets. It transforms unsupervised learning into supervised learning through a sliding window operation, thereby extracting high-dimensional features for time-series data prediction.

[0008] However, due to the fact that it only has deformation data as a feature, the current technology mainly suffers from problems such as lagging prediction results and poor prediction accuracy of the single sliding window model in the process of predicting deformation of surrounding rock in soft rock tunnels, which restricts the practicality of the technology in actual engineering. Summary of the Invention

[0009] Addressing the issue that the time-series data prediction model based on the recurrent neural network (RNN) model only has deformation data as a feature, the current technology for predicting the deformation of surrounding rock in soft rock tunnels suffers from problems such as lagging prediction results and poor prediction accuracy of single sliding window models. This invention uses BiGRU as the main prediction model, constructs a multi-sliding window prediction model through different sliding window scales, conducts a comprehensive prediction performance evaluation of the model, and finally uses the inverse error method to fuse the models, proposing a method and equipment for predicting the deformation of surrounding rock in soft rock tunnels by fusing multiple sliding window models.

[0010] To achieve the above-mentioned objectives, the present invention provides the following technical solution:

[0011] A method for predicting the deformation of surrounding rock in soft rock tunnels using a multi-sliding-window model fusion method, specifically including the following steps:

[0012] S1, Input soft rock tunnel surrounding rock deformation monitoring data;

[0013] S2, preprocess the soft rock tunnel surrounding rock deformation monitoring data according to the time series to obtain the test dataset;

[0014] S3 uses BiGRU as the main prediction model, constructs a multi-sliding window prediction model through different sliding window scales, evaluates the comprehensive prediction performance of the model, and uses the inverse error method to fuse the models to construct a multi-sliding window BiGRU model.

[0015] S4. Input the test dataset into the multi-window BiGRU model and output the deformation prediction results of the surrounding rock of the soft rock tunnel.

[0016] As a preferred embodiment of the present invention, in step S1, the soft rock tunnel surrounding rock deformation monitoring data is set to a data format of timestamp + deformation value.

[0017] As a preferred embodiment of the present invention, the preprocessing of the soft rock tunnel surrounding rock deformation monitoring data according to the time series in step S2 specifically refers to performing interpolation upsampling, stationarity test and first-order difference processing on the soft rock tunnel surrounding rock deformation monitoring data in sequence.

[0018] As a preferred embodiment of the present invention, S3 specifically includes the following steps:

[0019] S31, Construct a multi-sliding window BiGRU prediction model, which includes model A, model B and model C. Model A, model B and model C are BiGRU models with different sliding window sizes. Each BiGRU model includes an input layer, a BiGRU layer, a DENSE layer and an output layer, wherein the BiGRU layer is a single-layer hidden layer.

[0020] S32, train the models A, B and C respectively, and calculate the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient of the models A, B and C respectively;

[0021] S33, select a parameter as the error value from the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient, and calculate the weights of model A, model B and model C using the inverse error method;

[0022] S34, the predicted values ​​output by models A, B and C are multiplied by their respective weights and then summed to construct a multi-sliding window BiGRU model.

[0023] As a preferred embodiment of the present invention, in step S33, the weight calculation formulas for model A, model B, and model C are as follows:

[0024] ω1=(ε2+ε3) / 2(ε1+ε2+ε3)

[0025] ω2=(ε1+ε3) / 2(ε1+ε2+ε3)

[0026] ω3=(ε1+ε2) / 2(ε1+ε2+ε3)

[0027] In the formula, ω1, ω2, and ω3 represent the model weights of model A, model B, and model C, respectively; ε1, ε2, and ε3 represent the model error values ​​of model A, model B, and model C, respectively.

[0028] As a preferred embodiment of the present invention, in step S33, the mean absolute error is selected as the error value from the mean absolute error, the mean absolute error percentage, the root mean square error, and the correlation coefficient.

[0029] As a preferred embodiment of the present invention, the calculation formula for the multi-sliding window BiGRU model constructed in step S34 is as follows:

[0030] M = 0.3208M A +0.3491M B +0.3302M C

[0031] In the formula: M is the output value of the multi-sliding window BiGRU model after model fusion; M A M B M C These are the predicted output values ​​of model A, model B, and model C, respectively.

[0032] Based on the same concept, a soft rock tunnel surrounding rock deformation prediction system based on multi-sliding window model fusion is also proposed, including:

[0033] The data acquisition module is used to acquire monitoring data on the deformation of the surrounding rock in soft rock tunnels;

[0034] The soft rock tunnel surrounding rock deformation prediction module is used to input the soft rock tunnel surrounding rock deformation monitoring data into the multi-sliding window BiGRU model to obtain the soft rock tunnel surrounding rock deformation prediction results. The specific steps are as follows:

[0035] The soft rock tunnel surrounding rock deformation monitoring data were preprocessed according to time series to obtain a test dataset;

[0036] Input the test dataset into the multi-sliding window BiGRU model to output the prediction results of the surrounding rock deformation of soft rock tunnels;

[0037] The method for constructing the multi-sliding window BiGRU model is as follows: using BiGRU as the main prediction model, constructing a multi-sliding window prediction model through different sliding window scales, comprehensively evaluating the prediction performance of the model, and fusing the model using the inverse error method.

[0038] Based on the same concept, a soft rock tunnel surrounding rock deformation prediction device based on multi-sliding window model fusion is also proposed, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the soft rock tunnel surrounding rock deformation prediction method based on multi-sliding window model fusion described above.

[0039] Based on the same concept, a computer-readable storage medium is also proposed, on which a computer program is stored, which, when executed by a processor, implements a method for predicting the deformation of surrounding rock in soft rock tunnels by fusing multiple sliding window models as described above.

[0040] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0041] The multi-sliding window BiGRU fusion model constructed in this invention can fully consider the prediction effects of multiple models. The final comprehensive prediction result is often more reliable and has better performance than a single model. This technology has a wide range of application value. Attached image description:

[0042] Figure 1 This is a method for predicting the deformation of surrounding rock in soft rock tunnels using a multi-sliding-window model fusion, as described in Example 1.

[0043] Figure 2 The model structure of GRU in Example 1;

[0044] Figure 3 The model structure of BiGRU in Example 1;

[0045] Figure 4 This is the original data of the surrounding rock deformation dataset in Example 1;

[0046] Figure 5 The data is the first-order difference of the original surrounding rock deformation data in Example 1;

[0047] Figure 6 The following are the prediction results of the three types of sliding window BiGRU models in Example 1;

[0048] Figure 7 The images show the prediction results of each sliding window model and the prediction results of the fusion model in Example 2. Detailed Implementation

[0049] The present invention will be further described in detail below with reference to experimental examples and specific embodiments. However, this should not be construed as limiting the scope of the above-mentioned subject matter of the present invention to the following embodiments; all technologies implemented based on the content of the present invention fall within the scope of the present invention.

[0050] Example 1

[0051] A method for predicting surrounding rock deformation in soft rock tunnels using a multi-sliding window model fusion approach, flowchart shown below. Figure 1 As shown, the specific steps include:

[0052] S1, Input soft rock tunnel surrounding rock deformation monitoring data. Collect the original surrounding rock deformation monitoring data of the soft rock tunnel and organize it into a unified data format of timestamp + deformation value.

[0053] S2, the soft rock tunnel surrounding rock deformation monitoring data is preprocessed according to time series to obtain a test dataset. To ensure data integrity and mitigate the lag in prediction often seen in univariate time series data prediction, preprocessing measures such as interpolation upsampling, stationarity testing, and first-order differencing are performed on the data to improve dataset quality.

[0054] S3. Construct a multi-sliding-window BiGRU model, which has been pre-evaluated and fused. Using BiGRU as the main prediction model, a multi-sliding-window prediction model is constructed through different sliding window scales, and the comprehensive prediction performance of the model is evaluated. Finally, the models are fused using the inverse error method. Based on the volume of conventional soft rock tunnel surrounding rock deformation monitoring data and engineering requirements, the multi-sliding-window model constructed in this invention includes Model A (predicting the next day from 1 day prior), Model B (predicting the next day from 3 days prior), and Model C (predicting the next day from 5 days prior). Models A, B, and C are BiGRU models with time step sliding window sizes of 1, 3, and 5, respectively.

[0055] S4. Input the test dataset into the multi-window BiGRU model and output the deformation prediction results of the surrounding rock of the soft rock tunnel.

[0056] The detailed principles and steps are as follows:

[0057] 1. Introduction to the model and algorithm used

[0058] 1.1 BiGRU Time Series Forecasting Model

[0059] The Bidirectional Recurrent Gated Recurrent Unit Neural Network (BiGRU) model is a combination of the Bidirectional Recurrent Neural Network (BiRNN) proposed in 1997 and the Gated Recurrent Unit Neural Network (GRU) model proposed in 2014.

[0060] GRU is a further simplified variant of LSTM (LSTM is a complex variant of RNN). Compared to the three gates (input gate, forget gate, and output gate) of an LSTM neuron, the GRU model's neuron is simplified to only a reset gate and an update gate. In LSTM, there are two states: the cell state and the hidden state, while in GRU, there is only one state: the hidden state. Therefore, compared to LSTM, the GRU model retains the ability to transmit information from time-series data while requiring fewer parameters, making it easier to implement and compute, and mitigating the overfitting problem of LSTM.

[0061] The structure of the GRU model is shown in the diagram below. Figure 2 As shown. Where X t R represents the input at time t; t Indicates resetting the door; Z t Indicates an update gate; H t Indicates a hidden state; This represents the candidate hidden state. The reset gate and update gate are the input features X at the current time step. t And H at the previous moment t-1 The function, activated by the sigmoid function, sets the threshold between 0 and 1. It represents the candidate hidden state, which can also be represented as current information, and is derived from past information H. t-1 Determined jointly by the reset gate and the current information; H t It integrates long-term and short-term memory.

[0062] Based on the model structure of GRU, its calculation formulas are shown in equations (1) to (4):

[0063] R t =σ(X) t W xr +H t-1 W hr +b r (1)

[0064] Z t =σ(X) t W xz +H t-1 W hz +b z (2)

[0065]

[0066]

[0067] In the formula: σ represents the sigmoid activation function; W xr W hr W xz W hzW xh W hh b represents the trainable parameter matrix; r b z b h represents the bias; tanh represents the hyperbolic tangent activation function.

[0068] Regarding the bidirectional recurrent mechanism, earlier research found that unidirectional RNNs can only extract data from previous inputs to predict the current state, while bidirectional RNNs extract future data to improve their accuracy. BiRNNs use both forward and backward values ​​as inputs during prediction. This mechanism also applies to GRU, forming the basis of BiGRU. The BiGRU model structure is as follows: Figure 3 As shown in the figure, the solid lines represent forward propagation, and the dashed lines represent backward propagation.

[0069] Figure 3 In the middle, X i h represents the input value at time i. i y represents the value of the hidden layer at time i, while y represents the value of the hidden layer at time i. i This represents the value of the output layer at time i.

[0070] Meanwhile, in order to comprehensively measure the predictive performance of the model, the present invention uses the following indicators for evaluation in terms of specific performance assessment, as shown in equations (5) to (8):

[0071]

[0072]

[0073]

[0074]

[0075] In the formula: MAE represents the mean absolute error; MAPE represents the percentage of mean absolute error; RMSE represents the root mean square error; R 2 The correlation coefficient is represented by ; m represents the number of samples to be predicted. y represents the predicted value of the k-th sample; k This represents the true value of the k-th sample; This represents the average of all samples.

[0076] 1.2 ADF stationarity test and first-order difference

[0077] The main purpose of time series analysis is to use the historical and current characteristics of variables to predict future possible situations. This requires that the fundamental characteristics of the time series be maintained from the past to the predicted period; otherwise, predicting the future based on history and the current situation becomes unreliable. In other words, time series data must be stationary. Simply put, stationarity means that the fundamental characteristics of the time series remain unchanged. In other words, stationarity requires that the curve obtained from the sample time series continue along its existing pattern over a future period. Stationarity is a fundamental assumption of classical time series analysis; only predictions based on stationary time series are effective.

[0078] The lag in predictive performance of tunnel surrounding rock deformation under univariate data conditions may be due to the fact that the original tunnel surrounding rock deformation data is a non-stationary time series. Time series prediction is a classic regression problem, with the objective function being to minimize the error between the actual value and the predicted value at time t. If the time series dataset contains only univariate data, during the training of the prediction model, because it cannot learn data information from more features, the model usually selects the closest value from the features input to the regressor, i.e., the value at time t-1, to ensure its task safety, thus resulting in the lag phenomenon.

[0079] Therefore, before predicting the deformation of the surrounding rock in soft rock tunnels, the stationarity of the original deformation data should be checked. For non-stationary data, it is necessary to differ the data to a stationary state before making predictions.

[0080] (1) Stationarity test of ADF (unit root)

[0081] The ADF test is a type of unit root test used to determine the stationarity of a given time series. The ADF test is performed using the following three formulas, as shown in equations (9) to (11):

[0082]

[0083]

[0084]

[0085] Where: ΔX t X is the residual value at time t; δ is the parameter term; t-1 βt is the residual at time t-1; t is the time variable; βt is the trend term; α is the constant term; ε t This is the residual term.

[0086] The null hypothesis for the ADF test is always H0: δ = 0. The test starts with equation (11), then equation (11), and then equation (9). If the test rejects the null hypothesis, that is, the original series does not have a unit root, which is a stationary time series, the test can be stopped. Otherwise, the test continues until equation (9) is tested.

[0087] The results of the ADF test are usually determined by comparing the t-statistic with the parameters calculated based on the t-statistic. Commonly used secondary calculation parameters include the p-value (the probability value corresponding to the t-statistic) and critical values ​​(significance levels, commonly using 1%, 5%, and 10% levels, representing the critical ADF test values ​​at 99%, 95%, and 90% confidence intervals, respectively). The main criteria are as follows:

[0088] ① Comparison of the statistical values ​​and t-statistics for rejecting the null hypothesis at different levels of 1%, 5%, and 10%: If the t-statistic is less than 1%, 5%, and 10% at the same time, it means that the hypothesis is rejected very well, that is, there is no unit root. Otherwise, the null hypothesis is not valid and there is a unit root, and the time series data is not stationary.

[0089] ② Comparison of P-values: The closer the P-value is to 0, the better the stability of the data.

[0090] (2) Difference calculation

[0091] For non-stationary time series, difference operations are often used for data processing, and then the stationarity of the processed series is reassessed. The difference calculation formula is shown in equation (12):

[0092]

[0093] In the formula: x t x represents the sample value at time t; t-1 This represents the sample value at time t-1.

[0094] 1.4 Error Reciprocal Method

[0095] In forecasting practice, different forecasting methods are often used for the same problem. Combining these methods by assigning appropriate weighting coefficients yields a combined forecasting method. Combined forecasting reduces the influence of random factors in individual methods and comprehensively utilizes the information provided by various methods, thus improving forecast accuracy more effectively than individual methods. Determining the weighting coefficients of each individual forecasting method is crucial for combined forecasting.

[0096] The inverse error method is a method for determining weighting coefficients in combined prediction. In the inverse error method, prediction models with smaller errors are assigned larger weights, thus significantly reducing the overall error of the combined prediction model. To determine the weighting coefficients, this invention uses the inverse error method to assign weights to the model, as shown in equations (13) to (15):

[0097] ω1=(ε2+ε3) / 2(ε1+ε2+ε3) (13)

[0098] ω2=(ε1+ε3) / 2(ε1+ε2+ε3) (14)

[0099] ω3=(ε1+ε2) / 2(ε1+ε2+ε3) (15)

[0100] In the formula: ω1, ω2, and ω3 represent the model weights of models A, B, and C under different sliding window scales; ε1, ε2, and ε3 represent the model errors of models A, B, and C under different sliding window scales.

[0101] 2. Preprocessing of Time Series Surrounding Rock Deformation Data for Soft Rock Tunnels

[0102] 2.1 ADF Stationarity Test and First Difference

[0103] Based on actual tunnel data collected from a specific cross-section, real surrounding rock deformation monitoring data was collected. The time granularity was 1 day, meaning one deformation value was collected per day. A total of 40 days of data were collected, from March 14, 2022 to April 22, 2022, resulting in a dataset of 40 data samples. The specific raw data is as follows... Figure 4 As shown.

[0104] After ADF stationarity test and first-order differencing, the t-statistic and quadratic calculation parameter values ​​are shown in Table 1 below:

[0105] Table 1. Results of ADF stationarity test on original data and data after first-order differencing.

[0106]

[0107] Based on the ADF test data stationarity criterion described in section 1.3 above, the following conclusions can be drawn regarding this surrounding rock deformation dataset:

[0108] (1) For the original data of surrounding rock deformation, the t-statistic is -1.47, which is greater than the statistical values ​​of 1%, 5% and 10% to reject the null hypothesis to different degrees, and the P-value is 0.55. Therefore, the null hypothesis is rejected, that is, the original data of surrounding rock deformation is not stationary.

[0109] (2) For the first difference data of the original data of surrounding rock deformation, the t-statistic is -4.62, which is less than the statistical values ​​of 1%, 5% and 10% to reject the null hypothesis to different degrees, and the P-value is 0.00 (the original data is 0.000118, which has been rounded up). Therefore, the null hypothesis is accepted, that is, the first difference data of the original data of surrounding rock deformation is stationary.

[0110] Therefore, this invention transforms the prediction problem into a prediction problem using first-order difference data of the original surrounding rock deformation data. The first-order difference data is as follows: Figure 5 As shown.

[0111] 2.2 Dimensionless Data

[0112] To accelerate model convergence, time series data prediction problems typically involve dimensionless data processing before inputting the data into the model. This invention employs a standardization method, as shown in equation (16):

[0113] x * =(x-μ) / σ (16)

[0114] In the formula: μ represents the sample mean, and σ represents the sample standard deviation.

[0115] 3. Construction and Evaluation of Multi-Sliding-Window BiGRU Prediction Model

[0116] 3.1 Construction of Multi-Sliding-Window BiGRU Prediction Model

[0117] The BiGRU prediction model constructed in this invention consists of an input layer, a BiGRU layer, a DENSE layer, and an output layer, where the BiGRU layer is a single-layer hidden layer. The prediction model is a single-step prediction, meaning it only predicts data at the next time granularity of the input sequence data. The specific model architecture and data operation logic are as follows:

[0118] (1) Input layer: The first-order difference data of surrounding rock deformation obtained after preprocessing is used as the input of the prediction model. The time step sliding window size is set to T. The input sequence input into the prediction model at time t after preprocessing is shown in equation (17):

[0119] X = [x1, ... x t-1 ,x t ,…,x T ] T (17)

[0120] In the formula: X is the input sequence, x t Let be the sample data at time t. As described above, in this invention, T corresponds to 1, 3, and 5 for models A, B, and C, respectively.

[0121] (2) BiGRU layer: The input layer data is learned using forward and backward BiGRU. The computational logic of this layer is as shown in 1.1 above, and will not be repeated here;

[0122] (3) DENSE layer: This is a fully connected layer whose input is the output of the BiGRU layer. The layer uses Sigmoid as the activation function to predict the data value at the next time granularity. The prediction formula is shown in equation (18).

[0123] y*=Sigmoid(W o C t +b o (18)

[0124] In the formula: y * W represents the predicted value. o Represents the weight matrix; C t Represents the input value; b o This represents the bias term.

[0125] (4) Output layer: A fully connected layer without activation function is selected as the output layer to output the data at the next time granularity of the input sequence data.

[0126] Specifically, the parameter settings for each sliding window BiGRU prediction model are shown in Table 2:

[0127] Table 2 Parameter settings for each sliding window BiGRU prediction model

[0128] Model A 80 8 8 10 0.72：0.28 Model B 60 8 8 10 0.68：0.32 Model C 40 16 8 10 0.63：0.37

[0129] In Table 2: num_epochs represents the number of iterations; batch_size represents the batch size; h represents the number of hidden nodes in the BiGRU; f c This represents the number of nodes in the fully connected layer; train_test_split represents the ratio of the training set to the prediction set.

[0130] 3.2 Evaluation of Multi-Sliding-Window BiGRU Prediction Model

[0131] The multi-sliding-window BiGRU model constructed in section 3.1 was used to train and predict the data. Each model output deformation prediction results for a total of 9 consecutive days. The training and prediction results are as follows: Figure 6 As shown.

[0132] Figure 6The curves in the image, from light to dark, represent the original soft rock tunnel surrounding rock deformation data (Origin data), the prediction results of the model training set (Train data), and the prediction results of the model test set (Test data). The curves representing the prediction results of the model training set and the model test set are obtained by inversely standardizing the direct prediction output values ​​of the model and then adding the original soft rock tunnel surrounding rock deformation data for the corresponding dates.

[0133] Based on equations (5) to (8), a comprehensive performance evaluation of the three types of sliding window models is conducted, as shown in Table 3:

[0134] Table 3 Performance evaluation of each sliding window BiGRU prediction model

[0135] Model A 0.19 19％ 0.04 0.85 Model B 0.16 17％ 0.03 0.86 Model C 0.18 19％ 0.04 0.83

[0136] 4. Fusion of multi-sliding-window BiGRU prediction models

[0137] Using the MAE index in Table 3 as the error value, and substituting it into equations (13) to (15), the weights of each sliding window BiGRU pre-model can be obtained as 0.3208, 0.3491, and 0.3302, respectively. The final multi-sliding window BiGRU prediction model fusion calculation formula is shown in equation (19):

[0138] M = 0.3208M A +0.3491M B +0.3302M C (19)

[0139] In the formula: M is the final output value of the fusion model; M A M B M C These are the predicted output values ​​of model A, model B, and model C, respectively.

[0140] Example 2

[0141] The prediction results of the multi-sliding window BiGRU fusion model are output using equation (19). The final prediction effect and the performance evaluation of the fusion model are as follows: Figure 7 As shown.

[0142] like Figure 7 As shown, the constructed multi-sliding-window BiGRU fusion model has a MAE of 0.15, a MAPE of 16%, an RMSE of 0.03, and an R... 2 =0.92. Comparing the performance evaluation results of each sliding window BiGRU prediction model in Table 3, it can be seen that the multi-sliding window BiGRU fusion model constructed in this invention has better performance. This technology has sufficient application value, can fully consider the prediction effect of multiple models, and the final output comprehensive prediction result is often more reliable than that of a single model.

[0143] Example 3

[0144] Based on the same concept, a soft rock tunnel surrounding rock deformation prediction system based on multi-sliding window model fusion is also proposed, including:

[0145] The data acquisition module is used to acquire monitoring data on the deformation of the surrounding rock in soft rock tunnels;

[0146] The soft rock tunnel surrounding rock deformation prediction module is used to input the soft rock tunnel surrounding rock deformation monitoring data into the multi-sliding window BiGRU model to obtain the soft rock tunnel surrounding rock deformation prediction results. The specific steps are as follows:

[0147] The soft rock tunnel surrounding rock deformation monitoring data were preprocessed according to time series to obtain a test dataset;

[0148] The test dataset is input into the multi-window BiGRU model, and the deformation prediction results of the surrounding rock of the soft rock tunnel are output. The multi-window BiGRU model is constructed by using BiGRU as the main prediction model, constructing a multi-window prediction model with different sliding window scales, evaluating the comprehensive prediction performance of the model, and fusing the model using the inverse error method.

[0149] Based on the same concept, a soft rock tunnel surrounding rock deformation prediction device based on multi-sliding window model fusion is also proposed, including at least one processor and a memory communicatively connected to the at least one processor; the memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the soft rock tunnel surrounding rock deformation prediction method based on multi-sliding window model fusion described above.

[0150] Based on the same concept, a computer-readable storage medium is also proposed, on which a computer program is stored, which, when executed by a processor, implements a method for predicting the deformation of surrounding rock in soft rock tunnels by fusing multiple sliding window models as described above.

[0151] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the deformation of surrounding rock in soft rock tunnels using a multi-sliding-window model fusion approach, characterized in that, Specifically, the following steps are included: S1, Input soft rock tunnel surrounding rock deformation monitoring data; S2, preprocess the soft rock tunnel surrounding rock deformation monitoring data according to the time series to obtain the test dataset; The preprocessing specifically includes: performing a stationarity test on the soft rock tunnel surrounding rock deformation monitoring data; for non-stationary data, differentiating the data to stationary data before predicting the differentiated data; and transforming the soft rock tunnel surrounding rock deformation prediction problem into a first-order difference data prediction problem for the original surrounding rock deformation data. S3 uses BiGRU as the main prediction model, constructs a multi-sliding window prediction model through different sliding window scales, evaluates the comprehensive prediction performance of the model, and uses the inverse error method to fuse the models to construct a multi-sliding window BiGRU model. S4, input the test dataset into the multi-window BiGRU model, and output the prediction results of the surrounding rock deformation of the soft rock tunnel; S3 specifically includes the following steps: S31, Construct a multi-sliding window BiGRU prediction model, which includes model A, model B and model C. Model A, model B and model C are BiGRU models with different sliding window sizes. Each BiGRU model includes an input layer, a BiGRU layer, a DENSE layer and an output layer, wherein the BiGRU layer is a single-layer hidden layer. S32, train the models A, B and C respectively, and calculate the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient of the models A, B and C respectively; S33, select a parameter as the error value from the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient, and calculate the weights of model A, model B and model C using the inverse error method; S34, Multiply the predicted values ​​output by model A, model B and model C by their respective weights and sum them to construct a multi-sliding window BiGRU model; In step S33, the weight calculation formulas for model A, model B, and model C are as follows: ω1=(ε2+ε3) / 2(ε1+ε2+ε3) ω2=(ε1+ε3) / 2(ε1+ε2+ε3) ω3=(ε1+ε2) / 2(ε1+ε2+ε3) In the formula, ω1, ω2, and ω3 represent the model weights of model A, model B, and model C, respectively; ε1, ε2, and ε3 represent the model error values ​​of model A, model B, and model C, respectively.

2. The method for predicting surrounding rock deformation in soft rock tunnels using multi-sliding window model fusion as described in claim 1, characterized in that, In step S1, the soft rock tunnel surrounding rock deformation monitoring data is set to a data format of timestamp + deformation value.

3. The method for predicting surrounding rock deformation in soft rock tunnels using multi-sliding window model fusion as described in claim 1, characterized in that, In step S2, the preprocessing of the soft rock tunnel surrounding rock deformation monitoring data according to the time series specifically refers to performing interpolation upsampling, stationarity test, and first-order difference processing on the soft rock tunnel surrounding rock deformation monitoring data in sequence.

4. The method for predicting surrounding rock deformation in soft rock tunnels using multi-sliding window model fusion as described in claim 1, characterized in that, In step S33, the mean absolute error is selected as the error value from the mean absolute error, mean absolute error percentage, root mean square error, and correlation coefficient.

5. The method for predicting surrounding rock deformation in soft rock tunnels using multi-sliding window model fusion as described in claim 4, characterized in that, The calculation formula for the multi-sliding window BiGRU model constructed in step S34 is as follows: M=0.3208M A +0.3491M B +0.3302M C In the formula: M is the output value of the multi-sliding window BiGRU model after model fusion; M A M B M C These are the predicted output values ​​of model A, model B, and model C, respectively.

6. A soft rock tunnel surrounding rock deformation prediction system based on multi-sliding window model fusion, characterized in that, include: The data acquisition module is used to acquire monitoring data on the deformation of the surrounding rock in soft rock tunnels; The soft rock tunnel surrounding rock deformation prediction module is used to input the soft rock tunnel surrounding rock deformation monitoring data into the multi-sliding window BiGRU model to obtain the soft rock tunnel surrounding rock deformation prediction results. The specific steps are as follows: The soft rock tunnel surrounding rock deformation monitoring data is preprocessed according to time series to obtain a test dataset; the preprocessing specifically includes: performing a stationarity test on the soft rock tunnel surrounding rock deformation monitoring data, and for non-stationary data, differentiating the data to stationary data before predicting the differentiated data; the soft rock tunnel surrounding rock deformation prediction problem is transformed into a first-order difference data prediction problem of the original surrounding rock deformation data. Input the test dataset into the multi-sliding window BiGRU model to output the prediction results of the surrounding rock deformation of soft rock tunnels; The method for constructing the multi-sliding window BiGRU model is as follows: using BiGRU as the main prediction model, constructing a multi-sliding window prediction model through different sliding window scales, comprehensively evaluating the prediction performance of the model, and fusing the model using the inverse error method; The construction of a multi-sliding-window BiGRU model includes the following steps: S31, Construct a multi-sliding window BiGRU prediction model, which includes model A, model B and model C. Model A, model B and model C are BiGRU models with different sliding window sizes. Each BiGRU model includes an input layer, a BiGRU layer, a DENSE layer and an output layer, wherein the BiGRU layer is a single-layer hidden layer. S32, train the models A, B and C respectively, and calculate the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient of the models A, B and C respectively; S33, select a parameter as the error value from the mean absolute error, mean absolute error percentage, root mean square error and correlation coefficient, and calculate the weights of model A, model B and model C using the inverse error method; S34, Multiply the predicted values ​​output by model A, model B and model C by their respective weights and sum them to construct a multi-sliding window BiGRU model; In step S33, the weight calculation formulas for model A, model B, and model C are as follows: ω1=(ε2+ε3) / 2(ε1+ε2+ε3) ω2=(ε1+ε3) / 2(ε1+ε2+ε3) ω3=(ε1+ε2) / 2(ε1+ε2+ε3) In the formula, ω1, ω2, and ω3 represent the model weights of model A, model B, and model C, respectively; ε1, ε2, and ε3 represent the model error values ​​of model A, model B, and model C, respectively.

7. A device for predicting the deformation of surrounding rock in soft rock tunnels using a multi-sliding-window model fusion method, characterized in that, It includes at least one processor and a memory communicatively connected to at least one processor; The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform a method for predicting the deformation of surrounding rock in soft rock tunnels by fusing multiple sliding window models as described in any one of claims 1 to 5.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements a method for predicting the deformation of surrounding rock in soft rock tunnels by fusing multiple sliding window models as described in any one of claims 1 to 5.

Images