Tunnel surrounding rock deformation monitoring method based on time sequence neural network

Abstract

The invention discloses a tunnel surrounding rock deformation monitoring method based on a time sequence neural network, and the method comprises the steps: obtaining mountain tunnel surrounding rock deformation monitoring data, carrying out the preprocessing of the data through a time sequence preprocessing method, and obtaining a preprocessed time sequence data set; for the preprocessed time series data set, performing multi-scale decomposition on the data by adopting a wavelet transform method to obtain a decomposed multi-scale feature set; obtaining a plurality of feature matrixes according to the decomposed multi-scale feature set; carrying out matrix splicing by adopting a feature fusion method to generate a comprehensive feature matrix; according to the comprehensive characteristic matrix, time sequence modeling is carried out through a long and short term memory neural network model, the model is trained to monitor the surrounding rock deformation trend of the future time step, and a monitored deformation trend sequence is obtained. According to the method, through multi-scale feature extraction and fusion, the accuracy and reliability of surrounding rock deformation trend monitoring are improved by utilizing time sequence characteristics of monitoring data and combining external influence factors.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of tunnel engineering, and particularly relates to a tunnel surrounding rock deformation monitoring method based on a time sequence neural network. BACKGROUND

[0002] Tunnel surrounding rock deformation monitoring is a core field of mountain tunnel engineering safety management, and is directly related to the safety, stability and economy of engineering construction and operation. The surrounding rock deformation has significant time sequence characteristics, and its change law is complexly affected by geological conditions, construction disturbance and environmental factors. Accurate capture and monitoring of the deformation trend are crucial for preventing tunnel collapse and optimizing support design. At present, the traditional surrounding rock deformation monitoring method mainly relies on static analysis or simple regression model, and it is difficult to effectively process the nonlinear, dynamic and multi-scale characteristics of the deformation data. When facing the complex time sequence data of long-term trend, periodic fluctuation and random disturbance, these methods often show the defects of low monitoring accuracy and weak generalization ability, and it is difficult to meet the needs of real-time monitoring and early warning, especially in complex geological conditions.

[0003] In the surrounding rock deformation monitoring research based on the time sequence neural network, the core challenge is how to efficiently represent and extract the time sequence characteristics of the surrounding rock deformation. The multi-scale characteristics of the deformation data make it difficult for a single feature extraction method to capture long-term trends, periodic changes and random fluctuations at the same time, and the mixing of these features increases the difficulty of neural network learning of the time evolution law. In addition, the effectiveness of feature representation directly affects the quality of model input. If a time sequence feature matrix suitable for neural network cannot be constructed, it is difficult to achieve accurate modeling of the deformation dynamic process. Therefore, how to effectively separate the long-term trend, periodic change and random fluctuation of the surrounding rock deformation through multi-scale time sequence feature decomposition technology, and further through feature vectorization processing to construct a feature matrix that can fully reflect the time evolution law, has become a key problem to be solved in the tunnel surrounding rock deformation monitoring method based on the time sequence neural network. SUMMARY

[0004] To solve the problems existing in the prior art, the application provides a tunnel surrounding rock deformation monitoring method based on a time sequence neural network, which fully utilizes the time sequence characteristics of the monitoring data through multi-scale feature extraction and fusion, combines external influencing factors, improves the accuracy and reliability of the surrounding rock deformation trend monitoring, and provides effective technical support for tunnel construction safety management and risk assessment.

[0005] To achieve the above purpose, the application provides the following solutions.

[0006] A tunnel surrounding rock deformation monitoring method based on a time sequence neural network, the method comprising:

[0007] The mountain tunnel surrounding rock deformation monitoring data is acquired, time series preprocessing method is adopted to perform denoising and standardization processing on the data, and a preprocessed time series dataset is obtained;

[0008] For the preprocessed time series dataset, a wavelet transform method is adopted to perform multi-scale decomposition on the data, and a decomposed multi-scale feature set is obtained;

[0009] According to the decomposed multi-scale feature set, a plurality of feature matrices are obtained;

[0010] A feature fusion method is adopted to perform matrix splicing, and a comprehensive feature matrix is generated;

[0011] According to the comprehensive feature matrix, a long short-term memory neural network model is used for time series modeling, the model is trained to monitor the surrounding rock deformation trend at the future time step, and a monitored deformation trend sequence is obtained.

[0012] Preferably, the mountain tunnel surrounding rock deformation monitoring data is acquired, time series preprocessing method is adopted to perform denoising and standardization processing on the data, and a preprocessed time series dataset is obtained, including:

[0013] The displacement data and stress data containing time stamps are obtained from the tunnel surrounding rock by the sensor, and an original monitoring dataset is obtained;

[0014] The wavelet transform method is adopted to perform denoising processing on the original monitoring dataset, and a denoised monitoring dataset is obtained;

[0015] If the denoised monitoring dataset has outliers, a median filter method is used for secondary smoothing processing, and a smoothed monitoring dataset is obtained;

[0016] According to the smoothed monitoring dataset, the z-score standardization method is adopted to perform normalization processing on the displacement data and stress data, and a standardized time series dataset is obtained.

[0017] Preferably, for the preprocessed time series dataset, a wavelet transform method is adopted to perform multi-scale decomposition on the data, and a decomposed multi-scale feature set is obtained, including:

[0018] The wavelet transform is adopted to perform multi-scale decomposition on the preprocessed time series dataset, to obtain low-frequency components, medium-frequency components and high-frequency components, and to obtain a multi-scale feature set;

[0019] The low-frequency components are extracted from the multi-scale feature set, and the sliding window method is adopted to calculate the mean sequence of the low-frequency components, and a long-term trend sequence is obtained;

[0020] For the long-term trend sequence, if the absolute value of the slope of the sequence is greater than a preset threshold, it is judged as a significant trend, the starting point and the ending point of the significant trend are obtained, and a trend interval set is obtained;

[0021] According to the trend interval set, corresponding medium frequency components are extracted from the multi-scale feature set, the frequency distribution of the medium frequency components is analyzed by using Fourier transform, and a periodic change cycle parameter is obtained;

[0022] For the cycle parameter, if the parameter value is located in a preset cycle range, the medium frequency component is fitted by using an autoregressive model, a periodically changing monitoring sequence is obtained, and a cycle monitoring set is obtained;

[0023] The high frequency components are extracted from the multi-scale feature set, and the randomness of the high frequency components is judged by using a white noise test method, and a random fluctuation sequence is obtained;

[0024] The long-term trend sequence, the cycle monitoring set and the random fluctuation sequence are combined by weighting, a comprehensive feature sequence is generated, and the final multi-scale analysis result is obtained.

[0025] Preferably, according to the decomposed multi-scale feature set, a plurality of feature matrices are obtained, including:

[0026] The long-term trend component is extracted from the decomposed multi-scale feature set, the long-term trend feature vector is constructed in combination with the geological condition data, and the long-term trend feature matrix is obtained;

[0027] The periodic change component is extracted from the decomposed multi-scale feature set, the frequency and amplitude features of the periodic component are calculated by using the Fourier transform method, and the periodic change feature vector is constructed in combination with the construction disturbance data, and the periodic feature matrix is obtained;

[0028] The random fluctuation component is extracted from the decomposed multi-scale feature set, the fluctuation variance and skewness features are calculated by using a statistical analysis method, and the random fluctuation feature vector is constructed in combination with the environmental factor data, and the random fluctuation feature matrix is obtained.

[0029] Preferably, the long-term trend component is extracted from the decomposed multi-scale feature set, the long-term trend feature vector is constructed in combination with the geological condition data, and the long-term trend feature matrix is obtained, including:

[0030] CEEMDAN and VMD are used for double decomposition of multi-element time series data to extract internal modes and local features at different scales;

[0031] In the Transformer model, the NRBO algorithm is introduced to optimize the hyperparameters;

[0032] The optimized Transformer model is applied to the decomposed components and residual terms for monitoring, and the monitoring results of the components are reconstructed to obtain the final monitoring value, i.e. the long-term trend feature vector;

[0033] Based on the long-term trend feature vector, a long-term trend feature matrix is obtained.

[0034] Preferably, the periodic variation components are extracted from the decomposed multi-scale feature set, the Fourier transform method is used to calculate the frequency and amplitude characteristics of the periodic components, the periodic variation feature vector is constructed in combination with the construction disturbance data, the periodic feature matrix is obtained, including:

[0035] The periodic variation components are extracted from the multi-scale feature set using signal decomposition technology to obtain a set of periodic components;

[0036] The frequency domain analysis is performed on the set of periodic components by Fourier transform, the frequency and amplitude characteristics are calculated, and the frequency domain feature set is generated;

[0037] If the frequency characteristics in the frequency domain feature set exceed the preset threshold, the periodic variation feature vector is generated by using a weighted fusion method in combination with the construction disturbance data, and the feature vector set is obtained;

[0038] According to the feature vector set, the feature vectors are aggregated by using a matrix construction algorithm to construct a periodic feature matrix, and the feature matrix is obtained.

[0039] Preferably, the random fluctuation components are extracted from the decomposed multi-scale feature set, the fluctuation variance and skewness characteristics are calculated by statistical analysis method, the random fluctuation feature vector is constructed in combination with the environmental factor data, the random fluctuation feature matrix is obtained, including:

[0040] From the input multi-scale feature set, the signal decomposition algorithm is used to extract the random fluctuation components to obtain a set of fluctuation components;

[0041] The variance and skewness of each component are calculated by using statistical analysis method for the set of fluctuation components to obtain a fluctuation statistical feature set;

[0042] The environmental factor data is obtained from an external data source, and the environmental factor data and the fluctuation statistical feature set are combined by using a data fusion algorithm to generate a fusion feature set;

[0043] According to the fusion feature set, the random fluctuation feature vector is constructed by using a feature extraction method to obtain a feature vector set;

[0044] The correlation characteristics between the feature vectors are extracted from the feature vector set to generate a random fluctuation feature matrix.

[0045] Preferably, the feature fusion method is used for matrix splicing to generate a comprehensive feature matrix, including:

[0046] The long-term trend feature matrix, the periodic feature matrix and the random fluctuation feature matrix are obtained, and the corresponding feature representations are generated by using a feature extraction algorithm;

[0047] The long-term trend feature matrix, the periodic feature matrix and the random fluctuation feature matrix are spliced in a preset order by using a feature splicing method to generate a comprehensive feature matrix.

[0048] Preferably, the constructed time series monitoring model is a POA-Transformer-LSTM multivariate regression monitoring model.

[0049] The pelican algorithm (POA) is used to optimize the parameters of the Transformer-LSTM hybrid model, including the attention mechanism parameters of the Transformer layer and the weights and biases of the LSTM layer.

[0050] In the Transformer-LSTM hybrid model, the Transformer layer first extracts features from the input multivariate time series data to capture long-term patterns and global features in the data; the output of the Transformer is used as the input of the LSTM layer, and the LSTM layer further captures the short-term dynamic changes of the time series data; and the output of the LSTM layer is used as the monitoring result.

[0051] Compared with the prior art, the present application has the following advantages:

[0052] The present application discloses a tunnel surrounding rock deformation monitoring method based on a time series neural network, which collects displacement and stress data containing timestamps, pre-processes and multi-scale decomposes the data, extracts long-term trend, periodic change and random fluctuation features, and constructs a comprehensive feature matrix. The feature matrix is optimized in combination with geological conditions, construction disturbance and environmental factor data, and finally a long short-term memory neural network model is used for time series modeling to monitor the surrounding rock deformation trend at future time steps. The present application fully utilizes the time series characteristics of the monitoring data through multi-scale feature extraction and fusion, combines external influencing factors, improves the accuracy and reliability of the surrounding rock deformation trend monitoring, and provides effective technical support for tunnel construction safety management and risk assessment. BRIEF DESCRIPTION OF DRAWINGS

[0053] In order to more clearly illustrate the technical solutions of the present application, the following briefly introduces the drawings needed in the embodiments, and obviously, the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0054] Figure 1 A flowchart of a tunnel surrounding rock deformation monitoring method based on a time series neural network according to an embodiment of the present application. DETAILED DESCRIPTION

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0057] Example 1

[0058] like Figure 1 As shown in the figure, this embodiment provides a method for monitoring tunnel surrounding rock deformation based on a time-series neural network. The method includes:

[0059] Data on deformation monitoring of surrounding rock in mountain tunnels were acquired, and time series preprocessing methods were used to denoise and standardize the data to obtain a preprocessed time series dataset.

[0060] For the preprocessed time series dataset, wavelet transform is used to decompose the data into multi-scale features to obtain the decomposed multi-scale feature set.

[0061] Based on the decomposed multi-scale feature set, multiple feature matrices are obtained;

[0062] A feature fusion method is used to concatenate the matrices and generate a comprehensive feature matrix.

[0063] Based on the comprehensive feature matrix, a time series model is performed using a long short-term memory neural network model. The model is then trained to monitor the deformation trend of the surrounding rock in future time steps, resulting in a sequence of monitored deformation trends.

[0064] In this embodiment, monitoring data on the deformation of the surrounding rock in a mountain tunnel is acquired. A time-series preprocessing method is used to denoise and standardize the data, resulting in a preprocessed time-series dataset, including:

[0065] The raw monitoring dataset is obtained by acquiring displacement and stress data with timestamps from the surrounding rock of the tunnel using sensors. Among them, stress data refers to the stress state inside the surrounding rock measured by various sensors. The displacement data and stress data together constitute the deformation monitoring data of the surrounding rock of the mountain tunnel.

[0066] The original monitoring dataset was denoised using wavelet transform to obtain the denoised monitoring dataset.

[0067] If outliers are present in the denoised monitoring dataset, a second smoothing process is performed using median filtering to obtain a smoothed monitoring dataset.

[0068] According to the smooth monitoring data set, the displacement data and the stress data are normalized by using a z-score standardization method to obtain a standardized time series data set.

[0069] In this embodiment, for the preprocessed time series data set, a wavelet transform method is used to perform multi-scale decomposition on the data to obtain a decomposed multi-scale feature set, including:

[0070] The wavelet transform is used to perform multi-scale decomposition on the preprocessed time series data set to obtain low-frequency components, medium-frequency components and high-frequency components, thereby obtaining a multi-scale feature set;

[0071] The low-frequency components are extracted from the multi-scale feature set, and a sliding window method is used to calculate the mean sequence of the low-frequency components to obtain a long-term trend sequence;

[0072] For the long-term trend sequence, if the absolute value of the slope of the sequence is greater than a preset threshold, it is determined to be a significant trend, and the starting point and the ending point of the significant trend are obtained to obtain a trend interval set;

[0073] According to the trend interval set, the corresponding medium-frequency components are extracted from the multi-scale feature set, and a Fourier transform is used to analyze the frequency distribution of the medium-frequency components to obtain a periodic change cycle parameter;

[0074] For the cycle parameter, if the parameter value is within a preset cycle range, an autoregressive model is used to fit the medium-frequency components to obtain a periodic change monitoring sequence, thereby obtaining a cycle monitoring set;

[0075] The high-frequency components are extracted from the multi-scale feature set, and a white noise test method is used to determine the randomness of the high-frequency components to obtain a random fluctuation sequence;

[0076] The long-term trend sequence, the cycle monitoring set and the random fluctuation sequence are combined by weighting to generate a comprehensive feature sequence, thereby obtaining the final multi-scale analysis result.

[0077] In this embodiment, according to the decomposed multi-scale feature set, a plurality of feature matrices are obtained, including:

[0078] The long-term trend components are extracted from the decomposed multi-scale feature set, and a long-term trend feature vector is constructed in combination with the geological condition data to obtain a long-term trend feature matrix;

[0079] The periodic change components are extracted from the decomposed multi-scale feature set, and a Fourier transform method is used to calculate the frequency and amplitude features of the periodic components, and a periodic change feature vector is constructed in combination with the construction disturbance data to obtain a periodic feature matrix;

[0080] Random fluctuation components are extracted from the decomposed multi-scale feature set, and the fluctuation variance and skewness features are calculated by statistical analysis method, and the random fluctuation feature vector is constructed combined with environmental factor data, and the random fluctuation feature matrix is obtained.

[0081] Further, long-term trend components are extracted from the decomposed multi-scale feature set, and long-term trend feature vectors are constructed combined with geological condition data, and long-term trend feature matrices are obtained, including:

[0082] CEEMDAN and VMD are used for double decomposition of multivariate time series data to extract internal patterns and local features at different scales.

[0083] In the Transformer model, the NRBO algorithm is introduced to optimize the hyperparameters.

[0084] The optimized Transformer model is applied to the decomposed components and residual terms for monitoring, and the monitoring results of each component are reconstructed to obtain the final monitoring value, i.e., the long-term trend feature vector.

[0085] Based on the long-term trend feature vector, a long-term trend feature matrix is obtained.

[0086] Among them, the complementary ensemble empirical mode decomposition adaptive noise (CEEMDAN):

[0087] CEEMDAN is an improved empirical mode decomposition method aimed at solving the mode mixing and residual noise problems in EEMD. CEEMDAN adds adaptive noise to the original signal and performs EMD decomposition on the signal with added noise, then calculates the average of each decomposition as the final IMF (Intrinsic Mode Function) component. The main steps are as follows:

[0088] 1. Add a set of complementary positive and negative Gaussian white noise w i (t) to the original signal x(t) to get x + (t) = x(t) + ∈0w i (t) and x - (t) = x(t) - ∈0w i (t), where x + (t) and x - (t) represent the synthesized signal after adding Gaussian white noise, and ∈0 represents the initial value of the noise amplitude parameter.

[0089] 2. Decompose x + (t) and x -(t) performing Empirical Mode Decomposition (EMD) to obtain a first set of Intrinsic Mode Function (IMF) components and The average of the first layer of IMF is where N is the number of noise realizations.

[0090] 3. Calculate the first layer residual r1 = x(t) - IMF1.

[0091] 4. Add a new noise E1(∈ j w j (t)) to the residual r i j-1, where E1(·) denotes the first IMF component obtained by EMD decomposition of the signal, ∈ j is the standard deviation of the noise. Perform EMD decomposition on r j +E1(∈ j w i (t)) to obtain the IMF component IMF j+1,i of the j+1 layer. The average of the j+1 layer of IMF is

[0092] 5. Repeat step 4 until the residual term satisfies the stopping criterion (e.g., monotonicity of the residual), and finally the original signal can be decomposed into a series of IMF components and a residual term:

[0093]

[0094] where m is the number of layers of the residual.

[0095] Variational Mode Decomposition (VMD):

[0096] VMD is a time-frequency analysis method based on variational idea, which decomposes the signal into a set of mode components with finite bandwidth. VMD determines the center frequency and bandwidth of the mode components by constructing and solving a variational problem, thus realizing the best decomposition of the signal. The optimization goal of VMD is to minimize the sum of the bandwidths of each mode while ensuring that the sum of all modes can reconstruct the original signal. The objective function and constraint conditions can be expressed as:

[0097] Minimize:

[0098] Constraint condition:

[0099] where K is the number of decomposed modes, u k (t) is the kth mode component, ω kis the center frequency of the kth mode, δ(t) is the Dirac function, and * denotes convolution, represents L 2 norm. This optimization problem can be solved by the Alternating Direction Method of Multipliers (ADMM). Variational Modal Decomposition (VMD) has a solid mathematical foundation, is robust to noise, and has relatively few modal aliasing problems.

[0100] Newton-Raphson Optimization (NRBO):

[0101] Newton-Raphson method is a method for finding the zero of a function or optimizing a function. In the optimization problem, NRBO uses the gradient (first derivative) and Hessian matrix (second derivative) of the function to iteratively approach the optimal solution. For a target function f(θ), we want to find the parameter θ that makes it reach the extreme value. The iteration formula of Newton-Raphson method is:

[0102]

[0103] where θ n is the parameter value of the n-th iteration, is the gradient of the target function at θ n , [Hf(θ n )] -1 is the inverse of the Hessian matrix of the target function at θ n . In practical applications, calculating the Hessian matrix and its inverse may have a large amount of calculation, so the quasi-Newton method and other approximation methods are often used.

[0104] Transformer Model:

[0105] Transformer model is a neural network structure based on attention mechanism, composed of encoder (Encoder) and decoder (Decoder). Its core is multi-head self-attention mechanism (Multi-Head Self-Attention), which can calculate the correlation between any two positions in the input sequence, thus capturing long-distance dependencies.

[0106] The encoder is stacked by multiple identical layers, each containing a multi-head self-attention sublayer and a feedforward neural network sublayer. The decoder structure is similar to the encoder, but adds a sublayer that pays attention to the encoder output, and the self-attention sublayer uses a masking mechanism to prevent the decoder from seeing future positions when monitoring the current position.

[0107] Further, from the decomposed multi-scale feature set, the periodic variation component is extracted, the Fourier transform method is used to calculate the frequency and amplitude characteristics of the periodic component, combined with the construction disturbance data, a periodic variation feature vector is constructed, and a periodic feature matrix is obtained, including:

[0108] extracting periodic variation components from the multiscale feature set by using a signal decomposition technique to obtain a periodic component set;

[0109] performing frequency domain analysis on the periodic component set by using Fourier transform to calculate frequency features and amplitude features, and generating a frequency domain feature set;

[0110] if the frequency features in the frequency domain feature set exceed a preset threshold, generating a periodic variation feature vector by using a weighted fusion method in combination with construction disturbance data to obtain a feature vector set;

[0111] aggregating the feature vectors by using a matrix construction algorithm according to the feature vector set to construct a periodic feature matrix to obtain a feature matrix.

[0112] Further, random fluctuation components are extracted from the decomposed multiscale feature set, fluctuation variance and skewness features are calculated by using a statistical analysis method, and a random fluctuation feature vector is constructed in combination with environmental factor data to obtain a random fluctuation feature matrix, including:

[0113] extracting random fluctuation components from the input multiscale feature set by using a signal decomposition algorithm to obtain a fluctuation component set;

[0114] calculating the variance and skewness of each component by using a statistical analysis method for the fluctuation component set to obtain a fluctuation statistical feature set;

[0115] obtaining environmental factor data from an external data source, combining the environmental factor data with the fluctuation statistical feature set by using a data fusion algorithm to generate a fusion feature set;

[0116] constructing a random fluctuation feature vector according to the fusion feature set by using a feature extraction method to obtain a feature vector set;

[0117] extracting correlation features between the vectors from the feature vector set to generate a random fluctuation feature matrix.

[0118] In this embodiment, a feature fusion method is used for matrix splicing to generate a comprehensive feature matrix, including:

[0119] obtaining a long-term trend feature matrix, a periodic feature matrix and a random fluctuation feature matrix, and generating corresponding feature representations by using a feature extraction algorithm;

[0120] splicing the long-term trend feature matrix, the periodic feature matrix and the random fluctuation feature matrix in a preset order by using a feature splicing method to generate a comprehensive feature matrix.

[0121] wherein the long-term trend feature matrix, the periodic feature matrix and the random fluctuation feature matrix are obtained, and corresponding feature representations are generated by using a feature extraction algorithm, including:

[0122] 1. Long-term trend feature matrix: Moving average method: Extract long-term trends by calculating the moving average of time series data. Moving average method includes simple moving average method and weighted moving average method. Simple moving average method reflects long-term trends by calculating the average of a certain number of time series, while weighted moving average method gives more weight to recent data to more accurately reflect trends. Exponential smoothing method: Smooth time series data by giving higher weight to recent data and gradually reducing dependence on long-term data to extract long-term trends. Trend extrapolation method: Based on historical data of time series, a mathematical model is established to monitor future trends. 2. Periodic feature matrix: Fourier transform: By converting time series data from time domain to frequency domain, analyze its frequency components to identify periodic features. Seasonal decomposition method: Decompose time series data into trend, seasonality and residual parts, where the seasonal part reflects periodic characteristics. 3. Random fluctuation feature matrix: Residual analysis: After extracting long-term trends and periodic characteristics, the remaining part is random fluctuation characteristics. It can be obtained by subtracting long-term trends and periodic components from the original time series.

[0123] In this embodiment, the constructed time series monitoring model is: POA-Transformer-LSTM multivariate regression monitoring model;

[0124] Among them, the POA algorithm is used to optimize the parameters of the Transformer-LSTM hybrid model, including the attention mechanism parameters of the Transformer layer, the weights and biases of the LSTM layer;

[0125] In the Transformer-LSTM hybrid model, the Transformer layer first extracts features from the input multivariate time series data, capturing long-term patterns and global features in the data; The output of the Transformer is used as the input of the LSTM layer, and the LSTM layer further captures the short-term dynamic changes of the time series data; The output of the LSTM layer is the monitoring result.

[0126] Further, the original surrounding rock deformation data is acquired, an initial feature matrix is generated by a feature extraction technology, an optimized feature matrix is obtained, dimension reduction processing is performed on the initial feature matrix by using a principal component analysis algorithm, the optimized feature matrix is obtained, time series modeling is performed on the optimized feature matrix by using a long short-term memory neural network, the model is trained, a time series monitoring model is obtained, if the monitoring accuracy of the time series monitoring model is lower than a preset threshold, the model training parameters are adjusted, retraining is performed, and an optimized time series monitoring model is obtained, the future time steps are input according to the optimized time series monitoring model, the surrounding rock deformation trend is monitored, a monitoring trend sequence is obtained, the deviation of the monitoring trend sequence and actual data is calculated by using a monitoring accuracy evaluation method, an evaluation result is obtained, and if the deviation of the evaluation result exceeds the preset threshold, the feature extraction technology is updated, the optimized feature matrix is regenerated, the above steps are repeated, and a final monitoring trend sequence is obtained.

[0127] The above-described embodiments are only descriptions of the preferred modes of the present application and are not intended to limit the scope of the present application. Various modifications and improvements to the technical solutions of the present application made by those of ordinary skill in the art without departing from the design spirit of the present application shall fall within the protection scope of the present application defined by the claims.

Claims

1. A method for monitoring tunnel surrounding rock deformation based on a temporal neural network, characterized in that, The method includes: Data on deformation monitoring of surrounding rock in mountain tunnels were acquired, and time series preprocessing methods were used to denoise and standardize the data to obtain a preprocessed time series dataset. For the preprocessed time series dataset, wavelet transform is used to decompose the data into multi-scale features to obtain the decomposed multi-scale feature set. Based on the decomposed multi-scale feature set, multiple feature matrices are obtained; A feature fusion method is used to concatenate the matrices and generate a comprehensive feature matrix. Based on the comprehensive feature matrix, time series modeling is performed using a long short-term memory neural network model. The model is then trained to monitor the deformation trend of the surrounding rock in future time steps, resulting in a sequence of monitored deformation trends.

2. The method according to claim 1, characterized in that, Deformation monitoring data of the surrounding rock in mountain tunnels were acquired, and time series preprocessing methods were used to denoise and standardize the data to obtain a preprocessed time series dataset, including: The raw monitoring dataset is obtained by acquiring displacement and stress data, including timestamps, from the surrounding rock of the tunnel using sensors. The original monitoring dataset was denoised using wavelet transform to obtain the denoised monitoring dataset. If outliers are present in the denoised monitoring dataset, a second smoothing process is performed using median filtering to obtain a smoothed monitoring dataset. Based on the smoothed monitoring dataset, the displacement and stress data were normalized using the z-score normalization method to obtain a normalized time series dataset.

3. The method according to claim 1, characterized in that, For the preprocessed time-series dataset, wavelet transform is used to decompose the data at multiple scales, resulting in a multi-scale feature set, including: Wavelet transform is used to perform multi-scale decomposition on the preprocessed time series dataset to obtain low-frequency components, mid-frequency components and high-frequency components, and to obtain a multi-scale feature set. Low-frequency components are extracted from the multi-scale feature set, and the mean sequence of the low-frequency components is calculated using the sliding window method to obtain the long-term trend sequence. For a long-term trend sequence, if the absolute value of the slope of the sequence is greater than a preset threshold, it is judged as a significant trend. The starting point and ending point of the significant trend are obtained to obtain the trend interval set. Based on the trend interval set, the corresponding intermediate frequency components are extracted from the multi-scale feature set, and the frequency distribution of the intermediate frequency components is analyzed by Fourier transform to obtain the periodic parameters of the periodic change. For the periodic parameter, if the parameter value is within the preset periodic range, the mid-frequency component is fitted by an autoregressive model to obtain the periodic monitoring sequence and obtain the periodic monitoring set. High-frequency components are extracted from the multi-scale feature set, and the randomness of the high-frequency components is determined by the white noise test method to obtain a random fluctuation sequence. By weighting and combining long-term trend sequences, periodic monitoring sets, and random fluctuation sequences, a comprehensive feature sequence is generated, yielding the final multi-scale analysis results.

4. The method according to claim 3, characterized in that, Based on the decomposed multi-scale feature set, multiple feature matrices are obtained, including: Long-term trend components are extracted from the decomposed multi-scale feature set, and combined with geological condition data to construct a long-term trend feature vector, thus obtaining a long-term trend feature matrix. Periodic variation components are extracted from the decomposed multi-scale feature set, and the frequency and amplitude characteristics of the periodic components are calculated using the Fourier transform method. Combined with construction disturbance data, a periodic variation feature vector is constructed to obtain the periodic feature matrix. Random fluctuation components are extracted from the decomposed multi-scale feature set, and the fluctuation variance and skewness characteristics are calculated by statistical analysis methods. Combined with environmental factor data, a random fluctuation feature vector is constructed to obtain the random fluctuation feature matrix.

5. The method according to claim 4, characterized in that, Long-term trend components are extracted from the decomposed multi-scale feature set, and combined with geological condition data to construct a long-term trend feature vector, resulting in a long-term trend feature matrix, including: We use CEEMDAN and VMD to perform dual decomposition on multivariate time series data to extract intrinsic patterns and local features at different scales. In the Transformer model, the NRBO algorithm is introduced to optimize the hyperparameters; The optimized Transformer model is applied to monitor the decomposed components and residuals, and the monitoring results of each component are reconstructed to obtain the final monitoring value, i.e., the long-term trend feature vector. Based on the long-term trend feature vector, the long-term trend feature matrix is ​​obtained.

6. The method according to claim 4, characterized in that, Periodic variation components are extracted from the decomposed multi-scale feature set. The frequency and amplitude characteristics of these periodic components are calculated using Fourier transform. Combined with construction disturbance data, a periodic variation feature vector is constructed, resulting in a periodic feature matrix, including: Periodic component sets are obtained by extracting periodic variation components from a multi-scale feature set using signal decomposition techniques. Frequency domain analysis of the periodic component set is performed by Fourier transform to calculate frequency and amplitude characteristics and generate a frequency domain feature set. If the frequency features in the frequency domain feature set exceed the preset threshold, then a weighted fusion method is used to generate periodically changing feature vectors by combining construction disturbance data, thus obtaining a feature vector set. Based on the feature vector set, a matrix construction algorithm is used to aggregate the feature vectors and construct a periodic feature matrix to obtain the feature matrix.

7. The method according to claim 4, characterized in that, Random fluctuation components are extracted from the decomposed multi-scale feature set. The fluctuation variance and skewness characteristics are calculated using statistical analysis methods. Combined with environmental factor data, a random fluctuation feature vector is constructed, resulting in a random fluctuation feature matrix, including: Random fluctuation components are extracted from the input multi-scale feature set using a signal decomposition algorithm to obtain a set of fluctuation components; For the set of fluctuation components, statistical analysis methods are used to calculate the variance and skewness of each component, thus obtaining the set of fluctuation statistical characteristics. Environmental factor data is obtained from external data sources, and a data fusion algorithm is used to combine the environmental factor data with the fluctuation statistical feature set to generate a fused feature set. Based on the fused feature set, a random fluctuation feature vector is constructed using a feature extraction method to obtain a feature vector set. From the set of feature vectors, the correlation features between vectors are extracted to generate a random fluctuation feature matrix.

8. The method according to claim 1, characterized in that, A feature fusion method is used to concatenate the matrices to generate a comprehensive feature matrix, including: Obtain the long-term trend feature matrix, periodic feature matrix, and random fluctuation feature matrix, and generate corresponding feature representations using feature extraction algorithms; The feature splicing method is used to splice the long-term trend feature matrix, the periodic feature matrix, and the random fluctuation feature matrix in a preset order to generate a comprehensive feature matrix.

9. The method according to claim 1, characterized in that, The constructed time-series monitoring model is: POA-Transformer-LSTM multivariate regression monitoring model; Among them, the Pelican Algorithm (POA) is used to optimize the parameters of the Transformer-LSTM hybrid model, including the attention mechanism parameters of the Transformer layer, and the weights and biases of the LSTM layer. In the Transformer-LSTM hybrid model, the Transformer layer first extracts features from the input multivariate time series data to capture long-term patterns and global features in the data; the output of the Transformer is used as the input of the LSTM layer, which further captures the short-term dynamic changes of the time series data; the output of the LSTM layer serves as the monitoring result.

Images