A surrounding rock deformation trend prediction and multi-stage early warning determination method
By constructing an improved temporal neural network model and dynamically adjusting the early warning threshold, the problem of inaccurate prediction of surrounding rock deformation trends in existing technologies has been solved, achieving higher accuracy and more efficient early warning of surrounding rock deformation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA COAL RES INST
- Filing Date
- 2026-02-26
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for monitoring and early warning of surrounding rock deformation cannot accurately predict deformation trends, cannot adapt to changes in different geological conditions and construction stages, are prone to false alarms or missed alarms, and lack dynamic self-learning capabilities.
An improved temporal neural network model is constructed, which combines a bidirectional recurrent layer, an attention mechanism allocation layer, and a residual fusion layer to dynamically adjust model parameters and early warning thresholds. By constructing feature tensors and a multi-level early warning value database through real-time monitoring data, dynamic prediction and early warning of surrounding rock deformation are achieved.
It significantly improves the accuracy and stability of surrounding rock deformation trend prediction, enhances the environmental adaptability and response efficiency of early warning, and solves the problems of insufficient prediction capability and difficulty in adaptive threshold in existing technologies.
Smart Images
Figure CN122116596A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of surrounding rock deformation prediction and early warning technology, specifically to a method for predicting surrounding rock deformation trends and determining multi-level early warnings. Background Technology
[0002] During the construction of coal mine roadways, tunnels, and underground caverns, the deformation of surrounding rock typically exhibits strong nonlinearity, strong time-varying characteristics, and obvious stage-based abrupt changes. Especially in coal-bearing strata, fault fracture zones, and high-stress concentration areas, this type of deformation often manifests as a dangerous pattern of "slow evolution followed by rapid instability." Therefore, monitoring, predicting, and providing early warning of surrounding rock deformation is of great significance.
[0003] Existing methods for monitoring and early warning of surrounding rock mainly rely on static thresholds or empirical criteria. Their early warning logic is largely limited to post-event alarms based on whether limits have been exceeded, failing to provide forward-looking assessments of the development trend of surrounding rock deformation. Furthermore, the baseline state of surrounding rock deformation varies significantly under different geological conditions and construction stages, making it difficult for a single fixed threshold to adapt to complex and changing engineering environments, easily leading to false alarms or missed alarms. Currently, early warning of surrounding rock deformation primarily relies on the following existing technical solutions: (1) Fixed threshold early warning method: This method pre-sets a fixed deformation amount or rate limit (such as 30mm or 2mm / d) according to the design specifications. Its working principle is to perform a simple logical comparison between the real-time data returned by the sensor and this value. Its disadvantage is that it cannot adapt to different geological sections; in weak and fractured zones, the fixed threshold may be too high, leading to missed alarms; while in hard rock strata, normal excavation disturbances may frequently trigger false alarms.
[0004] (2) Simple statistical analysis and regression model: The least squares method or a simple grey model (GM) is used to fit the curve of the surrounding rock deformation monitoring data. Its disadvantage is that the surrounding rock deformation is affected by a variety of nonlinear factors such as blasting, excavation step distance, and support time. The linear statistical method cannot capture the fluctuation characteristics in the time series, and the prediction accuracy decreases rapidly as the time window extends.
[0005] (3) Static early warning logic: It mostly uses preset parameters and cannot automatically optimize the prediction algorithm based on real-time monitoring feedback during construction. Its disadvantage is that it is slow to respond when construction enters abrupt geological changes and lacks dynamic self-learning ability.
[0006] In summary, existing technologies generally suffer from shortcomings such as "insufficient predictive ability, difficulty in adaptive thresholds, and lack of quantifiable reserved time windows". Summary of the Invention
[0007] To address the aforementioned technical problems, this invention proposes a method for predicting surrounding rock deformation trends and determining multi-level early warning, comprising the following steps: S1. Constructing an improved temporal neural network model The system comprises an input layer, a bidirectional recurrent layer, an attention mechanism allocation layer, a residual fusion layer, and an output mapping layer, connected sequentially. The bidirectional recurrent layer includes a bidirectional recurrent neural network structure composed of forward and reverse long short-term memory units and a bidirectional feature fusion module. A residual connection structure is set outside the main prediction path, connecting the input layer and the residual fusion layer. The residual connection structure performs an identity mapping on the input features and fuses them with the attention-weighted features at the residual fusion node to obtain high-dimensional features. Using a feature tensor constructed based on the surrounding rock deformation time series as input, the output mapping layer maps the fused high-dimensional features to surrounding rock deformation data within a future preset time window. S2. Threshold Database Construction Configure a multi-level early warning value database for different geological conditions and different construction stages; S3. Real-time construction of time series data from surrounding rock deformation monitoring. During the construction of underground engineering projects, real-time monitoring of surrounding rock deformation data is conducted to obtain surrounding rock deformation monitoring data; feature tensors for model input are constructed based on the surrounding rock deformation time series. S4. Prediction of surrounding rock deformation trends and dynamic coordinated adjustment of model parameters and early warning thresholds S41. Input the feature tensor constructed in step S3 into the improved temporal neural network model constructed in step S1; S42. Compare and analyze the model's prediction results of surrounding rock deformation with the corresponding actual monitored surrounding rock deformation, and determine whether there is a mismatch between the model's current prediction results and the evolution trend of surrounding rock deformation based on the prediction residuals and confidence interval width; S43. If a match is found, select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold. If there is a mismatch, further determine whether the geological conditions or construction stage have changed. If there is no change, select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold. S44. If geological conditions or construction stages change, the model parameters should be adjusted, and the surrounding rock deformation trend should be predicted again after adjustment. The adjustment method is as follows: a. When the displacement rate or displacement acceleration of the surrounding rock increases significantly, reduce the time step weight attenuation coefficient α and shorten the sliding window time length L; b. When the surrounding rock deformation changes abruptly at a critical construction node, increase the attention weight amplification coefficient β and adjust the attention weight distribution smoothness coefficient γ. S45. Select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold; S5. Based on the new monitoring data of surrounding rock deformation, continuously repeat steps S3-S4.
[0008] Preferably, the improved temporal neural network model uses a feature tensor constructed based on the surrounding rock deformation time series as input; the input features are fed in parallel into the forward long short-term memory unit and the reverse long short-term memory unit in the bidirectional recurrent layer to simultaneously capture the time dependence of the surrounding rock deformation time series in both the forward and reverse directions. The temporal features output by the bidirectional recurrent neural network structure are spliced or weighted in the bidirectional feature fusion module and then input to the attention mechanism allocation layer.
[0009] Preferably, the attention mechanism allocation layer is used to evaluate the importance of the surrounding rock deformation time series at different time steps, and dynamically allocate weights according to their contribution to future surrounding rock deformation, so that the model can focus on the feature information corresponding to key nodes.
[0010] Preferably, the key nodes include special construction stages and special geological conditions.
[0011] Preferably, the different geological conditions include different surrounding rock combinations, different stress conditions, and fault geological structure conditions; the construction stage includes blasting excavation and support.
[0012] Preferably, the surrounding rock deformation monitoring data is uniformly time-aligned and serialized to construct a continuous surrounding rock deformation time series; the surrounding rock deformation monitoring data includes surrounding rock displacement and surrounding rock displacement velocity, surrounding rock displacement acceleration, convergence, and deep displacement obtained based on the surrounding rock displacement.
[0013] Preferably, a data cleaning method based on median filtering and the Laida criterion is used to identify and remove abnormal and sudden changes in surrounding rock deformation monitoring data; at the same time, missing surrounding rock deformation monitoring data is supplemented by a combination of interpolation and historical trend constraints.
[0014] Preferably, the surrounding rock deformation time series is subjected to multi-scale feature decomposition processing. Using the variational mode decomposition method, the surrounding rock deformation time series is decomposed into multiple components with clear physical meanings, including a trend component reflecting the long-term bearing state of the surrounding rock, a periodic component reflecting periodic construction disturbances, and a fluctuation component reflecting short-term instability characteristics. The surrounding rock deformation time series and its components are divided into fixed-length input sequences using a sliding window method, and a feature tensor for model input is constructed based on the surrounding rock deformation time series and its components.
[0015] Preferably, when any triggering condition is met, it is determined that there is a mismatch between the current prediction result of the model and the deformation evolution trend of the surrounding rock. The triggering conditions include: a. Triggering conditions for prediction residual adjustment: When the residual between the predicted rock deformation result and the corresponding actual monitored rock deformation exceeds the preset allowable range in multiple consecutive time steps, or when the residual shows a continuous increasing trend, it is determined that the model's ability to fit the current rock deformation behavior has decreased, triggering the dynamic collaborative adjustment mechanism of parameters and thresholds. b. Confidence interval width adjustment trigger condition: The predicted confidence interval width calculated based on Monte Carlo Dropout increases significantly.
[0016] Preferably, in step S44, the model parameter adjustment method further includes: c. When the surrounding rock is identified to change from stable deformation behavior to a significant increase in displacement rate or displacement acceleration or a sudden change in deformation, the model structure parameters are adjusted: more hidden layer neurons N are dynamically activated to synchronously increase the regularization intensity parameter D; after the surrounding rock recovers stable deformation, the model structure parameters are adjusted in reverse.
[0017] Beneficial technical effects: 1. By introducing an attention mechanism and a dynamic collaborative adjustment mechanism between model parameters and early warning thresholds, this invention enables the model to accurately characterize the nonlinear law of surrounding rock deformation, significantly improving the accuracy and stability of surrounding rock deformation trend prediction.
[0018] 2. While adjusting the early warning threshold, this invention dynamically optimizes some parameter structures of the temporal neural network, so that the model is always in the optimal state that adapts to the current geological conditions and construction stage.
[0019] 3. Compared with traditional early warning technologies, the core advantages of this invention are as follows: Breakthrough in prediction accuracy and realism: It solves the problem that existing technologies cannot handle nonlinear abrupt deformations, and captures key nodes through an attention mechanism, thereby improving prediction accuracy by more than 20%.
[0020] A leap forward in modeling and early warning efficiency: It has overcome the bottleneck that the model cannot adapt to the environment, greatly improved the environmental adaptability of early warning and the efficiency of dispatch response, and provided a more scientific basis for decision-making for underground engineering safety. Attached Figure Description
[0021] Figure 1 This is a diagram of the improved temporal neural network structure of the present invention.
[0022] Figure 2 This is a flowchart of the method for predicting the deformation trend of surrounding rock and determining multi-level early warning in this invention. Detailed Implementation
[0023] The specific embodiments of the present invention will now be described in conjunction with the accompanying drawings.
[0024] like Figure 1-2As shown, the present invention proposes a method for predicting the deformation trend of surrounding rock and determining multi-level early warning, which includes the following steps: S1. Constructing an improved temporal neural network model To address the engineering characteristics of surrounding rock deformation time series, which exhibit strong nonlinearity, strong time-varying nature, and significant influence from construction disturbances, this invention employs an improved time-series neural network model for predicting surrounding rock deformation time series and supporting dynamic adaptive parameter adjustment. This model uses bidirectional time-series modeling, key node importance characterization, and stability enhancement structures to uniformly model the nonlinear, time-varying, and abrupt changes in the evolution of surrounding rock deformation.
[0025] like Figure 1 As shown, the improved temporal neural network model includes an input layer, a bidirectional recurrent layer, an attention mechanism allocation layer, a residual fusion layer, and an output mapping layer connected in sequence. The bidirectional recurrent layer includes a bidirectional recurrent neural network structure composed of forward long short-term memory units and backward long short-term memory units, and a bidirectional feature fusion module. The improved temporal neural network model uses a feature tensor constructed based on the surrounding rock deformation time series as input. The input features are fed in parallel into the forward and backward long short-term memory units in the bidirectional recurrent layer to simultaneously capture the time-dependent relationships of the surrounding rock deformation time series, thereby enhancing the model's comprehensive ability to express long-term evolution trends and short-term disturbance features. The temporal features output by the bidirectional recurrent neural network structure are concatenated or weighted and fused in the bidirectional feature fusion module before being input to the attention mechanism allocation layer. The attention mechanism allocation layer is used to assess the importance of the surrounding rock deformation time series at different time steps and dynamically allocate weights according to their contribution to future surrounding rock deformation. This allows the model to focus on the feature information corresponding to key nodes such as special construction stages (e.g., blasting excavation) and special geological conditions (e.g., stress concentration), avoiding the problem of key information being averaged or attenuated during long-sequence modeling.
[0026] To improve the stability and generalization ability of the model during training, a residual connection structure (residual connection path) is set up in addition to the main prediction path. The residual connection structure connects the input layer and the residual fusion layer. The residual connection structure can perform identity mapping on the input features and fuse them with the attention-weighted features at the residual fusion node to obtain high-dimensional features, thereby effectively alleviating the gradient vanishing problem and enabling the model to maintain stable convergence when facing complex surrounding rock (or coal and rock) combinations and strong disturbance conditions.
[0027] The output mapping layer maps the fused high-dimensional features into surrounding rock deformation data within a future preset time window △t through a fully connected mapping layer. Based on the surrounding rock deformation data within the future preset time window △t, the characteristics of surrounding rock deformation development trend can be obtained, such as linear growth, accelerated growth or abrupt change trend.
[0028] S2. Threshold Database Construction A multi-level early warning value database is configured for different geological conditions and different construction stages; the different geological conditions generally include different combinations of surrounding rock (coal and rock), different stress conditions (such as stress concentration), faults and other geological structural conditions; the construction stages include blasting excavation and support, and can be further refined into various blasting parameters and various support parameters.
[0029] S3. Real-time construction and feature decomposition of time series data from surrounding rock deformation monitoring data During the construction of underground engineering projects, real-time monitoring of surrounding rock deformation data is conducted. These underground projects include, but are not limited to, coal mine roadways and tunnels. The surrounding rock deformation monitoring data includes surrounding rock displacement (generally referring to cumulative displacement) and displacement velocity, acceleration, convergence, and deep displacement obtained based on the displacement. The surrounding rock displacement and convergence data are obtained after excavation by monitoring the underground engineering wall using sensors and other measuring devices. The deep displacement data is obtained from monitoring data penetrating deep into the surrounding rock based on microseismic monitoring. Depending on the needs, the surrounding rock deformation monitoring data may also include stress and strain, anchor bolt axial force, groundwater level, and temperature. The surrounding rock deformation monitoring data is uniformly time-aligned and serialized to construct a continuous time series of surrounding rock deformation. To address common interference factors such as equipment vibration and communication anomalies during underground engineering construction, a data cleaning method combining median filtering and the Laida criterion is adopted to identify and remove abnormal and sudden changes in surrounding rock deformation monitoring data. At the same time, missing surrounding rock deformation monitoring data is supplemented by combining interpolation and historical trend constraints to ensure the continuity and consistency of the time series in both statistical and engineering sense.
[0030] Multi-scale feature decomposition is performed on the surrounding rock deformation time series. Specifically, the variational mode decomposition (VMD) method is used to decompose the surrounding rock deformation time series into multiple components with clear physical meanings, including a trend component reflecting the long-term bearing state of the surrounding rock, a periodic component reflecting periodic construction disturbances, and a fluctuation component reflecting short-term instability characteristics. A feature tensor for model input is constructed based on the surrounding rock deformation time series and its components. Specifically, a sliding window method is used to divide the surrounding rock deformation time series and its components into input sequences of fixed length. The length of the sliding window and its corresponding time coverage range can be dynamically adjusted according to the surrounding rock deformation state.
[0031] The feature tensors generated by the aforementioned time series construction and feature decomposition can directly support operations such as time weight allocation, attention mechanism adjustment, and model parameter optimization in the improved temporal neural network, providing a reliable data foundation for the prediction of surrounding rock deformation trends and subsequent dynamic threshold determination.
[0032] S4. Prediction of surrounding rock deformation trends and dynamic coordinated adjustment of model parameters and early warning thresholds To overcome the problems of fixed thresholds and insufficient environmental adaptability in existing rock deformation early warning methods, this invention introduces a dynamic collaborative adjustment mechanism for model parameters and thresholds. This allows the model to not only dynamically adjust the early warning threshold according to the rock deformation state, but also simultaneously adaptively optimize the parameters of the improved temporal neural network, thereby continuously maintaining prediction accuracy and the reliability of early warning judgment. Parameter adjustment can be performed on only key parameters. Specifically, the following steps are included: S41. Input the feature tensor constructed in step S3 into the improved temporal neural network model constructed in step S1; S42. Compare and analyze the model's predicted rock deformation (output results) with the corresponding actual monitored rock deformation. If any trigger condition is met, determine that the model's current prediction result does not match the rock deformation evolution trend; otherwise (if no trigger condition is met), determine that the model's current prediction result matches the rock deformation evolution trend. The trigger conditions include: a. Triggering conditions for prediction residual adjustment: When the residual between the predicted rock deformation result and the corresponding actual monitored rock deformation exceeds the preset allowable range in multiple consecutive time steps, or when the residual shows a continuous increasing trend, it is determined that the model's ability to fit the current rock deformation behavior has decreased, triggering the dynamic collaborative adjustment mechanism of parameters and thresholds. b. Triggering condition for confidence interval width adjustment: The predicted confidence interval width calculated based on Monte Carlo Dropout increases significantly; By setting trigger conditions, we can ensure that model parameter adjustments are only made when necessary, thus avoiding the impact of frequent parameter tuning on model stability.
[0033] S43. If a match is found (the triggering condition is not met), then the corresponding early warning threshold is selected based on the current geological conditions and construction stage, and it is determined whether an early warning is needed based on the early warning threshold. If there is no match (the triggering condition is met), further determine whether the geological conditions and construction stage have changed; if there is no change, select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold.
[0034] S44. If geological conditions or construction stages change, a dynamic collaborative adjustment mechanism for parameters and thresholds is triggered, and the surrounding rock deformation trend is re-predicted after parameter adjustment. The adjustment process can adopt a gradual parameter correction method to avoid unstable prediction results caused by a one-time large-scale modification. The specific adjustment method is as follows: a. Dynamic adjustment of weight parameters for time series of surrounding rock deformation When the displacement rate or displacement acceleration of the surrounding rock increases significantly, the following adjustment strategy shall be implemented: ① Reduce the time step weight decay coefficient α: This allows recent rock deformation monitoring data to account for a higher proportion of the total rock deformation monitoring data, thereby enhancing the model's sensitivity to sudden deformation trends. ② Shorten the sliding window time L: Switch from long-term window mode to short-term response mode to reduce the interference of surrounding rock deformation monitoring data on the current prediction results; Technical effect: The model can quickly focus on recent changes and avoid diluting risk signals with data from long-term stable periods.
[0035] b. Dynamic adjustment of attention mechanism parameters When the construction is in critical stages such as blasting excavation or stress concentration geological conditions (where sudden changes in surrounding rock deformation are likely to occur), the model will perform the following adjustments: ① Increase the attention weight amplification factor β: Automatically enhance the weight allocation of key node data, so that the model focuses on the surrounding rock deformation monitoring data that has a greater impact on the stability of the surrounding rock; ② Adjust the smoothing coefficient γ of attention weight distribution: In scenarios with multiple consecutive perturbations, appropriately relax the weight distribution to avoid excessive fluctuations in model prediction results caused by a single abnormal data.
[0036] Technical effect: By adjusting the parameters of the attention mechanism, the model can both capture the trend of sudden changes and suppress abnormal noise.
[0037] c. Dynamic adjustment of model structure parameters When the surrounding rock is identified to have changed from stable deformation behavior to a significant increase in displacement rate or displacement acceleration, or a sudden change in deformation, the following operations are performed: ① Dynamically activate more hidden layer neurons N to improve the model's expressive power and adapt to complex geological conditions or strong disturbance conditions such as blasting excavation.
[0038] ②Simultaneously increase the regularization strength parameter D to prevent the model from overfitting due to increased complexity and maintain the engineering stability of the prediction results.
[0039] After the surrounding rock has regained stable deformation, the model's structural parameters can be adjusted in reverse to reduce model complexity and improve operational efficiency.
[0040] S45. Select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold.
[0041] S5. Based on the new monitoring data of surrounding rock deformation, repeat steps S3-S4 continuously. That is, the above prediction process is carried out in a sliding window scrolling manner, so that the model can continuously update the prediction results as new monitoring data arrives.
[0042] This invention is not limited to the preferred embodiments described above. Anyone can derive other methods in various forms under the guidance of this invention. Any technical solution that is the same as or similar to this application falls within the protection scope of this invention.
Claims
1. A method for predicting surrounding rock deformation trends and determining multi-level early warning, characterized in that, Includes the following steps: S1. Constructing an improved temporal neural network model It includes an input layer, a bidirectional recurrent layer, an attention mechanism allocation layer, a residual fusion layer, and an output mapping layer connected in sequence. The bidirectional recurrent layer includes a bidirectional recurrent neural network structure composed of forward long short-term memory units and reverse long short-term memory units, and a bidirectional feature fusion module. A residual connection structure is set outside the main prediction path. The residual connection structure connects the input layer and the residual fusion layer. The residual connection structure performs an identity mapping on the input features and fuses them with the attention-weighted features at the residual fusion node to obtain high-dimensional features. The output mapping layer maps the fused high-dimensional features into surrounding rock deformation data within a future preset time window; S2. Threshold Database Construction Configure a multi-level early warning value database for different geological conditions and different construction stages; S3. Real-time construction of time series data from surrounding rock deformation monitoring. Real-time monitoring of surrounding rock deformation data is conducted during the construction of underground engineering projects to obtain surrounding rock deformation monitoring data. Construct feature tensors for model input based on the time series of surrounding rock deformation; S4. Prediction of surrounding rock deformation trends and dynamic coordinated adjustment of model parameters and early warning thresholds S41. Input the feature tensor constructed in step S3 into the improved temporal neural network model constructed in step S1; S42. Compare and analyze the model's prediction results of surrounding rock deformation with the corresponding actual monitored surrounding rock deformation, and determine whether there is a mismatch between the model's current prediction results and the evolution trend of surrounding rock deformation based on the prediction residuals and confidence interval width; S43. If a match is found, select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold. If there is a mismatch, further determine whether the geological conditions or construction stage have changed. If there is no change, select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold. S44. If changes occur, adjust the model parameters and re-predict the surrounding rock deformation trend. The adjustment methods are as follows: a. When the surrounding rock displacement rate or displacement acceleration increases significantly, reduce the time step weight attenuation coefficient α and shorten the sliding window time length L; b. When the surrounding rock deformation changes abruptly at a critical construction node, increase the attention weight amplification coefficient β and adjust the attention weight distribution smoothness coefficient γ. S45. Select the corresponding early warning threshold based on the current geological conditions and construction stage, and determine whether an early warning is needed based on the early warning threshold; S5. Based on the new monitoring data of surrounding rock deformation, continuously repeat steps S3-S4.
2. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 1, characterized in that, The improved temporal neural network model uses a feature tensor constructed based on the time series of surrounding rock deformation as input; Input features are fed in parallel into the forward long short-term memory (LSTM) and reverse LSM units in the bidirectional recurrent layer to simultaneously capture the time-dependent relationships of the surrounding rock deformation time series in both the forward and reverse directions. The time-series features output by the bidirectional recurrent neural network structure are spliced or weighted in the bidirectional feature fusion module and then input to the attention mechanism allocation layer.
3. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 1 or 2, characterized in that, The attention mechanism allocation layer is used to evaluate the importance of the surrounding rock deformation time series at different time steps, and dynamically allocate weights according to their contribution to future surrounding rock deformation, so that the model can focus on the feature information corresponding to key nodes.
4. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 3, characterized in that, The key nodes include special construction stages and special geological conditions.
5. The method for predicting and determining the deformation trend of surrounding rock according to claim 1 or 4, characterized in that, The different geological conditions include different surrounding rock combinations, different stress conditions, and fault geological structure conditions; the construction stages include blasting excavation and support.
6. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 1 or 4, characterized in that, The surrounding rock deformation monitoring data are uniformly time-aligned and serialized to construct a continuous surrounding rock deformation time series; the surrounding rock deformation monitoring data includes surrounding rock displacement and surrounding rock displacement velocity, surrounding rock displacement acceleration, convergence, and deep displacement obtained based on the surrounding rock displacement.
7. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 6, characterized in that, A data cleaning method based on median filtering and the Laida criterion was adopted to identify and remove abnormal and sudden changes in surrounding rock deformation monitoring data. At the same time, missing surrounding rock deformation monitoring data were supplemented by a combination of interpolation and historical trend constraints.
8. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 1 or 7, characterized in that, Multi-scale feature decomposition processing is performed on the time series of surrounding rock deformation. Using the variational mode decomposition method, the time series of surrounding rock deformation is decomposed into multiple components with clear physical meanings, including trend components reflecting the long-term bearing state of the surrounding rock, periodic components reflecting periodic construction disturbances, and fluctuation components reflecting short-term instability characteristics. The surrounding rock deformation time series and its components are divided into fixed-length input sequences using a sliding window method, and a feature tensor for model input is constructed based on the surrounding rock deformation time series and its components.
9. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 8, characterized in that, When any of the triggering conditions is met, it is determined that there is a mismatch between the current prediction result of the model and the deformation evolution trend of the surrounding rock. The triggering conditions include: a. Triggering conditions for prediction residual adjustment: When the residual between the predicted rock deformation result and the corresponding actual monitored rock deformation exceeds the preset allowable range in multiple consecutive time steps, or when the residual shows a continuous increasing trend, it is determined that the model's ability to fit the current rock deformation behavior has decreased, triggering the dynamic collaborative adjustment mechanism of parameters and thresholds. b. Confidence interval width adjustment trigger condition: The predicted confidence interval width calculated based on Monte Carlo Dropout increases significantly.
10. The method for predicting surrounding rock deformation trends and determining multi-level early warning according to claim 9, characterized in that, In step S44, the model parameter adjustment method also includes: c. When the surrounding rock is identified to change from stable deformation behavior to a significant increase in displacement rate or displacement acceleration or a sudden change in deformation, the model structure parameters are adjusted: more hidden layer neurons N are dynamically activated to synchronously increase the regularization intensity parameter D; after the surrounding rock recovers stable deformation, the model structure parameters are adjusted in reverse.