Stratum identification method and system for shield tunneling based on stress wave signal multi-modal model
By using a multi-modal model based on stress wave signals and the Transformer model for shield tunneling strata identification, the shortcomings of traditional methods in terms of real-time performance and accuracy are solved, and real-time identification and accurate classification of strata changes during shield tunneling are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA RAILWAY 11TH BUREAU GRP CORP LTD
- Filing Date
- 2025-01-21
- Publication Date
- 2026-07-21
AI Technical Summary
Traditional borehole logging and advanced geological prediction methods are difficult to reflect the changes in the strata in front of the cutterhead during shield tunneling in real time. Faced with complex geological environments, existing methods have limitations in terms of accuracy and real-time performance.
A multimodal model based on stress wave signals is adopted. By acquiring stress wave signals from the main drive of the tunnel boring machine in real time, the Transformer model is used to extract and fuse time and frequency domain features. Combined with the self-attention mechanism, the strata are classified. The system includes stress wave sensors, data acquisition cards and processing modules to achieve end-to-end signal processing.
It enables real-time identification of geological changes during tunnel boring, improves the accuracy and adaptability of geological classification, simplifies the feature extraction process, and enhances the robustness of the model and its ability to adapt to complex geological conditions.
Smart Images
Figure CN122432849A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of geological formation identification technology, specifically to a method and system for geological formation identification in shield tunneling based on a multi-modal model of stress wave signals. Background Technology
[0002] Rock and soil perception and identification is the key to the refined control of parameters in the shield tunneling process. Traditional borehole logging geological surveys are difficult to reflect the full picture of the geological type of the strata where the shield tunnel is located. Although advanced geological prediction technology has been developed in actual shield tunneling, it is difficult to apply it widely in shield tunneling in the face of complex geological environments. Moreover, due to the complexity of engineering geology and data quality limitations, in-depth research is still needed based on the real-time dynamic vibration response of the geological environment and the shield cutterhead.
[0003] Traditional borehole logging and advanced geological prediction methods are difficult to reflect the changes in the strata in front of the cutterhead during shield tunneling in real time. Faced with complex geological environments, existing methods have limitations in terms of accuracy and real-time performance. Therefore, this invention provides a method and system for shield tunneling strata identification based on a multi-modal model of stress wave signals. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a method and system for identifying geological formations during shield tunneling based on a multimodal model of stress wave signals. This solves the problem that traditional borehole logging and advanced geological prediction methods are unable to reflect the changes in the geological formations in front of the cutterhead during shield tunneling in real time. In the face of complex geological environments, existing methods have limitations in terms of accuracy and real-time performance.
[0005] To achieve the above objectives, the present invention provides the following technical solution: a method for identifying geological formations during shield tunneling based on a multi-modal model of stress wave signals, comprising the following specific steps:
[0006] S1. Real-time acquisition of stress wave signals from the main drive of the tunnel boring machine;
[0007] S2. Input the stress wave signal into the Transformer model to calculate the predicted value. To identify the geological strata information during shield tunneling;
[0008] The construction of the Transformer model includes the following steps:
[0009] Step 1: Preprocess the stress wave signal to obtain time-domain and frequency-domain datasets;
[0010] Step 2: Input the time-domain dataset into transformer 1 to extract time-domain features, and input the frequency-domain dataset into transformer 2 to extract frequency-domain features;
[0011] Step 3: Obtain the fused feature F by concatenating, weighting, or using cross-modal attention mechanisms on the time-domain and frequency-domain features. fused and F fused Flattened into a single vector F in the time dimension flatten , used for classification of fully connected layers;
[0012] Step 4: For the fusion feature F fused Prepare category labels, including: y1=1 for "hard rock layer across the entire cross section", y2=2 for "soft soil layer across the entire cross section", and y3=3 for "soft upper layer and hard lower layer".
[0013] Classification was performed using fully connected layers and the Softmax activation function.
[0014]
[0015] In formula ①, denoted by , where W represents the weight matrix and b represents the bias term;
[0016] The training process for a Transformer model includes: forward propagation, back propagation, and batch training.
[0017] Preferably, the preprocessing of the stress wave signal in step one to obtain the time-domain dataset and the frequency-domain dataset specifically includes:
[0018] Random noise and high-frequency interference are removed by five-point mean filtering or wavelet denoising methods.
[0019] The time-domain signal is converted into the frequency-domain signal using the Fast Fourier Transform, generating two datasets in the time and frequency domains.
[0020] Preferably, step two, in which the time-domain dataset is input into transformer 1 to extract time-domain features, specifically includes:
[0021] I. Input Processing:
[0022] The time-domain dataset signal is segmented, with each segment containing data points of a fixed length;
[0023] Each data point is embedded and mapped to a fixed dimension d to obtain a feature vector: Z time ,in, T is the sequence length, and d is the embedding dimension;
[0024] II. Location coding:
[0025] Add position encoding to time-domain signals to preserve the temporal order information of the sequence;
[0026] The generated matrix is: ,in, P represents the position encoding matrix;
[0027] III. Transformer 1 processing:
[0028] Time-domain signals with added position encoding after input embedding Extract global temporal features into the Transformer 1 encoder;
[0029] Transformer1 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs the temporal features: F time ,in ,
[0030] Step two, in which the frequency domain dataset is input into transformer 2 to extract frequency domain features, specifically includes:
[0031] I. Input Processing:
[0032] The frequency domain dataset signal is segmented, with each segment containing data points of a fixed length.
[0033] Embed the frequency domain data points and map them to a fixed dimension d to obtain the feature vector: Z fre ,in, T is the sequence length, and d is the embedding dimension;
[0034] II. Location coding:
[0035] Add position coding to frequency domain signals to preserve the temporal sequence information;
[0036] The generated matrix is: in, P represents the position encoding matrix;
[0037] III. Transformer 2 processing:
[0038] Input frequency domain signal with added position coding after embedding; embedded time domain signal In the Transformer encoder, global frequency domain features are extracted;
[0039] Transformer2 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs temporal features: F freq ,in, ,
[0040] Preferably, in step three, the time-domain features and frequency-domain features are spliced and fused to obtain feature F. fused ,in,
[0041] The third step will F fused Flattened into a single vector F in the time dimension flatten F flatten =Flatten(F fused ).
[0042] Preferably, the loss function used in the Transformer model is defined as follows:
[0043]
[0044] In Equation ②, N is the number of samples and C is the number of categories;
[0045] The AdamW optimizer is used to update the Transformer model parameters, set the initial learning rate, and dynamically adjust the learning rate using the learning rate scheduler.
[0046] Preferred options also include:
[0047] Model performance verification;
[0048] After the Transformer model is trained, accuracy and F1 score are used to verify the model's performance.
[0049] The definition of accuracy is as follows:
[0050]
[0051] The definition of F1 is as follows:
[0052]
[0053] In equations ③ and ④, TP is the number of true positives, TN is the number of true negatives, and FP is the number of false positives.
[0054] Another objective of this invention is to provide a shield tunneling stratum identification system based on a multi-modal model of stress wave signals, comprising:
[0055] The stress wave sensor is installed at the main drive position of the tunnel boring machine to collect stress wave signals generated by the interaction between the cutterhead and the stratum in real time.
[0056] The data acquisition card communicates with the stress wave sensor and is used to convert the analog signals acquired by the sensor into digital signals and transmit them to the processing module.
[0057] The processing module is connected to the data acquisition card for data transmission. The processing module is equipped with the above-mentioned method for identifying the strata in shield tunneling based on a multimodal model of stress wave signals.
[0058] The display and control terminal is communicatively connected to the processing module.
[0059] Preferably, the sampling frequency of the stress wave sensor is greater than 1 kHz.
[0060] This invention provides a method and system for identifying geological formations during shield tunneling based on a multi-modal model of stress wave signals. It offers the following advantages:
[0061] 1. This invention collects stress wave signals from the main drive position of the shield cutterhead in real time and directly inputs them into the Transformer model. It utilizes the self-attention mechanism to efficiently capture global features and ensure the real-time nature of stratum identification. During the shield tunneling process, data is collected in real time, and the original signals are processed end-to-end, which greatly reduces the delay time, simplifies the feature extraction process, and improves the adaptability of the model.
[0062] 2. This invention directly inputs time-domain and frequency-domain signals into the Transformer model, which uses its self-attention mechanism to automatically learn deep features in the signal without manual feature extraction. The Transformer model captures the global dependencies of the signal through a multi-head attention mechanism, adapting to different strata types and complex geological conditions.
[0063] 3. This invention extracts time-domain and frequency-domain signal features through independent Transformer modules, and utilizes a cross-modal attention mechanism in the fusion stage to achieve deep information interaction between modalities. The model has higher classification accuracy for complex geological environments and exhibits significant robustness in noisy environments.
[0064] 4. This invention automatically extracts key features through the encoder module of Transformer, eliminating the need for dimensionality reduction techniques such as principal component analysis (PCA), thus simplifying the data preprocessing process. Attached Figure Description
[0065] Figure 1 This is a schematic diagram of the shield tunneling stratum identification method based on stress wave signal multimodal model proposed in this invention;
[0066] Figure 2 This is a schematic diagram of the Transformer model for the shield tunneling strata identification method based on stress wave signal multimodal model proposed in this invention.
[0067] Figure 3 This is a schematic diagram of the three types of strata to be identified in the shield tunneling strata identification method based on stress wave signal multimodal model proposed in this invention;
[0068] Figure 4This is a schematic diagram of the stress wave signals corresponding to the three strata categories in the shield tunneling strata identification method based on stress wave signal multimodal model proposed in this invention;
[0069] Figure 5 This is a schematic diagram illustrating the splicing and fusion of time-domain and frequency-domain features in the shield tunneling stratum identification method based on stress wave signal multimodal model proposed in this invention.
[0070] Figure 6 This is a schematic diagram showing the installation position of the stress wave sensor in the shield tunneling stratum identification system based on a multi-modal model of stress wave signals proposed in this invention. Detailed Implementation
[0071] 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.
[0072] Example 1:
[0073] The method in this embodiment develops a real-time stratigraphic identification method that combines stress wave signals, large-model transformer feature extraction, multi-modal feature fusion, and large-model transformer classification, and solves at least the following problems.
[0074] 1. The difficulty of real-time stratigraphic identification:
[0075] Traditional borehole logging and advanced geological prediction methods are insufficient to reflect changes in the strata in front of the cutterhead during shield tunneling in real time.
[0076] Existing methods have limitations in terms of accuracy and real-time performance when faced with complex geological environments.
[0077] 2. Correlation between stress wave signals and formation conditions:
[0078] How to utilize the characteristics of the main driving stress wave signal during shield tunneling to extract effective features to reflect different geological conditions.
[0079] 3. Nonlinear stratigraphic classification problem:
[0080] The characteristics of different stratigraphic types are complex. How to fuse the features of different modes and input the fused features into a large model to achieve efficient stratigraphic classification?
[0081] like Figures 1-6As shown in the figure, this embodiment of the invention provides a method for identifying the strata during shield tunneling based on a multi-modal model of stress wave signals. The specific steps include:
[0082] S1. Real-time acquisition of stress wave signals from the main drive of the tunnel boring machine. Generally, stress wave sensors are used for monitoring, and data acquisition card modules are used to sample the signals at high frequencies (such as 1kHz or higher) to ensure signal integrity and high resolution.
[0083] S2. Input the stress wave signal into the Transformer model to calculate the predicted value. To identify the geological strata information during shield tunneling.
[0084] Traditional borehole logging and advanced geological prediction methods are difficult to reflect the dynamic changes of the strata in front of the cutterhead during shield tunneling in real time, and have a lag effect. The shield tunneling strata identification method based on stress wave signal multimodal model provided in this embodiment collects data in real time during shield tunneling, directly processes the original signals end-to-end, greatly reduces the delay time, simplifies the feature extraction process, and improves the adaptability of the model. The shield tunneling strata information output by the Transformer model can guide the refined control of the shield tunneling process.
[0085] Building a Transformer model involves the following steps:
[0086] Step 1: Preprocess the stress wave signal to obtain time-domain and frequency-domain datasets.
[0087] Preprocessing methods mainly include:
[0088] ■ Filtering: Five-point mean filtering or wavelet denoising method is used to remove random noise and high-frequency interference.
[0089] ■ Time-domain to frequency-domain conversion: The Fast Fourier Transform (FFT) is used to convert the time-domain signal into the frequency-domain signal, generating two datasets in the time and frequency domains.
[0090] Step 2: Input the time-domain dataset into transformer1 to extract time-domain features, and input the frequency-domain dataset into transformer2 to extract frequency-domain features.
[0091] Specifically, the process of inputting the time-domain dataset into transformer1 to extract time-domain features includes:
[0092] I. Input Processing:
[0093] The time-domain dataset signal is segmented, with each segment containing a fixed length of data points, such as 1024 points;
[0094] Each data point is embedded and mapped to a fixed dimension d to obtain a feature vector: Z time,in, T is the sequence length, and d is the embedding dimension;
[0095] II. Location coding:
[0096] Add position encoding to time-domain signals to preserve the temporal order information of the sequence;
[0097] The generated matrix is: ,in, P represents the position encoding matrix;
[0098] III. Transformer1 processing:
[0099] Time-domain signals with added position encoding after input embedding In the Transformer1 encoder, global temporal features are extracted;
[0100] Transformer1 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs the temporal features: F time ,in , .
[0101] Specifically, the process of inputting the frequency domain dataset into transformer2 to extract frequency domain features includes:
[0102] I. Input Processing:
[0103] The frequency domain dataset signal is segmented, with each segment containing a fixed length of data points, such as 1024 points;
[0104] Embed the frequency domain data points and map them to a fixed dimension d to obtain the feature vector: Z fre ,in, T is the sequence length, and d is the embedding dimension;
[0105] II. Location coding:
[0106] Add position coding to frequency domain signals to preserve the temporal sequence information;
[0107] The generated matrix is: ,in, P represents the position encoding matrix;
[0108] III. Transformer2 processing:
[0109] Input frequency domain signal with added position coding after embedding; embedded time domain signal In the Transformer encoder, global frequency domain features are extracted;
[0110] Transformer2 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs temporal features: F freq ,in, ,
[0111] Step 3: Obtain the fused feature F by concatenating, weighting, or using cross-modal attention mechanisms on the time-domain and frequency-domain features. fused and F fused Flattened into a single vector F in the time dimension flatten , used for classification of fully connected layers.
[0112] For example, feature F can be obtained by splicing and fusing time-domain features and frequency-domain features. fused ,in, , will F fused Flattened into a single vector F in the time dimension flatten F flatten =Flatten(F fused ).
[0113] Step 4: For the fusion feature F fused Prepare category labels, including: y1=1 for "hard rock layer across the entire cross section", y2=2 for "soft soil layer across the entire cross section", and y3=3 for "soft upper layer and hard lower layer".
[0114] Classification was performed using fully connected layers and the Softmax activation function.
[0115]
[0116] In formula ①, denoted by , where W represents the weight matrix and b represents the bias term;
[0117] The training process for the Transformer model includes: forward propagation, back propagation, and batch training.
[0118] Forward propagation: Input the fused features into the model and calculate the predicted value.
[0119] Backpropagation: Calculate the gradient based on the loss function and update the model parameters.
[0120] Batch training: Use a batch size (e.g., Batch_size=64) for mini-batch gradient descent.
[0121] The loss function used in the Transformer model is defined as follows:
[0122]
[0123] In Equation ②, N is the number of samples and C is the number of categories;
[0124] The AdamW optimizer is used to update the Transformer model parameters, setting an initial learning rate (e.g., 0.0002, 0.0002, 0.0002), and the learning rate is dynamically adjusted using a learning rate scheduler (e.g., a cosine annealing scheduler).
[0125] Model performance verification;
[0126] After the Transformer model is trained, accuracy and F1 score are used to verify the model's performance.
[0127] The definition of accuracy is as follows:
[0128]
[0129] The definition of F1 is as follows:
[0130]
[0131] In equations ③ and ④, TP is the number of true positives, TN is the number of true negatives, and FP is the number of false positives.
[0132] Real-time monitoring and early warning:
[0133] By deploying a multimodal transformer model on an edge computing device, the geological formation during large-diameter tunneling can be identified.
[0134] Example 2:
[0135] This invention provides a shield tunneling stratum identification system based on a multimodal model of stress wave signals, comprising: a stress wave sensor, a data acquisition card, a processing module, and a display and control terminal for human-machine interaction.
[0136] A stress wave sensor is installed at the main drive position of the tunnel boring machine (TBM) to collect stress wave signals generated by the interaction between the cutterhead and the strata in real time. The sampling frequency of the stress wave sensor is set to 1 kHz or higher to ensure signal integrity and high resolution. A data acquisition card is connected to the stress wave sensor to convert the analog signals collected by the sensor into digital signals and transmit them to the processing module. Preferably, a data acquisition card with multi-channel synchronous acquisition capability and support for real-time transmission of high-frequency data streams is used. The processing module is connected to the data acquisition card for data transmission. The processing module is equipped with the TBM strata identification method based on a multi-modal model of stress wave signals described in Example 1. The display and control terminal is connected to the processing module to communicate.
[0137] The working principle of the shield tunneling stratum identification system based on the stress wave signal multimodal model is as follows: During the shield tunneling process, the stress wave sensor collects the stress wave signal generated by the interaction between the cutterhead and the stratum in real time. The data acquisition card converts the analog signal collected by the sensor into a digital signal and transmits it to the processing module. The processing module processes the stress wave signal by using the shield tunneling stratum identification method based on the stress wave signal multimodal model and outputs the shield tunneling stratum information. This information is displayed through the display and control terminal.
[0138] The operation process includes the following steps:
[0139] (1) System initialization
[0140] Start the computation unit and load the Transformer model and related parameters.
[0141] Calibrate the stress wave sensor to ensure the accuracy of signal acquisition.
[0142] (2) Real-time stratigraphic identification
[0143] (2a) Data acquisition and transmission:
[0144] The stress wave sensor acquires signals in real time and transmits them to the computing unit via a data acquisition card.
[0145] (2b) Signal Processing and Feature Extraction:
[0146] The computing unit preprocesses and formats the received signals before inputting them into the Transformer model.
[0147] (2c) Model Reasoning and Classification:
[0148] The Transformer model extracts time-domain and frequency-domain features, fuses them, and outputs the stratigraphic category.
[0149] (3) Results Display and Feedback:
[0150] The geological formation identification results are displayed in real time on the control terminal and fed back to the tunnel boring machine's control system to optimize tunneling parameters.
[0151] (4) Data storage and analysis
[0152] The real-time acquired signals and classification results are stored in a database for subsequent analysis and model optimization.
[0153] It also provides a function to query historical data, supporting engineers to monitor and analyze stratigraphic information over a long period of time.
[0154] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.
Claims
1. A method for identifying geological formations during shield tunneling based on a multi-modal model of stress wave signals, characterized in that, The specific steps include: S1. Real-time acquisition of stress wave signals from the main drive of the tunnel boring machine; S2. Input the stress wave signal into the Transformer model to calculate the predicted value. Identify the geological strata information during shield tunneling; The construction of the Transformer model includes the following steps: Step 1: Preprocess the stress wave signal to obtain time-domain and frequency-domain datasets; Step 2: Input the time-domain dataset into transformer 1 to extract time-domain features, and input the frequency-domain dataset into transformer 2 to extract frequency-domain features; Step 3: Obtain the fused feature F by concatenating, weighting, or using cross-modal attention mechanisms on the time-domain and frequency-domain features. fused and F fused Flattened into a single vector F in the time dimension flatten , used for classification of fully connected layers; Step 4: For the fusion feature F fused Prepare category labels, including: y1=1 for "hard rock layer across the entire cross section", y2=2 for "soft soil layer across the entire cross section", and y3=3 for "soft upper layer and hard lower layer". Classification was performed using fully connected layers and the Softmax activation function. In formula ①, denoted by , where W represents the weight matrix and b represents the bias term; The training process for a Transformer model includes: forward propagation, back propagation, and batch training.
2. The method for identifying geological formations in shield tunneling based on a multi-modal model of stress wave signals according to claim 1, characterized in that, The preprocessing of the stress wave signal in step one to obtain the time-domain dataset and frequency-domain dataset specifically includes: Random noise and high-frequency interference are removed by five-point mean filtering or wavelet denoising methods. The time-domain signal is converted into the frequency-domain signal using the Fast Fourier Transform, generating two datasets in the time and frequency domains.
3. The method for identifying geological formations in shield tunneling based on a multi-modal model of stress wave signals according to claim 1, characterized in that, Step two, in which the time-domain dataset is input into transformer 1 to extract time-domain features, specifically includes: I. Input Processing: The time-domain dataset signal is segmented, with each segment containing data points of a fixed length; Each data point is embedded and mapped to a fixed dimension d to obtain a feature vector: Z time ,in, T is the sequence length, and d is the embedding dimension; II. Location coding: Add position encoding to time-domain signals to preserve the temporal order information of the sequence; The generated matrix is: ,in, P represents the position encoding matrix; III. Transformer 1 processing: Time-domain signals with added position encoding after input embedding In the Transformer encoder, global temporal features are extracted; Transformer 1 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs temporal features: F time ,in Step two, in which the frequency domain dataset is input into transformer 2 to extract frequency domain features, specifically includes: I. Input Processing: The frequency domain dataset signal is segmented, with each segment containing data points of a fixed length. Embed the frequency domain data points and map them to a fixed dimension d to obtain the feature vector: Z fre ,in, T is the sequence length, and d is the embedding dimension; II. Location coding: Add position coding to frequency domain signals to preserve the temporal sequence information; The generated matrix is: in, P represents the position encoding matrix; III. Transformer 2 processing: Input frequency domain signal with added position coding after embedding; embedded time domain signal In the Transformer encoder, global frequency domain features are extracted; Transformer2 learns the dependencies between different time points in the signal through a self-attention mechanism, and outputs temporal features: F freq ,in, , 4. The method for identifying strata in shield tunneling based on a multi-modal model of stress wave signals according to claim 3, characterized in that, Step three involves fusing the time-domain and frequency-domain features to obtain feature F. fused ,in, The third step will F fused Flattened into a single vector F in the time dimension flatten f flatten =Flatten(F fused ).
5. The method for identifying geological formations during shield tunneling based on a multi-modal model of stress wave signals according to claim 1, characterized in that: The loss function used in the Transformer model is defined as follows: In Equation ②, N is the number of samples and C is the number of categories; The AdamW optimizer is used to update the Transformer model parameters, set the initial learning rate, and dynamically adjust the learning rate using the learning rate scheduler.
6. The method for identifying geological formations in shield tunneling based on a multi-modal model of stress wave signals according to claim 1, characterized in that, Also includes: Model performance verification; After the Transformer model is trained, accuracy and F1 score are used to verify the model's performance. The definition of accuracy is as follows: The definition of F1 is as follows: In equations ③ and ④, TP is the number of true positives, TN is the number of true negatives, and FP is the number of false positives.
7. A shield tunneling stratum identification system based on a multi-modal model of stress wave signals, characterized in that, include: The stress wave sensor is installed at the main drive position of the tunnel boring machine to collect stress wave signals generated by the interaction between the cutterhead and the stratum in real time. The data acquisition card communicates with the stress wave sensor and is used to convert the analog signals acquired by the sensor into digital signals and transmit them to the processing module. The processing module is connected to the data acquisition card for data transmission, and the processing module is equipped with the shield tunneling stratum identification method based on stress wave signal multimodal model as described in any one of claims 1-6; The display and control terminal is communicatively connected to the processing module.
8. The shield tunneling stratum identification system based on stress wave signal multimodal model according to claim 7, characterized in that: The sampling frequency of the stress wave sensor is greater than 1 kHz.