Open tunnel boring machine tunneling parameter multi-step prediction method

Abstract

The application provides a kind of tunneling parameter multi-step prediction method for open type tunnel boring machine.The method is applied to TBM tunneling technical field, and includes collecting tunneling parameters such as boring machine advancing speed, cutterhead rotating speed, cutterhead torque, total advancing force, main machine belt machine speed, etc.; based on cutterhead torque and advancing speed, effective tunneling state is determined by time sequence logic, effective tunneling section data is extracted and numerical validity is cleaned, and a multi-step prediction time sequence training set is constructed according to preset input step span N and output step span M; a deep prediction model is constructed by fusing LSTM and large model architecture, trained based on the training set, and the optimal hyperparameters are determined by mean square error loss function and grid search to obtain the final prediction model; input the tunneling parameters to be predicted, and output the prediction results of the next M time steps. Thus, the accuracy of the multi-step prediction of the tunneling parameters of the open type tunnel boring machine is improved.

Description

Technical Field

[0001] This invention relates to the field of TBM tunneling technology, and more particularly to a multi-step prediction method for tunneling parameters of open-face tunnel boring machines. Background Technology

[0002] Open-face tunnel boring machines (TBMs) are the core equipment for long-distance hard rock tunnel construction. The accurate and advanced prediction of their tunneling parameters is the core prerequisite for realizing intelligent control of the construction process, ensuring tunneling safety, improving construction efficiency, and reducing equipment wear and tear.

[0003] Existing prediction schemes for TBM tunneling parameters are mostly based on single-step, short-term predictions, which are insufficient to meet the multi-step prediction requirements for long-term tunnel construction planning and advanced risk management. Current multi-step prediction technologies generally lack temporal logic judgment of tunneling conditions and fail to remove invalid data from ineffective tunneling sections such as shutdowns and idling, resulting in a large amount of noise in the training samples and directly reducing the model's prediction accuracy. Furthermore, existing prediction models based on a single long short-term memory network are insufficient in capturing deep features of long-term data, rely on manual experience for hyperparameter selection, lack a global optimality search mechanism, and experience significant error accumulation during multi-step prediction, resulting in low accuracy of long-term prediction results and making them unsuitable for the engineering requirements of real-time intelligent tunneling of open-face TBMs. Summary of the Invention

[0004] To address the problems existing in the prior art, this invention provides a multi-step prediction method for tunneling parameters of open-face tunnel boring machines. The method includes: The tunneling parameters of the open tunnel boring machine are collected, including the propulsion speed, cutterhead rotation speed, cutterhead torque, total propulsion force, and main machine belt conveyor speed. Based on the cutterhead torque and propulsion speed, a timing logic judgment is performed to determine whether the current operating state of the open tunnel boring machine is a valid tunneling state. If the operating state is determined to be a valid tunneling state, valid tunneling segment data is extracted based on the start and end times; the valid tunneling segment data is numerically cleaned to generate cleaned tunneling data that conforms to the rated range of the equipment; based on the cleaned tunneling data, a time series training set suitable for multi-step prediction is constructed according to the preset input time step span N and output time step span M. A deep prediction model based on the fusion of Long Short-Term Memory (LSTM) network and a large model architecture is constructed, and the deep prediction model is trained based on the time series training set. During the training process, the optimal hyperparameter combination of the deep prediction model is determined based on the mean squared error loss function and the grid search strategy to generate the final multi-step prediction model of the tunneling parameters. Input the tunneling parameters for the time period to be predicted into the final tunneling parameter multi-step prediction model, and output the tunneling parameter prediction results for the next M time steps.

[0005] Furthermore, the extraction of valid tunneling segment data based on start and end times includes: Determine whether the torque of the cutter head is greater than zero; If the cutter head torque is greater than zero, the propulsion speed is continuously monitored to obtain a continuous monitoring result; based on the continuous monitoring result, it is determined whether the propulsion speed is always greater than zero within a preset idling elimination time. When the cutter head torque is greater than zero and the feed speed is always greater than zero within the preset idling elimination time, it is determined that the start time condition is met, and the time point corresponding to the first data point that satisfies the condition that the cutter head torque is greater than zero is determined as the start time of this cycle segment.

[0006] Furthermore, the extraction of valid tunneling segment data based on start and end times includes: Based on the determined start time of the current cycle segment, the numerical change of the cutter head torque is retrieved backward along the time axis; Determine whether there is a data point where the cutter head torque is equal to zero; If there is a data point where the cutter head torque is equal to zero, then the time corresponding to the first data point where the cutter head torque is equal to zero after the start time is determined as the end time of this cycle segment. Based on the start time and the end time, the data between the two is extracted as the effective tunneling segment data.

[0007] Furthermore, the numerical validity cleaning of the effective tunneling section data includes: Obtain the rated values ​​of various tunneling parameters from the effective tunneling section data; Determine whether the values ​​of the propulsion speed, cutter head rotation speed, cutter head torque, total propulsion force, and main machine belt speed simultaneously satisfy the condition of being greater than or equal to zero and less than or equal to the corresponding equipment rated value; If the above conditions are met, the tunneling parameters of the corresponding item are retained; otherwise, the tunneling parameters of the corresponding item are deleted.

[0008] Furthermore, the construction is applicable to time series training sets for multi-step prediction, including: For each of the aforementioned effective tunneling segments, the current time is set to T; Data from N time steps prior to time T are selected to construct an input matrix of dimension N×5, and the time step sequence of the input matrix is ​​TN to T-1; A dimension M×5 output matrix is ​​constructed by selecting data from time T and M time steps thereafter, and the time step sequence of the output matrix is ​​from T to T+M-1. The input matrix and the output matrix are combined to form an input-output data pair; Based on all times of all the effective tunneling segments, multiple input-output data pairs are generated and summarized to form the time series training set.

[0009] Furthermore, the construction and data transfer process of the deep prediction model includes: The cleaned tunneling data is input into the Long Short-Term Memory (LSTM) network to extract the long and short-term dependency features of the time series. The feature vector output by the LSTM is input into the position encoding layer for position encoding to preserve the position information of the sequence; Self-attention is calculated by inputting the feature vector after position encoding into the self-attention layer in order to capture the global correlation within the tunneling parameters; The result of the self-attention calculation is input into the residual neural network, and the network degradation is prevented through the residual connection mechanism; The output of the residual neural network is fed into the feedforward neural network to generate the final prediction output.

[0010] Furthermore, determining the optimal hyperparameter combination of the depth prediction model based on the mean squared error loss function and the grid search strategy includes: The Adam optimizer is used as the parameter update algorithm for the depth prediction model. Construct a hyperparameter-based grid search space; The model is trained by traversing each set of hyperparameters in the grid search space; Calculate the mean squared error loss function value for each hyperparameter combination; Determine whether the mean squared error loss function value has reached its minimum value; The combination of hyperparameters corresponding to the minimum mean square error loss function value is determined as the optimal combination of hyperparameters.

[0011] Furthermore, the connection logic of each network layer in the deep prediction model is as follows: The location encoding layer is located after the Long Short-Term Memory (LSTM) network and before the Self-Attention layer. The residual neural network receives the output of the Self-Attention layer as input; The feedforward neural network receives the output of the residual neural network as input and outputs a prediction result with a dimension corresponding to the output time step span M.

[0012] This invention determines the effective tunneling state based on the time-series logic of cutterhead torque and propulsion speed, extracts effective tunneling segment data, and performs numerical validity cleaning to eliminate invalid working condition data and abnormal data exceeding the equipment's rated range. This ensures that the training data closely matches the actual tunneling conditions of the equipment, improving the effectiveness and working condition adaptability of the prediction model's input data. By constructing a time-series training set adapted for multi-step prediction according to a preset input and output time step span, and simultaneously building a deep prediction model that integrates a Long Short-Term Memory (LSTM) network with a large model architecture, this invention fully explores the long- and short-term dependencies and deep nonlinear correlations of the tunneling parameter time-series data, improving the model's ability to extract features from the time-series changes of tunneling parameters and its multi-step prediction fitting ability. By determining the optimal hyperparameter combination based on the mean squared error loss function and grid search strategy during model training, this invention achieves accurate convergence of model training and global optimization of hyperparameters, reducing model fitting bias and generalization error, and improving the output accuracy and stability of the final multi-step prediction model for tunneling parameters.

[0013] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0014] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. The drawings are provided for a better understanding of the invention and are not intended to limit the scope of the invention. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein: Figure 1 A flowchart illustrating a multi-step prediction method for tunneling parameters of an open tunnel boring machine according to an embodiment of the present invention is shown. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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.

[0016] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0017] Figure 1 A flowchart of a multi-step prediction method for tunneling parameters of an open-face tunnel boring machine according to an embodiment of the present invention is shown. The method includes: S101, Collect the tunneling parameters of the open tunnel boring machine, including the propulsion speed, cutterhead rotation speed, cutterhead torque, total propulsion force, and main machine belt conveyor speed; Based on the cutterhead torque and propulsion speed, perform timing logic judgment to determine whether the current operating state of the open tunnel boring machine belongs to the effective tunneling state; S102, if the operating state is determined to be an effective tunneling state, then extract the effective tunneling segment data based on the start time and end time; perform numerical validity cleaning on the effective tunneling segment data to generate cleaned tunneling data that meets the rated range of the equipment; based on the cleaned tunneling data, construct a time series training set suitable for multi-step prediction according to the preset input time step span N and output time step span M. S103, Construct a deep prediction model based on the fusion of Long Short-Term Memory Network (LSTM) and a large model architecture, and train the deep prediction model based on the time series training set; During the training process, determine the optimal hyperparameter combination of the deep prediction model based on the mean squared error loss function and the grid search strategy to generate the final multi-step prediction model for tunneling parameters. S104, input the tunneling parameters for the time period to be predicted into the final tunneling parameter multi-step prediction model, and output the tunneling parameter prediction results for the next M time steps.

[0018] In some embodiments, the step of extracting valid tunneling segment data based on start and end times includes: determining whether the cutterhead torque is greater than zero; if the cutterhead torque is greater than zero, continuously monitoring the propulsion speed to obtain a continuous monitoring result; based on the continuous monitoring result, determining whether the propulsion speed is always greater than zero within a preset idling elimination time; when the cutterhead torque is greater than zero and the propulsion speed is always greater than zero within the preset idling elimination time, determining that the start time condition is met, and determining the time point corresponding to the first data point that satisfies the condition of the cutterhead torque being greater than zero as the start time of this cycle segment. According to embodiments of the present invention, by pre-judging whether the cutterhead torque is greater than zero, preliminary filtering of non-tunneling condition data where the cutterhead has not started is achieved, improving the pre-screening efficiency for identifying effective tunneling sections; by continuously monitoring and judging the propulsion speed within a preset idling elimination time, invalid conditions where the cutterhead is idling and there is no actual tunneling action are eliminated, improving the accuracy of determining the effective tunneling start point; by using the timing logic judgment of the dual parameters of cutterhead torque and propulsion speed, the effective tunneling cycle start time is locked, providing a reliable time reference for the complete extraction of subsequent effective tunneling section data.

[0019] For example, for an open-face hard rock tunnel boring machine with a diameter of 8.03m, the data sampling frequency was set to 1Hz, and the preset idling time was 3s. In the collected time-series data, the cutterhead torque of 1280kN·m, which is greater than 0, first appeared at 08:30:00 on 2026-03-10. Subsequently, the advance speed was continuously monitored from 08:30:00 to 08:30:02 on 2026-03-10. The advance speeds at the three consecutive sampling points were 8.2mm / min, 8.5mm / min, and 8.7mm / min, respectively, all of which were always greater than 0, meeting the start time condition. Therefore, 08:30:00 on 2026-03-10 was determined as the start time of this cycle segment.

[0020] In some embodiments, the extraction of effective tunneling segment data based on start and end times includes: based on the determined start time of the current cycle segment, retrieving the numerical change of the cutterhead torque along the time axis; determining whether there is a data point where the cutterhead torque is equal to zero; if there is a data point where the cutterhead torque is equal to zero, determining the time corresponding to the first data point where the cutterhead torque is equal to zero after the start time as the end time of the current cycle segment; and extracting the data between the start and end times as the effective tunneling segment data. According to embodiments of the present invention, by retrieving the numerical change of the cutterhead torque along the time axis based on the determined start time, continuous time-series tracking of the tunneling condition's end node is achieved, improving the consistency of the tunneling cycle end time determination; by determining the first data point where the cutterhead torque is equal to zero after the start time as the end time, closed-loop locking of the complete time interval of a single-cycle tunneling operation is achieved, avoiding invalid shutdown data from being mixed into the effective tunneling segment; and by extracting the interval data between the start and end times, complete and non-redundant extraction of the effective tunneling segment data of a single cycle is achieved, providing a clean and effective data source for subsequent model training.

[0021] For example, taking 08:30:00 on March 10, 2026 as the starting time, the numerical changes of the cutterhead torque are retrieved point by point along the time axis. At 11:25:42 on March 10, 2026, the first sampled data with the cutterhead torque equal to 0 appears. Therefore, this time is determined as the end time of this cycle segment. All time-series data between 08:30:00 on March 10, 2026 and 11:25:42 on March 10, 2026, totaling 10,543 valid sampled data, are extracted as the valid tunneling segment data of this cycle segment.

[0022] In some embodiments, the numerical validity cleaning of the effective tunneling section data includes: obtaining the equipment rated values ​​of each tunneling parameter in the effective tunneling section data; determining whether the values ​​of the propulsion speed, cutterhead rotation speed, cutterhead torque, total propulsion force, and main conveyor belt speed simultaneously satisfy the condition of being greater than or equal to zero and less than or equal to the corresponding equipment rated value; if the above conditions are met, the corresponding tunneling parameter is retained; if not, the corresponding tunneling parameter is deleted. According to the embodiments of the present invention, by obtaining the equipment rated values ​​corresponding to each tunneling parameter, matching with the actual physical boundary of equipment operation is achieved, providing a compliant quantitative benchmark for data validity determination; by simultaneously performing compliance determination of multiple-dimensional tunneling parameters being greater than or equal to zero and less than or equal to the corresponding rated value, synchronous verification of the full-dimensional validity of a single data point is achieved, avoiding the mixing of single-parameter abnormal data into the training set; by deleting data that does not meet the numerical validity conditions, the invalid data such as equipment acquisition anomalies, parameter over-limits, and negative number errors are thoroughly cleaned, significantly improving the quality and reliability of training data.

[0023] For example, for the 10,543 valid tunneling data entries mentioned above, the rated values ​​of the open tunnel boring machine were first obtained: rated propulsion speed 120 mm / min, rated cutterhead speed 9 r / min, rated cutterhead torque 5200 kN·m, rated total propulsion force 22000 kN, and rated main conveyor speed 3.15 m / s. Each data entry was then evaluated line by line for its five tunneling parameters. One data entry had a propulsion speed of 128 mm / min > 120 mm / min, one data entry had a cutterhead torque of -120 kN·m < 0, and 13 data entries had a single parameter exceeding the rated range; all of these were deleted. The remaining 10,528 data entries all had five parameters that simultaneously met the condition of ≥0 and ≤ the corresponding equipment rated value; all of these were retained to generate cleaned tunneling data.

[0024] In some embodiments, the construction suitable for a time series training set for multi-step prediction includes: For each effective tunneling segment, the current time is set to T; data from N time steps before time T are selected to construct an input matrix of dimension N×5, with the time step sequence of the input matrix being TN to T-1; data from time T and M time steps thereafter are selected to construct an output matrix of dimension M×5, with the time step sequence of the output matrix being T to T+M-1; the input matrix and the output matrix are combined to form an input-output data pair; based on all times of all effective tunneling segments, multiple input-output data pairs are generated and summarized to form the time series training set. According to an embodiment of the present invention, an N×5-dimensional input matrix is ​​constructed by selecting multi-dimensional tunneling data from N time steps prior to the current moment, thereby achieving complete encapsulation of the historical temporal characteristics of tunneling parameters and providing input data that conforms to the temporal dependency logic for the multi-step prediction model. An M×5-dimensional output matrix is ​​constructed by selecting data from the current moment and M subsequent time steps, thereby achieving matching with the multi-step prediction target and clarifying the prediction output dimension and time span of the model. Input-output data pairs are generated and summarized by traversing all effective tunneling segments throughout all time steps, thereby maximizing the utilization of effective tunneling data and constructing a standardized and highly adaptable time series training set, laying the data foundation for model training.

[0025] For example, for the cleaned single effective tunneling segment data mentioned above, keeping the sampling frequency constant at 1Hz, the preset input time step span N=60 and output time step span M=10; for the current time T=60, select 5 tunneling parameter data from the 60 time steps before T to construct an input matrix with a dimension of 60×5; select 5 tunneling parameter data from time T and the 10 time steps after T to construct an output matrix with a dimension of 10×5; combine the above input matrix and output matrix into one input-output data pair; traverse along the time axis all times within the effective tunneling segment that meet the time series length requirements (T ranges from 60 to 10518), generating a total of 10459 input-output data pairs; summarize the input-output data pairs of all effective tunneling segments to form the final time series training set.

[0026] In some embodiments, the construction and data transfer process of the deep prediction model includes: inputting the cleaned tunneling data into the Long Short-Term Memory (LSTM) network to extract long and short-term dependency features of the time series; inputting the feature vector output by the LSTM into a position encoding layer for position encoding to preserve the position information of the sequence; inputting the position-encoded feature vector into a Self-Attention layer for self-attention calculation to capture the global correlation within the tunneling parameters; inputting the result of the self-attention calculation into a residual neural network to prevent network degradation through a residual connection mechanism; and inputting the output of the residual neural network into a feedforward neural network to generate the final prediction output. According to embodiments of the present invention, the Long Short-Term Memory (LSTM) network is used to extract features from the cleaned tunneling data, effectively capturing the long- and short-term dependency features of time series data, solving the problem of long-sequence information loss in traditional time series models, and improving the model's feature representation ability for tunneling time series data. Through cascaded processing of the position encoding layer and the Self-Attention layer, the complete preservation of the position information of the time series and the deep mining of the global correlation within the tunneling parameters are achieved, fully capturing the inherent coupling relationship between multi-dimensional tunneling parameters, and improving the model's global perception ability of the changing patterns of tunneling parameters. Through progressive processing of the residual neural network and the feedforward neural network, the degradation problem in the deep network training process is effectively avoided and the mapping from features to prediction results is achieved, improving the training stability and prediction output accuracy of the deep prediction model.

[0027] For example, the cleaned tunneling time-series data with a dimension of 60×5 is input into a Long Short-Term Memory (LSTM) network with a hidden layer dimension of 64 to extract the long-short-term dependency features of the time series, and the output feature vector has a dimension of 60×64. This 60×64 feature vector is then input into a position encoding layer, and position encoding is performed using a sinusoidal position encoding method consistent with the input feature dimension to preserve the temporal position information of the sequence, and the output position encoded feature vector still has a dimension of 60×64. The position encoded feature vector is then input into a Self-Attention layer with 8 heads to perform self-attention calculation, capturing the global correlation between the 5 tunneling parameters and within the time series, and the output attention feature matrix has a dimension of 60×64. The feature matrix obtained by self-attention calculation is then input into a residual neural network containing 2 residual blocks, and the residual connection mechanism alleviates the degradation problem of deep networks, and the output depth feature maintains a dimension of 60×64. The depth feature output by the residual neural network is then fed into a feedforward neural network containing 2 fully connected layers, and finally a prediction output with a dimension corresponding to the output time step span M=10 is generated.

[0028] In some embodiments, determining the optimal hyperparameter combination of the deep prediction model based on the mean squared error loss function and a grid search strategy includes: using the Adam optimizer as the parameter update algorithm for the deep prediction model; constructing a grid search space for hyperparameters; training the model by traversing each set of hyperparameter combinations in the grid search space; calculating the mean squared error loss function value corresponding to each set of hyperparameter combinations; determining whether the mean squared error loss function value has reached its minimum value; and determining the hyperparameter combination corresponding to the minimum mean squared error loss function value as the optimal hyperparameter combination. According to embodiments of the present invention, by employing the Adam optimizer as the model parameter update algorithm, adaptive and efficient parameter updates are achieved during the training of the deep prediction model, accelerating the model's convergence speed and improving its training efficiency and convergence stability. By constructing a grid search space for hyperparameters and completing full combinatorial traversal training, comprehensive optimization of hyperparameter value combinations is achieved, avoiding the subjectivity and limitations of manual parameter tuning and improving the comprehensiveness of hyperparameter optimization. By locking the optimal hyperparameter combination using the minimum mean squared error loss function value as the criterion, quantitative and controllable optimization of model prediction error is achieved, significantly reducing the model's prediction bias and improving the prediction accuracy of the final multi-step prediction model for tunneling parameters.

[0029] For example, using the Adam optimizer as the parameter update algorithm for a deep prediction model, a grid search space for hyperparameters is constructed. This search space includes: LSTM hidden layer dimension [32, 64, 128], model learning rate [1e-4, 5e-4, 1e-3], training batch size [16, 32, 64], and number of self-attention heads [4, 8, 16]. All 81 hyperparameter combinations within the grid search space are traversed, and the model is trained 100 times on the time-series training set for each combination. The mean squared error (MSE) loss function value on the validation set is calculated for each hyperparameter combination after training. The calculation shows that when the LSTM hidden layer dimension is 64, the learning rate is 5e-4, the batch size is 32, and the number of self-attention heads is 8, the corresponding MSE loss function value reaches its minimum of 0.0028. The MSE values ​​for the other combinations are distributed between 0.0031 and 0.0076. Therefore, this hyperparameter combination is determined as the optimal hyperparameter combination for the model.

[0030] In some embodiments, the connection logic of each network layer in the deep prediction model is as follows: the position encoding layer is located after the long short-term memory network LSTM and before the self-attention layer; the residual neural network receives the output of the self-attention layer as input; the feedforward neural network receives the output of the residual neural network as input and outputs a prediction result with a dimension corresponding to the output time step span M.

[0031] It should be understood that the various processes described above can be used to rearrange, add, or delete steps. For example, the steps described in this invention can be executed in parallel, sequentially, or in different orders, as long as the desired result of the technical solution of this invention can be achieved, and this is not limited herein.

[0032] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A multi-step prediction method for tunneling parameters of open-face tunnel boring machines, characterized in that, include: The tunneling parameters of the open tunnel boring machine are collected, including the propulsion speed, cutterhead rotation speed, cutterhead torque, total propulsion force, and main machine belt conveyor speed. Based on the cutterhead torque and propulsion speed, a timing logic determination is performed to determine whether the current operating state of the open tunnel boring machine is an effective tunneling state. If the operating state is determined to be a valid tunneling state, then valid tunneling segment data is extracted based on the start time and end time; Numerical validity cleaning is performed on the effective tunneling segment data to generate cleaned tunneling data that meets the rated range of the equipment; based on the cleaned tunneling data, a time series training set suitable for multi-step prediction is constructed according to the preset input time step span N and output time step span M. A deep prediction model based on the fusion of Long Short-Term Memory (LSTM) network and a large model architecture is constructed, and the deep prediction model is trained based on the time series training set. During training, the optimal hyperparameter combination of the depth prediction model is determined based on the mean squared error loss function and the grid search strategy to generate the final multi-step prediction model of the tunneling parameters. Input the tunneling parameters for the time period to be predicted into the final tunneling parameter multi-step prediction model, and output the tunneling parameter prediction results for the next M time steps.

2. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 1, characterized in that, The extraction of valid tunneling segment data based on start and end times includes: Determine whether the torque of the cutter head is greater than zero; If the cutter head torque is greater than zero, the propulsion speed is continuously monitored to obtain a continuous monitoring result; based on the continuous monitoring result, it is determined whether the propulsion speed is always greater than zero within a preset idling elimination time. When the cutter head torque is greater than zero and the feed speed is always greater than zero within the preset idling elimination time, it is determined that the start time condition is met, and the time point corresponding to the first data point that satisfies the condition that the cutter head torque is greater than zero is determined as the start time of this cycle segment.

3. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 2, characterized in that, The extraction of valid tunneling segment data based on start and end times includes: Based on the determined start time of the current cycle segment, the numerical change of the cutter head torque is retrieved backward along the time axis; Determine whether there is a data point where the cutter head torque is equal to zero; If there is a data point where the cutter head torque is equal to zero, then the time corresponding to the first data point where the cutter head torque is equal to zero after the start time is determined as the end time of this cycle segment. Based on the start time and the end time, the data between the two is extracted as the effective tunneling segment data.

4. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 3, characterized in that, The numerical validity cleaning of the effective tunneling section data includes: Obtain the rated values ​​of various tunneling parameters from the effective tunneling section data; Determine whether the values ​​of the propulsion speed, cutter head rotation speed, cutter head torque, total propulsion force, and main machine belt speed simultaneously satisfy the condition of being greater than or equal to zero and less than or equal to the corresponding equipment rated value; If the above conditions are met, the tunneling parameters of the corresponding item are retained; otherwise, the tunneling parameters of the corresponding item are deleted.

5. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 4, characterized in that, The construction is suitable for time series training sets for multi-step prediction, including: For each of the aforementioned effective tunneling segments, the current time is set to T; Data from N time steps prior to time T are selected to construct an input matrix of dimension N×5, and the time step sequence of the input matrix is ​​TN to T-1; A dimension M×5 output matrix is ​​constructed by selecting data from time T and M time steps thereafter, and the time step sequence of the output matrix is ​​from T to T+M-1. The input matrix and the output matrix are combined to form an input-output data pair; Based on all times of all the effective tunneling segments, multiple input-output data pairs are generated and summarized to form the time series training set.

6. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 5, characterized in that, The construction and data transfer process of the depth prediction model includes: The cleaned tunneling data is input into the Long Short-Term Memory (LSTM) network to extract the long and short-term dependency features of the time series. The feature vector output by the LSTM is input into the position encoding layer for position encoding to preserve the position information of the sequence; Self-attention is calculated by inputting the feature vector after position encoding into the self-attention layer in order to capture the global correlation within the tunneling parameters; The result of the self-attention calculation is input into the residual neural network, and the network degradation is prevented through the residual connection mechanism; The output of the residual neural network is fed into the feedforward neural network to generate the final prediction output.

7. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 6, characterized in that, The process of determining the optimal hyperparameter combination of the depth prediction model based on the mean squared error loss function and the grid search strategy includes: The Adam optimizer is used as the parameter update algorithm for the depth prediction model. Construct a hyperparameter-based grid search space; The model is trained by traversing each set of hyperparameters in the grid search space; Calculate the mean squared error loss function value for each hyperparameter combination; Determine whether the mean squared error loss function value has reached its minimum value; The combination of hyperparameters corresponding to the minimum mean square error loss function value is determined as the optimal combination of hyperparameters.

8. The multi-step prediction method for tunneling parameters of an open tunnel boring machine according to claim 7, characterized in that, The connection logic of each network layer in the deep prediction model is as follows: The location encoding layer is located after the Long Short-Term Memory (LSTM) network and before the Self-Attention layer. The residual neural network receives the output of the Self-Attention layer as input; The feedforward neural network receives the output of the residual neural network as input and outputs a prediction result with a dimension corresponding to the output time step span M.

Images