TBM dynamic attitude prediction method and system based on ALA hybrid model

Abstract

The invention discloses a TBM dynamic attitude prediction method and system based on an ALA hybrid model, and relates to the technical field of TBM dynamic attitude prediction, and the method comprises the steps: collecting time sequence data, and carrying out the preprocessing; performing parameter estimation and optimization based on an ARIMA model; predicting results are fused and verified. The TBM dynamic attitude prediction method based on the ALA hybrid model provided by the invention has the advantages of high prediction result precision and high accuracy, realizes multi-attitude prediction of the TBM under front horizontal, front vertical, rear horizontal and rear vertical dynamic attitudes, provides richer information for efficient and safe tunneling of a shield tunneling machine, and improves the prediction efficiency. According to the method, better effects are achieved in the aspects of fusing various data characteristic processing capability, long-term prediction capability and model adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of TBM dynamic posture prediction, and in particular to a TBM dynamic posture prediction method and system based on an ALA hybrid model. Background Art

[0002] With the widespread application of TBM technology, dynamic posture prediction has gradually become an important research direction to ensure the accuracy and safety of TBM construction. As an important data processing tool, time series analysis technology has wide application value in the field of TBM dynamic posture prediction. Among them, the autoregressive integrated moving average (ARIMA) model is widely used to describe the complex data change trends in the TBM construction process due to its accurate modeling ability for time series trends and periodicity. At the same time, the long short-term memory network (LSTM) model based on deep learning further improves the accuracy of dynamic prediction by capturing nonlinear and long-term dependencies. In recent years, hybrid prediction methods that integrate traditional statistical models with deep learning technology have become a research hotspot, providing a new solution for improving the accuracy of TBM dynamic posture prediction.

[0003] Existing TBM dynamic posture prediction methods mainly rely on a single model, such as the ARIMA model or the LSTM model. However, a single model has limitations in processing complex nonlinear data and capturing long-term trends. Specifically, although the ARIMA model can better capture the linear change trend of the time series, it is insufficient in the performance of nonlinear characteristics in the data. Especially in the TBM construction process, due to the dynamic changes in geological conditions and mechanical states, there are usually a large number of nonlinear fluctuations in the data. On the other hand, although the LSTM model can handle nonlinear characteristics, its accuracy for short-term prediction is more dependent on the optimization of model parameters, and it is easy to cause insufficient generalization ability due to overfitting problems. In addition, the existing methods lack effective correction methods for model prediction errors, which limits the stability and accuracy of the prediction results. Although the optimization based on the Attention mechanism can weight the data of key time steps, the dynamic weighting method of Attention in existing studies has not been fully utilized when fusing multi-model results. To address these shortcomings, the present invention introduces a hybrid model based on ARIMA and LSTM (ALA hybrid model) and combines the Attention mechanism to dynamically optimize the model output, which not only effectively improves the accuracy and robustness of the prediction, but also overcomes the limitations of the existing technology in dealing with complex dynamic scenarios. Summary of the invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problem solved by the present invention is: the existing dynamic posture prediction method has insufficient ability to process nonlinear features, the prediction accuracy of a single model is not high, there is a lack of an effective correction mechanism for prediction errors, and how to achieve high-precision and robust dynamic posture prediction.

[0006] To solve the above technical problems, the present invention provides the following technical solutions: a TBM dynamic posture prediction method based on an ALA hybrid model, comprising obtaining time series data of the TBM dynamic posture from TBM working data and performing data preprocessing; selecting an ARIMA model based on the trend of the data, determining the regression order and the sliding average order; obtaining parameter estimates of the ARIMA model, running the ARIMA model to output prediction results and a residual sequence; using the residual sequence as input data of an LSTM model, using the LSTM model for prediction, and outputting an output sequence of the LSTM model; introducing an Attention mechanism, dynamically weighting the output sequence of the LSTM, and outputting a final LSTM-Attention output; fusing the prediction results of the ARIMA model with the prediction results of the LSTM-Attention model; and comparing the prediction results of the model with actual monitoring results and evaluating indicators.

[0007] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, wherein: the obtaining of time series data under the TBM dynamic posture from the TBM working data includes obtaining the time series data under the TBM dynamic posture from the TBM working data, and performing data preprocessing on the time series data.

[0008] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, wherein: the data preprocessing includes determining the trend of the data, determining the regression order and the sliding average order, and selecting the ARIMA model; Set the time series to And the regression equation is set as: ; ; in, Represents the value of the time series at time t, that is, the value at time point t. Represents the value of the time series at time t-1, that is, the value at the previous time point, represents the first-order difference of the time series y at time t, represents the coefficient to be tested, represents the lag term coefficient, represents the lag order, represents the error term; The estimated results are used to calculate the test statistic and compared with the critical value of the ADF test; If the test statistic is less than the critical value, the null hypothesis is rejected and the sequence is considered to be stationary; otherwise, the null hypothesis is not rejected and the probability of the sequence being non-stationary is high. Take the time series that is considered to be stable and combine it with the ACF image to determine and Order; When the ACF value gradually decays and approaches 0 with the increase of the lag order, that is, the order when the ACF graph shows a gradual decay but no truncation is the selected autoregressive order p; In the process of gradual decay, due to the periodicity of the time series, there is another peak value smaller than the maximum peak value after the lag q order, and the moving average order q is obtained; The calculation formula of Bayesian Information Criterion BIC is expressed as: ; in, represents the number of samples, Indicates the number of parameters to be estimated in the model. For the ARIMA model, , represents the autoregressive order, represents the moving average order, 1 is a constant term, represents the maximum likelihood estimate of the model.

[0009] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, wherein: the selection of the ARIMA model based on the trend of the data, determining the regression order and the sliding average order includes outputting the parameter estimation value of the ARIMA model, running the ARIMA model to output the prediction result and the residual sequence; In the ARIMA model, the output AR part and MA part are expressed as: ; ; in, represents a constant term, represents the AR parameter, represents the MA parameters, represents the error term; The AR and MA parameters are fitted using the least squares method to minimize the sum of squared errors { } to estimate AR parameters , that is, Regress the lagged terms and output AR parameter estimates; To minimize the residual sum of squares { } to estimate MA parameters , that is, regression is performed based on the error term and output to MA parameter estimation; According to the estimated parameters of AR and MA parts and , using the model's prediction formula Calculate the model's predictions , denotes the estimated parameters of the autoregressive part, Indicates that the time series is The autoregressive part predicts the current value by weighted summing of past values ​​of the time series. represents the estimated parameters of the moving average part, Indicates that the time series is The moving average part adjusts the predicted value by weighted summation of past error terms. Output residual sequence To further diagnose whether the model is suitable, the residual sequence is input into the LSTM model.

[0010] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, wherein: the residual sequence is used as the input data of the LSTM model, the LSTM model is used for prediction, and the output sequence of the LSTM model is output; the Attention mechanism is introduced, the output sequence of the LSTM is dynamically weighted, and the final LSTM-Attention output includes the residual sequence of the ARIMA model After integration, continue to build the input data set of the LSTM model based on 80% of the TBM training data and 20% of the test data; The information flow is controlled by the gate mechanism. The data processed by the forget gate is passed to the input gate and mapped into the interval [0, 1] by the sigmoid function. The tanh layer is used to create new candidate values. , and multiply the input gate and the candidate value as the new memory unit Assignment; The LSTM neural network consists of 4 parts: , Input Gate , candidate memory , Unit Status , output gate and hidden state ; LSTM is represented as: ;

[0011] ;

[0012] ;

[0013] ;

[0014] ;

[0015] in, and Indicates that the hidden layer is Moment and The state of the memory unit at the moment, and Respectively The unit input and The unit output vector at time , and Respectively represent the weight matrix of the forget gate, input gate, tanh unit and input gate, and Respectively represent the bias vectors of the corresponding layer weight matrix; Use the training set to train the LSTM model; During the training process, the model is verified using the validation set, and the changes in the model's loss value and evaluation indicators on the validation set are monitored to prevent overfitting; When the loss value on the validation set no longer decreases or shows an upward trend, stop training and save the LSTM model; Add Attention mechanism to the output layer of LSTM model; For the sequence of LSTM output , represents the sequence length, Indicates LSTM output of the time step, and the Attention weight of each time step is calculated through the fully connected layer , The weight is expressed as: ;

[0016] ;

[0017] in, , represents a learnable parameter; The Attention weight With LSTM output Perform weighted summation and output LSTM-Attention output .

[0018] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, the prediction result fusion includes fusing the prediction result of the ARIMA model and the prediction result of the LSTM-Attention model, and outputting the final hybrid model prediction result.

[0019] As a preferred solution of the TBM dynamic posture prediction method based on the ALA hybrid model described in the present invention, wherein: the comparison of the prediction results of the model with the actual monitoring results and the index evaluation include using the test set data to evaluate the prediction results of the hybrid model; Evaluation index determination coefficient , mean absolute error MAE, and root mean square error RMSE are expressed as: ;

[0020] ;

[0021] ;

[0022] in, The test set True value, represents the predicted value of the mixed model, Indicates the number of test set samples.

[0023] Another object of the present invention is to provide a TBM dynamic posture prediction system based on the ALA hybrid model, which can solve the problem that the current dynamic posture prediction method lacks an effective correction mechanism for prediction errors by performing parameter estimation and optimization based on the ARIMA model.

[0024] As a preferred solution of the TBM dynamic posture prediction system based on the ALA hybrid model described in the present invention, it includes: a data acquisition and preprocessing module, a model prediction and optimization module, and a prediction fusion and result verification module; the data acquisition and preprocessing module is used to obtain time series data under dynamic posture from TBM working data, and perform data preprocessing, after determining the data trend, set the regression order and the sliding average order, and select the optimal ARIMA model; the model prediction and optimization module is used to use the ARIMA model to perform parameter estimation, obtain prediction results and residual sequences, use the residual sequences as inputs of the LSTM model, combine the Attention mechanism for dynamic weighted optimization, and generate LSTM-Attention prediction results; the prediction fusion and result verification module is used to fuse the ARIMA prediction results and the LSTM-Attention prediction results, generate hybrid model prediction results, compare the prediction results with the actual monitoring results, and evaluate and verify the model performance.

[0025] A computer device includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement a step of a TBM dynamic posture prediction method based on an ALA hybrid model.

[0026] A computer-readable storage medium stores a computer program, which, when executed by a processor, implements the steps of a TBM dynamic posture prediction method based on an ALA hybrid model.

[0027] Beneficial effects of the invention: The TBM dynamic posture prediction method based on the ALA hybrid model provided by the invention has the advantages of high prediction result precision and high accuracy, and realizes multi-posture prediction of TBM in front horizontal, front vertical, rear horizontal and rear vertical dynamic postures, providing more abundant information for efficient and safe excavation of shield machines. The invention achieves better results in terms of integration of multiple data feature processing capabilities, long-term prediction capabilities and model adaptability. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0029] Figure 1 A diagram of the implementation steps of a TBM dynamic posture prediction method based on an ALA hybrid model provided in the first embodiment of the present invention.

[0030] Figure 2 A flowchart of a TBM dynamic posture prediction method based on the ALA hybrid model is provided for the first embodiment of the present invention.

[0031] Figure 3 A deviation diagram of the front horizontal posture of a TBM dynamic posture prediction method based on the ALA hybrid model provided in the second embodiment of the present invention.

[0032] Figure 4 A deviation diagram of a front vertical posture of a TBM dynamic posture prediction method based on an ALA hybrid model provided in the second embodiment of the present invention.

[0033] Figure 5 A deviation diagram of a TBM dynamic posture prediction method based on an ALA hybrid model in a rear horizontal posture provided in the second embodiment of the present invention.

[0034] Figure 6A deviation diagram of a TBM dynamic posture prediction method based on an ALA hybrid model in a rear vertical posture provided in the second embodiment of the present invention.

[0035] Figure 7 An overall flow chart of a TBM dynamic posture prediction system based on the ALA hybrid model provided in the third embodiment of the present invention. DETAILED DESCRIPTION

[0036] In order to make the above-mentioned purposes, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the drawings of the specification. Obviously, the described embodiments are part of the embodiments of the present invention, but not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary persons in the art without creative work should fall within the scope of protection of the present invention.

[0037] Example 1, reference Figure 1 - Figure 2 , as an embodiment of the present invention, provides a TBM dynamic posture prediction method based on the ALA hybrid model, comprising: S1: Obtain the time series data of the TBM dynamic posture from the TBM working data and perform data preprocessing.

[0038] Furthermore, the time series data of the TBM in the dynamic postures of front horizontal deviation, front vertical deviation, rear horizontal deviation and rear vertical deviation are collected from the TBM working data, and the collection frequency is Considering the characteristics of the data and related research, the extracted data is extracted with a time interval of 60 seconds to extract real-time monitoring data. In addition, 80% of the pre-processed TBM dynamic posture time series data is used as a training set; 20% is used as a test set.

[0039] S2: Select the ARIMA model based on the trend of the data, determine the regression order and the moving average order.

[0040] Furthermore, after determining the trend of the data, the regression order and the sliding average order are determined, and a more accurate ARIMA model is selected. By setting the time series to and setting the regression equation, the estimated results are calculated to test the statistic and compare it with the critical value of the ADF test. Take a stationary time series and determine the order in combination with the autocorrelation function (ACF) image. The goodness of fit and complexity of the model are considered. After obtaining ARIMA (p, q) of different groups under different postures, we use the evaluation index BIC to calculate the BIC value of each model and find the model order combination ARIMA (p, q) that minimizes BIC.

[0041] After determining the trend of the data, determine the regression order and sliding average order, and select the ARIMA model; Set the time series to And the regression equation is set as: ;

[0042] ;

[0043] in, Represents the value of the time series at time t, that is, the value at time point t. Represents the value of the time series at time t-1, that is, the value at the previous time point, represents the first-order difference of the time series y at time t, represents the coefficient to be tested, represents the lag term coefficient, represents the lag order, represents the error term; The estimated results are used to calculate the test statistic and compared with the critical value of the ADF test; If the test statistic is less than the critical value, the null hypothesis is rejected and the sequence is considered to be stationary; otherwise, the null hypothesis is not rejected and the probability of the sequence being non-stationary is high. Take the time series that is considered to be stable and combine it with the ACF image to determine and Order; When the ACF value gradually decays and approaches 0 with the increase of the lag order, that is, the order when the ACF graph shows a gradual decay but no truncation is the selected autoregressive order p; In the process of gradual decay, due to the periodicity of the time series, there is another peak value smaller than the maximum peak value after the lag q order, and the moving average order q is obtained; The calculation formula of Bayesian Information Criterion BIC is expressed as: ;

[0044] in, represents the number of samples, Indicates the number of parameters to be estimated in the model. For the ARIMA model, , represents the autoregressive order, represents the moving average order, 1 is a constant term, represents the maximum likelihood estimate of the model.

[0045] S3: Get the parameter estimates of the ARIMA model, run the ARIMA model to output the forecast results and residual sequence.

[0046] Furthermore, the parameter estimates of the ARIMA model are obtained, and the ARIMA model is run to obtain the forecast results and residual sequence. Output the parameter estimates of the ARIMA model, and run the ARIMA model to output the forecast results and residual sequence; In the ARIMA model, the output AR part and MA part are expressed as: is the AR parameter) ( is the MA parameter, is the error term) in, represents a constant term, represents the AR parameter, represents the MA parameter, represents the error term; The AR and MA parameters are fitted using the least squares method to minimize the sum of squared errors { } to estimate AR parameters , that is, Regress the lagged terms and output AR parameter estimates; To minimize the residual sum of squares { } to estimate MA parameters , that is, regression is performed based on the error term and output to MA parameter estimation; According to the estimated parameters of AR and MA parts and , using the model's prediction formula Calculate the model's predictions , denotes the estimated parameters of the autoregressive part, Indicates that the time series is The autoregressive part predicts the current value by weighted summing of past values ​​of the time series. represents the estimated parameters of the moving average part, Indicates that the time series is The moving average part adjusts the predicted value by weighted summation of past error terms. Output residual sequence To further diagnose whether the model is suitable, the residual sequence is input into the LSTM model.

[0047] S4: Use the residual sequence as the input data of the LSTM model, use the LSTM model to predict, and obtain the output sequence of the LSTM model; introduce the Attention mechanism, dynamically weight the output sequence of the LSTM. Obtain the final LSTM-Attention output. Use the residual sequence as the input data of the LSTM model, use the LSTM model to predict, and obtain the output sequence of the LSTM model.

[0048] Furthermore, the residual sequence of the ARIMA model After integration, continue to build the input data set of the LSTM model according to 80% of the TBM training data and 20% of the test data.

[0049] The core of LSTM is to control the information flow through the gate mechanism (input gate, forget gate, output gate). The forget gate is the first step after the data is input. Its main function is to evaluate the importance of the information in the previous memory cell, so as to purposefully forget some information. The data processed by the forget gate is passed to the input gate and mapped into the interval [0, 1] through the sigmoid function. Then the tanh layer is used to create new candidate values. , and multiply the input gate and the candidate value as the new memory unit Assignment. Figure 2 It can be seen that the LSTM neural network is mainly composed of 4 parts: the forget gate , Input Gate , candidate memory , Unit Status , output gate and hidden state The formula for LSTM is as follows: ;

[0050] ;

[0051] ;

[0052] ;

[0053] ;

[0054] in, and Indicates that the hidden layer is Moment and The state of the memory unit at the moment, and Respectively The unit input and The unit output vector at time , and Respectively represent the weight matrix of the forget gate, input gate, tanh unit and input gate, and Respectively represent the bias vectors of the corresponding layer weight matrix; Use the training set to train the LSTM model; During the training process, the model is verified using the validation set, and the changes in the model's loss value and evaluation indicators on the validation set are monitored to prevent overfitting; When the loss value on the validation set no longer decreases or shows an upward trend, stop training and save the LSTM model; Add Attention mechanism to the output layer of LSTM model; For the sequence of LSTM output , represents the sequence length, Indicates LSTM output of the time step, and the Attention weight of each time step is calculated through the fully connected layer , The weight is expressed as: ;

[0055] ;

[0056] in, , represents a learnable parameter; The Attention weight With LSTM output Perform weighted summation and output LSTM-Attention output .

[0057] S5: Fusion the prediction results of the ARIMA model and the LSTM-Attention model.

[0058] Furthermore, the prediction results of the ARIMA model and the LSTM-Attention model are combined to obtain the final prediction results of the hybrid model. The LSTM model is trained using the training set. During the training process, the model is verified using the validation set, and the loss value and evaluation index changes of the model on the validation set are monitored to prevent overfitting. When the loss value on the validation set no longer decreases or shows an upward trend, stop training and save the trained LSTM model.

[0059] Add an Attention mechanism to the output layer of the LSTM model. The Attention mechanism is a technology widely used in deep learning models. Its core idea is to dynamically assign different weights based on the importance of different parts of the input data to the current task. When processing time series data, the Attention mechanism calculates the weight of each position, allowing the model to pay more attention to those elements that are more important for decision-making or prediction.

[0060] For the sequence of LSTM output ( is the sequence length, Indicates The LSTM output of each time step) is used to calculate the Attention weight of each time step through the fully connected layer. , the calculation formula is ,in ( , is a learnable parameter). With LSTM output Perform weighted summation to obtain the LSTM-Attention output (

[0061] S6: Compare the model's prediction results with the actual monitoring results and evaluate the indicators.

[0062] Furthermore, the prediction results of the model are compared with the actual monitoring results and the indicators are evaluated. The prediction results of the hybrid model are evaluated using the test set data. The determination coefficient of the evaluation indicator ( ), mean absolute error (MAE), and root mean square error (RMSE).

[0063] ;

[0064] ;

[0065] ;

[0066] in, The test set True value, is the predicted value of the mixed model, is the number of samples in the test set.

[0067] Example 2, reference Figure 3 - Figure 6, which is an embodiment of the present invention, provides a TBM dynamic posture prediction method based on the ALA hybrid model. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0068] First, the time series data of the TBM in the dynamic postures of front horizontal deviation, front vertical deviation, rear horizontal deviation and rear vertical deviation are collected from the TBM working data. The collection frequency is Considering the characteristics of the data and related research, the extracted data is extracted with a time interval of 60 seconds to extract real-time monitoring data. In addition, 80% of the pre-processed TBM dynamic posture time series data is used as a training set; 20% is used as a test set.

[0069] The estimated results are calculated by setting the time series and regression equation to compare with the critical value of the ADF test. Take a stable time series and combine it with the autocorrelation function (ACF) graph to determine and Order. After obtaining the ARIMA (p, q) of different groups under different excavation postures, this information is provided by the model fitting software Python. Using the BIC evaluation index, the optimal ARIMA model order of TBM under different excavation postures is obtained, which are (2, 2) for the front horizontal deviation, (2, 4) for the front vertical deviation, (2, 0) for the rear horizontal deviation, and (1, 2) for the rear vertical deviation. In the ARIMA model, the AR part and the MA part are calculated. Then, the AR and MA parameters are fitted using the least squares method to minimize the sum of squared errors to obtain the parameter estimates of AR and MA. Predict the predicted values ​​under the ARIMA prediction model. With the residual sequence .

[0070] The residual series of the ARIMA model Integrate and use the training set to train the LSTM model. During the training process, use the validation set to validate the model, monitor the changes in the model's loss value and evaluation indicators on the validation set, and prevent overfitting. When the loss value on the validation set no longer decreases or shows an upward trend, stop training and save the trained LSTM model. Add the Attention mechanism to the output layer of the LSTM model. With LSTM output Perform weighted summation to get the LSTM-Attention output.

[0071] The prediction results of the ARIMA model and the LSTM-Attention model are combined to obtain the final hybrid model prediction results. ; The prediction results of the hybrid model are evaluated using the test set data. The evaluation index determination coefficient ( ), mean absolute error (MAE), and root mean square error (RMSE).

[0072] Table 1 Evaluation index table of ARIMA-LSTM-Attention hybrid model under different dynamic postures

[0073] Table 1 shows the results of the evaluation indicators. Figure 3 to Figure 4 These are the data for each time period in different postures output by the ARIMA-LSTM-Attention hybrid model, which are the y values ​​output by the training set and the test set.

[0074] Example 3, reference Figure 7 , which is an embodiment of the present invention, provides a TBM dynamic posture prediction system based on the ALA hybrid model, including a data acquisition and preprocessing module, a model prediction and optimization module, and a prediction fusion and result verification module.

[0075] The data acquisition and preprocessing module is used to obtain time series data under dynamic posture from TBM working data, and perform data preprocessing. After determining the data trend, the regression order and sliding average order are set, and the optimal ARIMA model is selected; the model prediction and optimization module is used to use the ARIMA model for parameter estimation, obtain the prediction results and residual sequence, and use the residual sequence as the input of the LSTM model, combined with the dynamic weighted optimization of the Attention mechanism, to generate the LSTM-Attention prediction results; the prediction fusion and result verification module is used to fuse the ARIMA prediction results and the LSTM-Attention prediction results, generate the hybrid model prediction results, compare the prediction results with the actual monitoring results, and evaluate and verify the model performance.

[0076] If the function is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or the part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium, including several instructions for a computer device (which can be a personal computer, server, or network device, etc.) to perform all or part of the steps of the methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), disk or optical disk, etc. Various media that can store program codes.

[0077] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as an ordered list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by an instruction execution system, device or apparatus (such as a computer-based system, a system including a processor, or other system that can fetch instructions from an instruction execution system, device or apparatus and execute instructions), or in conjunction with such instruction execution systems, devices or apparatuses. For the purposes of this specification, "computer-readable medium" can be any device that can contain, store, communicate, propagate or transmit a program for use by an instruction execution system, device or apparatus, or in conjunction with such instruction execution systems, devices or apparatuses.

[0078] More specific examples of computer-readable media (a non-exhaustive list) include the following: an electrical connection with one or more wires (electronic devices), a portable computer disk case (magnetic devices), a random access memory (RAM), a read-only memory (ROM), an erasable and programmable read-only memory (EPROM or flash memory), an optical fiber device, and a portable compact disk read-only memory (CDROM). In addition, the computer-readable medium may even be a paper or other suitable medium on which the program is printed, since the program may be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, deciphering or, if necessary, processing in another suitable manner, and then stored in a computer memory.

[0079] It should be understood that the various parts of the present invention can be implemented by hardware, software, firmware or a combination thereof. In the above-mentioned embodiments, multiple steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. For example, if implemented by hardware, as in another embodiment, it can be implemented by any one of the following technologies known in the art or their combination: a discrete logic circuit having a logic gate circuit for implementing a logical function for a data signal, a dedicated integrated circuit having a suitable combination of logic gate circuits, a programmable gate array (PGA), a field programmable gate array (FPGA), etc. It should be noted that the above embodiments are only used to illustrate the technical solution of the present invention and are not limited. Although the present invention is described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that the technical solution of the present invention can be modified or replaced by equivalents without departing from the spirit and scope of the technical solution of the present invention, which should be included in the scope of the claims of the present invention.

[0080] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A TBM dynamic posture prediction method based on ALA hybrid model, characterized in that: include: Obtain the time series data of the dynamic posture of the TBM from the TBM working data and perform data preprocessing; Select the ARIMA model based on the trend of the data, determine the regression order and the moving average order; Get the parameter estimates of the ARIMA model, run the ARIMA model to output the forecast results and residual sequence; The residual sequence is used as the input data of the LSTM model, and the LSTM model is used for prediction, and the output sequence of the LSTM model is output; the Attention mechanism is introduced, the output sequence of the LSTM is dynamically weighted, and the final LSTM-Attention output is output; Combine the prediction results of the ARIMA model and the LSTM-Attention model; The prediction results of the model are compared with the actual monitoring results and the indicators are evaluated.

2. The TBM dynamic posture prediction method based on the ALA hybrid model according to claim 1, characterized in that: The obtaining of the time series data of the TBM in a dynamic posture from the TBM working data includes obtaining the time series data of the TBM in a dynamic posture from the TBM working data and performing data preprocessing on the time series data.

3. The TBM dynamic posture prediction method based on the ALA hybrid model as claimed in claim 2, characterized in that: The data preprocessing includes determining the trend of the data, determining the regression order and the sliding average order, and selecting the ARIMA model; Set the time series to And the regression equation is set as: ; ; in, Represents the value of the time series at time t, that is, the value at time point t. Represents the value of the time series at time t-1, that is, the value at the previous time point, represents the first-order difference of the time series y at time t, represents the coefficient to be tested, represents the lag term coefficient, represents the lag order, represents the error term; The estimated results are used to calculate the test statistic and compared with the critical value of the ADF test; If the test statistic is less than the critical value, the null hypothesis is rejected and the sequence is considered to be stationary; otherwise, the null hypothesis is not rejected and the probability of the sequence being non-stationary is high. Take the time series that is considered to be stable and combine it with the ACF image to determine and Order; When the ACF value gradually decays and approaches 0 with the increase of the lag order, that is, the order when the ACF graph shows a gradual decay but no truncation is the selected autoregressive order p; In the process of gradual decay, due to the periodicity of the time series, there is another peak value smaller than the maximum peak value after the lag q order, and the moving average order q is obtained; The calculation formula of Bayesian Information Criterion BIC is expressed as: ; in, represents the number of samples, Indicates the number of parameters to be estimated in the model. For the ARIMA model, , represents the autoregressive order, represents the moving average order, 1 is a constant term, represents the maximum likelihood estimate of the model.

4. The TBM dynamic posture prediction method based on the ALA hybrid model as claimed in claim 3, characterized in that: The selecting of the ARIMA model based on the trend of the data, determining the regression order and the sliding average order includes outputting the parameter estimation value of the ARIMA model, running the ARIMA model to output the prediction result and the residual sequence; In the ARIMA model, the output AR part and MA part are expressed as: ; ; in, represents a constant term, represents the AR parameter, represents the MA parameters, represents the error term; The AR and MA parameters are fitted using the least squares method to minimize the sum of squared errors { } to estimate AR parameters , that is, Regress the lagged terms and output AR parameter estimates; To minimize the residual sum of squares { } to estimate MA parameters , that is, regression is performed based on the error term and output to MA parameter estimation; According to the estimated parameters of AR and MA parts and , using the model's prediction formula Calculate the model's predictions , denotes the estimated parameters of the autoregressive part, Indicates that the time series is The autoregressive part predicts the current value by weighted summing of past values ​​of the time series. represents the estimated parameters of the moving average part, Indicates that the time series is The moving average part adjusts the predicted value by weighted summation of past error terms. Output residual sequence To further diagnose whether the model is suitable, the residual sequence is input into the LSTM model.

5. The TBM dynamic posture prediction method based on the ALA hybrid model according to claim 4, characterized in that: The residual sequence is used as the input data of the LSTM model, and the LSTM model is used for prediction, and the output sequence of the LSTM model is output; the Attention mechanism is introduced, the output sequence of the LSTM is dynamically weighted, and the final LSTM-Attention output includes the residual sequence of the ARIMA model. After integration, continue to build the input data set of the LSTM model based on 80% of the TBM training data and 20% of the test data; The information flow is controlled by the gate mechanism. The data processed by the forget gate is passed to the input gate and mapped into the interval [0, 1] by the sigmoid function. The tanh layer is used to create new candidate values. , and multiply the input gate and the candidate value as the new memory unit Assignment; The LSTM neural network consists of 4 parts: , Input Gate , candidate memory , Unit Status , output gate and hidden state ; LSTM is represented as: ; ; ; ; ; in, and Indicates that the hidden layer is Moment and The state of the memory unit at the moment, and Respectively The unit input and The unit output vector at time , and Respectively represent the weight matrix of the forget gate, input gate, tanh unit and input gate, and Respectively represent the bias vectors of the corresponding layer weight matrix; Use the training set to train the LSTM model; During the training process, the model is verified using the validation set, and the changes in the model's loss value and evaluation indicators on the validation set are monitored to prevent overfitting; When the loss value on the validation set no longer decreases or shows an upward trend, stop training and save the LSTM model; Add Attention mechanism to the output layer of LSTM model; For the sequence of LSTM output , represents the sequence length, Indicates LSTM output of the time step, and the Attention weight of each time step is calculated through the fully connected layer , The weight is expressed as: ; ; in, , represents a learnable parameter; The Attention weight With LSTM output Perform weighted summation and output LSTM-Attention output .

6. The TBM dynamic posture prediction method based on the ALA hybrid model according to claim 5, characterized in that: The prediction result fusion includes fusing the prediction result of the ARIMA model and the prediction result of the LSTM-Attention model, and outputting the final hybrid model prediction result.

7. The TBM dynamic posture prediction method based on the ALA hybrid model according to claim 6, characterized in that: The comparison of the prediction results of the model with the actual monitoring results and the index evaluation include using the test set data to evaluate the prediction results of the hybrid model; Evaluation index determination coefficient , mean absolute error MAE, and root mean square error RMSE are expressed as: ; ; ; in, The test set True value, represents the predicted value of the mixed model, Indicates the number of test set samples.

8. A system using the TBM dynamic posture prediction method based on the ALA hybrid model as claimed in any one of claims 1 to 7, characterized in that: It includes data acquisition and preprocessing module, model prediction and optimization module, prediction fusion and result verification module; The data acquisition and preprocessing module is used to acquire time series data under dynamic posture from TBM working data, and perform data preprocessing. After determining the data trend, the regression order and sliding average order are set, and the optimal ARIMA model is selected; The model prediction and optimization module is used to use the ARIMA model to perform parameter estimation, obtain prediction results and residual sequences, use the residual sequences as inputs of the LSTM model, and combine the Attention mechanism for dynamic weighted optimization to generate LSTM-Attention prediction results; The prediction fusion and result verification module is used to fuse the ARIMA prediction results and the LSTM-Attention prediction results, generate a hybrid model prediction result, compare the prediction result with the actual monitoring result, and evaluate and verify the model performance.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the TBM dynamic posture prediction method based on the ALA hybrid model described in any one of claims 1 to 7 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the TBM dynamic posture prediction method based on the ALA hybrid model described in any one of claims 1 to 7 are implemented.

Images