Device thrust prediction method, apparatus, computer equipment, and storage medium based on fractional-order GRU networks

Abstract

This application discloses a method, apparatus, computer device, and storage medium for predicting equipment thrust based on a fractional-order GRU network, relating to the field of neural network technology, and capable of achieving accurate prediction of thrust in multiple regions. The method includes: constructing a GRU network by discretizing the fractional derivative of continuous RL and using an adjustable memory length parameter; determining a preset dynamic adjustment strategy; training the fractional-order GRU network using historical time-series data of the target equipment during its historical operation; updating the memory length parameter and the order of the fractional order according to the dynamic adjustment strategy to obtain a trained fractional-order GRU network; during equipment thrust prediction, using a correlation coefficient analysis algorithm to analyze the correlation between various parameters in the original excavation parameters and the thrust, and selecting the excavation parameters to be input; and inputting the excavation parameters to be input into the trained fractional-order GRU network for prediction.

Description

Technical Field

[0001] This application relates to the field of neural network technology, and in particular to a device thrust prediction method, apparatus, computer device, and storage medium based on a fractional-order GRU network. Background Technology

[0002] Recurrent Neural Networks (RNNs) and their variants have demonstrated significant advantages in time series data modeling. In particular, Gated Recurrent Units (GRUs) effectively alleviate the gradient vanishing and exploding problems of traditional RNNs by introducing a gating mechanism, thus enabling efficient processing of long-term sequence tasks. In the field of underground engineering, the thrust of an Earth Pressure Balanced Tunnel Boring Machine (EPB-TBM) is a key parameter for maintaining the stability and attitude control of the excavation face. Accurately predicting thrust changes is crucial for preventing face collapse or excessive surface subsidence.

[0003] In related technologies, there are several schemes for predicting the thrust of TBMs (Tunnel Boring Machines). For example, some schemes use the operating tunnel boring machine as the seismic source and achieve long-distance estimation of thrust based on passive seismic wave detection technology; other schemes construct prediction models based on fuzzy logic, or model the thrust under different geological conditions through mechanical decoupling; or with the popularization of deep learning technology in recent years, GRU networks have been introduced into the TBM thrust prediction task to achieve short-term prediction of total thrust using real-time collected operating data.

[0004] In the process of implementing the relevant technology, the applicant recognized that the relevant technology has at least the following technical problems: The hidden state update of GRU relies on the recursive form obtained by discretizing integer-order differential equations. Its memory range is essentially limited to the most recent information within a fixed time window. For thrust evolution processes that need to retain correlations over long time steps, this can easily lead to the rapid decay of historical information, thereby weakening the network's ability to capture long-distance time dependencies. Secondly, there are complex spatial coupling and temporal lag relationships between multi-region thrusts. In the standard GRU, the hidden state of each region is updated independently, which makes the model exhibit limited state representation capabilities when facing nonlinear and non-stationary thrust sequences. It is difficult to achieve accurate prediction of multi-region thrust and cannot provide a relatively accurate control basis during equipment attitude adjustment. Summary of the Invention

[0005] In view of this, this application provides a method, apparatus, computer device and storage medium for predicting device thrust based on fractional-order GRU networks. The main purpose is to solve the problems that it is currently difficult to achieve accurate prediction of thrust in multiple regions and to provide relatively accurate control basis during the process of device attitude adjustment.

[0006] According to a first aspect of this application, a method for predicting device thrust based on a fractional-order GRU network is provided, the method comprising: A GRU network is constructed. Based on the discretization of the fractional derivative of continuous RL, the difference equation of the hidden state in the GRU network is transformed, and an adjustable memory length parameter is introduced into the difference equation of the hidden state after transformation to obtain a fractional GRU network. A preset dynamic adjustment strategy is determined, and the fractional-order GRU network is trained using historical time-series data of the target device during its historical operation. During the training process, the memory length parameter and the order of the fractional order are updated according to the dynamic adjustment strategy to obtain the trained fractional-order GRU network. When predicting equipment thrust, the original excavation parameters of the target equipment at the time to be predicted are collected. The correlation coefficient analysis algorithm is used to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. The excavation parameters to be input are input into the trained fractional-order GRU network, and the multiple thrust prediction values ​​of the target equipment in multiple partitions are obtained from the trained fractional-order GRU network to complete the equipment thrust prediction of the target equipment.

[0007] According to a second aspect of this application, a device thrust prediction apparatus based on a fractional-order GRU network is provided, the apparatus comprising: A fractional-order GRU network construction module is used to construct a GRU network. Based on the discretization of the fractional derivative of continuous RL, the difference equation of the hidden state in the GRU network is transformed, and an adjustable memory length parameter is introduced into the difference equation of the hidden state after transformation to obtain the fractional-order GRU network. The training module is used to determine a preset dynamic adjustment strategy, train the fractional-order GRU network using historical time-series data of the target device during its historical operation, and update the memory length parameter and the order of the fractional order according to the dynamic adjustment strategy during the training process to obtain the trained fractional-order GRU network. The filtering module is used to collect the original excavation parameters of the target equipment at the time to be predicted when predicting equipment thrust, and use the correlation coefficient analysis algorithm to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. The prediction module is used to input the excavation parameters to be input into the trained fractional-order GRU network, and obtain multiple thrust prediction values ​​of the target equipment in multiple partitions output by the trained fractional-order GRU network, so as to complete the equipment thrust prediction of the target equipment.

[0008] According to a third aspect of this application, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the method described in any of the first aspects above.

[0009] According to a fourth aspect of this application, a storage medium is provided that stores a computer program thereon, which, when executed by a processor, implements the steps of the method described in any one of the first aspects above.

[0010] By employing the above technical solutions, this application provides a method, apparatus, computer device, and storage medium for predicting equipment thrust based on a fractional-order GRU network. This application constructs a fractional-order GRU network based on the discretization of continuous RL fractional-order derivatives and introduces an adjustable memory length parameter into the hidden state update equation. This allows the hidden state at the current moment to integrate state information from all historical moments rather than relying solely on the previous moment, thereby enhancing the ability to capture long-term time dependencies during thrust evolution. Simultaneously, using a dynamic adjustment strategy, the memory length parameter can be dynamically adjusted according to the training cycle number, and the fractional-order order can be dynamically switched according to the update gate output. This enables the model to adaptively balance the weights of long-term and short-term memories during training, thereby improving the ability to represent the state of nonlinear and non-stationary thrust sequences. Furthermore, by using correlation coefficients to select multi-dimensional excavation parameters highly correlated with thrust as inputs, and combining the fusion characteristics of the fractional-order GRU for global historical information, the model can implicitly model the spatial coupling and time-delay relationships between different regions when predicting thrust in multiple regions. This achieves accurate prediction of thrust in multiple regions, providing accurate and reliable control basis for the attitude adjustment of tunnel boring machines under complex geological conditions.

[0011] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, the following are specific embodiments of this application. Attached Figure Description

[0012] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 This paper illustrates a flowchart of a device thrust prediction method based on a fractional-order GRU network provided in an embodiment of this application. Figure 2 This paper illustrates a schematic diagram of a GRU network structure provided in an embodiment of this application. Figure 3 This illustration shows a schematic diagram of a device thrust prediction device based on a fractional-order GRU network according to an embodiment of this application. Figure 4 A schematic diagram of the device structure of a computer device provided in an embodiment of this application is shown. Detailed Implementation

[0013] Exemplary embodiments of the present application will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present application are shown in the drawings, it should be understood that the present application may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this application will be thorough and complete, and will fully convey the scope of the present application to those skilled in the art.

[0014] This application provides a device thrust prediction method based on a fractional-order GRU network, such as... Figure 1 As shown, the method includes: S10: Construct a GRU network. Based on the discretization of the fractional derivative of continuous RL, transform the difference equation of the hidden state in the GRU network, and introduce an adjustable memory length parameter into the difference equation of the transformed hidden state to obtain a fractional GRU network.

[0015] In this embodiment, a traditional GRU network needs to be constructed first. The GRU network controls the flow of information through reset and update gates, and its hidden state update depends only on the state information from the previous time step. Then, discretization of the fractional derivative of continuous RL (Riemann-Liouville) transforms the difference equation of the hidden state in the GRU network, enabling the current hidden state to integrate all historical state information from the initial time step to the previous time step. Simultaneously, an adjustable memory length parameter is further introduced into the transformed hidden state update equation. The truncation range of historical states is limited to the time before the current moment. At that moment, exceeding The historical state information at each time step is discarded, thus obtaining a fractional-order GRU network with dynamic memory capabilities.

[0016] In this way, during the above process, by utilizing the nonlocal memory property of fractional calculus, the hidden state update equation can include the cumulative contribution of all historical moments, which can overcome the long-distance dependency decay problem caused by the traditional GRU relying only on the previous moment. At the same time, an adjustable memory length parameter is introduced to control the computational cost while retaining effective historical information, laying the foundation for subsequent dynamic adjustment.

[0017] S20: Determine the preset dynamic adjustment strategy, use the historical time series data of the target device in the historical operation process to train the fractional GRU network, and update the memory length parameter and the order of fractional order according to the dynamic adjustment strategy during the training process to obtain the trained fractional GRU network.

[0018] In this embodiment, a preset dynamic adjustment strategy needs to be determined. This strategy includes two sets of adjustment rules: one for the memory length parameter and the other for the fractional order. The dynamic adjustment strategy in this embodiment instructs the memory length parameter to be adjusted based on the ratio of the current training cycle number to the total number of cycles, while the fractional order is adjusted based on the comparison between the updated gate output at the current moment and the threshold of 0.5. This embodiment uses historical time-series data from the target device's historical operation to train the fractional-order GRU network. During training, the memory length parameter is updated according to the dynamic adjustment strategy described above in each cycle, and the fractional order is updated according to the same strategy at each time step, until all training cycles are completed, resulting in a trained fractional-order GRU network.

[0019] In this way, during the above process, using a larger memory length parameter in the early stage of training can learn long-term dependencies, while reducing the memory length parameter in the later stage of training can focus on recent patterns. At the same time, the order of the fractional order is dynamically switched according to the real-time output of the update gate, so that the model can automatically reduce the order of the fractional order when it needs to enhance historical memory and switch to the integer order when it needs to focus on the current state. This adaptively balances the weights of long-term and short-term memory and improves the model's ability to represent the state of nonlinear and non-stationary thrust sequences.

[0020] S30: When predicting equipment thrust, the original excavation parameters of the target equipment at the time to be predicted are collected. The correlation coefficient analysis algorithm is used to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters.

[0021] In this embodiment, when predicting equipment thrust, it is necessary to first collect the original excavation parameters of the target equipment at the time to be predicted. These original excavation parameters may specifically include multi-point sealed chamber pressure, rotation angle, screw conveyor speed, cutterhead torque, pitch angle, hinge angle, and propulsion speed. Then, a correlation coefficient analysis algorithm is used to calculate the correlation coefficient between each parameter in the original excavation parameters and the thrust. The absolute value of the correlation coefficient calculated for each parameter is compared with a preset parameter screening threshold, thereby selecting parameters whose absolute correlation coefficient value is greater than the preset threshold as the input excavation parameters.

[0022] In this way, by eliminating redundant parameters that are weakly correlated with thrust through quantitative correlation analysis, the input dimensionality can be reduced, the computational burden on the model can be decreased, and the statistical significance of the input features and the prediction target can be ensured, thereby improving the accuracy and robustness of the prediction.

[0023] For example, taking TBM as an example, assuming that the raw parameters collected at the time to be predicted include to Sealed chamber pressure, rotation angle Screw conveyor speed Cutter head torque ,thrust Pitch angle Angle measurement Hinge Angle Speed ​​of advancement Side measurement difference and cutter head speed There are 11 parameters, including the tool shroud speed. Analysis using a correlation coefficient algorithm revealed this. If the absolute value of the correlation coefficient is less than 0.3 of the preset parameter screening threshold, it is rejected; while the absolute values ​​of the correlation coefficients of the remaining 10 parameters are all greater than the preset parameter screening threshold, these 10 parameters are used as the input vector to be used as the excavation parameters.

[0024] S40: Input the excavation parameters to be input into the trained fractional-order GRU network, and obtain the target equipment's multiple thrust prediction values ​​in multiple partitions output by the trained fractional-order GRU network to complete the equipment thrust prediction for the target equipment.

[0025] In this embodiment of the application, the selected excavation parameters to be input need to be input into the trained fractional GRU network. The trained fractional GRU network will perform forward calculation step by step according to the truncated hidden state update equation, and finally output multiple thrust prediction values ​​of the target device in multiple partitions. Each partition corresponds to a thrust output component, thereby completing the device thrust prediction of the target device.

[0026] In the above process, the fractional-order GRU network can implicitly learn the spatial coupling and time delay relationship between the thrust of each region by fusing global historical information, realize the coordinated and accurate prediction of thrust in multiple regions, and provide independent and coordinated control basis for equipment attitude adjustment, thereby avoiding the defect that it is impossible to guide fine control of each region when only predicting the total thrust.

[0027] Taking the four-zone thrust prediction of TBM as an example, the selected 10-dimensional input parameters are fed into the trained fractional GRU network, and the network outputs the thrust of the four zones A, B, C and D as F1, F2, F3 and F4 respectively.

[0028] Optionally, in this embodiment, a GRU network is constructed by discretizing the continuous RL fractional derivative, transforming the difference equation of the hidden state in the GRU network, and introducing an adjustable memory length parameter into the transformed difference equation of the hidden state to obtain a fractional GRU network. This includes: constructing a GRU network including a reset gate and an update gate, and determining the continuous RL fractional derivative, wherein the expression of the continuous RL fractional derivative is shown in the following formula. , in, Represents the fractional derivative operator; Indicates the starting time for calculating the fractional derivative; This indicates the order of the fraction, and ; This represents a function whose fractional derivative needs to be calculated. Indicates the current time point; Indicates greater than The smallest integer, and ; Represents the gamma function; Representation function In the integral variable The value at; Indicates the integral variable The differential; Indicates to of First derivative; express Time to the current time The contribution weight of the derivative; determine the difference equation of the hidden state in the GRU network as shown in the following formula. , in, Indicates the current hidden state. This indicates the hidden state at the previous moment. This indicates an update to the gate output. Represent the candidate hidden state at the previous time step; convert the continuous RL fractional derivative to a step size. Discretization is then performed to obtain the discretized approximate expression shown in the following formula.

[0029] in, , By combining the discretized approximation expression with the difference equation of the hidden state, we obtain the expression for the hidden state as shown in the following formula. ; Transforming the expression for the hidden state yields the difference equation for the transformed hidden state, as shown in the following formula.

[0030] in, An adjustable memory length parameter is introduced into the difference equation of the transformed hidden state. The truncation range of historical moments is limited to those before the current moment. At that moment, exceeding Historical state information at each time step is discarded to obtain the truncated hidden state update equation, and the current GRU network containing the truncated hidden state update equation is used as a fractional-order GRU network.

[0031] In this embodiment, firstly, a conventional GRU network containing a reset gate and an update gate is constructed. The reset gate is used to control the degree to which the hidden state of the previous time step is ignored, and the update gate is used to determine the fusion ratio between the hidden state of the previous time step and the current candidate hidden state. The GRU network provides the basic structure for the subsequent introduction of a fractional-order memory mechanism.

[0032] Secondly, the fractional derivative of continuous RL is determined. The expression for the fractional derivative of continuous RL is shown in Formula 1 below: Formula 1: , In formula 1, Represents the fractional derivative operator; Indicates the starting time for calculating the derivative; The order of the fractional order is represented in the embodiments of this application. ,when When equal to 1, it degenerates into an integer derivative; Denotes the function to be differentiated; Indicates the current time; Indicates satisfaction The smallest integer; Represents the gamma function; Representation function In the integral variable The value at; Indicates the integral variable The differential; Indicates to of First derivative; express The function value at time t is relative to the current time. The contribution weight of the derivative decreases power-lawfully with increasing time interval, reflecting the nonlocal memory characteristic of fractional calculus. The memory effect of fractional calculus aligns well with the requirements of GRU for processing time series data, and the fractional derivative of RL has lower requirements for function smoothness. Therefore, it is selected to modify GRU in this embodiment.

[0033] Next, based on the GRU gating update mechanism, this embodiment will determine the difference equation expression of the hidden state shown in Equation 2 below: Formula 2: , In formula 2, This indicates the hidden state at the current moment. This indicates the hidden state at the previous moment. This indicates an update to the gate output. This represents the candidate hidden state at the previous time step. The difference equation for the hidden state described in Equation 2 indicates that the change in the hidden state from the previous time step to the current time step is adjusted by the update gate. Furthermore, in order to apply the continuous fractional derivative to the discrete time series, this embodiment discretizes the RL fractional derivative; specifically, it takes a step size of... Then, we can obtain the discretized approximation expression shown in Formula 3 below: Formula 3:

[0034] In formula 3, Indicates the first A discrete time point; For the first The hidden state at a historical moment; in Formula 3, the coefficient , These coefficients are determined by the order of the fractional order. and index Uniquely certain, and following It increases and monotonically decreases, thus ensuring that information from further back in history contributes less to the current derivative.

[0035] Next, combining the above discretization approximation with the difference equation of the hidden state, we can obtain the expression for the hidden state as shown in Equation 4 below: Formula 4: .

[0036] By performing an algebraic transformation on the expression for the hidden state described by Formula 4, the hidden state at the current time can be solved. The transformed hidden state update equation is shown in Formula 5 below: Formula 5:

[0037] In formula 5, The transformed hidden state update equation described by Equation 5 shows that the current hidden state depends not only on the state at the previous time step, but also on the state from the initial time step. arrive The hidden states of all historical moments are represented, and all coefficients are non-negative, thus achieving true global memory fusion.

[0038] Considering the number of time steps As the number of historical items increases linearly, in order to balance memory capacity and computational efficiency, an adjustable memory length parameter needs to be introduced into the transformed equation. The upper limit of the summation of historical terms is increased from... Cut off as That is, only the closest time before the current moment is retained. The state information at each historical moment is introduced, along with a truncation coefficient. Thus, the truncated hidden state update equation is obtained.

[0039] After completing the above process, the GRU network containing the truncated hidden state update equation can be used as the final fractional-order GRU network. Thus, in the above process, global historical memory is introduced through the discretization of the fractional derivative of RL, enabling the network to capture long-term time dependencies and overcome the problem of rapid decay of historical information caused by the first-order Markov assumption in traditional GRUs. Simultaneously, an adjustable memory length parameter... It allows for flexible truncation based on the actual sequence length and computing resources, avoiding redundant storage and computing overhead caused by infinitely growing historical items, and providing a structural basis for subsequent dynamic adjustment strategies.

[0040] Optionally, in this embodiment of the application, constructing a GRU network including a reset gate and an update gate includes: combining the input parameters at the current moment. and the hidden state of the previous moment Construct the reset gate that performs computation operations according to the following expression. , in, This represents the weight matrix of the reset gate. This represents the bias matrix of the reset gate. This represents the Sigmoid activation function. This involves concatenating the input parameters with the previous hidden state as a vector; constructing the candidate hidden state calculation expression as shown in the following formula, where the candidate hidden state calculation expression obtains the result of the reset gate filtering the previous hidden state, and combines the result with the current input parameters to calculate the candidate hidden state. , in, The weight matrix representing the candidate hidden state. The bias matrix represents the candidate hidden state. This represents element-wise multiplication. Construct an update gate that performs computation operations according to the following expression. , in, This represents the weight matrix of the updated gate. The bias matrix represents the update gate; the following expression is constructed for the current hidden state, which is used to perform a weighted fusion of the previous hidden state and the candidate hidden state according to the update gate, to obtain the GRU network including the reset gate and the update gate. .

[0041] In this embodiment, firstly, it is necessary to combine the input parameters at the current moment. and the hidden state of the previous moment Construct a reset gate that performs computation operations according to the expression described in Formula 6 below: Formula 6: , In formula 6, This represents the weight matrix of the reset gate. This represents the bias matrix of the reset gate. This means concatenating the input vector with the hidden state from the previous time step along the feature dimension. The sigmoid activation function, with an output range of (0,1), is used to control the opening and closing degree of the gating. The reset gate determines the hidden state from the previous moment. The value indicates how much information needs to be reset or ignored; the closer the value is to 0, the higher the degree of forgetting.

[0042] Next, the candidate hidden state calculation expression shown in Formula 7 is constructed, wherein the candidate hidden state calculation expression obtains the result of the reset gate filtering the hidden state of the previous time step, and combines the result with the input parameters of the current time step to calculate the candidate hidden state: Formula 7: , In formula 7, The weight matrix representing the candidate hidden state. This represents the corresponding bias matrix. This indicates element-wise multiplication. The hyperbolic tangent activation function has an output range of (-1, 1) and is used to compress candidate states to the normal range. Equation 7 represents the process of first passing through the reset gate. Hidden state from the previous moment Element-wise scaling is performed, then concatenated with the current input, followed by linear transformation and non-linear activation to generate a candidate hidden state that integrates the current input and filtered historical information.

[0043] Then, construct the update gate that performs the computation operation according to the expression shown in Formula 8 below: Formula 8: , In formula 8, and These represent the weight matrix and bias matrix of the update gate, respectively. The update gate also uses a sigmoid activation to output a value between (0,1), which controls the fusion ratio between the previous hidden state and the current candidate hidden state.

[0044] Finally, the expression for the current hidden state, shown in Equation 9 below, is constructed to perform a weighted fusion of the previous hidden state and the candidate hidden state based on the update gate, resulting in a GRU network including the reset gate and the update gate: Formula 9: , In Formula 9, when updating the gate When the value is close to 1, the current hidden state mainly comes from the candidate hidden states. That is, the network tends to adopt new input information; when When the value is close to 0, the hidden state from the previous moment is mainly retained. That is, to keep historical information unchanged.

[0045] Thus, in the above process, the gating mechanism of reset and update gates makes the GRU structure more streamlined and computationally less expensive, while also enabling flexible control over information forgetting and updating. This lays a compatible foundation for the subsequent introduction of fractional-order global memory. Because the update equation of a traditional GRU is essentially still a first-order Markov process, meaning the current hidden state only depends on the state of the previous time step and cannot explicitly utilize earlier historical information, the embodiments of this application subsequently modify the hidden state update equation on this basic network using fractional-order calculus, thereby overcoming this limitation.

[0046] In practical applications, the internal structure of the GRU network constructed in this embodiment is as follows: Figure 2 As shown, the input at the current moment The hidden state of the previous moment They jointly participate in the calculation of two gating units. First, the input... and hidden state The inputs are fed into the reset gate and the update gate, respectively, both of which are activated by the Sigmoid function. Compress the output value to the (0,1) interval to obtain the reset gate output. and update gate output Reset the door Used to control the hidden state of the previous moment. The degree to which it is ignored will and Multiply element by element, then multiply with the current input. splicing, after The activation function obtains the candidate hidden states. Update Gate This determines the hidden state at the previous moment. With candidate hidden state The fusion ratio, through calculation We obtain the weights that retain historical information, and finally the hidden state at the current moment. Calculated as .

[0047] Optionally, in an embodiment of this application, an adjustable memory length parameter is introduced into the difference equation of the transformed hidden state. The truncation range of historical moments is limited to those before the current moment. At that moment, exceeding Historical state information at each time step is discarded to obtain the truncated hidden state update equation, including: setting an adjustable memory length parameter. For positive integers, define the truncation coefficient as shown in the following expression, where the memory length parameter is... This indicates the number of historical moments retained.

[0048] In the difference equation of the transformed hidden state, the upper bound for the summation of the history terms is increased from... 1. Cut off as At the same time, a cutoff factor is introduced. To control the initial term By preserving the specified method, we obtain the truncated hidden state update equation shown in the following formula. .

[0049] In this embodiment, an adjustable memory length parameter needs to be introduced into the transformed hidden state difference equation. Specifically, the first step is to set an adjustable memory length parameter. and This is a positive integer, representing the maximum number of historical states that can be traced back and retained from the current time step. This is used to control the initial state. In this embodiment of the application, a cutoff coefficient is also defined to determine whether an element is included in the update equation. , The expression is shown in Formula 10 below: Formula 10:

[0050] In formula 10, This is the index of the current discrete time step. Corresponding to the At that moment. Cutoff coefficient. Its function is to set the memory length Less than the index minus 1 at the current time, that is At that time, the initial term The coefficient is set to zero, that is, the distant initial state is discarded; only when equal In other words, the initial item is only retained when the memory length is sufficient to cover the entire history from the initial moment to the previous moment.

[0051] Next, in the previously derived hidden state update equation that includes all historical items, the upper bound for the summation of historical items is increased from the original... 1. Cut off as At the same time, using the above cutoff coefficient By controlling the initial terms, we obtain the truncated hidden state update equation shown in Equation 11 below: Formula 11:

[0052] Similarly, in Formula 11, as described above, The order of the fractional order. ; This is the discretization step size; The table shows the gamma function; The update gate output for the current moment; The state was hidden in the previous moment; For the first The hidden state of a historical moment, among which Take from 2 That is, keep the most recent A historical state, plus common indivual; The initial hidden state; The candidate hidden state at the previous time step; coefficient Depend on Definition, and , These coefficients vary with the index. The increasing and monotonically decreasing trend reflects the principle that historical information contributes less to distant information, while its contribution decreases over time.

[0053] Thus, in the above process, when processing short-term sequences or when computational resources are limited, a smaller [specification value] can be set. Only the most recent historical states are retained, significantly reducing the number of summation terms and computational cost; when dealing with long-term sequences and needing to fully capture long-range dependencies, a larger value can be set. This application's embodiment preserves more historical information. Compared to the ideal fractional-order GRU that infinitely retains all historical states, the truncation operation in this embodiment sacrifices the small contribution of very early information in a controllable approximation, in exchange for feasibility in practical engineering applications. It avoids the linear expansion of the number of historical items as the time step increases, thus effectively controlling memory usage and computational complexity. At the same time, since the historical coefficients themselves decay, the error caused by truncation is limited to an acceptable range.

[0054] In this embodiment of the application, optionally, a preset dynamic adjustment strategy is determined, and a fractional-order GRU network is trained using historical time-series data of the target device during its historical operation. During training, the memory length parameter and the order of the fractional order are updated according to the dynamic adjustment strategy to obtain a trained fractional-order GRU network. This includes setting the total number of training cycles to... and setting constants The basic scaling is used to control the memory length parameter; the fractional-order GRU network is trained using historical time-series data, and in the current training cycle... In this process, a dynamic adjustment strategy is employed according to the following expression, based on the current cycle number. Proportional relationship with the total number of cycles Dynamically determine memory length parameters The possible values ​​of , where, Increasing from 1 to ,

[0055] At the same time, at each time step of the current training cycle In the middle, obtain the update gate output at the current time. And according to the following expression, based on The order of the fractional order at the current time is dynamically determined by comparing the result with the preset threshold of 0.5. The value of ,

[0056] The memory length parameter will be dynamically determined during the current training cycle. and the order of fractions The values ​​are substituted into the truncated hidden state update equation in the fractional-order GRU network for forward computation and backpropagation to complete the training of the current training cycle; iterative training continues until completion. After one training cycle, a fractional-order GRU network is obtained.

[0057] In this embodiment, firstly, the total number of training cycles needs to be set to... ,For example The value can be 50 or 70, and a constant is set. The base scaling factor is used to control the memory length parameter. Then, the fractional-order GRU network is trained using historical time-series data from the target device during its historical operation, with each training cycle... In the middle, based on the current cycle number and half of the total number of cycles... The proportional relationship is used to determine the memory length parameter according to the dynamic adjustment strategy shown in Formula 12 below. The possible values ​​of , where, Increasing from 1 to : Formula 12:

[0058] In formula 12, This represents the current training epoch number, i.e., the number of batches completed or the epoch number. This is the preset total number of training cycles; It is a constant used to scale the absolute value of the memory length; Indicates taking 2 and The larger value in the formula ensures that the memory length is at least 2, avoiding truncation that would prevent the capture of any historical dependencies. The rule described in Formula 12 applies to the first half of training, i.e. The length of memory increases linearly with the number of cycles, allowing the model to gradually learn increasingly longer historical dependencies; in the latter half, that is... The memory length decreases linearly with the number of cycles, allowing the model to focus on recent states in the later stages of training, thereby improving convergence stability and avoiding overfitting.

[0059] Furthermore, the embodiments of this application will also include each time step in each training cycle. In the middle, obtain the update gate output at the current time. The order of the fractional order at the current moment is dynamically determined according to the rules shown in Formula 13 below. The possible values ​​of: Formula 13:

[0060] In formula 13, It is the order of the fractional order for enhancing memory, which is preset according to the characteristics of the task. It can be a positive number less than 1, such as 0.7 or 0.8. The output value of the update gate at the current moment reflects whether the network tends to retain historical states or adopt candidate new states. According to the rule described in Equation 13, when the update gate output is less than or equal to 0.5, it indicates that the network currently relies more on historical information, and in this case, a fractional order less than 1 is used. This method leverages the nonlocal memory property of fractional calculus to enhance the retention of accumulated historical states over long periods. When the update gate output is greater than 0.5, it indicates that the current input information is more critical, and at this point... Setting it to 1 degenerates the fractional-order GRU network into a traditional integer-order GRU, enabling it to respond quickly to the latest changes and avoid excessive historical interference.

[0061] At each time step of each training cycle, the dynamically determined memory length parameter mentioned above will be... and the order of fractions Substituting the values ​​into the truncated hidden state update equation of the fractional-order GRU network, forward computation is performed to obtain the current hidden state and predicted output. Then, backpropagation is performed based on the prediction error to update the weights and bias parameters in the network. After one training cycle is completed, the next cycle begins, and the update continues according to the above rules. and until all are completed. After several training cycles, a fractional-order GRU network is finally obtained.

[0062] In this way, through the above process, the memory length parameter is stored. The training cycle is adaptive, in the early stages of training. When the initial latency is small, the model first learns short-term dependencies, quickly reducing the loss; as training progresses... Gradually increase the size of the sequence, exposing the model to long-term dependencies at an earlier stage, thus avoiding gradient instability caused by handling extremely long sequences from the beginning; in the later stages of training... Further reduction allows the model to focus on recent patterns, improving generalization ability. On the other hand, the order of the fractional order is then considered. Time-step adaptation, updating the gate in real time by monitoring This determines whether to enable fractional-order memory enhancement, allowing the model to automatically switch memory modes based on the rate of change of the data itself. Fractional-order memory is enabled in stages where the trend is stable and long-term information accumulation is required, while fractional-order memory is disabled and switched to integer-order memory in stages of sudden changes or high-frequency fluctuations, thereby achieving the best balance between memory strength and response speed.

[0063] Optionally, in this embodiment, iterative training can continue until completion. After obtaining the trained fractional-order GRU network in one training cycle, the method further includes: obtaining a test sample set, which includes... One sample, In the nth sample The true thrust value of each sample is The test sample set is input into the trained fractional-order GRU network, which then predicts the thrust value for each sample. The mean absolute percentage error, root mean square error, and coefficient of determination of the trained fractional-order GRU network in thrust value prediction are calculated using the following formulas.

[0064] in, Indicates the mean absolute percentage error. This indicates the calculation of the root mean square error. The coefficient of determination is represented by the mean absolute percentage error, root mean square error, and coefficient of determination. These are used as performance evaluation metrics for the trained fractional-order GRU network on the multi-region thrust prediction task, and the performance evaluation metrics are output.

[0065] In this embodiment, after completing all After training for one epoch and obtaining the trained fractional-order GRU network, the network's prediction performance will be further quantitatively evaluated. Specifically, this first requires obtaining a test sample set, which includes... Each sample is independent of the training data, and each sample records a set of real thrust values. These true values ​​can come from thrust data collected and stored by the TBM programmable logic controller in actual engineering projects. In this embodiment, input parameters from the test sample set, such as 10-dimensional features like chamber pressure, cutterhead torque, and propulsion speed, are input into a trained fractional-order GRU network. The network outputs a corresponding thrust prediction value for each sample. .

[0066] To comprehensively evaluate the prediction accuracy and fitting effect, this application uses three regression evaluation indicators, specifically including mean absolute percentage error, root mean square error, and coefficient of determination, calculated as shown in Formula 14 below: Formula 14:

[0067] In formula 14, The total number of test samples; For the first The true thrust value of each sample; This represents the network's predicted thrust value for this sample. It is the arithmetic mean of all actual thrust values.

[0068] in, The prediction error is measured as a percentage relative to the true value. The smaller the value, the more accurate the prediction. It is also unaffected by dimensions and facilitates comparison between different operating conditions. Square root of the error can amplify the impact of larger deviations, is sensitive to outliers in the prediction, and reflects the overall degree of deviation between the predicted value and the true value. This indicates the degree to which the predicted value explains the variation in the true value. The value ranges from [0,1]. The closer it is to 1, the better the model fit and the stronger the predictive ability.

[0069] After calculating the above three indicators, this application embodiment will use them as performance evaluation indicators of the trained fractional-order GRU network on the multi-region thrust prediction task, and output them for engineers to refer to. In this way, through the three complementary quantitative indicators of MAPE, RMSE and R2R2, the performance of the prediction model is comprehensively measured from three dimensions: relative error, absolute error and goodness of fit. This provides an objective basis for the effectiveness of the model under different geological conditions and different engineering sections, and also provides a quantitative comparison benchmark for subsequent model hyperparameter tuning and engineering deployment.

[0070] In this embodiment of the application, optionally, the original excavation parameters of the target equipment at the time to be predicted are collected, and the correlation coefficient analysis algorithm is used to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. This includes: collecting the parameters of the target equipment at the time to be predicted to obtain the original excavation parameters, wherein the original excavation parameters include multi-point sealed chamber pressure, rotation angle, etc. Screw conveyor speed Cutter head torque ,thrust Pitch angle Angle measurement Hinge Angle Speed ​​of advancement Side measurement difference and cutter head speed Multi-point sealed chamber pressure refers to the pressure values ​​at different spatial locations within the target equipment, including... to Using the following formula, the correlation coefficient analysis algorithm is used to calculate the relationship between each parameter in the original excavation parameters and the thrust. Correlation coefficient between , , in, and Representing parameters respectively and thrust The One observation value, and Representing parameters respectively and thrust The sample mean, To determine the number of observation samples, the absolute values ​​of the correlation coefficients corresponding to each calculated parameter are compared with preset parameter screening thresholds. Parameters with absolute values ​​of correlation coefficients greater than the preset parameter screening thresholds are selected as input excavation parameters. These input excavation parameters include the multi-point sealed chamber pressure. to Rotation angle Screw conveyor speed Cutter head torque Pitch angle Hinge Angle Speed ​​of advancement .

[0071] In this embodiment, firstly, it is necessary to collect various operating parameters of the target equipment at the time to be predicted to obtain the original excavation parameters. The target equipment can be an earth pressure balance tunnel boring machine (TBM). The original excavation parameters are initially selected based on engineering experience and equipment sensor configuration, and may specifically include parameters denoted as... to Multi-point sealed chamber pressure, to The pressure values ​​correspond to different spatial locations within the sealed chamber; it also includes the rotation angle. Screw conveyor speed Cutter head torque ,thrust Pitch angle Angle measurement Hinge Angle Speed ​​of advancement Side measurement difference and cutter head speed In practical applications, the original excavation parameters can be read from a programmable logic controller (PLC) or an industrial computer database.

[0072] In order to filter out the input variables that are truly related to the thrust and have statistical significance from the above-mentioned original excavation parameters, the embodiments of this application will use a correlation coefficient analysis algorithm to calculate the relationship between each parameter and the thrust. The degree of linear correlation between them, where the correlation coefficient analysis algorithm can be the Pearson algorithm, and the specific calculation formula is shown in Formula 15 below: Formula 15: , In formula 15, This refers to a specific parameter, such as the pressure in the sealed chamber. The One observation value, The first represents the thrust. One observation value; and These are the sample mean values ​​for the parameter and the thrust, respectively. This represents the number of observation samples used to calculate the correlation coefficient. The range of values ​​for is [ [1,1], the closer the absolute value is to 1, the stronger the linear correlation, and the closer it is to 0, the weaker the linear correlation.

[0073] In this embodiment, the absolute values ​​of the correlation coefficients corresponding to the calculated parameters are compared with a preset screening threshold, for example, 0.3. Correlation coefficients with an absolute value greater than the preset screening threshold are considered to have a moderate or high correlation and are retained; those less than or equal to 0.3 are considered weakly correlated parameters and are discarded. In actual analysis, it is known that thrust and multi-point sealed chamber pressure... to Rotation angle Screw conveyor speed Cutter head torque Pitch angle Hinge Angle Speed ​​of advancement The absolute values ​​of the correlation coefficients were all greater than 0.3, while the correlation coefficients with the side measurement difference were all greater than 0.3. and cutter head speed The correlation between these parameters is relatively weak. Therefore, the final selected excavation parameters in this application embodiment can be combined into a 10-dimensional vector as described in Formula 16 below: Formula 16: .

[0074] In other words, in this embodiment, the input at each moment consists of the 10 parameters described in Formula 16 above. The output variable is the thrust of each region, denoted as... ,in .

[0075] In this way, through quantitative Pearson correlation coefficient analysis, the redundancy or noise that may be introduced by relying on subjective experience to select input variables can be avoided, and parameters that are weakly correlated with thrust can be eliminated. This effectively reduces the input dimension of the fractional-order GRU network, which can not only reduce the computational cost and overfitting risk of the model, but also improve the model's focus on the core driving factors, making subsequent thrust predictions more accurate and robust.

[0076] It should be noted that the technical solutions in this application embodiment have been verified through field experiments. Specifically, actual operating data from a certain subway line in a certain region were used for experimental verification. The geological conditions of the subway line in that region are typical sandy soft rock. All data were recorded by the programmable logic controller of the tunnel boring machine and then stored in a database after being read by an industrial computer. The experimental environment was configured with an Intel i9-14900HX processor, an RTX 4060 graphics card with 8 GB of video memory, and the software was based on PyCharm, programmed in Python under the TensorFlow framework. The comparison models included the fractional-order GRU network proposed in this application embodiment, the traditional GRU, and LSTM (Long Short-Term Memory). The multi-region thrust prediction performance was evaluated through controlled experiments.

[0077] In the prediction performance experiment under the same geological conditions, the learning rate was set to 0.001, the training period to 50, the number of hidden layers to 1, and the number of neurons to 15. During training, the loss value of the fractional-order GRU network steadily decreased and gradually approached zero, verifying the rationality of the parameter selection. Compared with LSTM and GRU, RLGRU consistently maintained a lower loss value, converged faster, and demonstrated superior early prediction performance. The fractional-order GRU network showed the best consistency between the predicted thrust value and the measured value, with a more stable error curve and a significantly reduced overall error, demonstrating good stability and accuracy.

[0078] Furthermore, in experiments under different geological conditions, two types of non-clay geological data were selected, designated Soil1 and Soil2, with a training cycle count of 70. Other parameters remained the same as in the previous experiments. The fractional-order GRU network accurately fitted the measured values ​​in both the drastic and stable thrust fluctuation ranges, demonstrating higher prediction accuracy than LSTM and GRU. Furthermore, the percentage error curve exhibited a smaller fluctuation range, and the overall error value was significantly lower than the other two models. The thrust in the first partition of Soil1 was used as an example. For example, the mean absolute percentage error (MAPE) of the fractional-order GRU network is 1.18%, which is 78.1% higher than that of GRU (5.41%) and 79.3% higher than that of LSTM (5.70%). Similar performance improvements were observed in other partitions and in Soil2. Furthermore, the coefficient of determination (R²) of the fractional-order GRU network consistently remains close to 1, indicating that its predictive ability is significantly stronger than that of GRU and LSTM under different geological conditions.

[0079] In addition, to verify the model's generalization ability in different engineering projects, data from a subway line in other regions were selected for comparative experiments. The training parameters were consistent with different geological experiments, and the total thrust was used as the prediction target. The results showed that the fractional-order GRU network tracked the measured thrust curve most closely, while LSTM and GRU generally exhibited response delays and biases. The relative error of the fractional-order GRU network remained consistently below 3%, with small overall fluctuations and a stable error distribution, demonstrating strong prediction robustness. The errors of LSTM and GRU varied significantly and were amplified with changes in the operating scenario. Quantitatively, the root mean square error (RMSE) of the fractional-order GRU network was 181.35, and the MAPE was only 0.97%. 2 The value is close to 1. Compared to LSTM, the fractional-order GRU network reduces RMSE by approximately 49.1% and MAPE by 14.2%; compared to GRU, it reduces RMSE by approximately 67.2% and MAPE by 51.3%. These results demonstrate that the fractional-order GRU network significantly outperforms LSTM and GRU in error control and prediction accuracy.

[0080] In summary, the fractional-order GRU network proposed in this application has successfully captured state information at multiple historical time points by introducing a global history memory mechanism based on continuous RL fractional derivatives, an adjustable memory length parameter, and a dynamic order adjustment strategy. The model exhibits good prediction accuracy and stability under the same geological conditions, different geological conditions, and different engineering scenarios, providing a reliable technical solution for accurate prediction of multi-region thrust of tunnel boring machines.

[0081] The method provided in this application constructs a fractional-order GRU network based on the discrete fractional derivative of continuous RL, and introduces an adjustable memory length parameter into the hidden state update equation. This allows the hidden state at the current moment to integrate the state information of all historical moments rather than relying solely on the previous moment, thereby enhancing the ability to capture long-term time dependencies in the thrust evolution process. Simultaneously, by utilizing a dynamic adjustment strategy, the memory length parameter can be dynamically adjusted according to the training cycle number, and the order of the fractional order can be dynamically switched according to the update gate output. This enables the model to adaptively balance the weights of long-term and short-term memories during training, thereby improving the ability to represent the state of nonlinear and non-stationary thrust sequences. Furthermore, by using correlation coefficients to select multi-dimensional excavation parameters highly correlated with thrust as inputs, and combining the fusion characteristics of the fractional-order GRU for global historical information, the model can implicitly model the spatial coupling and time delay relationships between different regions when predicting thrust in multiple regions. This achieves accurate prediction of thrust in multiple regions and provides accurate and reliable control basis for the attitude adjustment of tunnel boring machines under complex geological conditions.

[0082] Furthermore, as Figure 1 In a specific implementation of the method, this application provides a device thrust prediction apparatus based on a fractional-order GRU network, such as... Figure 3 As shown, the device includes: a fractional-order GRU network construction module 301, a training module 302, a screening module 303, and a prediction module 304.

[0083] The fractional-order GRU network construction module 301 is used to construct a GRU network. Based on the discretization of the fractional derivative of continuous RL, the difference equation of the hidden state in the GRU network is transformed, and an adjustable memory length parameter is introduced into the difference equation of the hidden state after transformation to obtain the fractional-order GRU network. Training module 302 is used to determine a preset dynamic adjustment strategy, train the fractional GRU network using historical time series data of the target device during historical operation, and update the memory length parameter and the order of fractional order according to the dynamic adjustment strategy during the training process to obtain the trained fractional GRU network. The filtering module 303 is used to collect the original excavation parameters of the target equipment at the time to be predicted when performing equipment thrust prediction, and use the correlation coefficient analysis algorithm to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. The prediction module 304 is used to input the excavation parameters to be input into the trained fractional-order GRU network, and obtain multiple thrust prediction values ​​of the target equipment in multiple partitions output by the trained fractional-order GRU network, so as to complete the equipment thrust prediction of the target equipment.

[0084] In specific application scenarios, the fractional-order GRU network construction module 301 is used to construct the GRU network including reset gates and update gates, and to determine the continuous RL fractional-order derivative, wherein the expression for the continuous RL fractional-order derivative is shown in the following formula. , in, Represents the fractional derivative operator; Indicates the starting time for calculating the fractional derivative; This indicates the order of the fraction, and ; This represents a function whose fractional derivative needs to be calculated. Indicates the current time point; Indicates greater than The smallest integer, and ; Represents the gamma function; Representation function In the integral variable The value at; Indicates the integral variable The differential; Indicates to of First derivative; express Time to the current time The contribution weight of the derivative; determine the difference equation of the hidden state in the GRU network as shown in the following formula. , in, Indicates the current hidden state. This indicates the hidden state at the previous moment. This indicates an update to the gate output. Represents the candidate hidden state at the previous time step; the continuous RL fractional derivative is expressed in step size... Discretization is then performed to obtain the discretized approximate expression shown in the following formula.

[0085] in, , Combining the discretized approximation expression with the difference equation of the hidden state yields the expression for the hidden state shown in the following formula. ; Transforming the expression for the hidden state yields the transformed difference equation for the hidden state, as shown in the following formula.

[0086] in, An adjustable memory length parameter is introduced into the difference equation of the transformed hidden state. The truncation range of historical moments is limited to those before the current moment. At that moment, exceeding Historical state information at each time step is discarded to obtain a truncated hidden state update equation, and the GRU network currently containing the truncated hidden state update equation is taken as the fractional-order GRU network.

[0087] In specific application scenarios, the fractional-order GRU network construction module 301 is used to combine the input parameters at the current moment. and the hidden state of the previous moment Construct the reset gate that performs computation operations according to the following expression. , in, This represents the weight matrix of the reset gate. This represents the bias matrix of the reset gate. This represents the Sigmoid activation function. This involves concatenating the input parameters with the previous hidden state as a vector; constructing the candidate hidden state calculation expression as shown in the following formula, wherein the candidate hidden state calculation expression obtains the result of the reset gate filtering the previous hidden state, and combines the result with the current input parameters to calculate the candidate hidden state. , in, The weight matrix representing the candidate hidden state. The bias matrix represents the candidate hidden state. This represents element-wise multiplication. ; Construct the update gate that performs computation operations according to the following expression. , in, This represents the weight matrix of the update gate. The bias matrix represents the update gate; the following expression is constructed for the current hidden state, which is obtained by weighted fusion of the previous hidden state and the candidate hidden state according to the update gate, to obtain the GRU network including the reset gate and the update gate. .

[0088] In specific application scenarios, the fractional-order GRU network construction module 301 is used to set the adjustable memory length parameter. For positive integers, define the truncation coefficient as shown in the following expression, where the memory length parameter... This indicates the number of historical moments retained.

[0089] In the difference equation of the transformed hidden state, the upper bound for the summation of the history terms is increased from... 1. Cut off as At the same time, the cutoff coefficient is introduced. To control the initial term By preserving the specified method, we obtain the truncated hidden state update equation shown in the following formula. .

[0090] In a specific application scenario, the training module 302 is used to set the total number of training cycles to... and setting constants The base scaling used to control the memory length parameter; the fractional-order GRU network trained using the historical time-series data, and in the current training cycle. In this process, according to the dynamic adjustment strategy shown in the following expression, based on the current cycle number... Proportional relationship with the total number of cycles Dynamically determine memory length parameters The possible values ​​of , where, Increasing from 1 to ,

[0091] At the same time, at each time step of the current training cycle In the middle, obtain the update gate output at the current time. And according to the following expression, based on The order of the fractional order at the current time is dynamically determined by comparing the result with the preset threshold of 0.5. The value of ,

[0092] The memory length parameter will be dynamically determined during the current training cycle. and the order of fractions The values ​​are substituted into the truncated hidden state update equation in the fractional-order GRU network for forward calculation and backpropagation to complete the training of the current training cycle; iterative training continues until completion. After one training cycle, the fractional-order GRU network is obtained.

[0093] In specific application scenarios, the training module 302 is also used to acquire a test sample set, the test sample set including One sample, In the nth sample The true thrust value of each sample is The test sample set is input into the trained fractional-order GRU network, which then predicts the thrust value for each sample. The mean absolute percentage error, root mean square error, and coefficient of determination of the trained fractional-order GRU network in thrust value prediction are calculated using the following formulas.

[0094] in, This represents the mean absolute percentage error. This indicates the calculated root mean square error. The determination coefficient is represented by the mean absolute percentage error, the root mean square error, and the determination coefficient. These are used as performance evaluation metrics for the trained fractional-order GRU network on the multi-region thrust prediction task, and the performance evaluation metrics are output.

[0095] In a specific application scenario, the screening module 303 is used to collect parameters of the target equipment at the time to be predicted to obtain the original excavation parameters, wherein the original excavation parameters include multi-point sealing chamber pressure and rotation angle. Screw conveyor speed Cutter head torque ,thrust Pitch angle Angle measurement Hinge Angle Speed ​​of advancement Side measurement difference and cutter head speed The multi-point sealed chamber pressure refers to the pressure values ​​at different spatial locations within the target equipment and includes... to According to the following formula, the correlation coefficient analysis algorithm is used to calculate the relationship between each parameter in the original excavation parameters and the thrust. Correlation coefficient between , , in, and Representing parameters respectively and thrust The One observation value, and Representing parameters respectively and thrust The sample mean, To determine the number of observation samples, the absolute values ​​of the correlation coefficients corresponding to each calculated parameter are compared with a preset parameter screening threshold. Parameters whose absolute values ​​of correlation coefficients are greater than the preset parameter screening threshold are selected as the excavation parameters to be input. These excavation parameters include the multi-point sealed chamber pressure. to Rotation angle Screw conveyor speed Cutter head torque Pitch angle Hinge Angle Speed ​​of advancement .

[0096] The apparatus provided in this application constructs a fractional-order GRU network based on the discrete fractional derivative of continuous RL, and introduces an adjustable memory length parameter into the hidden state update equation. This allows the hidden state at the current moment to integrate the state information of all historical moments rather than relying solely on the previous moment, thereby enhancing the ability to capture long-term time dependencies in the thrust evolution process. Simultaneously, by utilizing a dynamic adjustment strategy, the memory length parameter can be dynamically adjusted according to the training cycle number, and the order of the fractional order can be dynamically switched according to the update gate output. This enables the model to adaptively balance the weights of long-term and short-term memories during training, thereby improving the ability to represent the state of nonlinear and non-stationary thrust sequences. Furthermore, by using correlation coefficients to select multi-dimensional excavation parameters highly correlated with thrust as inputs, and combining the fusion characteristics of the fractional-order GRU for global historical information, the model can implicitly model the spatial coupling and time delay relationships between different regions when predicting thrust in multiple regions. This achieves accurate prediction of thrust in multiple regions and provides accurate and reliable control basis for the attitude adjustment of tunnel boring machines under complex geological conditions.

[0097] It should be noted that other corresponding descriptions of the functional units involved in the device thrust prediction device based on a fractional-order GRU network provided in this application embodiment can be found in the following references. Figures 1 to 2 The corresponding description in [the document] will not be repeated here.

[0098] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0099] It should be noted that any AI models, software tools, or components not belonging to this company appearing in the embodiments of this application are merely illustrative examples and do not represent actual use. All user personal information involved in the embodiments of this application has been authorized (with the knowledge and consent) by the relevant parties or has been fully authorized by all parties, and the executing entity may obtain it through various legal and compliant means. The collection, storage, use, processing, transmission, provision, and disclosure of the information, data, and signals involved all comply with relevant laws and regulations and do not violate public order and good morals.

[0100] The above embodiments and the technical features in the embodiments can be combined arbitrarily. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0101] The embodiments described above are merely examples of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these modifications and improvements all fall within the protection scope of this application.

[0102] In an exemplary embodiment, see Figure 4 The invention also provides a computer device including a bus, a processor, a memory, and a communication interface. It may also include an input / output interface and a display device, wherein the various functional units can communicate with each other via the bus. The memory stores a computer program, and the processor executes the program stored in the memory to perform the device thrust prediction method based on a fractional-order GRU network described in the above embodiments.

[0103] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the device thrust prediction method based on a fractional-order GRU network.

[0104] Through the above description of the embodiments, those skilled in the art can clearly understand that this application can be implemented in hardware or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solution of this application can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, external hard drive, etc.) and includes several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0105] Those skilled in the art will understand that the accompanying drawings are merely schematic diagrams of a preferred embodiment, and the modules or processes shown in the drawings are not necessarily essential for implementing this application.

[0106] Those skilled in the art will understand that the modules in the apparatus of the implementation scenario can be distributed within the apparatus of the implementation scenario as described, or they can be located in one or more apparatuses different from this implementation scenario, with corresponding changes. The modules of the above-described implementation scenario can be combined into one module, or they can be further divided into multiple sub-modules.

[0107] The serial numbers in this application are for descriptive purposes only and do not represent the superiority or inferiority of the implementation scenario.

[0108] The above disclosures are only a few specific implementation scenarios of this application. However, this application is not limited to these. Any variations that can be conceived by those skilled in the art should fall within the protection scope of this application.

Claims

1. A device thrust prediction method based on fractional-order GRU networks, characterized in that, include: A GRU network is constructed. Based on the discretization of the fractional derivative of continuous RL, the difference equation of the hidden state in the GRU network is transformed, and an adjustable memory length parameter is introduced into the difference equation of the hidden state after transformation to obtain a fractional GRU network. A preset dynamic adjustment strategy is determined, and the fractional-order GRU network is trained using historical time-series data of the target device during its historical operation. During the training process, the memory length parameter and the order of the fractional order are updated according to the dynamic adjustment strategy to obtain the trained fractional-order GRU network. When predicting equipment thrust, the original excavation parameters of the target equipment at the time to be predicted are collected. The correlation coefficient analysis algorithm is used to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. The excavation parameters to be input are input into the trained fractional-order GRU network, and the multiple thrust prediction values ​​of the target equipment in multiple partitions are obtained from the trained fractional-order GRU network to complete the equipment thrust prediction of the target equipment.

2. The method according to claim 1, characterized in that, The construction of the GRU network involves discretizing the hidden states based on the fractional derivatives of continuous RL, transforming the difference equations of the hidden states in the GRU network, and introducing an adjustable memory length parameter into the transformed difference equations of the hidden states to obtain a fractional GRU network, including: Construct the GRU network including reset gates and update gates, and determine the continuous RL fractional derivative, wherein the expression for the continuous RL fractional derivative is shown in the following formula. ， in, Represents the fractional derivative operator; Indicates the starting time for calculating the fractional derivative; This indicates the order of the fraction, and ; This represents a function whose fractional derivative needs to be calculated. Indicates the current time point; Indicates greater than The smallest integer, and ; Represents the gamma function; Representation function In the integral variable The value at; Indicates the integral variable The differential; Indicates to of First derivative; express Time to the current time The contribution weight of the derivative; Determine the difference equations for the hidden states in the GRU network shown in the following formula. in, Indicates the current hidden state. This indicates the hidden state at the previous moment. This indicates an update to the gate output. This indicates the candidate hidden state at the previous moment; The continuous RL fractional derivative is expressed in step size Discretization is then performed to obtain the discretized approximate expression shown in the following formula. in, , ; Combining the discretized approximation expression with the difference equation of the hidden state yields the expression for the hidden state shown in the following formula. ； Transforming the expression for the hidden state yields the transformed difference equation for the hidden state, as shown in the following formula. in, ; An adjustable memory length parameter is introduced into the difference equation of the transformed hidden state. The truncation range of historical moments is limited to those before the current moment. At that moment, exceeding Historical state information at each time step is discarded to obtain a truncated hidden state update equation, and the GRU network currently containing the truncated hidden state update equation is taken as the fractional-order GRU network.

3. The method according to claim 2, characterized in that, The construction of the GRU network, which includes reset gates and update gates, includes: Combined with the input parameters at the current moment and the hidden state of the previous moment Construct the reset gate that performs computation operations according to the following expression. ， in, This represents the weight matrix of the reset gate. This represents the bias matrix of the reset gate. This represents the Sigmoid activation function. This means concatenating the input parameters with the hidden state from the previous time step into a vector; Construct the candidate hidden state calculation expression as shown in the following formula, wherein the candidate hidden state calculation expression obtains the result of the reset gate filtering the hidden state at the previous time step, and combines the result with the input parameters at the current time step to calculate the candidate hidden state. ， in, The weight matrix representing the candidate hidden state. The bias matrix represents the candidate hidden state. This represents element-wise multiplication. ; Construct the update gate that performs computation operations according to the following expression. ， in, This represents the weight matrix of the update gate. This represents the bias matrix of the update gate; Construct the following expression for the current hidden state, which is used to perform a weighted fusion of the previous hidden state and the candidate hidden state according to the update gate, to obtain the GRU network including the reset gate and the update gate. 。 4. The method according to claim 2, characterized in that, An adjustable memory length parameter is introduced into the difference equation of the transformed hidden state. The truncation range of historical moments is limited to those before the current moment. At that moment, exceeding Historical state information at each time step is discarded to obtain the truncated hidden state update equation, including: The memory length parameter can be set to be adjustable. For positive integers, define the truncation coefficient as shown in the following expression, where the memory length parameter... This indicates the number of historical moments retained. In the difference equation of the transformed hidden state, the upper bound for the summation of the history terms is increased from...

1. Cut off as At the same time, the cutoff coefficient is introduced. To control the initial term By preserving the specified method, we obtain the truncated hidden state update equation shown in the following formula. 。 5. The method according to claim 1, characterized in that, The step of determining a preset dynamic adjustment strategy, training the fractional-order GRU network using historical time-series data from the target device's historical operation, and updating the memory length parameter and the order of the fractional order according to the dynamic adjustment strategy during training to obtain the trained fractional-order GRU network includes: Set the total number of training cycles to [number]. and setting constants Basic scaling used to control the memory length parameter; The fractional-order GRU network is trained using the historical time-series data, and in the current training cycle... In this process, according to the dynamic adjustment strategy shown in the following expression, based on the current cycle number... Proportional relationship with the total number of cycles Dynamically determine memory length parameters The possible values ​​of , where, Increasing from 1 to , At the same time, at each time step of the current training cycle In the middle, obtain the update gate output at the current time. , and according to the following expression, based on The order of the fractional order at the current time is dynamically determined by comparing the result with the preset threshold of 0.

5. The value of , The memory length parameter will be dynamically determined during the current training cycle. and the order of fractions The value of is substituted into the truncated hidden state update equation in the fractional GRU network for forward calculation and backpropagation to complete the training of the current training cycle. Continue iterative training until completion. After one training cycle, the fractional-order GRU network is obtained.

6. The method according to claim 5, characterized in that, The training process continues iteratively until completion. After obtaining the trained fractional-order GRU network in one training cycle, the method further includes: Obtain a test sample set, the test sample set including One sample, In the nth sample The true thrust value of each sample is ; The test sample set is input into the trained fractional-order GRU network, which then predicts the thrust value for each sample. The mean absolute percentage error, root mean square error, and coefficient of determination of the trained fractional-order GRU network in thrust value prediction are calculated using the following formulas. in, This represents the mean absolute percentage error. This indicates the calculated root mean square error. The determination coefficient is represented by the coefficient of determination. The mean absolute percentage error, the root mean square error, and the coefficient of determination are used as performance evaluation metrics for the trained fractional-order GRU network on the multi-region thrust prediction task, and the performance evaluation metrics are output.

7. The method according to claim 1, characterized in that, The process involves collecting the original excavation parameters of the target equipment at the predicted time, and using a correlation coefficient analysis algorithm to analyze the correlation between various parameters in the original excavation parameters and the thrust, in order to filter the excavation parameters to be input from the original excavation parameters, including: The parameters of the target equipment at the predicted time are collected to obtain the original excavation parameters, wherein the original excavation parameters include the multi-point sealing chamber pressure and the rotation angle. Screw conveyor speed Cutter head torque ,thrust Pitch angle Angle measurement Hinge Angle Speed ​​of advancement Side measurement difference and cutter head speed The multi-point sealed chamber pressure refers to the pressure values ​​at different spatial locations within the target equipment and includes... to ; According to the following formula, the correlation coefficient analysis algorithm is used to calculate the relationship between each parameter in the original excavation parameters and the thrust. Correlation coefficient between , ， in, and Representing parameters respectively and thrust The One observation value, and Representing parameters respectively and thrust The sample mean, The number of observed samples; The absolute values ​​of the correlation coefficients corresponding to each calculated parameter are compared with a preset parameter screening threshold. Parameters whose absolute values ​​of correlation coefficients are greater than the preset parameter screening threshold are selected as the excavation parameters to be input. These excavation parameters include the multi-point sealed chamber pressure. to Rotation angle Screw conveyor speed Cutter head torque Pitch angle Hinge Angle Speed ​​of advancement .

8. A device thrust prediction device based on a fractional-order GRU network, characterized in that, include: A fractional-order GRU network construction module is used to construct a GRU network. Based on the discretization of the fractional derivative of continuous RL, the difference equation of the hidden state in the GRU network is transformed, and an adjustable memory length parameter is introduced into the difference equation of the hidden state after transformation to obtain the fractional-order GRU network. The training module is used to determine a preset dynamic adjustment strategy, train the fractional-order GRU network using historical time-series data of the target device during its historical operation, and update the memory length parameter and the order of the fractional order according to the dynamic adjustment strategy during the training process to obtain the trained fractional-order GRU network. The filtering module is used to collect the original excavation parameters of the target equipment at the time to be predicted when predicting equipment thrust, and use the correlation coefficient analysis algorithm to analyze the correlation between each parameter in the original excavation parameters and the thrust, so as to filter the excavation parameters to be input from the original excavation parameters. The prediction module is used to input the excavation parameters to be input into the trained fractional-order GRU network, and obtain multiple thrust prediction values ​​of the target equipment in multiple partitions output by the trained fractional-order GRU network, so as to complete the equipment thrust prediction of the target equipment.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 7.

10. A storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 7.

Images