A tbm cutter wear prediction method fusing physical information neural network and cross-condition transfer learning

Abstract

The application discloses a TBM cutter wear prediction method fusing a physical information neural network and cross-condition transfer learning, and comprises the following steps: constructing a source domain sample data set according to geological parameters, tunneling parameters and cutter configuration parameters in a historical TBM construction process; embedding a cutter wear evolution law as a physical constraint term into a physical information neural network model, pre-training the physical information neural network model embedded with the physical constraint term based on the source domain sample data set, and obtaining a pre-training model; migrating the pre-training model to a target project and fixing part of network layer parameters, training and updating unfrozen parameters by using a small amount of sample data of the target project, and obtaining a target project cutter wear prediction model; obtaining real-time tunneling data of a to-be-predicted project, inputting the real-time tunneling data of the to-be-predicted project into the target project cutter wear prediction model, and obtaining a cutter wear prediction result. The application can realize accurate cutter wear prediction and intelligent decision-making of cutter replacement timing.

Description

Technical Field

[0001] This invention belongs to the field of intelligent monitoring technology for TBM hob wear, and particularly relates to a TBM hob wear prediction method that integrates physical information neural network and cross-working condition transfer learning. Background Technology

[0002] Full-face hard rock tunnel boring machines (TBMs) are core equipment for constructing long and deep tunnels, widely used in railway, highway, and urban underground space development projects. During hard rock tunneling, the cutterhead, as a key component directly breaking the rock, has its wear condition directly determining tunneling efficiency, construction costs, and project safety. Excessive cutterhead wear will lead to a sharp decline in rock-breaking efficiency and an abnormal increase in cutterhead torque, potentially causing catastrophic accidents such as cutterhead jamming and face instability. Therefore, there is an urgent need to develop intelligent sensing and early warning technology for cutterhead wear conditions to reduce downtime and improve excavation efficiency.

[0003] Existing hob wear prediction technologies mainly include theoretical model methods and pure data-driven methods. However, theoretical model methods rely on a large number of idealized assumptions and simplified conditions, which limits their engineering applicability and prediction accuracy. Pure data-driven methods are prone to producing non-physical predictions due to the lack of physical constraints, and have poor applicability to new engineering projects that have not been trained. They require the re-collection of a large number of samples for training, making it difficult to meet the real-time requirements of prediction.

[0004] Physical information neural networks (PENs) achieve an end-to-end fusion of data-driven approaches and physical mechanisms by embedding partial differential equations or algebraic physical constraints into the neural network loss function. However, applying PSN technology to hob wear prediction still faces the following shortcomings: (1) The cross-domain migration capability is weak. Different engineering geological conditions vary significantly. The source domain model is difficult to apply directly to the target domain. In addition, only a small amount of opening detection data can usually be obtained in the initial tunneling stage of a new project. Traditional physical information neural networks have difficulty converging under small sample conditions, resulting in poor applicability to new projects. (2) There is a lack of engineering application value. Existing models only output the wear amount and fail to provide a clear cutter replacement decision scheme, which is difficult to directly guide on-site construction. (3) The physical constraints are not fully covered. The wear of the cutter involves the coupling of multiple physical fields. A single physical constraint is difficult to characterize all wear patterns. (4) Existing methods mainly determine the timing of cutter replacement based on the wear amount or remaining life of the cutter. They do not further consider the wear differences of different cutter positions on the cutterhead, the maximum number of cutter replacements, and the surrounding rock conditions in front, making it difficult to determine the appropriate cutter replacement mileage position and combined cutter replacement scheme. Summary of the Invention

[0005] To address the aforementioned technical issues, this invention proposes a TBM cutter wear prediction method that integrates physical information neural networks and cross-condition transfer learning. This method enables accurate prediction of cutter wear and intelligent decision-making regarding cutter replacement timing, reducing the risk of unplanned downtime and improving the model's adaptability and predictive reliability under complex geological conditions.

[0006] To achieve the above objectives, this invention provides a TBM hobbing tool wear prediction method that integrates physical information neural networks and cross-working-condition transfer learning, comprising: Based on the geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes, a source domain sample dataset was constructed. The wear evolution law of hobbing cutter is embedded as a physical constraint term into the physical information neural network model. The physical information neural network model with the embedded physical constraint term is pre-trained based on the source domain sample dataset to obtain the pre-trained model. The pre-trained model is transferred to the target project and some network layer parameters are fixed. The unfrozen parameters are trained and updated using a small amount of sample data from the target project to obtain the hob wear prediction model for the target project. Acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result.

[0007] Optionally, constructing the source domain sample dataset includes: Based on the continuous tunneling parameters during historical TBM construction, the parameters are sorted by timestamp and abnormal data such as equipment downtime, communication failures, and data outside the normal working range are removed to obtain cleaned tunneling parameters. Based on the cleaned tunneling parameters, a correspondence between time and tunneling distance is established as a unified alignment benchmark for multi-source data; Based on the geological survey data and the unified alignment benchmark of the multi-source data, the uniaxial compressive strength of rock and Cerchar abrasive index are mapped to the continuous tunneling time series according to the stratum mileage interval to obtain the time series data of geological parameters. Based on the cutter replacement record, the current service cycle is terminated when a cutter replacement occurs, and the newly installed cutter is used as the starting point for wear evolution. The source domain sample dataset is constructed by combining the geological parameter time series data, the cleaning tunneling parameters, and the cutter configuration parameters.

[0008] Optionally, obtaining the physical constraint terms includes: Based on the radial wear theory of hobs, the theoretical radial wear rate of the hob under the corresponding working conditions is determined, and wear evolution constraints are constructed. Based on the irreversible cumulative characteristics of normal hob wear, a monotonic evolution constraint for radial wear is constructed; Based on the radial wear range and the maximum allowable radial wear during normal service of the hob, physical boundary constraints are constructed; The physical constraint term is constructed by weighted combination of the wear evolution constraint, the radial wear monotonic evolution constraint, and the physical boundary constraint.

[0009] Optionally, the physical information neural network model embedding the physical constraint terms is pre-trained based on the source domain sample dataset to obtain a pre-trained model, including: Based on the actual radial wear data of the hobbing cutter obtained from the opening detection in the source domain sample dataset, a data loss function is constructed. Based on the aforementioned physical constraints, a physical constraint loss function is constructed; Based on the data loss function and the physical constraint loss function, construct the total loss function for the source domain pre-training stage; Based on the total loss function, the network weights and bias parameters are updated using the backpropagation algorithm for iterative training to obtain the pre-trained model.

[0010] Optionally, training and updating the unfrozen parameters using a small amount of sample data from the target project includes: The network parameters of the pre-trained model are loaded into the target engineering physical information neural network as initial parameters. A preset number of network layer parameters near the input end are frozen, and only the rear hidden layer and output layer near the output end are trained for rapid adaptive training. Based on the model after rapid adaptation training, the frozen state of the frozen network layers is gradually unfrozen in the order from the output end to the input end and from the deep layer to the shallow layer, so that the newly unfrozen layers and the original trainable network layers participate in training together, and progressive unfreezing training is carried out. Based on the results of the progressive unfreezing training, the frozen state of all network layer parameters is released, allowing all network parameters to participate in the optimization of the target domain loss function. Fine-grained convergence training of the entire network is then performed to obtain the target engineering hob wear prediction model.

[0011] Optionally, the real-time tunneling data of the project to be predicted is input into the cutter wear prediction model of the target project to obtain the cutter wear prediction results, including: Based on the real-time tunneling data of the project to be predicted, the radial wear prediction value of each cutter at the corresponding time is obtained using the cutter wear prediction model of the target project. Calculate the wear degree coefficient based on the predicted radial wear value and the corresponding maximum allowable radial wear of the hob; Based on the wear coefficient, the hob wear prediction result is obtained.

[0012] Optionally, the method further includes: Based on the hob wear prediction results, the hob wear state is evaluated, and a tool replacement decision scheme is generated.

[0013] Optionally, evaluating the hob wear condition based on the hob wear prediction results includes: Based on the hob wear prediction results, the hob wear state is divided into slight wear, moderate wear, and severe wear; Based on the wear condition of the hobs, hobs that are in a severely worn state are added to the set of hobs to be replaced; Based on the set of cutterheads to be replaced, and in conjunction with the geological conditions of the current working face and the section to be excavated, determine whether the conditions for opening the cutterhead for replacement are met.

[0014] This invention also provides a TBM hobbing tool wear prediction system that integrates physical information neural networks and cross-condition transfer learning, comprising: The data acquisition module is used to construct a source domain sample dataset based on geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes. The model pre-training module is used to embed the hob wear evolution law as a physical constraint term into the physical information neural network model, and to pre-train the physical information neural network model with the embedded physical constraint term based on the source domain sample dataset to obtain the pre-trained model; The cross-working-condition transfer module is used to transfer the pre-trained model to the target project and fix some network layer parameters. It uses a small amount of sample data from the target project to train and update the unfrozen parameters to obtain the hob wear prediction model for the target project. The cutter wear prediction module is used to acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result. The evaluation module is used to evaluate the wear state of the hob based on the hob wear prediction results and generate a tool replacement decision scheme.

[0015] The present invention also provides a computer-readable storage medium storing a computer program that, when running, executes the steps of a TBM hobbing tool wear prediction method that integrates physical information neural networks and cross-condition transfer learning.

[0016] Compared with the prior art, the present invention has the following advantages and technical effects: (1) This invention, through source domain pre-training and target domain transfer fine-tuning, can quickly adapt to new projects with only a small number of open-cell samples, significantly reducing the dependence on a large amount of measured data and effectively solving the problem of poor model applicability under different geological and tunneling conditions. (2) This invention embeds wear evolution law, monotonicity and boundedness constraints into the loss function, so that the prediction results strictly meet the physical nature of wear irreversibility, effectively suppressing non-physical anomalies in non-sample areas and improving the model's generalization ability and engineering credibility. (3) This invention combines cutter wear state, front geological conditions generation and maximum cutter replacement amount constraint to generate combined cutter replacement schemes, realizing the extension from cutter state prediction to cutter maintenance decision-making, reducing the risk of unplanned downtime. Attached Figure Description

[0017] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of a TBM hobbing tool wear prediction method that integrates physical information neural network and cross-working condition transfer learning according to an embodiment of the present invention. Figure 2 This is a schematic diagram of a source domain pre-trained model based on a physical information neural network according to an embodiment of the present invention; Figure 3 This is a flowchart of cross-condition transfer learning and tool change decision-making according to an embodiment of the present invention; Figure 4 This is a distribution diagram of the wear state of the cutter head hob according to an embodiment of the present invention. Detailed Implementation

[0018] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] It should be noted that the steps shown in the flowchart in the accompanying drawings can be executed in a computer system such as a set of computer-executable instructions, and although a logical order is shown in the flowchart, in some cases the steps shown or described may be executed in a different order than that shown here.

[0020] This embodiment proposes a TBM hobbing tool wear prediction method that integrates physical information neural networks and cross-working-condition transfer learning, such as... Figure 1 As shown, the specific steps include: Based on the geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes, a source domain sample dataset was constructed. The wear evolution law of hobbing cutter is embedded as a physical constraint term into the physical information neural network model. The physical information neural network model with the embedded physical constraint term is pre-trained based on the source domain sample dataset to obtain the pre-trained model. The pre-trained model is transferred to the target project and some network layer parameters are fixed. The unfrozen parameters are trained and updated using a small amount of sample data from the target project to obtain the hob wear prediction model for the target project. Acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result.

[0021] Specifically, S1, obtain the geological parameters, tunneling parameters, and cutter configuration parameters from the historical TBM construction process to construct a source domain sample dataset; wherein, the source domain sample dataset includes the geological parameters, tunneling parameters, cutter configuration parameters, and each opening record from the historical TBM construction process; The geological parameters include: uniaxial compressive strength of rock. s c (MPa), Cerchar abrasion index (CAI); the tunneling parameters include: cutterhead thrust. F n (kN), cutter head torque T q (kN·m), cutter head speed RPM (rpm), penetration h (mm / r) tunneling speed v (mm / min); The hob configuration parameters include: hob number i Hob installation radius R (m), hob diameter D (mm), blade width T (mm), blade spacing S (mm), Yield strength of the cutter ring s y (MPa), Hardness of hobbing cutter material H m (HV); The record for each opening includes: radial wear of the hobbing cutter. V B (mm), tool change record, opening mileage.

[0022] S2. Embed the wear evolution law of the hob as a physical constraint into the physical information neural network model, and pre-train the model based on the source domain dataset to obtain the pre-trained model; S3. Transfer the pre-trained model to the target project, fix some network layer parameters, and use a small amount of sample data from the target project to train and update the unfrozen parameters to obtain the hob wear prediction model for the target project. S4. Input the real-time tunneling data of the target project to obtain the predicted wear amount and wear rate of each cutter; S5. Establish a hob wear condition assessment model and generate a tool replacement decision scheme.

[0023] Furthermore, constructing the source domain sample dataset includes: Based on the continuous tunneling parameters during historical TBM construction, the parameters are sorted by timestamp and abnormal data such as equipment downtime, communication failures, and data outside the normal working range are removed to obtain cleaned tunneling parameters. Based on the cleaned tunneling parameters, a correspondence between time and tunneling distance is established as a unified alignment benchmark for multi-source data; Based on the geological survey data and the unified alignment benchmark of the multi-source data, the uniaxial compressive strength of rock and Cerchar abrasive index are mapped to the continuous tunneling time series according to the stratum mileage interval to obtain the time series data of geological parameters. Based on the cutter replacement record, the current service cycle is terminated when a cutter replacement occurs, and the newly installed cutter is used as the starting point for wear evolution. The source domain sample dataset is constructed by combining the geological parameter time series data, the cleaning tunneling parameters, and the cutter configuration parameters.

[0024] Specifically, the preprocessing includes: sorting continuous tunneling parameters by timestamp, and removing abnormal data such as equipment downtime, communication anomalies, and data exceeding normal operating range; establishing a correspondence between time t and tunneling distance L as a unified alignment benchmark for multi-source data; and determining the uniaxial compressive strength σ of rock based on geological survey data. c The Cerchar abrasive index (CAI) is mapped to a continuous tunneling time series according to the stratum mileage interval; when a cutter replacement occurs, the current service cycle is terminated and the newly installed cutter is used as the starting point for wear evolution.

[0025] Furthermore, the preprocessing includes: The continuous tunneling parameters, such as cutterhead thrust, cutterhead torque, cutterhead rotation speed, penetration depth, and tunneling speed, are sorted according to timestamps, and abnormal data such as equipment downtime, communication failures, and those that are significantly outside the normal operating range of the equipment are removed. Establish the correspondence between time t and tunneling distance L, and use the time-mileage relationship as a unified alignment benchmark for multi-source data so that continuous tunneling parameters at each moment can correspond to a unique tunneling position. Based on geological survey data and rock test results, the uniaxial compressive strength σc and Cerchar abrasion index CAI of the rock are mapped to the continuous tunneling time series according to their corresponding stratum mileage intervals. When a hob is replaced, the current service cycle is terminated, and the newly installed hob is used as the new starting point for wear evolution to avoid continuous accumulation of wear before and after the replacement. For the radial wear of the cutter obtained from the opening inspection, instead of continuous interpolation over time or tunneling ring, the discrete measured values ​​at the actual opening location are retained.

[0026] The source domain sample dataset includes: effective measured radial wear data of cutters from historical hard rock TBM projects, and physical constraint points selected from the corresponding continuous construction time series; wherein, the measured radial wear data of cutters is obtained by opening detection, and the physical constraint points are composed of geological parameters, tunneling parameters and cutter configuration parameters at the corresponding time.

[0027] Furthermore, obtaining the physical constraint terms includes: Based on the radial wear theory of hobs, the theoretical radial wear rate of the hob under the corresponding working conditions is determined, and wear evolution constraints are constructed. Based on the irreversible cumulative characteristics of normal hob wear, a monotonic evolution constraint for radial wear is constructed; Based on the radial wear range and the maximum allowable radial wear during normal service of the hob, physical boundary constraints are constructed; The physical constraint term is constructed by weighted combination of the wear evolution constraint, the radial wear monotonic evolution constraint, and the physical boundary constraint.

[0028] Furthermore, based on the source domain sample dataset, a physical information neural network model embedding the physical constraint terms is pre-trained to obtain a pre-trained model, including: Based on the actual radial wear data of the hobbing cutter obtained from the opening detection in the source domain sample dataset, a data loss function is constructed. Based on the aforementioned physical constraints, a physical constraint loss function is constructed; Based on the data loss function and the physical constraint loss function, construct the total loss function for the source domain pre-training stage; Based on the total loss function, the network weights and bias parameters are updated using the backpropagation algorithm for iterative training to obtain the pre-trained model.

[0029] Furthermore, training and updating the unfrozen parameters using a small amount of sample data from the target project includes: The network parameters of the pre-trained model are loaded into the target engineering physical information neural network as initial parameters. A preset number of network layer parameters near the input end are frozen, and only the rear hidden layer and output layer near the output end are trained for rapid adaptive training. Based on the model after rapid adaptation training, the frozen state of the frozen network layers is gradually unfrozen in the order from the output end to the input end and from the deep layer to the shallow layer, so that the newly unfrozen layers and the original trainable network layers participate in training together, and progressive unfreezing training is carried out. Based on the results of the progressive unfreezing training, the frozen state of all network layer parameters is released, allowing all network parameters to participate in the optimization of the target domain loss function. Fine-grained convergence training of the entire network is then performed to obtain the target engineering hob wear prediction model.

[0030] Specifically, the source domain sample dataset is divided into a source domain training set and a source domain test set in an 8:2 ratio. The source domain training set is used for parameter learning of the physical information neural network model, and the source domain test set is used to evaluate the radial wear prediction performance of the pre-trained model. Data from the same hob within the same service cycle are only assigned to the same data subset to avoid data leakage caused by continuous time series data entering the training set and test set at the same time. like Figure 2 As shown, continuous time t The geological parameters, TBM tunneling parameters, and cutterhead configuration parameters corresponding to the time are used as inputs to the physical information neural network model; with the first i Put the hobbing cutter at all times t The predicted radial wear value is used as the output of the physical information neural network model, that is: ; In the formula, The parameter is i The physical information neural network, Indicates the first i Put the hobbing cutter at all times t The predicted radial wear is given in mm.

[0031] Based on the theory of radial wear of hobs, determine the first i Set the cutting tool at the tunneling distance L Radial wear rate: ; In the formula, oh ( L ) is the first i Set the cutting tool at the tunneling distance L Theoretical radial wear under the corresponding working conditions; Ks The abrasive wear coefficient; Yes The distance between the blades; σc It represents the uniaxial compressive strength of the rock. Ri The hob installation radius; Di The diameter of the hob; Ti The width of the blade; σy,i The yield strength of the blade ring; h ( L () represents the penetration degree.

[0032] Based on the chain relationship: ; Substituting, we get: ; In the formula, oh ( t ) is the first i During the tunneling process, the cutting rollers are used to cut the tunneling machine. t Theoretical radial wear under the corresponding working conditions; Ks The abrasive wear coefficient; Yes The distance between the blades; σc It represents the uniaxial compressive strength of the rock. Ri The hob installation radius; Di The diameter of the hob; Ti The width of the blade; σy,i The yield strength of the blade ring; h ( t ) represents penetration degree. v ( t (time) t The corresponding TBM tunneling speed.

[0033] Specifically, a physical constraint loss is constructed based on the aforementioned time-domain hob radial wear evolution relationship. An automatic differentiation method is used to calculate the predicted hob radial wear value output by the physical information neural network. Regarding time t The first derivative of , and construct the . i Put the hobbing cutter at all times t Physical residuals: ; Construct a wear evolution loss function based on the wear evolution residual: ; In the formula, Nphy This refers to the number of physical constraint points selected from the continuous construction time series data in the source domain. tj For the first j The time corresponding to each physical constraint point.

[0034] Specifically, based on the irreversible cumulative characteristic of normal hob wear, a monotonic evolution constraint for radial wear is constructed, namely: ; The corresponding wear monotonicity constraint loss function is: ; Specifically, physical boundary constraints are constructed based on the radial wear range during the normal service process of the hob, so that the radial wear prediction value satisfies: ; in, This represents the allowable radial wear limit for the i-th hob.

[0035] The corresponding physical boundary constraint loss function is: ; Specifically, the wear evolution constraint, wear monotonicity constraint, and physical boundary constraint are weighted and combined to construct the hob wear physical constraint loss function: ; In the formula, σ1, σ2, and σ3 are the weight coefficients corresponding to the wear evolution constraint, wear monotonicity constraint, and physical boundary constraint, respectively.

[0036] Specifically, a data loss function is constructed based on the actual radial wear data of the hobbing cutter obtained from the opening detection in the source domain sample dataset: ; In the formula, Nobs The number of valid measured data on radial wear of hobs in the source domain training set; For the first i Put the hob in the first k The actual radial wear amount obtained during the initial warehouse inspection; tk This corresponds to the opening detection time.

[0037] Specifically, the physical constraint loss function and the data loss function are weighted and combined to construct the total loss function for the source domain pre-training stage: ; In the formula, L data and L phy represents the data loss term and the physical constraint loss term, respectively, with corresponding weighting coefficients of . l data and l phy; The source domain training set is input into the physical information neural network model. The radial wear prediction value of each hob is obtained through forward propagation, and the total loss is calculated according to the data loss function and the physical constraint loss function. The network weights and bias parameters are updated using the backpropagation algorithm. The training is repeated iteratively until the total loss function meets the preset convergence condition or reaches the preset maximum number of iterations, thus obtaining the source domain pre-trained model.

[0038] The source domain pre-trained model is used to extract the common mapping relationship between cutter wear and geological conditions, tunneling status and cutter configuration parameters under different TBM engineering conditions, and serves as the initial model for subsequent transfer learning of the target domain cutter wear prediction model.

[0039] Furthermore, the real-time tunneling data of the project to be predicted is input into the cutter wear prediction model of the target project to obtain the cutter wear prediction results, including: Based on the real-time tunneling data of the project to be predicted, the radial wear prediction value of each cutter at the corresponding time is obtained using the cutter wear prediction model of the target project. Calculate the wear degree coefficient based on the predicted radial wear value and the corresponding maximum allowable radial wear of the hob; Based on the wear coefficient, the hob wear prediction result is obtained.

[0040] Specifically, such as Figure 3 As shown, further, the source domain pre-trained model is transferred to the target TBM project, and the source domain pre-trained model is transferred and trained using a small amount of measured radial wear data of the hobbing cutter in the target domain and continuous construction time series data to obtain the target domain hobbing cutter radial wear prediction model, specifically including: Geological parameters, TBM tunneling parameters, cutter configuration parameters, and actual radial wear data of the cutters obtained from the opening inspection during the construction process of the target TBM project are acquired. Outlier processing and time alignment are performed using the same data preprocessing method as the source domain sample data to construct the target domain sample dataset.

[0041] Source domain pre-trained model parameters θs Load the target domain physical information neural network and use it as the initial parameters of the target domain model: ; In the formula, θs For the parameters of the pre-trained model in the source domain, These are the initial model parameters at the start of target domain transfer training.

[0042] Suppose that the physical information neural network contains M trainable network layers, and the parameters of each layer are expressed as follows: ; in, θm Indicates the first m The weights and bias parameters of each network layer.

[0043] The first stage is rapid adaptation training. In this embodiment, approximately 70% of the network layer parameters near the input are frozen, leaving only the rear hidden layers and the output layer near the output as trainable. For the frozen network layers, their parameters remain unchanged in this stage, i.e.: ; In the formula, This represents the set of frozen network layers at the current stage, and q represents the number of training iterations.

[0044] For unfrozen network layers, based on the total loss function of the target domain... Update parameters: ; In the formula, This represents the set of trainable network layers at the current stage. or 1 represents the learning rate in the first stage.

[0045] This embodiment uses the AdamW optimization algorithm for fast adaptive training, with the initial learning rate set as follows: ; Reduce the learning rate if the loss function does not decrease significantly over several consecutive training rounds.

[0046] The second stage is progressive unfreezing training. After completing the first stage of training, the frozen network layers are gradually unfrozen in the order of output to input and from deep to shallow layers. In this embodiment, approximately 50% of the network layers closest to the output are unfrozen in the first stage, allowing the newly unfrozen layers to participate in the target domain training together with the original trainable network layers.

[0047] In this embodiment, the progressive unfreezing phase involves 200 training rounds; training is stopped early when the target domain loss fails to improve for 25 consecutive rounds. After completing the training of the current network layer, the next set of frozen network layers is unfrozen at the input, until all hidden layers participate in the training.

[0048] The third stage is fine-tuning and convergence training of the entire network. After gradual unfreezing, the frozen state of all network layer parameters is lifted, so that: ; All network parameters are involved in the optimization of the target domain loss function.

[0049] Training stops when the preset convergence condition is met or the maximum number of iterations is reached, thus obtaining the final target domain hob wear prediction model.

[0050] Specifically, based on the target domain hob radial wear prediction model, the first... i Put the hobbing cutter at all times t The radial wear prediction value is used, and combined with the maximum allowable radial wear of the hob, the hob wear coefficient is calculated: ; In the formula, or ᵢ ( t ) is the first i The wear coefficient of the hob, Current radial wear prediction value, This represents the maximum allowable radial wear of the corresponding hob.

[0051] Furthermore, the method also includes: Based on the hob wear prediction results, the hob wear state is evaluated, and a tool replacement decision scheme is generated.

[0052] Furthermore, based on the hob wear prediction results, the hob wear condition is evaluated, including: Based on the hob wear prediction results, the hob wear state is divided into slight wear, moderate wear, and severe wear; Based on the wear condition of the hobs, hobs that are in a severely worn state are added to the set of hobs to be replaced; Based on the set of cutterheads to be replaced, and in conjunction with the geological conditions of the current working face and the section to be excavated, determine whether the conditions for opening the cutterhead for replacement are met.

[0053] Specifically, as shown in Table 1, based on the wear coefficient or ᵢ ( t The wear condition of the hob is classified into slight wear, moderate wear, and severe wear: Table 1 Specifically, a tool change timing judgment module is established. First, it determines whether the hob has reached the tool change condition based on the hob wear level. For hobs in a severely worn state, they are added to the hob set to be replaced. When the set is not empty, it indicates that there are hobs on the cutter head that have reached the tool change warning condition, and further geological condition judgment is performed.

[0054] Specifically, when there are cutterheads to be replaced, the geological conditions of the current working face and the section to be excavated are used to determine whether the conditions for opening the cutterhead for replacement are met. If the current geological conditions are not suitable for opening the cutterhead for replacement, the replacement is postponed and the wear condition of the cutterheads is continuously monitored; if the current geological conditions meet the requirements for opening the cutterhead for replacement, the process proceeds to the cutterhead replacement plan generation stage.

[0055] When the current geological conditions permit the opening of the cutter head for cutter replacement, the wear status of all the cutters on the cutter head should be uniformly screened; cutters that have reached a severely worn state should be replaced first; for cutters that have entered the cutter replacement warning state but have not yet reached the maximum allowable radial wear, if the geological conditions permit and the opening operation has already been carried out, they can be included in the same cutter replacement plan based on their wear degree and installation position, so as to achieve the centralized replacement of multiple high-wear cutters.

[0056] Specifically, a tool replacement plan is generated based on the set of hobs to be replaced under the current opening conditions. The tool replacement plan includes at least the hob number, hob installation position, current radial wear prediction value, wear degree coefficient, wear level, tool replacement priority, and whether replacement is recommended for this opening.

[0057] The tool replacement scheme is generated in the following order: First, it is determined whether the wear of the hob reaches the tool replacement warning condition; second, it is determined whether the current geological conditions allow for tool replacement by opening the chamber; after the above conditions are met, multiple hobs on the cutter head that have reached severe wear are screened in a unified manner, and a tool replacement scheme is formed by taking the method of opening the chamber once to replace multiple high-wear hobs in one go.

[0058] The following is a detailed description of this embodiment: Step S1, Data Acquisition and Preprocessing: We selected 760 sets of effective cutter radial wear measurement data from four historical hard rock TBM projects, along with 10,000 physical constraint points selected from the corresponding continuous construction time series. Each measurement data sample includes: uniaxial compressive strength of rock σ. c (MPa), Cerchar abrasion index (CAI), cutter head thrust (F) n (kN), cutter head torque T q (kN·m), cutterhead rotation speed RPM (rpm), penetration h (mm / r), tunneling speed v (mm / min), cutter number i, cutter installation radius R (m), cutter diameter D (mm), cutter width T (mm), cutter spacing S (mm), cutter ring yield strength σ y (MPa), Hardness H of hobbing cutter material m (HV), hob radial wear V B (mm), tool change record, opening mileage; output is hob radial wear V. B (mm).

[0059] The preprocessing process includes: sorting continuous tunneling parameters by timestamp, removing abnormal values ​​such as equipment downtime, communication anomalies, and values ​​outside the normal operating range; establishing the correspondence between time t and tunneling distance L as a benchmark for multi-source data alignment; terminating the current service cycle after each cutter replacement, and using the newly installed cutter as the starting point for wear evolution.

[0060] The 760 sets of measured data and 10,000 physical constraint points selected from the continuous time series were divided into a source domain training set (608 sets of measured data + 8,000 constraint points) and a source domain test set (152 sets of measured data + 2,000 constraint points) in an 8:2 ratio.

[0061] The target domain is a newly started section of a tunnel, traversing a composite stratum of moderately to slightly weathered granite with a uniaxial compressive strength of 60–150 MPa. The first 50 rings (approximately 75 m) have been inspected, yielding 25 wear samples. Outlier handling and time alignment were performed using the same data preprocessing methods as the source domain samples to construct the target domain sample dataset.

[0062] Step S2, Source Domain Pre-training: Construct a physical information neural network, with continuous time t and corresponding geological parameters, tunneling parameters, and cutter configuration parameters as inputs, and output the predicted radial wear of the i-th cutter at time t. The network adopts a 3-layer fully connected structure (input layer → 128 → 64 → 1), with Tanh activation function and linear activation for the output layer.

[0063] The hob wear evolution equation, monotonicity constraints, and physical boundary constraints are embedded into the loss function. Weights λ are set for data loss and physical constraint loss. data =1.0, λ phy =0.5; where the wear evolution constraint weight σ1=1.0, the monotonicity constraint weight σ2=0.5, and the physical boundary constraint weight σ3=0.3. The abrasive wear coefficient K... s As a learnable parameter, the initial value was set to 0.15, and it was jointly optimized with the network weights. The Adam optimizer was used, with a batch size of 64 and an initial learning rate of 1×10⁻⁶. -3 The training process lasted for 200 epochs. During the training, the total loss gradually converged from the initial 1.83 to 0.029.

[0064] Step S3, Target Domain Transfer Training: Source domain pre-trained model parameters Loaded into the target domain network as initial parameters A three-stage migration strategy was adopted for fine-tuning: Rapid adaptation phase: Freeze the first 70% of the network layers at the input end, and train only the later hidden and output layers, with a learning rate of [missing information]. Training for 50 epochs; Gradual unfreezing phase: Gradually unfreeze 50% of the frozen layer from the output to the input, reducing the learning rate to... Training for 200 epochs, with an early stop patience value of 25 rounds; Network convergence phase: Unfreeze all data, overall learning rate Fine convergence until loss stability.

[0065] Step S4, Wear Prediction and Result Analysis: As shown in Table 2, the prediction accuracy on the target domain test set (opening measurement data on rings 51-400) is compared as follows: Table 2 This embodiment significantly outperforms the comparative method in both MAE and RMSE, and the prediction curve is strictly monotonically increasing with no physical anomalies. Further verification of robustness with small samples: by gradually reducing the number of fine-tuned samples in the target domain from 25 to 10, the MAE of this invention only increased to 0.68 mm, verifying the strong robustness of this invention in extremely small sample cross-condition scenarios.

[0066] Step S5, Tool Change Scheme Decision: like Figure 4 As shown, based on the target domain prediction results, the condition of the cutter head hob at the 200th ring is evaluated: the wear state of the cutter head hob is distinguished by color, where green indicates slight wear, yellow indicates moderate wear, and red indicates severe wear. There are 8 wear degree coefficients. The rock has entered a state of severe wear. Based on the geological prediction that the uniaxial compressive strength of the rock within 30 m ahead suddenly increases from 110 MPa to 155 MPa, a cutter replacement warning is generated: it is recommended to open the chamber before reaching this hard rock section and replace the 8 cutters and 2 moderately worn cutters in the surrounding area.

[0067] This invention also provides a TBM hobbing tool wear prediction system that integrates physical information neural networks and cross-condition transfer learning, comprising: The data acquisition module is used to construct a source domain sample dataset based on geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes. The model pre-training module is used to embed the hob wear evolution law as a physical constraint term into the physical information neural network model, and to pre-train the physical information neural network model with the embedded physical constraint term based on the source domain sample dataset to obtain the pre-trained model; The cross-working-condition transfer module is used to transfer the pre-trained model to the target project and fix some network layer parameters. It uses a small amount of sample data from the target project to train and update the unfrozen parameters to obtain the hob wear prediction model for the target project. The cutter wear prediction module is used to acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result. The evaluation module is used to evaluate the wear state of the hob based on the hob wear prediction results and generate a tool replacement decision scheme.

[0068] The present invention also provides a computer-readable storage medium storing a computer program that, when running, executes the steps of a TBM hobbing tool wear prediction method that integrates physical information neural networks and cross-condition transfer learning.

[0069] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting TBM hobbing tool wear by integrating physical information neural networks and cross-working-condition transfer learning, characterized in that, include: Based on the geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes, a source domain sample dataset was constructed. The wear evolution law of hobbing cutter is embedded as a physical constraint term into the physical information neural network model. The physical information neural network model with the embedded physical constraint term is pre-trained based on the source domain sample dataset to obtain the pre-trained model. The pre-trained model is transferred to the target project and some network layer parameters are fixed. The unfrozen parameters are trained and updated using a small amount of sample data from the target project to obtain the hob wear prediction model for the target project. Acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result.

2. The TBM hobbing tool wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 1, characterized in that, Constructing the source domain sample dataset includes: Based on the continuous tunneling parameters during historical TBM construction, the parameters are sorted by timestamp and abnormal data such as equipment downtime, communication failures, and data outside the normal working range are removed to obtain cleaned tunneling parameters. Based on the cleaned tunneling parameters, a correspondence between time and tunneling distance is established as a unified alignment benchmark for multi-source data; Based on the geological survey data and the unified alignment benchmark of the multi-source data, the uniaxial compressive strength of rock and Cerchar abrasive index are mapped to the continuous tunneling time series according to the stratum mileage interval to obtain the time series data of geological parameters. Based on the cutter replacement record, the current service cycle is terminated when a cutter replacement occurs, and the newly installed cutter is used as the starting point for wear evolution. The source domain sample dataset is constructed by combining the geological parameter time series data, the cleaning tunneling parameters, and the cutter configuration parameters.

3. The TBM hob wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 1, characterized in that, Obtaining the physical constraint terms includes: Based on the radial wear theory of hobs, the theoretical radial wear rate of the hob under the corresponding working conditions is determined, and wear evolution constraints are constructed. Based on the irreversible cumulative characteristics of normal hob wear, a monotonic evolution constraint for radial wear is constructed; Based on the radial wear range and the maximum allowable radial wear during normal service of the hob, physical boundary constraints are constructed; The physical constraint term is constructed by weighted combination of the wear evolution constraint, the radial wear monotonic evolution constraint, and the physical boundary constraint.

4. The TBM hobbing tool wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 3, characterized in that, Based on the source domain sample dataset, a physical information neural network model embedding the physical constraint terms is pre-trained to obtain a pre-trained model including: Based on the actual radial wear data of the hobbing cutter obtained from the opening detection in the source domain sample dataset, a data loss function is constructed. Based on the aforementioned physical constraints, a physical constraint loss function is constructed; Based on the data loss function and the physical constraint loss function, construct the total loss function for the source domain pre-training stage; Based on the total loss function, the network weights and bias parameters are updated using the backpropagation algorithm for iterative training to obtain the pre-trained model.

5. The TBM hob wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 1, characterized in that, Training and updating unfrozen parameters using a small amount of sample data from the target project includes: The network parameters of the pre-trained model are loaded into the target engineering physical information neural network as initial parameters. A preset number of network layer parameters near the input end are frozen, and only the rear hidden layer and output layer near the output end are trained for rapid adaptive training. Based on the model after rapid adaptation training, the frozen state of the frozen network layers is gradually unfrozen in the order from the output end to the input end and from the deep layer to the shallow layer, so that the newly unfrozen layers and the original trainable network layers participate in training together, and progressive unfreezing training is carried out. Based on the results of the progressive unfreezing training, the frozen state of all network layer parameters is released, allowing all network parameters to participate in the optimization of the target domain loss function. Fine-grained convergence training of the entire network is then performed to obtain the target engineering hob wear prediction model.

6. The TBM hobbing tool wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 1, characterized in that, The real-time tunneling data of the project to be predicted is input into the cutter wear prediction model of the target project to obtain the cutter wear prediction results, including: Based on the real-time tunneling data of the project to be predicted, the radial wear prediction value of each cutter at the corresponding time is obtained using the cutter wear prediction model of the target project. Calculate the wear degree coefficient based on the predicted radial wear value and the corresponding maximum allowable radial wear of the hob; Based on the wear coefficient, the hob wear prediction result is obtained.

7. The TBM hob wear prediction method integrating physical information neural network and cross-working condition transfer learning according to claim 1, characterized in that, The method further includes: Based on the hob wear prediction results, the hob wear state is evaluated, and a tool replacement decision scheme is generated.

8. The TBM hobbing tool wear prediction method according to claim 7, characterized in that, The evaluation of the hob wear condition based on the hob wear prediction results includes: Based on the hob wear prediction results, the hob wear state is divided into slight wear, moderate wear, and severe wear; Based on the wear condition of the hobs, hobs that are in a severely worn state are added to the set of hobs to be replaced; Based on the set of cutterheads to be replaced, and in conjunction with the geological conditions of the current working face and the section to be excavated, determine whether the conditions for opening the cutterhead for replacement are met.

9. A TBM hob wear prediction system integrating physical information neural networks and cross-condition transfer learning, used to implement the method as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to construct a source domain sample dataset based on geological parameters, tunneling parameters, and cutter configuration parameters from historical TBM construction processes. The model pre-training module is used to embed the hob wear evolution law as a physical constraint term into the physical information neural network model, and to pre-train the physical information neural network model with the embedded physical constraint term based on the source domain sample dataset to obtain the pre-trained model; The cross-working-condition transfer module is used to transfer the pre-trained model to the target project and fix some network layer parameters. It uses a small amount of sample data from the target project to train and update the unfrozen parameters to obtain the hob wear prediction model for the target project. The cutter wear prediction module is used to acquire real-time tunneling data of the project to be predicted, input the real-time tunneling data of the project to be predicted into the cutter wear prediction model of the target project, and obtain the cutter wear prediction result. The evaluation module is used to evaluate the wear state of the hob based on the hob wear prediction results and generate a tool replacement decision scheme.

10. A computer-readable storage medium storing a computer program thereon, characterized in that, The computer program executes the steps of the method according to any one of claims 1-8 when it runs.

Images