TBM surrounding rock quality real-time sensing method based on physical guidance online transfer learning

Abstract

The invention discloses a TBM surrounding rock quality real-time sensing method based on physical guidance online transfer learning. The TBM surrounding rock quality real-time sensing method comprises the steps that 1, source domain and target domain tunnel engineering objects are divided; step 2, feature parameter extraction based on a source domain field monitoring data sharing parameter space; 3, constructing a thrust prediction model based on LSTM (Long Short Term Memory); 4, constructing a surrounding rock state prediction model based on a similarity index; and step 5, target domain surrounding rock quality online prediction. According to the method, characteristic values of working section data streams are extracted based on shared space parameters in different tunnel engineering field monitoring data in a source domain, and a surrounding rock quality online prediction module based on shared characteristics is constructed for real-time identification of a cross-engineering surrounding rock state. The migration application feasibility of a related intelligent model is broadened, the online prediction module can achieve rapid diagnosis of the surrounding rock state by means of real-time monitoring data flow in the construction process of the newly-built tunnel TBM, and an important basis is provided for guaranteeing efficient and safe construction of the TBM.

Description

Technical Field

[0001] The present invention relates to the field of intelligent application of tunnel boring machine construction monitoring, and in particular to a real-time perception method of TBM surrounding rock quality based on physics-guided online transfer learning. Background Art

[0002] As an efficient, safe and environmentally friendly mechanized excavation method, tunnel boring machines are widely used in underground space development and construction. The performance of tunnel boring machines is affected by changes in the geological conditions of the surrounding rock. If the deterioration of the surrounding rock cannot be identified in time during the excavation process, it will cause equipment jamming, surrounding rock collapse and other engineering disasters, resulting in equipment damage and even serious threats to the lives of on-site personnel. However, due to the influence of the machine shield, it is inconvenient to implement methods such as advance exploration and geological mapping during TBM construction, and the analysis time of related results is long, which cannot meet the needs of rapid and real-time diagnosis on the engineering site. Therefore, tunnel engineering needs to develop relevant TBM surrounding rock quality real-time perception methods.

[0003] In recent years, the advancement of sensor technology has laid the foundation for the collection of data based on field monitoring. The on-site excavation data of the TBM excavation process is the result of the interaction between the machine and the rock mass. Therefore, it is possible to realize the real-time intelligent evaluation of the surrounding rock quality during the TBM construction process based on the on-site monitoring excavation data. However, related studies mainly focus on the training and testing of intelligent diagnosis of surrounding rock quality based on the monitoring data of a single tunnel project, and lack the field application verification of cross-project transfer learning of related models in newly built tunnels. In addition, the initial construction process of the newly built tunnel faces the problem of insufficient monitoring learning sample data, and the field monitoring data is collected in the form of real-time data streams. Therefore, how to realize the prediction of the intelligent diagnosis model of surrounding rock quality under online transfer learning under the conditions of TBM construction site monitoring data stream is the key to ensuring the cross-project application of related methods.

[0004] The development of online transfer learning methods for on-site monitoring data streams for new tunnel projects still needs to solve the following main problems. First, how to effectively process the data streams obtained in real time for new tunnel projects to extract the effective working sections of TBMs and extract features that can achieve transfer learning for different tunnels; then, how to develop online learning models suitable for new projects without supervised learning labels to achieve surrounding rock quality diagnosis based on real-time monitoring data streams; finally, how to achieve transfer learning on-site application of online diagnosis models during the construction of new tunnels. The establishment of relevant methods can enable the application of surrounding rock quality evaluation models trained based on completed engineering monitoring data during the construction of new tunnel projects, realize real-time evaluation of rock mass stability, reduce the cost of geological exploration for new tunnel projects, and ensure safe and efficient TBM construction. Summary of the invention

[0005] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide a real-time perception method for TBM surrounding rock quality based on physics-guided online transfer learning. The real-time perception method for TBM surrounding rock quality based on physics-guided online transfer learning uses the characteristic parameters widely existing in different tunnel projects extracted from the physical process of TBM rock breaking as the basis of transfer learning, and realizes real-time perception of surrounding rock quality across tunnel projects by constructing a standardized online data stream processing and physics-guided feature extraction module, a surrounding rock category online prediction module and a transfer learning application module.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0007] A real-time perception method of TBM surrounding rock quality based on physics-guided online transfer learning includes the following steps.

[0008] Step 1: Divide the source domain and target domain tunnel engineering objects: The source domain refers to the tunnel engineering that has been excavated. Its on-site monitoring data space is represented by X S It means that the data space of the source domain contains the on-site monitoring data of M tunnel projects, where M≥1; the surrounding rock quality status of the M tunnel projects in the source domain is divided into S types, where S≥2.

[0009] The target domain refers to the N tunnel projects that have not yet been excavated, and the field monitoring data space is represented by X T Represents, where N≥1;

[0010] Step 2: Extracting characteristic parameters based on the shared parameter space of source domain field monitoring data, including the following steps:

[0011] Step 2-1: Segment the data streams of effective rock breaking working stages of TBMs in different tunnel projects in the source domain to obtain n effective rock breaking working section data streams.

[0012] Step 2-2: Determine the shared parameter space X in the source domain Shard : A shared parameter space X is established for each effective rock breaking working section data stream Shard ;X Shard It is determined according to the number M of tunnel engineering field monitoring data in the source domain, specifically:

[0013] A. When M = 1, X Shard Rock breaking parameters.

[0014] B. When M>1, X Shard It is the union of rock breaking parameters and non-rock breaking parameters.

[0015] Step 2-3, extract shared parameter characteristic values: shared parameter characteristic values ​​include rock breaking parameter characteristic values ​​and non-rock breaking parameter characteristic values; wherein, rock breaking parameter characteristic values ​​include rock breaking parameter statistical characteristic values ​​and excavability index; wherein, excavability index is an index related to the rock breaking process and guided by physics; non-rock breaking parameter characteristic value is the mean value of each non-rock breaking parameter.

[0016] Step 3. Construct a thrust prediction model based on LSTM: Construct a thrust prediction model based on the LSTM neural network algorithm; suppose the current working segment to be predicted is the t-th working segment, then the input layer of the thrust prediction model is the shared parameter feature values ​​of the L historical working segments before the t-th working segment; the output layer is the thrust prediction value of the t-th working segment; according to the source domain shared parameter feature values ​​extracted in step 2, the constructed thrust prediction model is trained to obtain the trained thrust prediction model.

[0017] Step 4, construct a surrounding rock state prediction model based on similarity index: calculate the similarity error between the thrust prediction value of the t-th working section and the actual thrust value of the t-th working section; according to the source domain shared parameter characteristic value extracted in step 2, obtain the similarity error sequence SI corresponding to the n effective rock breaking working sections; and then combine the S surrounding rock quality states divided in step 1 to construct a surrounding rock state prediction model based on the similarity index sequence SI and the corresponding surrounding rock classification sequence S.

[0018] Step 5: Online prediction of target domain surrounding rock quality, including the following steps:

[0019] Step 5-1, initial excavation of the target area: The TBM excavates the target area to the hth working section and collects tunnel engineering on-site monitoring data for each working section; at the same time, a field survey is conducted on the surrounding rock types of the 1st to hth excavated working sections.

[0020] Step 5-2, continue to excavate the target area: the TBM excavates the target area from the h+1th working section to the tth working section, and conducts tunnel engineering on-site monitoring data for each working section; where t≥h+1.

[0021] Step 5-3, extracting the target domain shared parameter characteristic values: For each working segment in step 5-1 and step 5-2, the shared parameter characteristic values ​​are extracted.

[0022] Step 5-4, thrust prediction for the tth working section: Input the shared parameter feature values ​​of the L historical working sections before the tth working section extracted in step 3 into the thrust prediction model based on LSTM constructed in step 3 to obtain the thrust prediction value of the tth working section in the target domain.

[0023] Step 5-5, calculate similarity: the actual thrust value of the target domain t-th working section obtained by monitoring in step 5-2 The thrust prediction value obtained in step 5-4 Calculate the similarity error SI of the tth working segment in the target domain t .

[0024] Step 5-6: Online prediction of surrounding rock quality of the t+1th working section: The surrounding rock type of the t-1th working section is taken as the surrounding rock type s before migration. t-1 , will s t-1 and SI t Substitute it into the surrounding rock state prediction model constructed in step 4 and find t-1 and SI t The probability values ​​of different surrounding rock states under the following conditions; the surrounding rock state corresponding to the maximum probability value is the predicted surrounding rock quality state of the t+1th working section; the surrounding rock type of the t-1th working section is directly surveyed according to step 5-1 or obtained through online quality prediction based on the survey data of step 5-1.

[0025] In step 1, the surrounding rock quality status of the M tunnel projects in the source domain adopts the national standard "Code for Geological Investigation of Water Conservancy and Hydropower Engineering GB50487-2008", and the tunnel surrounding rock is divided into stable I, basically stable II, partially stable III, unstable IV and extremely unstable V, with a total of S = 5 types.

[0026] In step 2-2, the rock breaking parameters include cutter head thrust, cutter head torque, propulsion speed, cutter head rotation speed and penetration; wherein the penetration is the ratio of propulsion speed to cutter head rotation speed.

[0027] In step 2-3, the excavability indicators include the linear fitting slope a and thrust intercept b of the single-pole thrust and penetration data streams, and the linear fitting slope TPI of the single-pole torque and penetration.

[0028] In steps 2-3, the statistical characteristic values ​​of rock breaking parameters include statistical characteristic values ​​of rock breaking operation parameters and statistical characteristic values ​​of rock breaking non-operation parameters; in step 3, the LSTM neural network includes an LSTM layer and a fully connected layer; the input of the LSTM layer is the statistical characteristic values ​​of rock breaking non-operation parameters, excavability indicators and non-rock breaking parameter characteristic values ​​of L historical working sections before the t-th working section; the input of the fully connected layer is the output of the LSTM layer and the statistical characteristic values ​​of rock breaking operation parameters; the output of the fully connected layer is the thrust prediction value of the t-th working section.

[0029] In step 3, the thrust prediction model is trained according to the source domain shared parameter feature values ​​extracted in step 2, and during the training process, the hyperparameters of the thrust prediction model are optimized. The specific optimization method includes the following steps:

[0030] Step 3B-1, sample library division: divide the n source domain working segment data streams loaded in step 3A-2 into a training set and a validation set according to a set ratio.

[0031] Step 3B-2, design hyperparameters: perform several combination designs on the hyperparameters involved in the thrust prediction model.

[0032] Step 3B-3, model training: Use the training set data to train each hyperparameter combination in the thrust prediction model to obtain several thrust prediction models that have been trained.

[0033] Step 3B-4, hyperparameter optimization: Use the validation set to validate each thrust prediction model trained in step 3B-3, and select a hyperparameter combination with the best validation performance as the optimal hyperparameter. The optimal hyperparameter is frozen and fixed. The thrust prediction model with the optimal hyperparameter is also called the optimal thrust prediction model.

[0034] In step 4, let the similarity error sequence of the source domain be SI = {si1,...,si i ,...,si n}, and according to the surrounding rock quality status classification results in the source domain, the surrounding rock classification sequence S = {s0, s1, ..., s i ,...,s n}; Based on the similarity index sequence SI and the corresponding surrounding rock classification sequence, a rock mass classification state transition probability model is constructed, as shown in the following formula:

[0035] M(s,si)={p0(s0|s,si),...,p i (s i |s,si),...,p n (s n |s,si)}

[0036] In the formula, p i (s i |s,si) is the conditional probability density distribution function when the excavated state is s and the similarity index is si when transferring to any other state. This distribution function establishes a non-parametric probability density statistical model by the similarity index of different rock mass categories in the surrounding rock classification sequence when the state is transferred.

[0037] In step 5-1, h=20.

[0038] The present invention has the following beneficial effects:

[0039] (1) The present invention can segment the data stream monitored during the TBM construction process, extract characteristic parameters, and calculate similarity indicators, thereby realizing real-time prediction of the surrounding rock state probability. The method can complete online intelligent prediction of surrounding rock quality. Compared with traditional geological exploration and advanced geological drilling, it does not affect the normal excavation process of TBM, has high time efficiency and reduces related exploration costs.

[0040] (2) The extraction of feature parameters in the source domain and target domain transfer learning methods takes into account the parameters related to the rock breaking process based on physical guidance and other shared feature parameters. Among them, the rock breaking parameter characteristics based on the interaction between TBM and rock mass are the main parameters in the shared feature space, which are convenient and feasible for transfer application. The reason is that although the dimensions of the monitoring parameters in different tunnel engineering sites are different, the main rock breaking parameters such as thrust, torque, propulsion speed and cutter head speed in the TBM excavation process are simple and easy to monitor in different tunnels. This provides a data feature basis for the establishment of transfer learning related models and cross-engineering applications.

[0041] (3) The present invention realizes an online surrounding rock state intelligent perception method that constructs an online learning module based on shared features in the source domain and applies it to cross-project migration in the target domain. The established online learning module includes LSTM similarity index prediction and surrounding rock state prediction model, which does not require the acquisition of monitoring data and supervised learning labels of the target domain in advance. Compared with the traditional offline learning method, it does not require the accumulation of monitoring data of the new tunnel project, and can realize the real-time prediction of the surrounding rock state in the early and middle stages of the new tunnel project, significantly expanding the cross-project application scope of the migration learning module, making it possible to predict the surrounding rock state of more mechanized new tunnel projects, thereby ensuring the safety of the TBM construction process. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The present invention is a flow chart of the real-time perception method of TBM surrounding rock quality based on physics-guided online transfer learning.

[0043] Figure 2 The schematic diagram of rock breaking process and feature extraction of TBM cutting rock mass based on working section data stream. Sub-figure (a) is the extracted representative working section data stream, sub-figure (b) is the schematic diagram of the effective rock breaking process of the cutter intruding into the rock mass, sub-figure (c) is the schematic diagram of the linear fitting of single-cutter thrust and penetration, and sub-figure (d) is the schematic diagram of single-cutter torque and penetration.

[0044] Figure 3 Schematic diagram of the prediction process of the online prediction module for surrounding rock quality.

[0045] Figure 4 Schematic diagram of online learning module parameter freezing and target domain migration application based on shared feature parameter space.

[0046] Figure 5 This is a structural diagram of the LSTM-based thrust prediction model.

[0047] Figure 6 This is an example diagram of the probability density distribution of surrounding rock state transition based on the similarity error index.

[0048] Figure 7 This is an example diagram of the target domain application results of the online transfer learning TBM surrounding rock state real-time perception method, where sub-figure (a) is the single-blade thrust prediction value and measured value of the working section data stream between pile numbers 65334 and 65325 of the target domain Yinchuojiliao No. 6 tunnel obtained by using the LSTM prediction model; sub-figure (b) is the change process of the average percentage error of the similarity index within the corresponding pile number range; sub-figure (c) is an example diagram of the change curve of the probability prediction result of different surrounding rock state categories within the corresponding pile number range.

[0049] Figure 8 Figure 3 is the surrounding rock state transfer and classification prediction results of the online working section data stream of the Jinan-Jiliao Tunnel in the target domain, where sub-figure (a) shows the surrounding rock state transfer accuracy, surrounding rock category prediction accuracy and F1 index result example diagram of the invented method in the example tunnel sections of the source domain and the target domain; sub-figure (b) shows the confusion matrix results of the data stream of all working sections in the target domain under different surrounding rock states. DETAILED DESCRIPTION

[0050] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

[0051] The present invention uses the on-site monitoring data of the open TBM construction process of Section 4 between Chalu River and Yinma River of the Yinsong Project in Jilin Province as the source domain data for transfer learning, and the real-time data stream of on-site monitoring of the open TBM construction process of the tunnel of Section 6 of Yinchuojiliao in Inner Mongolia as the target domain for transfer learning, and explains the present invention in detail.

[0052] like Figure 1 As shown, a real-time perception method of TBM surrounding rock quality based on physical guided online transfer learning includes the following steps.

[0053] Step 1: Divide source domain and target domain tunnel engineering objects

[0054] The source domain refers to the tunnel project that has been excavated, and its on-site monitoring data space is represented by .

[0055] The source domain data space may include M tunnel engineering field monitoring data, where M≥1; in this embodiment, M=1 is preferred. The source domain in the present invention relies on the field monitoring data of Section 4 of the Yinsong Water Diversion Tunnel in Jilin Province, which has been excavated and is totaling 19.77 km.

[0056] Furthermore, it is required that the surrounding rock quality assessment method of the M tunnel projects in the source domain adopts the same specification standard, and the relevant specifications divide the surrounding rock quality status into S types, where S≥2. In the example of the present invention, the national standard "Specification for Geological Survey of Water Conservancy and Hydropower Engineering GB50487-2008" is adopted to divide the tunnel surrounding rock into S=5 types, namely stable (I), basically stable (II), partially stable (III), unstable (IV) and extremely unstable (V).

[0057] The target domain refers to N tunnel projects that have not yet been excavated, and its on-site monitoring data space is represented by , where N ≥ 1; in this example, N = 1 is preferred, and the target domain relies on the Inner Mongolia Yinchoujiliao No. 6 Tunnel Project that has not yet been excavated.

[0058] The same type of TBM is used in the excavation of the source and target tunnels, including open type, single shield type and double shield type.

[0059] Step 2: Extract feature parameters based on the shared parameter space of the source domain on-site monitoring data. This step mainly extracts the shared features of different tunnel engineering data obtained by monitoring in the source domain for establishing an online transfer learning model. The steps include:

[0060] Step 2-1: Segmentation of the data streams of effective rock-breaking working stages of different tunnel projects in the source domain. The TBM shutdown status data in the source domain is zero-value data, which is invalid data and needs to be segmented to extract the effective data of the working stage. The segmentation of the effective working section is based on the cutter head speed (RPM) data stream, and the data stream data [D1, D2, ..., D n-1 ,D n ] as an example, the working segment extraction formula is as follows:

[0061]

[0062] According to the above formula, the working section data stream of the on-site monitoring data of the Yinsong water diversion tunnel is segmented, and a total of n = 13,200 TBM working section data are obtained.

[0063] The cutterhead speed is used as the online segmentation indicator for the real-time data stream of the working section. The reason is that when the TBM starts working, the TBM driver will first set the cutterhead speed value. In addition, the real-time segmentation criterion for the working section is required to identify that the product of the cutterhead speed RPM values ​​in the online data stream is greater than zero.

[0064] Step 2-2: Determine the shared parameter space in the source domain

[0065] Compare the dimensional characteristics of field monitoring data of different tunnel projects in the source domain and determine the shared parameter space included in different tunnel projects in the source domain in

[0066] The shared parameter space consists of two parts. One part is the rock breaking parameters that are closely related to TBM cutting rock mass during tunnel construction. These parameters can be monitored and obtained in tunnel construction using the TBM method. The other part is the parameters that are not related to the rock breaking process but are collected in each tunnel project in the source domain.

[0067] A shared parameter space is established for each effective rock breaking working section data stream It is determined according to the number M of tunnel engineering field monitoring data in the source domain.

[0068] When the number of TBM-constructed tunnel projects in the source domain is M=1, the rock breaking parameters guided by physical processes are used as the shared parameter space. When the number of tunnel projects in the source domain is M>1, the union of the rock breaking parameters and other parameters is used as the shared parameter space.

[0069] In this example, M = 1 field monitoring of the Yinsong water diversion tunnel is used as the data source for transfer learning in the source domain, so the rock breaking parameters guided by physical processes are used as the shared parameter space.

[0070] The shared parameters in the source domain that are closely related to the rock breaking process include: single blade thrust, single blade torque, propulsion speed, and cutterhead speed. These parameters are the performance and control parameters of TBM cutting rock mass. Among them, the penetration can be obtained by dividing the propulsion speed by the cutterhead speed.

[0071] Therefore, the shared parameter space X of the Songshui water diversion tunnel is Shard It includes cutter disc thrust, cutter disc torque, propulsion speed, cutter disc rotation speed and penetration parameters, among which the penetration can be obtained by dividing the propulsion speed by the cutter disc rotation speed.

[0072] Step 2-3: Extract the shared parameter space X Shard Characteristic values ​​of various parameters in the data flow of the TBM working section.

[0073] The shared parameter characteristic values ​​include rock breaking parameter characteristic values ​​and non-rock breaking parameter characteristic values; wherein the rock breaking parameter characteristic values ​​include rock breaking parameter statistical characteristic values ​​and excavability indicators.

[0074] The rock breaking parameter statistical characteristic values ​​include the mean values ​​of single blade thrust, cutter head torque, propulsion speed, and cutter head rotation speed of each working section data stream. Furthermore, the rock breaking parameter statistical characteristic values ​​include the rock breaking operation parameter statistical characteristic values ​​and the rock breaking non-operation parameter statistical characteristic values; among them, the propulsion speed mean, cutter head rotation speed mean, and penetration rate belong to the rock breaking operation parameter statistical characteristic values; the single blade thrust mean and cutter head torque mean belong to the rock breaking non-operation parameter statistical characteristic values.

[0075] The digging index is an index related to the rock breaking process and guided by physics; the calculation of the digging index uses the effective rock breaking stage data in the working section data stream. Figure 2 As shown in Figure 2, the TBM rock breaking stage corresponds to the second stage in sub-figure (a), that is, the process in which the cutter invades the rock mass and the cracks extend to form slag pieces as shown in Figure (b); among them, the excavability index includes the linear fitting slope a and thrust intercept b of the single-knife thrust and penetration data flow, as shown in sub-figure (c), and the linear fitting slope TPI of the single-knife torque and penetration, as shown in sub-figure (d).

[0076] The characteristic value of non-rock breaking parameters is the mean value of each non-rock breaking parameter such as cutter shield pressure, cutter shield displacement, belt conveyor speed, slag particle size statistics, etc.

[0077] Step 3: Build a thrust prediction model based on LSTM

[0078] like Figure 3 and Figure 5 As shown, a thrust prediction model is constructed based on the LSTM neural network algorithm; assuming that the current working segment to be predicted is the t-th working segment, the input layer of the thrust prediction model is the shared parameter feature values ​​of the L historical working segments before the t-th working segment; the output layer is the thrust prediction value of the t-th working segment.

[0079] Furthermore, the above-mentioned LSTM neural network preferably includes an LSTM layer and a fully connected layer; the input of the LSTM layer is the statistical characteristic values ​​of rock breaking non-operational parameters, excavability indicators and non-rock breaking parameter characteristic values ​​of L historical working sections before the t-th working section; the input of the fully connected layer is the output of the LSTM layer and the statistical characteristic values ​​of rock breaking operation parameters; the output of the fully connected layer is the thrust prediction value of the t-th working section.

[0080] The constructed thrust prediction model is trained according to the source domain shared parameter feature values ​​extracted in step 2 to obtain a trained thrust prediction model. The thrust prediction model training method preferably includes the following steps.

[0081] Step 3A-1, source domain working section data stream preparation: Arrange the shared characteristic working section data streams in each tunnel project in the source domain in the order of excavation time, and then obtain n tunnel working section data streams.

[0082] Step 3A-2, source domain working section data stream loading. The data stream and features of each tunnel engineering working section are loaded in turn to build a surrounding rock similarity quantification model based on LSTM. The implementation method is to read the working section data stream and shared features in turn. When the TBM excavates to the tth working section, the shared parameter data stream and features of the current data stream and the previous L working sections are loaded.

[0083] Step 3A-3: Use the loaded source domain working segment data to train the constructed thrust prediction model.

[0084] Furthermore, according to the source domain shared parameter feature values ​​extracted in step 2, during the training process of the thrust prediction model, hyperparameter optimization of the thrust prediction model can be performed. The specific optimization method includes the following steps.

[0085] Step 3B-1, sample library division: the n source domain working segment data streams loaded in step 3A-2 are divided into a training set and a validation set according to a set ratio. In this embodiment, the set ratio is preferably 90%:10%.

[0086] Step 3B-2, design hyperparameters: perform several combinations of hyperparameters involved in the thrust prediction model. The specific design methods are shown in Table 1.

[0087] Step 3B-3, model training: Use the training set data to train each hyperparameter combination in the thrust prediction model to obtain several thrust prediction models that have been trained.

[0088] Step 3B-4, hyperparameter optimization: Use the validation set to validate each thrust prediction model trained in step 3B-3, and select a hyperparameter combination with the best validation performance as the optimal hyperparameter. The optimal hyperparameter is frozen and fixed. The thrust prediction model with the optimal hyperparameter is also called the optimal thrust prediction model.

[0089] In this embodiment, the optimal hyperparameters after freezing are shown in Table 2.

[0090] Table 1 LSTM model structure parameters and online prediction process hyperparameter ranges

[0091]

[0092] Table 2 Optimal hyperparameter combinations for freezing

[0093]

[0094] Step 4: Construct a surrounding rock state prediction model based on similarity index

[0095] Calculate the similarity error between the thrust prediction value of the tth working section and the actual thrust value of the tth working section.

[0096] The rock mass similarity index in this example uses the mean percentage error MPE between the predicted thrust and the actual monitoring data stream t, and its calculation formula is as follows:

[0097]

[0098] In the formula, A irepresents the real thrust value monitoring sequence of the t-th working section data stream, P i It represents the thrust prediction value sequence of the LSTM model of the t-th working section data stream, and n is the number of samples in the t-th working section data stream sequence.

[0099] According to the source domain shared parameter eigenvalues ​​extracted in step 2, the similarity error sequence SI corresponding to the n effective rock breaking working sections is obtained; combined with the S surrounding rock quality states divided in step 1, a surrounding rock state prediction model based on the similarity index sequence SI and the corresponding surrounding rock classification sequence S is constructed.

[0100] Assume that the similarity error sequence of the source domain is SI = {si1,...,si i ,...,si n}, and according to the surrounding rock quality status classification results in the source domain, the surrounding rock classification sequence S = {s0, s1, ..., s i ,...,s n}; Based on the similarity index sequence SI and the corresponding surrounding rock classification sequence, a rock mass classification state transition probability model is constructed, as shown in the following formula:

[0101] M(s,si)={p0(s0|s,si),...,p i (s i |s,si),...,p n (s n |s,si)}

[0102] In the formula, p i (s i |s,si) is the conditional probability density distribution function when the state is s and the similarity index is si when it is transferred to any other state. This distribution function establishes a non-parametric probability density statistical model by the similarity index of different rock mass categories in the surrounding rock classification sequence when the state is transferred.

[0103] In this embodiment, the average percentage MPE error index of the surrounding rock state category transfer corresponding to the data stream of 13200 working sections of the source domain Yinsong water diversion tunnel is statistically analyzed, and the probability density distribution of different surrounding rock category transfers is obtained by using the density estimation method. Figure 6 As shown in the figure, the row direction is the state of the surrounding rock before the transfer, and the column direction is the state of the surrounding rock after the transfer. The prediction process is shown as follows: when the TBM excavates to the tth working section, the MPE is obtained based on the LSTM similarity prediction model. t = -70%, and s is known t-1 =III, input it into Figure 6 In the probability density distribution function of different columns under the middle row label III, the probability p of the surrounding rock state where the current data stream is located is obtained35 >p 34 >p 33 >p 32 , so it is predicted that the surrounding rock where the current data stream is located is Class V rock mass. In the formula, p 35 、p 34 、p 33 and p 32 Respectively indicate that the labels are Figure 6 When the second row is Class III surrounding rock, the probability density values ​​when the surrounding rock categories after transfer are Class V, Class IV, Class III and Class II respectively.

[0104] Step 5: Online prediction of target domain surrounding rock quality, including the following steps.

[0105] Step 5-1, initial excavation of the target area: The TBM excavates the target area to the hth working section, and conducts on-site monitoring data of the tunnel engineering for each working section; at the same time, conducts field surveys of the surrounding rock types of the first to h excavated working sections. Preferably, h = 20.

[0106] Step 5-2, continue to excavate the target area: the TBM excavates the target area from the h+1th working section to the tth working section, and conducts tunnel engineering on-site monitoring data for each working section; where t≥h+1.

[0107] Step 5-3, extracting the target domain shared parameter characteristic values: For each working segment in step 5-1 and step 5-2, the shared parameter characteristic values ​​are extracted.

[0108] The method for extracting the characteristic value of the target domain shared parameter specifically refers to the segmentation of the working phase data stream in step 2-1, the determination of the shared parameter space in the target domain in step 2-2, and the extraction of the shared parameter characteristics in the target domain in step 2-3.

[0109] In this embodiment, the field monitoring data dimension of the target domain Yinchuojiliao No. 6 water diversion tunnel is not completely consistent with the field monitoring dimension of the source domain Yinsong water diversion tunnel No. 4. Share feature space with the source domain of Section 4 of the Yinsong Water Diversion Tunnel The intersection of: As the migration application parameter space. There are two types of parameters: characteristic parameters related to and unrelated to the rock breaking process. Contains parameters related to rock breaking such as cutter head thrust, cutter head torque, advance speed, cutter head speed and penetration parameters. These parameters directly related to cutting rock can be easily obtained on site, thus ensuring migration Is a non-empty collection.

[0110] Then, based on steps 2-3, extract the target domain The characteristic values ​​related to the rock breaking process in the real-time working section data stream. The characteristic values ​​contain the characteristics related to the rock breaking process to physically guide the digging index, where the digging index includes the linear fitting slope a and thrust intercept b of the single-knife thrust and penetration data stream, and the linear fitting slope TPI of the single-knife torque and penetration.

[0111] The statistical characteristic values ​​of each real-time working segment data stream in the target domain, such as the mean thrust of the cutter head, the mean thrust speed, and the mean cutter head speed, are also extracted. Share feature space with source domain The intersection of: As the migration application parameter space.

[0112] Step 5-4, thrust prediction for the tth working section: Input the shared parameter feature values ​​of the L historical working sections before the tth working section extracted in step 3 into the thrust prediction model based on LSTM constructed in step 3 to obtain the thrust prediction value of the tth working section in the target domain.

[0113] Step 5-5, calculate similarity: the actual thrust value of the target domain t-th working section obtained by monitoring in step 5-2 The thrust prediction value obtained in step 5-4 Calculate the similarity error SI of the tth working segment in the target domain t .

[0114] Step 5-6: Online prediction of surrounding rock quality of the t+1th working section: The surrounding rock type of the t-1th working section is taken as the surrounding rock type s before migration. t-1 , will s t-1 and SI t Substitute the surrounding rock state prediction model P(s) constructed in step 4 into t )=p ij (s t-1 |s t ,MPE i ), find out in s t-1 and SI t The probability values ​​of different surrounding rock states under the following conditions; the surrounding rock state corresponding to the maximum probability value is the predicted surrounding rock quality state of the t+1th working section; the surrounding rock type of the t-1th working section is directly surveyed according to step 5-1 or obtained through online quality prediction based on the survey data of step 5-1.

[0115] When the TBM excavates to the next working section, repeat the above steps 5-3 to 5-6 to obtain the data of the t+1th working section and predict the surrounding rock category status.

[0116] Based on the above steps, 861 real-time data streams in the target domain Yinchuojiliao No. 6 TBM construction process were predicted online, and the single-blade thrust prediction value sequence, similarity error MPE sequence and surrounding rock state probability prediction sequence of each online data stream were obtained. This example demonstrates the prediction performance of the online transfer learning method by taking the example of reducing the surrounding rock state category from pile number 65334 to 65325 in the target domain.

[0117] The actual category of the surrounding rock state in the field near the target domain pile numbers 65334 to 65325 was determined by the geological engineer according to the relevant national standards, and it changed from Class II to Class IV, such as Figure 7 As shown. Sub-figure (a) shows the change process of the predicted thrust and actual thrust of the online data stream, where the actual thrust values ​​in data streams 1 and 2 are consistent with the predicted values, while data stream 3 shows a large relative error, and the actual thrust is 19.12% lower than the predicted value. The subsequent three data streams also showed similar phenomena, which indicates that the thrust required for the excavated surrounding rock is less than that of the excavation section, which means that the strength of the surrounding rock is reduced. Sub-figure (b) gives the change trend of the rock mass similarity index MPE calculated by the online data stream. The MPE value of data stream 3 in the transition zone is -19.12%, and when the surrounding rock state changes from II to IV, the MPE values ​​of data streams 4, 5 and 6 drop to -33.53%, -31.43% and -25.59%, respectively. After the above MPE is input into the surrounding rock state prediction model, the probability changes under different states are obtained.

[0118] As shown in sub-figure (c), Class II shows a trend of decreasing first and then increasing, MPE increases first and then decreases, and Classes III, IV, and V show a trend of increasing first and then stabilizing, but the rock mass with the largest probability value is Class IV, so the online prediction model predicts that the surrounding rock after pile number 65331 is Class IV. The results show that the real-time prediction method of surrounding rock state based on online transfer learning can accurately identify the change process of surrounding rock quality.

[0119] In addition, a real-time prediction method of surrounding rock state based on online transfer learning is used to statistically analyze the prediction results of surrounding rock state of all working section data streams in the target domain, such as Figure 8 As shown in the figure. Among them, the accuracy of identifying surrounding rock state transition in the source domain of Yinsong water diversion tunnel is 86.23%, and the accuracy of surrounding rock classification in the working section is 86.70%; the accuracy of surrounding rock state transition in the target domain Yinchuojiliao No. 6 section is 70.83%, and the accuracy of surrounding rock classification in the real-time working section data stream is 83.39%. In addition, the confusion matrix of tunnel surrounding rock classification in the target domain is shown in sub-figure (b), and its microscopic and macroscopic F1 values ​​are 62.09% and 84.62% respectively. It can be seen that the online transfer learning method can achieve ideal prediction results in the target domain.

[0120] The preferred embodiments of the present invention are described in detail above; however, the present invention is not limited to the specific details in the above embodiments. Within the technical concept of the present invention, various equivalent transformations can be made to the technical solutions of the present invention, and these equivalent transformations all belong to the protection scope of the present invention.

Claims

1. A real-time perception method of TBM surrounding rock quality based on physics-guided online transfer learning, characterized by: The steps include: Step 1: Divide the source domain and target domain tunnel engineering objects: The source domain refers to the tunnel engineering that has been excavated. Its on-site monitoring data space is represented by X S Indicates that the data space of the source domain contains M tunnel engineering field monitoring data, where M≥1; the surrounding rock quality status of the M tunnel engineering in the source domain is divided into S types, where S≥2; The target domain refers to the N tunnel projects that have not yet been excavated, and the field monitoring data space is represented by X T Represents, where N≥1; Step 2: Extracting characteristic parameters based on the shared parameter space of source domain field monitoring data, including the following steps: Step 2-1, segmenting the data streams of effective rock breaking working stages of TBMs in different tunnel projects in the source domain to obtain n effective rock breaking working section data streams; Step 2-2: Determine the shared parameter space X in the source domain Shard : A shared parameter space X is established for each effective rock breaking working section data stream Shard ;X Shard It is determined according to the number M of tunnel engineering field monitoring data in the source domain, specifically: A. When M = 1, X Shard is the rock breaking parameter; B. When M>1, X Shard is the union of rock-breaking parameters and non-rock-breaking parameters; Step 2-3, extracting shared parameter characteristic values: the shared parameter characteristic values ​​include rock breaking parameter characteristic values ​​and non-rock breaking parameter characteristic values; wherein the rock breaking parameter characteristic values ​​include rock breaking parameter statistical characteristic values ​​and digging index; wherein the digging index is an index related to the rock breaking process and guided by physics; the non-rock breaking parameter characteristic values ​​are the mean values ​​of each non-rock breaking parameter; Step 3, constructing a thrust prediction model based on LSTM: constructing a thrust prediction model based on the LSTM neural network algorithm; assuming that the current working segment to be predicted is the t-th working segment, the input layer of the thrust prediction model is the shared parameter feature values ​​of the L historical working segments before the t-th working segment; the output layer is the thrust prediction value of the t-th working segment; according to the source domain shared parameter feature values ​​extracted in step 2, the constructed thrust prediction model is trained to obtain the trained thrust prediction model; Step 4: Construct a surrounding rock state prediction model based on similarity index: calculate the similarity error between the thrust prediction value of the t-th working section and the actual thrust value of the t-th working section; obtain the similarity error sequence SI corresponding to the n effective rock breaking working sections according to the source domain shared parameter characteristic value extracted in step 2; and then combine the S surrounding rock quality states divided in step 1 to construct a surrounding rock state prediction model based on the similarity index sequence SI and the corresponding surrounding rock classification sequence S; Step 5: Online prediction of target domain surrounding rock quality, including the following steps: Step 5-1, initial excavation of the target area: The TBM excavates the target area to the hth working section, and conducts on-site monitoring data of the tunnel engineering for each working section; at the same time, a field survey is conducted on the surrounding rock types of the first to h excavated working sections; Step 5-2, the target area continues to be excavated: the TBM excavates the target area from the h+1th working section to the tth working section, and conducts tunnel engineering on-site monitoring data for each working section; where t≥h+1; Step 5-3, extracting the target domain shared parameter characteristic value: for each working segment in step 5-1 and step 5-2, the shared parameter characteristic value is extracted; Step 5-4, thrust prediction for the tth working section: Input the shared parameter feature values ​​of the L historical working sections before the tth working section extracted in step 3 into the thrust prediction model based on LSTM constructed in step 3 to obtain the thrust prediction value of the tth working section in the target domain. Step 5-5, calculate similarity: the actual thrust value of the target domain t-th working section obtained by monitoring in step 5-2 The thrust prediction value obtained in step 5-4 Calculate the similarity error SI of the tth working segment in the target domain t ; Step 5-6: Online prediction of surrounding rock quality of the t+1th working section: The surrounding rock type of the t-1th working section is taken as the surrounding rock type s before migration. t-1 , will s t-1 and SI t Substitute it into the surrounding rock state prediction model constructed in step 4 and find t-1 and SI t The probability values ​​of different surrounding rock states under the following conditions; the surrounding rock state corresponding to the maximum probability value is the predicted surrounding rock quality state of the t+1th working section; the surrounding rock type of the t-1th working section is directly surveyed according to step 5-1 or obtained through online quality prediction based on the survey data of step 5-1.

2. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In step 1, the surrounding rock quality status of the M tunnel projects in the source domain adopts the national standard "Code for Geological Investigation of Water Conservancy and Hydropower Engineering GB50487-2008", and the tunnel surrounding rock is divided into stable I, basically stable II, partially stable III, unstable IV and extremely unstable V, with a total of S = 5 types.

3. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In step 2-2, the rock breaking parameters include cutter head thrust, cutter head torque, propulsion speed, cutter head rotation speed and penetration; wherein the penetration is the ratio of propulsion speed to cutter head rotation speed.

4. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 3 is characterized by: In step 2-3, the excavability indicators include the linear fitting slope a and thrust intercept b of the single-pole thrust and penetration data streams, and the linear fitting slope TPI of the single-pole torque and penetration.

5. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In steps 2-3, the rock breaking parameter statistical characteristic values ​​include the rock breaking operation parameter statistical characteristic values ​​and the rock breaking non-operation parameter statistical characteristic values; in step 3, the LSTM neural network includes an LSTM layer and a fully connected layer; the input of the LSTM layer is the rock breaking non-operation parameter statistical characteristic values, digging index and non-rock breaking parameter characteristic values ​​of the L historical working sections before the t-th working section; the input of the fully connected layer is the output of the LSTM layer and the rock breaking operation parameter statistical characteristic values; The output of the fully connected layer is the thrust prediction value of the tth working section.

6. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In step 3, the thrust prediction model is trained according to the source domain shared parameter feature values ​​extracted in step 2, and during the training process, the hyperparameters of the thrust prediction model are optimized. The specific optimization method includes the following steps: Step 3B-1, sample library division: the n source domain working segment data streams loaded in step 3A-2 are divided into a training set and a validation set according to a set ratio; Step 3B-2, designing hyperparameters: performing several combinations of hyperparameters involved in the thrust prediction model; Step 3B-3, model training: using the training set data to train each hyperparameter combination in the thrust prediction model, to obtain several thrust prediction models that have been trained; Step 3B-4, hyperparameter optimization: Use the validation set to validate each thrust prediction model trained in step 3B-3, and select a hyperparameter combination with the best validation performance as the optimal hyperparameter. The optimal hyperparameter is frozen and fixed. The thrust prediction model with the optimal hyperparameter is also called the optimal thrust prediction model.

7. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In step 4, let the similarity error sequence of the source domain be SI = {si1,...,si i ,...,si n }, and according to the surrounding rock quality status classification results in the source domain, the surrounding rock classification sequence S = {s0, s1, ..., s i ,...,s n }; Based on the similarity index sequence SI and the corresponding surrounding rock classification sequence, a rock mass classification state transition probability model is constructed, as shown in the following formula: M(s,si)={p0(s0|s,si),...,p i (with i |s,si),...,p n (with n |s,si)} In the formula, p i (s i |s,si) is the conditional probability density distribution function when the excavated state is s and the similarity index is si when transferring to any other state. This distribution function establishes a non-parametric probability density statistical model by the similarity index of different rock mass categories in the surrounding rock classification sequence when the state is transferred.

8. The real-time perception method of TBM surrounding rock quality based on physical guidance online transfer learning according to claim 1 is characterized by: In step 5-1, h=20.

Images