Tunnel surrounding rock stability diagnosis method based on TBM multi-modal data fusion

Abstract

The invention discloses a tunnel surrounding rock stability diagnosis method based on TBM multi-modal data fusion. The method comprises the steps of 1, dividing surrounding rock stability types; step 2, preparing TBM (Tunnel Boring Machine) multi-modal data; step 3, carrying out space-time alignment and fusion on the TBM multi-modal data; step 4, multi-modal data null value and duplicate value filling; 5, multi-modal data feature extraction based on the excavation and tunneling section; step 6, constructing a TBM multi-modal feature database based on the tunneling section; 7, constructing a tunnel surrounding rock stability prediction model; and step 8, tunnel surrounding rock stability field diagnosis application. According to the method, the alignment accuracy of different modal data is guaranteed based on time and space information, the modal data with low data null value degree and high data null value degree are repaired and complemented through the nearest neighbor null value filling and linear interpolation method, and the method can be suitable for fusion of different monitoring modals in the tunnel engineering construction process; rapid diagnosis of on-site surrounding rock stability can be realized, and construction environment safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of tunnel boring machine construction monitoring informatization, and in particular to a tunnel surrounding rock stability diagnosis method based on TBM multi-modal data fusion. Background Art

[0002] The tunnel boring machine excavation method has the advantages of high mechanization, fast excavation speed and safe construction environment, so it is widely used in tunnel engineering excavation. Due to the constraints of the working space during the construction process, traditional advanced geological exploration methods such as microseismic and geological radar have high costs, affect the normal excavation process and the diagnosis results are not timely. This leads to the diagnosis of surrounding rock stability during mechanized tunnel construction mainly relying on the incomplete geological survey report in the early stage. However, tunnel boring machines are sensitive to changes in geological conditions. When encountering extremely broken and unstable sections, they will cause major engineering accidents such as landslides and machine jams, which seriously affect the life and property safety of on-site workers. Therefore, it is urgent to develop a real-time prediction method that can be used for on-site surrounding rock stability diagnosis.

[0003] In recent years, intelligent prediction models for surrounding rock stability during construction based on field monitoring data have been developed. Most existing methods are based on the same modal data source for data processing, feature extraction and model construction, which have the disadvantages of single data source, high correlation and lack of complementarity, resulting in limited model prediction performance. However, with the advancement of sensing technology, it has become possible to monitor and obtain data of different modes during TBM excavation. For example, the machine operating parameters of the tunnel boring machine, the radial wear value of the cutter under different cutter positions on the cutterhead, the vibration data of different monitoring points on the cutterhead, and the images of surrounding rock or slag on the belt conveyor during construction are all collected and stored as the tunnel excavation process. The establishment of a real-time diagnosis model for surrounding rock stability based on the data fusion processing of different modes during machine excavation can greatly reduce the cost of on-site advanced geological exploration, timely and accurately predict the stability of the surrounding rock, and realize rapid diagnosis and early warning of unstable sections, thereby ensuring the safe and efficient on-site excavation process.

[0004] Real-time diagnosis of surrounding rock stability based on multimodal data requires the collection of detailed construction data of different modal dimensions under different geological conditions on site. It also requires the construction of an associated fusion database of information from different data sources, the establishment of a standardized fusion processing technology system for multi-source data, the realization of multimodal data alignment and completion, and the extraction of characteristic parameters at a reasonable frequency for the tunneling process of the tunnel boring machine. This provides an important data basis for the construction of a real-time diagnosis model for surrounding rock stability, which can not only realize rapid diagnosis of surrounding rock, but also reduce the time processing cost of high-frequency modal data. This method can effectively reduce the cost of on-site advanced geological surveys, efficiently and accurately evaluate the surrounding rock status during the construction of the tunnel boring machine, and thus ensure the safety of the construction environment. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion in view of the deficiencies of the above-mentioned prior art. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion further ensures the accuracy of alignment between different modal data based on time and space information, and repairs and completes modal data with low and high null value degrees through nearest neighbor null value filling and linear interpolation methods. It can be applied to different monitoring modal fusion databases in the tunnel engineering construction process, reduces the time processing cost of different modal data, and at the same time realizes rapid diagnosis of on-site surrounding rock stability to ensure the safety of the construction environment.

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

[0007] A tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion includes the following steps.

[0008] Step 1: Classify the surrounding rock stability types: divide the tunnel surrounding rock into five types: stable, basically stable, partially stable, unstable and extremely unstable.

[0009] Step 2, TBM multimodal data preparation: For each type of stability state of the tunnel surrounding rock, the TBM mechanized construction method is used for excavation, and during the tunnel excavation construction process, the modal data of different surrounding rock stability types are continuously monitored and collected; therefore, n types of modal data are collected under each type of surrounding rock stability state condition; then, the different modal data under each type of stability state condition are divided into days to form n types of modal data blocks.

[0010] Step 3: Spatiotemporal alignment of TBM multimodal data: For each stability state type, the n modal data blocks of the tunnel surrounding rock are sorted from high to low according to the sampling frequency. Taking the modal data block corresponding to the highest sampling frequency as the benchmark, the remaining n-1 modal data blocks are traversed based on the time and space pile number positions, and copied to the modal data block corresponding to the highest sampling frequency to achieve alignment and fusion processing.

[0011] Step 4, data filling: Identify the null values ​​and duplicate values ​​in the TBM multimodal data after spatiotemporal alignment in step 4, and use the nearest neighbor filling or linear interpolation method to repair the identified null values ​​and duplicate values.

[0012] Step 5, feature extraction: for each stability state type, the n types of modal data blocks filled with tunnel surrounding rock data are all extracted using the excavation section as a unit; for each modal data in an excavation section, several values ​​are extracted as feature values. If n modal values ​​can be obtained in the same excavation section, ifeatures, then n modal data blocks in the same excavation section correspond to n sum =n1+n2+…+n n characteristic values; among them, the excavation section refers to the stage data in which the TBM cutter head speed changes from zero to non-zero and finally drops to zero during the process from the start of the cutter head to the shutdown when the maximum stroke is reached.

[0013] Step 6: Construct a TBM multimodal feature database based on the excavation section: All characteristic values ​​of all excavation sections extracted based on different surrounding rock stability conditions in step 5 together constitute a TBM multimodal feature database based on the excavation section.

[0014] Step 7: Construct a tunnel surrounding rock stability prediction model: The tunnel surrounding rock stability prediction model is a machine learning model. The input of the tunnel surrounding rock stability prediction model is the characteristic value of n types of modal data of a tunneling section, and the output of the tunnel surrounding rock stability prediction model is the surrounding rock stability type corresponding to the tunneling section. Use the TBM multimodal feature database based on the tunneling section constructed in step 6 to train the tunnel surrounding rock stability prediction model.

[0015] Step 8: Tunnel surrounding rock stability diagnosis application: Monitor the n types of modal data of the TBM during tunnel excavation construction, and obtain a total of n different modal data of the current excavation section. sum Then, the n eigenvalues ​​of the current excavation section are sum The characteristic values ​​are input into the tunnel surrounding rock stability prediction model after training in step 7 to predict the surrounding rock stability state of the current tunnel excavation section under construction.

[0016] In step 2, the modal data includes TBM operation data, cutterhead cutter wear data, cutterhead vibration data, and surrounding rock or slag images.

[0017] In step 2, the TBM operation data includes thrust, torque, propulsion speed, rotation speed and penetration; the cutterhead cutter wear data includes the radial wear data of the cutter at each position on the cutterhead; and the cutterhead vibration data includes the vibration acceleration data at the monitoring position.

[0018] In step 3, the spatiotemporal alignment method of TBM multimodal data includes the following steps:

[0019] Step 3-1, inter-modal sorting: sort the n modal data blocks after data segmentation from high to low according to the sampling frequency, namely: the first modal data block, the second modal data block, ..., the nth modal data block.

[0020] Step 3-2, Intra-modal sorting: Sort each modal data block in the order of monitoring time or spatial stake position.

[0021] Step 3-3, alignment of the second modal data block: Based on the time and space pile number position, taking the time information of the first modal data block as a reference, through the time index, the data at each acquisition frequency in the second modal data block is inserted into the first modal data block with the same time accuracy, so as to realize data alignment and fusion of the second modal data block and the first modal data block.

[0022] Step 3-4: Repeat step 3-3 until the data alignment and fusion of the nth modal data block and the first modal data block are completed.

[0023] When filling the data in step 4, the degree type of null values ​​and duplicate values ​​must be determined first; for modal data blocks with a small degree of null values ​​and duplicate values, the KNN nearest neighbor filling method is used to repair the null values ​​and duplicate values; for modal data blocks with a large degree of null values ​​and duplicate values, the nearest neighbor linear interpolation method is used to repair the null values ​​and duplicate values.

[0024] In step 4, the specific method of data filling includes the following steps:

[0025] Step 4-1, identifying null values ​​and duplicate values: identifying null values ​​and duplicate values ​​in the second modality data block after alignment and fusion processing.

[0026] Step 4-2, determine the degree of null values ​​and repeated values: compare the number of null values ​​and repeated values ​​of each mode in each day with the set ratio value. When the number of null values ​​and repeated values ​​in the i-th day is less than the set ratio value, all null values ​​and repeated values ​​in the i-th day are judged as null values ​​or repeated values ​​with a small degree; otherwise, they are judged as null values ​​or repeated values ​​with a large degree.

[0027] Step 4-3, repairing empty values ​​and repeated values: When the empty value or repeated value degree in the second modal data block determined in step 4-2 is small, the KNN nearest neighbor filling method is used for repair; when the empty value or repeated value degree in the second modal data block determined in step 4-2 is large, the nearest neighbor linear interpolation method is used for repair, such as Figure 3 shown.

[0028] Step 4-4, repeat step 4-1 to step 4-3, and complete the repair of empty values ​​and duplicate values ​​of the third modal data block to the nth modal data block in sequence.

[0029] In step 5, feature extraction is performed on different modal data in the form of excavation sections using statistical features (such as mean) or methods with physical indicators, such as Figure 4 For example, in the same excavation section, the ith mode obtains n i features, then for all n modal data blocks, n sum =n1+n2+…+n nThe excavation section refers to the stage data when the TBM cutter head speed changes from zero to non-zero and finally drops to zero during the process from the start of the cutter head to the maximum stroke stop. Then, the characteristic values ​​of n modes are extracted for all the excavation section data under different working conditions.

[0030] In step 7, the deep learning model is an estimation model based on a gradient-based boosting ensemble learning algorithm.

[0031] In step 4-2, the ratio threshold for determining the null value and duplicate value types is set to 5% or 10% of the total number of samples in the corresponding mode within the corresponding day.

[0032] The present invention has the following beneficial effects:

[0033] (1) For the prediction of surrounding rock stability during TBM tunnel construction, the traditional method based on geological survey and advanced geological drilling has high construction cost and long prediction and analysis time, which is not conducive to efficient and timely diagnosis in mechanized rapid construction. The present invention uses multi-modal monitoring data from the tunnel construction site to construct a feature fusion method based on the excavation section, effectively realizes the alignment of different modal data, reduces the time processing cost of high-frequency data, and provides a basis for timely diagnosis of surrounding rock stability on site, thereby ensuring the safe and efficient construction process.

[0034] (2) The present invention proposes a multimodal data alignment method based on spatiotemporal sorting, which ensures the accuracy of data block alignment between different modalities from the two dimensions of time information and spatial position, and realizes the effective fusion of multimodal data.

[0035] (3) The multimodal surrounding rock stability diagnostic model proposed in the present invention can be applied to timely diagnose the surrounding rock stability status during the construction process based on field monitoring data, significantly reducing the cost of field testing such as on-site advance geological drilling, while avoiding the impact on tunnel boring machine construction and improving tunnel excavation efficiency. The established diagnostic model can also provide early warning of tunnel instability risks, thereby greatly avoiding the occurrence of major engineering accidents and ensuring construction safety. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] Figure 1 The present invention is a flow chart of the tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion.

[0037] Figure 2 This is a schematic diagram of the data structure of TBM monitoring mode 1, mode 2, mode n itself and the fused data structure after spatiotemporal alignment when the sampling frequency is arranged from high to low.

[0038] Figure 3Schematic diagram of the principle of filling in null values ​​after TBM multimodal fusion data, where Figure (a) is a schematic diagram of the principle of filling using the nearest neighbor KNN, which is mainly suitable for repairing modal data with a low degree of null values; Figure (b) is a schematic diagram of the principle of filling using the nearest neighbor linear interpolation, which is mainly suitable for modal data with a high degree of null values.

[0039] Figure 4 Schematic diagram of the TBM multimodal feature extraction method based on the excavation section. Figure (a) shows the principle diagram of feature extraction of different modal data of each excavation section; Figure (b) shows the database diagram after the feature fusion of different modal data of each excavation section.

[0040] Figure 5 This is an example diagram of the Jiliao TBM vibration, machine and cutter wear modal fusion database.

[0041] Figure 6 This is an example diagram of the fusion database of vibration, machine and cutter wear modal features of the Yinchoujiliao TBM based on the excavation section.

[0042] Figure 7 This is the prediction result of surrounding rock stability category based on multi-modal data fusion of Yinchuojiliao TBM. DETAILED DESCRIPTION

[0043] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific preferred embodiments.

[0044] The present invention is described in detail by taking the on-site multimodal monitoring data of the open TBM construction process of the No. 6 section of the tunnel in Yinchuojiliao, Inner Mongolia as an example.

[0045] like Figure 1 As shown in FIG, the tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion includes the following steps.

[0046] Step 1. Classify the surrounding rock stability types: According to the national standard for on-site surrounding rock stability evaluation of this tunnel project "Code for Geological Investigation for Water Conservancy and Hydropower Engineering GB50487-2008", the tunnel surrounding rock is divided into five types according to the categories divided by known standards: stable, basically stable, partially stable, unstable and extremely unstable.

[0047] Step 2: TBM multimodal data preparation

[0048] A. Data Collection

[0049] For each type of tunnel surrounding rock in a stable state, TBM is used for excavation, and different types of modal data are continuously monitored and collected during the tunnel excavation construction process; therefore, n types of modal data M=[M1,M2,...,Mn ].

[0050] The above modal data include TBM operation data, cutterhead cutter wear data, and cutterhead vibration data. Among them, TBM operation data include thrust, torque, propulsion speed, rotation speed and penetration, etc., which are measured by different types of sensors such as displacement, velocity and force sensors; cutterhead cutter wear data include radial wear data of 34 cutters on the cutterhead, and the wear value is measured by the on-site cutterhead maintenance construction team; cutterhead vibration data are measured by vibration sensors installed behind the cutterhead.

[0051] In this embodiment, taking the on-site monitoring data of the No. 6 branch tunnel boring machine of Yinchuojiliao as an example, three modal data of cutter vibration, machine operation and roller wear within the pile number range of 66058.81 to 55855.69 were collected, and the rock stability within this pile number range includes four states: basically stable, partially stable, unstable and extremely unstable.

[0052] In this embodiment, the machine operation modal data within the pile number range of 66058.81 to 55855.69 is mainly acquired through sensors for parameters such as total thrust, total torque, propulsion speed, and rotation speed, and the sampling frequency is preferably 1 Hz / s.

[0053] In this embodiment, the cutter disc wear data between the pile number range of 66058.81 and 55855.69 includes radial wear values ​​of the roller cutters at 34 positions on the cutter disc, and the sampling frequency is preferably 1 Hz / day.

[0054] In this embodiment, the cutter disc vibration data within the pile number range of 66058.81 to 55855.69 mainly includes 6 sensors installed behind the cutter disc, and the sampling frequency is preferably 240 Hz / s.

[0055] B. Data segmentation

[0056] Each modal data of tunnel surrounding rock of each stability state type is divided into n modal data blocks by day. The data of the first day is taken as an example.

[0057] In this embodiment, the modal data of the three dimensions between the pile number range of 66058.81 and 55855.69 is divided into data blocks in units of days by writing a time segmentation function, and the data of December 19, 2020 is taken as an example. The data representation is

[0058] Step 3: Spatiotemporal alignment of TBM multimodal data: For each type of stability state, the n modal data blocks of tunnel surrounding rock are sorted from high to low according to the sampling frequency. The cutterhead vibration modal data block corresponding to the highest sampling frequency is used as the benchmark. Based on the time and space pile number position, the remaining n-1 modal data blocks are traversed and copied and pasted into the modal data block corresponding to the highest sampling frequency to achieve alignment and fusion processing.

[0059] The above-mentioned method for spatiotemporal alignment of TBM multimodal data preferably includes the following steps.

[0060] Step 3-1, sorting between modalities: sort the n modal data blocks after data segmentation from high to low according to the sampling frequency, namely: the first modal data block (also called modality 1), the second modal data block (also called modality 2), ..., the nth modal data block (also called modality n). That is, the sampling frequency of modality 1 is the highest.

[0061] In this embodiment, the three types of modal data are first sorted according to the size of the sampling frequency to obtain: vibration modal sampling frequency>machine operation modal sampling frequency>hob wear modal sampling frequency.

[0062] Step 3-2, sorting within the mode: sort each mode data block in the order of monitoring time and spatial pile number position. For example, first sort the monitoring data of the first day in mode 1. The monitoring data under the same monitoring mode in this data block are arranged in sequence according to the time record and pile number displacement information.

[0063] Then the monitoring data of the first day in mode 2 Similarly, the data in mode 2 are sorted according to the time and pile number record information.

[0064] Then the monitoring data of the first day in mode 3 Similarly, the data in mode 3 are sorted according to the time and pile number record information.

[0065] By analogy, the monitoring data of the first day in mode n is All records are sorted according to time and pile number.

[0066] In this embodiment, the cutterhead vibration data with the highest sampling frequency is selected, and the data of December 19, 2020 are sorted in chronological order to obtain the data shown in Table 1. During the construction of Tunnel No. 6 of Yinchuojiliao Section 2, a total of 6 sensors were installed on the TBM cutterhead, and the sampling frequency of each sensor was 240Hz / s. Therefore, the 1st to 240Hz data of the first sensor are shown in the vibration 1-240 column data.

[0067] Table 1 Cutterhead vibration modal data of Yinchuojiliao Section 6

[0068]

[0069] Select the machine operation data with the second sampling frequency, and sort it according to the data of December 19, 2020 or the sequence of pile numbers to obtain the data shown in Table 2.

[0070] Table 2 Machine operation modal data of Section 6 of Yinchuojiliao

[0071]

[0072] The cutter wear data with the third highest sampling frequency is selected, and it is also sorted according to the time or pile number sequence, and the data shown in Table 3 are obtained.

[0073] Table 3 Radial wear value data of hobs at different cutter positions in Section 6 of Yinchuojiliao (unit: mm)

[0074]

[0075] Step 3-3, alignment of the second modal data block: Based on the time and space pile number position, taking the time information of the first modal data block as a reference, through the time index, the data at each acquisition frequency in the second modal data block is inserted into the first modal data block with the same time accuracy, so as to realize data alignment and fusion of the second modal data block and the first modal data block.

[0076] Since the data monitoring frequency in mode 2 is less than the frequency of the data in mode 1, the method to achieve the fusion of the two modal data is to map the data in mode 2 to mode 1. Therefore, it is necessary to traverse the time and pile number index in mode 2, and after retrieving the same spatiotemporal position in mode 1, insert the data in mode 2 into the corresponding position of mode 1 and add a new column to save it. Finally, the spatiotemporal position alignment of the two modal data is achieved.

[0077] That is, after sorting in the modal Insert it to the corresponding position of mode 1 and add a new column to save it. For example, based on the time information 2020 / 12 / 19 0:30:31 as the index, extract the data of the corresponding row of the machine operation modal data, copy the corresponding data and insert it into the corresponding cutter vibration modal data.

[0078] Step 3-4, repeat step 3-3 until the data alignment and fusion of the nth modal data block and the first modal data block at each time and stake number are completed; that is, for the last modal data Perform time and space alignment with the modal 1 data with the highest sampling frequency respectively, so as to establish a mapping relationship between the existing modal data and the modal 1 with the highest sampling frequency.

[0079] Among them, based on the time 2020 / 12 / 19, the information of the pile number position 63926.3, the wear data at different tool positions on that day are copied and inserted into the fusion data closest to the pile number 63926.3 in the cutter head vibration and machine fusion database to obtain the multimodal fusion database as follows Figure 5 shown.

[0080] Finally, repeat steps 3-2 to 3-4 to complete the alignment and fusion of multimodal data of all working days data blocks.

[0081] Step 4, data filling: Identify the null values ​​and duplicate values ​​in the TBM multimodal data after spatiotemporal alignment in step 4, and use the nearest neighbor filling or linear interpolation method to repair the identified null values ​​and duplicate values.

[0082] When filling the above data, it is necessary to first determine the degree of null values ​​and duplicate values; for modal data blocks with a small degree of null values ​​and duplicate values, the KNN nearest neighbor filling method is used to repair the null values ​​and duplicate values; for modal data blocks with a large degree of null values ​​and duplicate values, the nearest neighbor linear interpolation method is used to repair the null values ​​and duplicate values.

[0083] The specific method of filling the above data preferably includes the following steps.

[0084] Step 4-1, identifying null values ​​and duplicate values: identifying null values ​​and duplicate values ​​in the second modality data block after alignment and fusion processing.

[0085] Step 4-2, determine the degree of null values ​​and repeated values: compare the number of null values ​​and repeated values ​​of each mode in each day with the set ratio value. When the number of null values ​​and repeated values ​​in the i-th day is less than the set ratio value, all null values ​​and repeated values ​​in the i-th day are judged as null values ​​or repeated values ​​with a small degree; otherwise, they are judged as null values ​​or repeated values ​​with a large degree.

[0086] The above-mentioned set ratio value is preferably 5% or 10% of the total number of the corresponding modes in the corresponding day.

[0087] Step 4-3: Fix empty and duplicate values

[0088] When the null value or repeated value in the second modal data block determined in step 4-2 is small, the KNN nearest neighbor filling method is used for repair. The KNN nearest neighbor filling method is preferably: querying k non-null value data before and after the null value and using the average value of these non-null value data to fill the current null value.

[0089] When the null value or repeated value in the second modal data block determined in step 4-2 is large, the nearest neighbor linear interpolation method is used for repair. The nearest neighbor linear interpolation method preferably performs linear interpolation by retrieving the null value time or the non-null data before and after the pile number and the pile number position relationship.

[0090] Step 4-4, repeat step 4-1 to step 4-3, and complete the repair of empty values ​​and duplicate values ​​of the third modal data block to the nth modal data block in sequence.

[0091] In this case, the sampling frequency of the machine operation modal data is high, and its null value data is small and less than the set threshold, so it belongs to low null value modal data. The sampling frequency of the machine operation modal data and the vibration frequency can be aligned at the second level, so the nearest neighbor KNN filling method is used to repair it. The specific implementation technology is to fill the null values ​​in the machine operation modal data with the average of its nearest k non-null values. In this case, k=1 is selected.

[0092] In this case, the sampling rate of the hob modal data is low, and its null value data is large and greater than the set threshold, so it belongs to high null value modal data. Therefore, the nearest neighbor linear interpolation method is used for repair. The specific implementation method is to query the wear value closest to the null value position time or pile position, perform linear interpolation based on the non-null value and pile number spatial position, and fill the interpolation result into the null value position of the multimodal fusion data.

[0093] For example, the radial wear value of the tool increases from 2mm to 8mm between the pile number range of 63926.3 and 63898.3 at tool position 32. The wear rate of the roller within the pile number range can be calculated as (8mm-2mm) / (63926.3-63898.3)=0.2143mm / m. The wear value of the roller within the pile number range can be calculated by the interpolation formula based on the wear rate and the corresponding pile number position. For example, the wear value at the pile number position 63098.55 is calculated by 2+(63926.3-63098.55)×0.2143=5.80mm.

[0094] Step 5, feature extraction: for each stability state type, the n modal data blocks filled with tunnel surrounding rock data are all subjected to feature extraction in units of excavation sections; the same modal data block in the same excavation section extracts several eigenvalues ​​based on statistical or fitting methods. Therefore, the n modal data blocks in the same excavation section correspond to n sum =n1+n2+…+n n characteristic values; among them, the excavation section refers to the stage data when the cutter head speed value of the TBM drops from a non-zero value to zero.

[0095] Step 6: Construct a TBM multimodal feature database based on the excavation section: All characteristic values ​​of all excavation sections extracted from the four types of tunnel surrounding rocks in step 5 together constitute a TBM multimodal feature database based on the excavation section.

[0096] In this embodiment, feature extraction is performed on multimodal data in units of excavation sections. Taking the three modal monitoring data on December 19, 2020 as an example, 9 effective excavation excavation sections were identified based on the cutter head speed that reflects the working status of the cutter head. Feature extraction is performed on the different modal data within each excavation section. For the vibration modal data, the maximum vibration amplitude, minimum vibration amplitude, effective vibration amplitude and other parameters of each excavation section are calculated; for the excavation data, the excavation index, thrust and penetration fitting slope a and intercept b, torque and penetration fitting slope TPI and other parameters are calculated for each excavation section. For the cutter wear modal characteristics, the tool wear speed parameters of 34 cutter positions are calculated based on the cutter wear value and excavation distance. Finally, a database of cutter vibration, machine operation and cutter wear characteristics based on the excavation section is obtained as follows: Figure 6 shown.

[0097] In this embodiment, the TBM multimodal feature database based on the excavation section is constructed, including feature data of 5317 excavation sections.

[0098] Step 7: Construct a tunnel surrounding rock stability prediction model: The tunnel surrounding rock stability prediction model is a machine learning model. The input of the tunnel surrounding rock stability prediction model is n different modes of a tunneling section. sum The output of the tunnel surrounding rock stability prediction model is the tunnel surrounding rock stability type. Among them, the machine learning model is preferably an estimation model based on the gradient boosting ensemble learning algorithm.

[0099] Step 8: Prediction model training: Use the TBM multimodal feature database based on the excavation section constructed in step 6 to train the tunnel surrounding rock stability prediction model constructed in step 7.

[0100] First, the database is divided into training set and test set according to the four surrounding rock stability states, and 4785 sets of data containing four surrounding rock stability states are used to build a prediction model based on the gradient boosting ensemble learning method. Figure 7 As shown in the figure, the prediction accuracy and F1 score of the classification diagnosis of surrounding rock stability using only machine operation modal data are 0.90 and 0.56, respectively, which indicates that many categories of surrounding rock stability in the excavation area are missed or misreported. When the two modal data features of machine excavation and cutter wear are considered, the accuracy and F1 score reach 0.97 and 0.92, respectively, which indicates that the false alarm rate and missed alarm rate have been reduced while the accuracy is significantly improved. However, after considering the three modal features of machine excavation, cutter wear and cutter vibration, the accuracy of surrounding rock stability reaches the highest, which indicates that the prediction method based on multi-modal fusion can significantly improve the prediction performance. Finally, the surrounding rock stability prediction model trained based on multi-modal data is exported and deployed in the intelligent excavation platform of the host in the on-site TBM main control room.

[0101] Step 9: Tunnel surrounding rock stability diagnosis: Monitor n types of modal data of the TBM during tunnel excavation and obtain n types of modal data of the current excavation section. sum Then, the n eigenvalues ​​of the current excavation section are sum The characteristic values ​​are input into the tunnel surrounding rock stability prediction model after training in step 8 to predict the stability state of the current tunnel surrounding rock.

[0102] In this example, the multimodal data monitored during the excavation of the new working section are aligned and repaired, and feature extraction is performed on a tunneling section basis, thereby achieving real-time diagnosis of the surrounding rock stability.

[0103] The present invention realizes the standardization and normalization of the multimodal data processing flow of independently collecting different modal data at the construction site, further ensures the accuracy of alignment between different modal data based on time and space information, and repairs and completes the modal data with low and high null values ​​through the nearest neighbor null value filling and linear interpolation methods. This method is applicable to the fusion database of different monitoring modes in the construction process of tunnel engineering, reduces the time processing cost of different modal data, and realizes the rapid diagnosis of the stability of the surrounding rock on site, ensuring the safety of the construction environment.

[0104] 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 tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion, characterized by: The steps include: Step 1: Classify surrounding rock stability types: Classify surrounding rock stability into N types of stability states; where N ≥ 2; Step 2, TBM multimodal data preparation: TBM is used to excavate the surrounding rock of each stability state type of tunnel, and different types of modal data are continuously monitored and collected during the tunnel excavation construction process; thus, n types of modal data are collected for each type of stability state of the tunnel surrounding rock; then, each modal data of each type of stability state of the tunnel surrounding rock is divided into days to form n types of modal data blocks; Step 3: Time-space alignment of TBM multimodal data: sort the n modal data blocks of tunnel surrounding rock of each stability state type from high to low according to the sampling frequency, take the modal data block corresponding to the highest sampling frequency as the benchmark, rely on the time and space pile number position, traverse the remaining n-1 modal data blocks, and copy them to the modal data block corresponding to the highest sampling frequency to achieve alignment and fusion processing; Step 4, data filling: Identify the null values ​​and duplicate values ​​in the TBM multimodal data after spatiotemporal alignment in step 4, and repair the identified null values ​​and duplicate values ​​by using the KNN nearest neighbor filling or linear interpolation method by distinguishing the percentage types of null values ​​and duplicate values; Step 5, feature extraction: for each stability state type, the n types of modal data blocks filled with tunnel surrounding rock data are all extracted with the excavation section as the unit; for the same modal data block in the same excavation section, several values ​​are extracted as feature values ​​by statistical or calculation of physical indicators. Therefore, n modal values ​​can be obtained in the same excavation section. i features, then for all n modal data blocks corresponding to n sum =n1+n2+…+n n characteristic values; among them, the excavation section refers to the stage data in which the cutter head speed of the TBM changes from zero to non-zero and finally drops to zero during the process from the start of the cutter head to the shutdown at the maximum stroke; Step 6: Construct a TBM multimodal feature database based on the excavation section: The feature values ​​of different modal data of all excavation sections extracted based on different surrounding rock stability working conditions in step 5 together constitute a TBM multimodal feature database based on the excavation section; Step 7, construct a tunnel surrounding rock stability prediction model: the tunnel surrounding rock stability prediction model is a machine learning model, the input of the tunnel surrounding rock stability prediction model is the characteristic value of n modes in a tunneling section, and the output of the tunnel surrounding rock stability prediction model is the tunnel surrounding rock stability category; use the TBM multimodal feature database based on the tunneling section constructed in step 6 to train the tunnel surrounding rock stability prediction model; Step 8: Tunnel surrounding rock stability diagnosis: Monitor n types of modal data of the TBM during tunnel excavation and obtain a total of n different modal data of the current excavation section. sum Then, the n eigenvalues ​​of the current excavation section are sum The characteristic values ​​are input into the tunnel surrounding rock stability prediction model trained in step 7 to predict the stability state of the tunnel surrounding rock in the current excavation section.

2. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1 is characterized by: In step 2, the modal data includes TBM operation data, cutterhead cutter wear data and cutterhead vibration data.

3. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 2 is characterized by: In step 2, the TBM operation data includes thrust, torque, propulsion speed, rotation speed and penetration; the cutter disc cutter wear data includes the radial wear data of the cutter disc at each cutter position on the cutter disc; and the cutter disc vibration data includes the vibration acceleration data at the cutter disc monitoring position.

4. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1 is characterized by: In step 3, the spatiotemporal alignment method of TBM multimodal data includes the following steps: Step 3-1, sorting between modalities: sorting the n modal data blocks after data segmentation from high to low according to the sampling frequency, namely: the first modal data block, the second modal data block, ..., the nth modal data block; Step 3-2, intra-modal sorting: sort each modal data block in the order of monitoring time or spatial stake position; Step 3-3, second modal data block alignment: according to the time and spatial stake position, taking the time information of the first modal data block as the reference, through the time index, insert the data at each acquisition frequency in the second modal data block into the first modal data block with the same time accuracy, so as to realize the data alignment and fusion of the second modal data block and the first modal data block; Step 3-4: Repeat step 3-3 until the data alignment and fusion of the nth modal data block and the first modal data block are completed.

5. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 4 is characterized by: When filling the data in step 4, it is necessary to first determine the degree type of null values ​​and repeated values ​​of the modal data; for modal data blocks with a small degree of null values ​​and repeated values, the KNN nearest neighbor filling method is used to repair the null values ​​and repeated values; for modal data blocks with a large degree of null values ​​and repeated values, the nearest neighbor linear interpolation method is used to repair the null values ​​and repeated values.

6. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 5 is characterized by: In step 4, the specific method of data filling includes the following steps: Step 4-1, identifying null values ​​and duplicate values: identifying null values ​​and duplicate values ​​in the second modality data block after alignment and fusion processing; Step 4-2, determining the degree of null values ​​and duplicate values: comparing the number of null values ​​and duplicate values ​​of each modality in each day with the set ratio value, when the number of null values ​​and duplicate values ​​in the i-th day is less than the set ratio value, all null values ​​and duplicate values ​​in the i-th day are judged as null values ​​or small degree of duplicate values, otherwise, they are judged as null values ​​or large degree of duplicate values; Step 4-3, repairing null values ​​and duplicate values: when the null value or duplicate value degree in the second modal data block determined in step 4-2 is small, the KNN nearest neighbor filling method is used for repair; when the null value or duplicate value degree in the second modal data block determined in step 4-2 is large, the nearest neighbor linear interpolation method is used for repair; Step 4-4, repeat step 4-1 to step 4-3, and complete the repair of empty values ​​and duplicate values ​​of the third modal data block to the nth modal data block in sequence.

7. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1 is characterized by: In step 5, feature extraction is performed on different modal data in the same excavation section. For the machine operation mode, statistical characteristics and linear fitting methods are used to extract the mean and rock excavation indicators; for the wear mode, the radial wear velocity value of the roller cutter at different cutter positions is calculated as the characteristic parameter; for the cutterhead vibration mode, the statistical characteristic value of the vibration acceleration is extracted as the characteristic parameter.

8. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1 is characterized by: In step 7, the deep learning model is an evaluation model based on a gradient-based boosting ensemble learning algorithm.

9. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1, characterized in that: In step 1, N=5, which are stable, basically stable, partially stable, unstable and extremely unstable.

10. The tunnel surrounding rock stability diagnosis method based on TBM multimodal data fusion according to claim 1, characterized in that: In step 4-2, the proportion value is set to 5% or 10% of the total number of corresponding modes in the corresponding day.

Images