Rock-soil body instability early warning method and system based on physical constraint neural network

Abstract

The application discloses a geotechnical body instability early warning method and system based on a physical constraint neural network, relates to the technical fields of foundation pit geotechnical body monitoring data processing and neural network prediction, and is executed by a computer and comprises the following steps: S1, determining the support number and support activation time of each support to divide a construction stage according to adjacent time points; S2, screening physical connections and writing the reference stiffness of each physical connection; S3, constructing a monitoring response window and generating a standardized response increment and a physical dimension response increment; S4, adopting a gated recurrent graph neural network training algorithm to establish a stage connection prediction model; and S5, calculating node balance error and support axial force consistency error and generating an effective dynamic connection record. Through the method provided in the application, the connection relationship between monitoring points can be changed along with the layered excavation and support activation process, and the dynamic connection weight entering the risk early warning judgment corresponds to the force transmission path of the construction stage.

Description

Technical Field

[0001] This application relates to the fields of monitoring data processing and neural network prediction technology for foundation pit soil and rock masses, and in particular to an early warning method and system for soil and rock instability based on a physically constrained neural network. Background Technology

[0002] With the expansion of foundation pit engineering scale and the development of digital monitoring devices, data on retaining structure displacement, surface settlement, deep horizontal displacement, and support axial force can be continuously collected during foundation pit construction. Computers use neural networks to learn the relationship between monitoring data and time changes, and determine whether the monitored quantities have reached engineering control limits based on prediction results. This shifts foundation pit risk assessment from alarm processing after measured values ​​exceed limits to predicting future monitoring responses.

[0003] Existing neural network prediction methods use historical monitoring data and data from adjacent monitoring points as model inputs, learning the mapping relationship between monitoring quantities and future deformations through the neural network. Existing graph neural network methods further treat the monitored objects as graph nodes and the spatial or construction relationships between the monitored objects as graph connections, aggregating features of adjacent nodes through graph connections, thereby utilizing the spatial correlation of monitoring data in the prediction process.

[0004] For example, the invention patent with announcement number CN115238850B discloses a method for predicting slope displacement in mountainous areas based on MI-GRA and improved PSO-LSTM, including: (1) collecting and constructing original data for slope displacement prediction; (2) establishing a MI-GRA slope displacement feature selection model based on the constructed original data for slope displacement prediction; (3) using the data after feature selection as the input of the optimal feature set for slope displacement prediction to establish an improved PSO-LSTM slope displacement prediction model; and (4) performing model prediction and testing on the established slope prediction model.

[0005] For example, the invention patent with announcement number CN109711530B discloses a landslide prediction method and system, which includes: fusing the time series and spatial series of landslide body sensors to form a sensor network spatiotemporal sequence; inputting the time series into a long short-term memory network and inputting the network spatiotemporal sequence into a radial basis function neural network; and weighting and fusing the outputs of the two neural networks to obtain landslide prediction data for future times.

[0006] In existing technologies, existing systems establish connection relationships based on the spatial distance of monitoring points, initial layout relationships, pre-selected adjacent monitoring points, or correlation of monitoring data. This has the following shortcomings: After excavation is completed, soil connections passing through the excavated void still participate in the monitoring response propagation; when the support is installed but the support axial force has not yet been generated, the support connection is easily written into the connection relationship in advance; after the support generates axial force, the newly formed support connection cannot be added to the connection relationship according to the support activation time; when the construction stage changes, all connection weights are used as a whole or recalculated as a whole, and it is impossible to distinguish between connections that disappear due to excavation, connections that are added due to support activation, and connections that have not undergone structural changes. Summary of the Invention

[0007] This application provides a method and system for early warning of soil and rock instability based on physical constraint neural networks. It can update the physical connection relationship between monitoring points as the layered excavation and support activation process progresses, reducing the propagation of monitoring response along the excavated void or the corresponding connection of the inactive support. At the same time, it makes the dynamic connection weights entering the risk warning judgment correspond to the support axial force monitoring results.

[0008] This application provides an early warning method for soil and rock instability based on a physical constraint neural network, comprising: S1, determining the support number and activation time of each support, forming a construction stage transition timetable, dividing the construction stages by adjacent times, and outputting the construction stage sequence; S2, filtering physical connections by combining the list of excavated voids and activated support numbers for each construction stage, and writing the reference stiffness of each physical connection; S3, constructing a monitoring response window and generating standardized response increments and physical dimension response increments; S4, using a gated cyclic graph neural network training algorithm to establish a stage connection prediction model, and obtaining dynamic connection weights to form connection identifiers, forming a construction stage connection update record; S5, calculating the node balance error and the consistency error of the support axial force, generating an effective dynamic connection record, calculating the predicted over-limit amplitude, determining the earliest predicted over-limit time, and outputting the monitoring quantity over-limit risk warning and abnormal response transmission path.

[0009] Another embodiment of this application provides an early warning system for soil and rock instability based on a physically constrained neural network, comprising: a construction stage determination module, used to determine the support number and activation time of each support, form a construction stage transition timetable, divide the construction stages by adjacent times, and output the construction stage sequence; a physical connection generation module, used to filter physical connections by combining the list of excavated voids and activated support numbers of each construction stage, and write the reference stiffness of each physical connection; a monitoring response window construction module, used to construct a monitoring response window and generate standardized response increments and physical dimension response increments; a dynamic connection weight update module, used to establish a stage connection prediction model using a gated cyclic graph neural network training algorithm, and obtain dynamic connection weights to form connection identifiers and form a construction stage connection update record; and a risk warning output module, used to calculate the node balance error and the support axial force consistency error, generate effective dynamic connection records, calculate the predicted over-limit amplitude, determine the earliest predicted over-limit time, and output the monitoring quantity over-limit risk warning and abnormal response transmission path.

[0010] The beneficial effects achieved by the technical solutions provided in this application include at least the following: (1) In this invention, by writing the completion time of layered excavation and the activation time of support into the construction stage transition time table, and generating a physical connection record of the construction stage based on the excavated empty area of ​​the stage, the remaining soil of the stage, and the list of activated support numbers of the stage, the connection relationship between monitoring points corresponds to the force transmission range of the corresponding construction stage.

[0011] (2) In this invention, the support activation time is determined by comparing the support design connection point number and the support connection point installation acceptance record item by item, and by comparing the B adjacent axial force sampling values ​​after the installation end time with the support activation threshold. This allows the support connection to be written into the physical connection record of the construction stage after the support generates axial force, thereby reducing the situation where the support connection is established in advance before the support force transmission relationship is formed.

[0012] (3) In this invention, by inputting the timing response characteristics of the monitoring point, the unique thermal vector of the connection type and the stiffness ratio into the connection weight generation layer, the dynamic connection weight of each physical connection is obtained, and the incremental force transmitted by the support connection is checked by the measured support axial force increment, so that the dynamic connection weight entering the risk warning judgment of the monitoring quantity exceeding the limit corresponds to the monitoring response change, reference stiffness and support axial force monitoring results.

[0013] (4) In this invention, by comparing the physical connection records of adjacent construction stages, the weights of deleted connections, newly added connections, endpoint associated connections and inherited connections are cleared, re-obtained or reused respectively. When the same monitoring point and the same monitoring quantity are predicted to exceed the limit in adjacent C monitoring response windows, the monitoring quantity exceedance risk warning and abnormal response transmission path are output, reducing the situation where a warning is directly issued when a single monitoring response window is predicted to exceed the limit.

[0014] It should be understood that the content described in this section is not intended to identify key or important features of the embodiments of this application, nor is it intended to limit the scope of this application. Attached Figure Description

[0015] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of an early warning method for soil and rock instability based on a physical constraint neural network, provided in Embodiment 1 of this application. Figure 2 This is a structural diagram of an early warning system for soil and rock instability based on a physical constraint neural network, provided in Embodiment 2 of this application. Figure 3 It is based on the construction stage division diagram provided in Embodiment 1 of this application; Figure 4 It is a physical connection plan of the construction stage according to Embodiment 1 of this application; Figure 5 This is a schematic diagram of dynamic connection weight distribution provided in Embodiment 1 of this application. Detailed Implementation

[0017] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0018] The following is a brief explanation of the relevant technical terms used in this application.

[0019] Please see Figures 1-5Embodiment 1 of this invention provides an early warning method for soil and rock instability based on a physical constraint neural network, comprising the following steps: S1, determining the support number and activation time of each support, forming a construction stage transition timetable, dividing the construction stages by adjacent times, and outputting the construction stage sequence; S2, filtering physical connections by combining the list of excavated voids and activated support numbers for each construction stage, and writing the reference stiffness of each physical connection; S3, constructing a monitoring response window, and generating standardized response increments and physical dimension response increments; S4, using a gated cyclic graph neural network training algorithm to establish a stage connection prediction model, and obtaining dynamic connection weights to form connection identifiers, forming a construction stage connection update record; S5, calculating the node balance error and the support axial force consistency error, generating effective dynamic connection records, calculating the predicted over-limit amplitude, determining the earliest predicted over-limit time, and outputting the monitoring quantity over-limit risk warning and abnormal response transmission path.

[0020] Specifically, determining the support number and activation time for each support includes: The system receives the following data: layered excavation completion record, initial ground elevation, initial soil area, support design connection point list, support connection point installation and acceptance record, support axial force monitoring sequence, sensor minimum resolution, axial force determination sampling quantity B, support activation threshold, sampling interval, and axial force sensor range. The layered excavation completion record includes the excavation completion time, excavation bottom elevation, and excavation space area for that layer. The support connection point installation and acceptance record includes the support number, connection point number, and acceptance time. The axial force determination sampling quantity B, support activation threshold, and sampling interval are taken from the monitoring plan, and the axial force sensor range is taken from the sensor calibration record. B is a positive integer greater than or equal to 3. The monitoring plan includes the support stress confirmation time and axial force sampling interval. The computer divides the support stress confirmation time by the axial force sampling interval and rounds the result up to obtain the axial force determination sampling quantity B. If the result is less than 3, B is written as 3.

[0021] For the same support structure, the design connection point numbers in the support design connection point list are compared item by item with the connection point numbers in the installation and acceptance records. If each design connection point has an installation and acceptance record with the same number, the latest acceptance time in the installation and acceptance record is recorded as the support installation completion time. If any design connection point lacks a corresponding installation and acceptance record, no support installation completion time is generated.

[0022] Read B adjacent axial force sampling values ​​after the end of support installation. If the B axial force sampling values ​​are not empty and are within the range of the axial force sensor, the difference between adjacent sampling times is not greater than the sampling interval, and the B axial force sampling values ​​are all greater than or equal to the support activation threshold, the Bth axial force sampling time is recorded as the support activation time. If the end of support installation does not exist, or if there are no B adjacent axial force sampling values ​​that meet the judgment rules in the support axial force monitoring sequence, no support activation time is generated.

[0023] (i) Determination of support number: The design connection point numbers in the support design connection point list of the same support are compared item by item with the connection point numbers in the installation and acceptance records. When there are installation and acceptance records with the same number for each design connection point, the support number of that support is determined to be valid, and the latest acceptance time in the installation and acceptance records is recorded as the support installation end time. When any design connection point lacks a corresponding installation and acceptance record, the support number of that support is determined to be invalid, and no support installation end time is generated.

[0024] (II) Determination of Support Activation Time: For each support with a valid support number, B adjacent axial force sampling values ​​are read after the end of support installation to determine activation. (1) Validity check of axial force sampling values ​​- All B axial force sampling values ​​are not empty and are all within the range of the axial force sensor; (2) Sampling continuity check - The difference between any two adjacent sampling times in the B axial force sampling values ​​is not greater than the sampling interval; (3) Activation condition determination - When all B axial force sampling values ​​are greater than or equal to the support activation threshold, the support activation threshold is the lower limit of axial force recorded according to the support type in the monitoring plan, and the unit is the same as the support axial force sampling value; the computer reads a support activation threshold according to the support type corresponding to the support number, and compares the axial force sampling value with the support activation threshold, and records the Bth axial force sampling time as the support activation time; (4) Handling invalid situations: If the end time of support installation does not exist, or if there are no B adjacent axial force sampling values ​​in the support axial force monitoring sequence that meet the judgment rules in items (1) to (3) above, the support activation time will not be generated; the support number and support activation time will be output.

[0025] In this implementation plan, the support design connection point numbers are checked item by item against the installation and acceptance records. The support activation time is confirmed by B consecutive valid, time-continuous axial force sampling values ​​that reach the support activation threshold. This establishes a continuous judgment relationship between support number confirmation, installation completion confirmation, and actual stress confirmation, reducing false support activation caused by missing acceptance records, single axial force fluctuations, sampling interruptions, or abnormal readings. It also ensures that subsequent construction stage divisions and support connection establishment adopt time nodes corresponding to the actual stress process of the support.

[0026] Specifically, the process of forming a construction phase transition timetable, dividing construction phases by adjacent times, and outputting a construction phase sequence includes: The system receives the start and end times of the current analysis as specified by the user. The start time of the current analysis is no earlier than the earliest time of the monitoring data record, and the end time of the current analysis is no later than the latest time of the monitoring data record. If the user does not specify, the earliest time of the monitoring data record will be used as the start time of the current analysis, and the latest time of the monitoring data record will be used as the end time of the current analysis.

[0027] The data is compiled by retrieving the completion time of layered excavation and the activation time of support between the start and end times of this analysis. These times are then sorted in ascending order and duplicate times are removed to create a construction stage transition timetable. The times in this timetable serve as the boundary points between construction stages. Construction stages are divided by adjacent times, with the preceding time being the stage start time and the following time being the stage end time. Stage intervals follow a left-closed, right-open rule: the stage start time belongs to the current stage, while the stage end time does not belong to the current stage and is assigned to the next stage. The start time of this analysis is the first time in the construction stage transition timetable, and the end time of this analysis is the last time in the timetable. Data between the start time of this analysis and the first completion time of layered excavation or the activation time of support belongs to the first construction stage, and data between the last construction event time and the end time of this analysis belongs to the last construction stage.

[0028] For each construction stage, all layered excavation completion records with an excavation completion time no later than the stage start time are read. If multiple records exist, the minimum excavation bottom elevation among the records is taken as the stage excavation bottom elevation, and the excavation space ranges in each record are geometrically unioned to obtain the stage excavated empty area. If no layered excavation completion records exist, the stage excavation bottom elevation is taken as the project's initial ground elevation, and the stage excavated empty area is an empty set. The stage excavated empty area is subtracted from the project's initial soil range to obtain the stage remaining soil. The support numbers with an activation time no later than the stage start time are written into the stage activated support number list, forming a construction stage sequence. Each construction stage includes the stage start time, stage end time, stage excavation bottom elevation, stage excavated empty area, stage remaining soil, and stage activated support number list.

[0029] like Figure 3As shown, the construction stage division diagram illustrates the correspondence between the completion time of layered excavation, the activation time of support, and the construction stages. The horizontal axis represents construction time, the blue bars represent the duration of each layered excavation, the green bars represent the effective range after each support is activated, and the red dashed lines represent the construction stage transition times. The computer arranges the completion time of layered excavation and the activation time of support in ascending order of time, and divides the construction stages by the time interval between two adjacent construction stage transition times. In the diagram, stages 1 to 6 correspond to different excavated empty areas and activated support number lists for each stage, respectively. Subsequent physical connection filtering and dynamic connection weight calculation are performed according to the corresponding construction stages. Figure 3 This does not indicate the specific project duration, number of excavation layers, or number of supports. The actual time, event type, and number of events are read from the layer excavation completion record and the support activation time.

[0030] In this implementation plan, the completion time of layered excavation and the activation time of support are used as the dividing points of construction stages. Each construction stage corresponds to a determined list of excavation bottom elevation, excavated voids, remaining soil, and activated support numbers. This avoids the same monitoring data belonging to two construction stages at the same time and ensures that subsequent physical connection screening, dynamic connection weight calculation, and monitoring response prediction have consistent construction time boundaries.

[0031] Specifically, the process of selecting physical connections by combining the list of excavated voids and activated support numbers from each construction stage, and writing the reference stiffness of each physical connection, includes: The system receives the construction phase sequence, monitoring point layout records, and design calculation models. The monitoring point layout records include the monitoring point number, three-dimensional coordinates of the monitoring point, monitoring quantity type, monitoring direction unit vector, and monitoring object type. The design calculation models include adjacent records of retaining structure segments, common boundary records of soil units, soil-retaining structure contact surface records, support connection point records, geometric range, and reference stiffness of each calculation object.

[0032] The monitoring point layout record also includes the monitoring object number, which can be the retaining structure segment number, soil unit number, or support connection point number. The computer matches the monitoring points with the corresponding calculation objects in the design calculation model based on the monitoring object type and monitoring object number. Adjacent retaining structure segment records include two retaining structure segment numbers; common boundary records for soil units include two soil unit numbers; soil-retaining structure contact surface records include soil unit numbers, retaining structure segment numbers, and the geometric range of the contact surface; and support connection point records include the support number and two support connection point numbers.

[0033] When monitoring point numbers are used as graph nodes, and the retaining structure segment numbers corresponding to two monitoring points are both written into the adjacent records of the same retaining structure segment in the design calculation model, a retaining structure connection is established between the two monitoring points. When the soil element numbers corresponding to two monitoring points are both written into the common boundary record of the same soil element in the design calculation model, a candidate soil connection is established between the two monitoring points. When a soil element corresponding to one monitoring point and a retaining structure segment corresponding to another monitoring point both exist in the same soil-retaining structure contact surface record, a candidate soil-retaining structure connection is established. When two monitoring points correspond to two support connection points of the same support, and the support number is within the list of activated support numbers for the stage, a support connection is established. The computer compares the monitoring object number with the calculation object number in each record to determine whether two monitoring points correspond to the same record. The same monitoring point does not establish a connection with itself; when there are duplicate connections of the same connection type between two monitoring points, only one connection is retained.

[0034] Spatially compare the internal regions of the two soil units corresponding to the two ends of the candidate soil connection, the corresponding contact surface of the candidate soil-retaining structure connection, and the excavated void area of ​​the stage. If there is an intersection of internal regions, the connection is deleted; if only the boundary is connected, the connection is retained. Existing support connections are directly retained. The corresponding reference stiffness is read from the design calculation model according to the connection type, and connections with empty or non-zero reference stiffness are deleted. The two monitoring points at both ends of the physical connection are numbered in ascending order of value. Taking the first monitoring point as the starting point and the second monitoring point as the ending point, the coordinate difference vector is obtained by subtracting the three-dimensional coordinates of the first monitoring point from the three-dimensional coordinates of the second monitoring point. When the length of the coordinate difference vector is equal to zero, the corresponding physical connection is deleted. When the length of the coordinate difference vector is greater than zero, the coordinate difference vector is divided by the length of the coordinate difference vector to obtain the unit vector of the connection direction, forming the physical connection sequence of the construction stage. If the intersection volume of any soil unit corresponding to the candidate soil connection and the excavated void area of ​​the stage is greater than zero, it is determined to be an internal region intersection; if the intersection area of ​​the corresponding contact surface of the candidate soil-retaining structure connection and the excavated void area of ​​the stage is greater than zero, it is determined to be an internal region intersection. For retaining structure connections, the reference stiffness of the retaining structure segment along the connection direction is read. For candidate soil connections, the reference stiffness in the normal direction of the common boundary of the two soil elements is read. For candidate soil-retaining structure connections, the reference stiffness in the normal direction of the contact surface is read. For support connections, the axial reference stiffness of the corresponding support is read. The connection identifier consists of two monitoring point numbers arranged in ascending numerical order followed by the connection type. Dividing the coordinate difference by the vector length of the coordinate difference yields the unit vector of the connection direction. The physical connection sequence during the construction phase includes the start time of the construction phase, the first monitoring point number, the second monitoring point number, the connection type, the reference stiffness, and the unit vector of the connection direction.

[0035] like Figure 4As shown, the physical connection plan diagram for the construction phase is used to represent the force transmission connection relationship between monitoring points within the same construction phase. The horizontal and vertical axes represent the X and Y coordinates of the monitoring points in the engineering plane coordinate system, respectively. Blue dots represent monitoring points of the retaining structure; orange dots represent monitoring points of the soil; green dots represent monitoring points of the support; and purple square dots represent monitoring points of the support axial force. Black solid lines represent retaining structure connections, used to connect monitoring points corresponding to adjacent structural segments within the same retaining structure; gray dashed lines represent soil connections, used to connect monitoring points corresponding to soil units with a common boundary; gray dotted lines represent soil-retaining structure connections, or simply soil-retaining connections, used to connect soil units with contact surfaces to monitoring points corresponding to retaining structure segments; red solid lines represent support connections, established only after the support number list has been activated during the corresponding support number writing phase. The various connections in the diagram constitute the physical connection relationship for the corresponding construction phase. Soil connections that overlap with the excavated voids in the stage and soil-retaining structure connections are not included in the drawing. Support connections only correspond to activated supports, so that the connection relationships in the drawing correspond to the remaining soil range and support stress relationship of the construction stage.

[0036] In this implementation plan, monitoring points are mapped to retaining structure segments, soil units, contact surfaces, and support connection points in the design calculation model by the monitoring object number. Soil connections and soil-retaining structure connections that have lost their soil force transmission conditions are deleted based on the excavated voids of each stage. Support connections are written according to the list of activated support numbers for each stage. This ensures that the physical connection sequence of each construction stage corresponds to the remaining soil range of the stage, the adjacent relationship of the retaining structure, and the force relationship of the support. It also provides connection direction and reference stiffness for dynamic connection weight calculation, connection force increment calculation, and abnormal response transmission path backtracking.

[0037] Specifically, the construction of the monitoring response window and the generation of standardized response increments and physical dimension response increments include: For each construction phase, the common sampling times of all activated monitoring data from all activated monitoring points between the start and end times of the phase are extracted, and the last W sampling times are selected to construct a monitoring response window; if there are fewer than W common sampling times, no window is generated; the monitoring data includes monitoring point number, monitoring quantity type, sampling time, and monitoring value; the length W of the monitoring response window is taken from the monitoring scheme, and the minimum sensor resolution is taken from the sensor calibration record. All activated monitoring points mentioned in this section refer to all monitoring points that serve as connection endpoints in the physical connection sequence of the corresponding construction phase; the data set participating in the intersection of sampling times is the monitoring sequence composed of each monitoring point number and monitoring quantity type, and the axial force monitoring sequence corresponding to each support in the list of activated support numbers for the phase. The monitoring data also includes sensor numbers, and the sensor calibration record includes the sensor number and the corresponding minimum sensor resolution. The computer uses the unified sampling time specified in the monitoring plan as the reference time, and assigns each monitoring data point to the nearest reference time whose absolute time difference is no greater than half of the unified sampling interval. When two monitoring data points in the same monitoring sequence correspond to the same reference time, the monitoring data with the smaller absolute time difference from the reference time is retained; if the absolute time differences are the same, the monitoring data with the earlier sampling time is retained. If any monitoring sequence participating in the intersection calculation lacks a monitoring value at a certain reference time, or if the monitoring value is empty or not a numerical value, that reference time is deleted from the shared sampling times. The monitoring response window length W is a positive integer not less than 2. The monitoring plan also records the window coverage duration and the unified sampling interval. The window coverage duration is divided by the unified sampling interval, rounded up, and then 1 is added to obtain W. If the obtained value is less than 2, W is set to 2.

[0038] Subtracting the corresponding monitoring value at the first sampling time from each monitoring value within the window yields the physical dimension response increment. Dividing this increment by the minimum resolution of the sensor corresponding to the monitoring value yields the standardized response increment. The physical dimension response increment and standardized response increment of the axial force of the activated support are then calculated. The standardized response increment is used as input to the neural network, while the physical dimension response increment is used to connect the transmission force increment and the calculation of the consistency error of the support axial force. Based on the sensor number in the monitoring data, the computer reads the minimum resolution of the sensor corresponding to the monitoring value from the sensor calibration record, ensuring that each monitoring value corresponds to a single minimum resolution. If the minimum resolution of the sensor is empty, equal to zero, or less than zero, the corresponding monitoring sequence is recorded as an invalid monitoring sequence, and the monitoring response window containing that sequence is deleted. If any monitoring value within the window is empty or not a numerical value, the corresponding reference time is deleted, and the last W common sampling times are reselected. If the number of common sampling times after reselection is less than W, no monitoring response window is generated. The physical dimension response increment and standardized response increment of the axial force of the activated support are read according to the sensor number in the support axial force monitoring data, and the minimum resolution of the axial force sensor is read, and the missing value handling and time alignment rules are the same as those of the monitoring point monitoring value.

[0039] In this implementation scheme, time alignment is performed on data from different monitoring points, different monitoring quantities, and support axial force by unifying the sampling time, and missing values, non-numerical values, and invalid data with minimum resolution are deleted, so that each monitoring sequence in the monitoring response window has the same sampling time; the physical dimension response increment is calculated based on the first sampling time of the window, and a standardized response increment is generated according to the minimum resolution of the corresponding sensor, so that the neural network input adopts a unified scale, while retaining the physical dimension data required for calculating the consistency error of the connection transmission force increment and the support axial force.

[0040] Specifically, the step of using a gated cyclic graph neural network training algorithm to establish a stage connection prediction model and to obtain dynamic connection weights includes: A stage connectivity prediction model is established using a gated recurrent graph neural network training algorithm. The model comprises a temporal coding layer, a connectivity weight generation layer, a physical connectivity weighted propagation layer, and a response prediction layer. The temporal coding layer employs a gated recurrent unit, the connectivity weight generation layer uses a two-layer fully connected network, the physical connectivity weighted propagation layer performs multiplication and summation, and the response prediction layer uses a fully connected network. The gated recurrent unit processes the standardized response increments of the same monitoring point according to the sampling time sequence. The connectivity weight generation layer generates a dynamic connectivity weight based on the temporal response characteristics, connectivity type, and stiffness ratio of the monitoring points at both ends of the physical connection. The physical connectivity weighted propagation layer aggregates the temporal response characteristics of adjacent monitoring points according to the dynamic connectivity weight and stiffness ratio. The response prediction layer outputs the standardized response increments of each monitoring point over the next H sampling times.

[0041] For each construction stage, the standardized response increments at W sampling times within the monitoring response window are input into a gated loop unit, outputting the temporal response features of the monitoring points. For each physical connection in a construction stage, the reference stiffness is divided by the maximum reference stiffness of the same connection type within the same construction stage to obtain the stiffness ratio. Connection types are converted into four-dimensional one-hot vectors in the order of retaining structure connection, soil connection, soil-retaining structure connection, and support connection. The temporal response features of the monitoring points at both ends of the connection, the one-hot vector of the connection type, and the stiffness ratio are concatenated and input into the connection weight generation layer. The first fully connected network uses the ReLU function, and the second fully connected network uses the Sigmoid function to output dynamic connection weights between zero and one. The temporal response features of the monitoring points at both ends of the connection are concatenated in the order of the first and second monitoring points in the connection identifier. The concatenation result includes the temporal response features of the first monitoring point, the temporal response features of the second monitoring point, the four-dimensional one-hot vector, and the stiffness ratio. Each physical connection corresponds to only one dynamic connection weight within a monitoring response window.

[0042] For each monitoring point, the temporal response features of adjacent monitoring points are weighted and summed according to dynamic connection weights and stiffness ratios to obtain physical connection propagation features. The temporal response features of the monitoring points are concatenated with the physical connection propagation features and input into the response prediction layer to obtain standardized response increments for the next H sampling times. These are then converted into physical dimension response increments based on the sensor's minimum resolution. The multilayer perceptron consists of two fully connected layers in the connection weight generation layer, and its input also includes a one-hot vector of connection type. The parameter H is a positive integer not less than 1. The monitoring scheme records the predicted coverage duration and sampling interval. The predicted coverage duration is divided by the sampling interval, and the result is rounded up to obtain H. If the result is less than 1, H is set to 1. The response prediction layer outputs H standardized response increments corresponding to each monitoring point number and monitoring quantity type. The computer determines the corresponding sensor number based on the monitoring point number and monitoring quantity type, reads the sensor's minimum resolution, and multiplies each standardized response increment by the corresponding sensor's minimum resolution to obtain the corresponding physical dimension response increment.

[0043] For each physical connection, the temporal response characteristics of adjacent monitoring points are multiplied by the dynamic connection weight and stiffness ratio to obtain the connection propagation vector. The absolute values ​​of each component of the connection propagation vector are then summed to obtain the connection propagation contribution value. This value is recorded in the connection propagation contribution record according to the start time of the construction phase, the end time of the monitoring response window, and the monitoring point number. For the two endpoints of a physical connection, the connection propagation vector and connection propagation contribution value from the monitoring point at the other end to the current monitoring point are calculated respectively. The connection identifier, the monitoring point number of the feature source, the monitoring point number of the feature receiving, and the connection propagation contribution value are then recorded in the connection propagation contribution record.

[0044] For physical connections where the unit vector in the connection direction is not empty, the physical dimension displacement increments at both ends of the connection are converted into displacement increment vectors. The physical dimension displacement increment at the first monitoring point is multiplied by the unit vector in the monitoring direction of the first monitoring point to obtain the first displacement increment vector; the physical dimension displacement increment at the second monitoring point is multiplied by the unit vector in the monitoring direction of the second monitoring point to obtain the second displacement increment vector; the projection of the difference between the two displacement increment vectors in the connection direction is calculated, and the projection result is multiplied by the reference stiffness and dynamic connection weight to obtain the connection transmission force increment; for support connections, the axial unit vector of the support is determined based on the three-dimensional coordinates of the two support connection points. The unit vector of the support connection direction is set as the unit vector of the support axis; the incremental connection force calculated according to the support connection direction is directly recorded as the predicted incremental support axial force; the monitoring response window is read from the construction stage that has been sampled; if there are H sampling times after the end of the monitoring response window, and all H sampling times are earlier than the end of the construction stage, the monitoring response window is written into the training sample, and the measured standardized response increments of the H sampling times are written as the prediction target; the predicted standardized response increments are subtracted from the measured standardized response increments one by one, and the average of the squared differences is calculated to calculate the prediction error.

[0045] The physical dimension displacement increment conversion is performed only on monitoring points where the monitored quantity type is displacement and the monitoring direction unit vector is not empty. The displacement includes horizontal displacement, vertical displacement, and displacement along a predetermined direction. The monitoring direction unit vector is taken from the monitoring point layout record and uses the same coordinate system as the physical dimension displacement increment. For a displacement-type monitoring point, the scalar form of the physical dimension displacement increment is multiplied by the three coordinate components of the corresponding monitoring direction unit vector to obtain the three-dimensional displacement increment vector. If the monitored quantity type at either end of a physical connection is not displacement, or if either end lacks a monitoring direction unit vector, the connection force increment of that physical connection is not calculated. The support axial direction is determined by the three-dimensional coordinates of the two support connection points in the support connection point record. When constructing training samples, if any monitoring point or any monitored quantity lacks a measured standardized response increment at any of the H predicted target sampling times, the corresponding monitoring response window is not written into the training samples. The prediction error is calculated by subtracting each predicted value from its corresponding measured value and calculating the square of the difference. The sum of all squared differences is then divided by the total number of predicted values ​​used in the calculation to obtain the prediction error.

[0046] The predicted increment of support axial force is subtracted from the increment of the physical dimension response of support axial force at the same support and sampling time. The difference is divided by the minimum resolution of the axial force sensor, and the average of the squares of the resulting values ​​is calculated to obtain the normalized support axial force training error. When there is no support connection in the physical connection sequence during the construction phase, the support axial force consistency error is set to zero. The calculation range of the support axial force consistency error is the support connections in all training samples and the next H sampling times. The computer matches each predicted increment of support axial force with the corresponding measured increment of the physical dimension response of support axial force according to the support number and sampling time, and averages the squares of all matching results. If the minimum resolution of any support axial force sensor is empty, equal to zero, or less than zero, the monitoring response window containing the corresponding support connection is not written into the training samples.

[0047] The connection weight timing error is obtained by averaging the squared differences of the corresponding dynamic connection weights between adjacent monitoring response windows within the same construction phase. When only one monitoring response window exists within a construction phase, the connection weight timing error is zero. When two monitoring response windows belong to different construction phases, no connection weight timing error is calculated. The corresponding dynamic connection weight refers to the dynamic connection weight of physical connections with the same connection identifier under two adjacent monitoring response windows. The computer calculates the squared differences of the dynamic connection weights corresponding to each identical connection identifier, and divides the sum of all squared differences by the number of identical connection identifiers to obtain the connection weight timing error. When two adjacent monitoring response windows do not have identical connection identifiers, the connection weight timing error is zero.

[0048] The training error is obtained by summing the prediction error, the support axis force consistency error, and the connection weight temporal error. The Adam algorithm and backpropagation are then used to update the network parameters until the training error change is less than or equal to the training termination threshold, or the number of training epochs reaches the training epoch threshold. No parameters are set to be updated in the physical connection weighted propagation layer. Adam is an abbreviation for Adaptive Moment Estimation. The Adam algorithm determines the update amount of each network parameter based on the first-order and second-order moment estimates of the gradient obtained from the training error with respect to the network parameters. Backpropagation is used to calculate the gradient of the training error relative to each network parameter in the gated recurrent unit, the connection weight generation layer, and the response prediction layer. The Adam algorithm updates the corresponding network parameters based on these gradients.

[0049] After training, the monitoring response window with the latest end time and its corresponding construction stage physical connection record are input into the stage connectivity prediction model. The model outputs the dynamic connection weights of each physical connection, the standardized response increments and physical dimension response increments for the next H sampling times, and the connection propagation contribution record. The monitoring response window with the latest end time refers to the last monitoring response window in chronological order among all monitoring response windows whose end time is no later than the calculation time of this prediction within the construction stage to which this prediction is being performed. The computer reads the construction stage physical connection sequence of the construction stage to which this monitoring response window belongs and inputs it into the stage connectivity prediction model along with this monitoring response window. The construction stage to which this prediction is being performed refers to a construction stage whose calculation time is no earlier than the stage start time and earlier than the stage end time.

[0050] In this implementation scheme, the temporal response characteristics of each monitoring point are extracted through a gated loop unit, and dynamic connection weights are generated by combining the temporal response characteristics of both ends of the connection, the connection type, and the stiffness ratio. This ensures that the monitoring response is weighted and propagated only along the physical connection of the corresponding construction stage. The model training is constrained by prediction error, support axial force consistency error, and connection weight temporal error. This ensures that the model output simultaneously corresponds to the temporal continuity of monitoring response changes, support force relationship, and connection weights. It also provides dynamic connection weights, future monitoring responses, and connection propagation contribution values ​​for subsequent physical consistency checks, prediction limit judgments, and abnormal response propagation path backtracking.

[0051] Specifically, the components of the connection identifier, forming the connection update record during the construction phase, include: The two monitoring point numbers are arranged numerically and, together with the connection type, form a connection identifier. Connection identifiers from adjacent construction stages are compared; connections existing only in the previous construction stage are marked as deleted connections with a weight of zero. Deleted connections include soil connections and soil-retaining structure connections that were not retained due to the screening of excavated empty areas in the stage. The monitoring point number with the smaller value is designated as the first monitoring point number, and the monitoring point number with the larger value is designated as the second monitoring point number. The connection identifier is composed of the first monitoring point number, the second monitoring point number, and the connection type in sequence. When the same connection identifier exists within the same construction stage, only one physical connection is retained. The dynamic connection weight of the deleted connection is set to zero, and the corresponding connection identifier is written into the construction stage connection update record.

[0052] Connections existing only in the next construction stage are designated as new connections; connections existing in two construction stages and sharing monitoring points with deleted or new connections are designated as endpoint-associated connections. New connections and endpoint-associated connections, along with the monitoring response window of the next construction stage, are input into the stage connection prediction model. The temporal response characteristics of the monitoring points at both ends are extracted by the temporal coding layer, and then combined with the stiffness ratio. The dynamic connection weights are then recalculated through the connection weight generation layer, without using the dynamic connection weights from the previous construction stage. New connections include support connections established based on the list of activated support numbers in the next construction stage. Shared monitoring points refer to at least one identical number among the first or second monitoring point numbers of two physical connections. The monitoring response window of the next construction stage is the earliest monitoring response window that ends within the next construction stage; if a monitoring response window has not yet been formed in the next construction stage, the dynamic connection weights for new connections and endpoint-associated connections are not calculated. When recalculating the dynamic connection weights, the standardized response increments of the monitoring points at both ends of the connection, the one-hot vector of the connection type, and the stiffness ratio are input into the stage connection prediction model, and a dynamic connection weight is output for each new connection and endpoint-associated connection.

[0053] Connections existing simultaneously in both construction phases are denoted as inherited connections, and their dynamic connection weights from the end of the previous construction phase are retained. When the second monitoring response window is formed in the subsequent construction phase, inherited connections, new connections, and endpoint-related connections are input into the phase connection prediction model, and the dynamic connection weights are recalculated based on the second standardized monitoring response window. This generates a construction phase connection update record. Each construction phase connection update record includes the start time of the subsequent construction phase, the first monitoring point number, the second monitoring point number, the connection type, the connection update type, and the dynamic connection weight. The dynamic connection weight at the end of the previous construction phase refers to the dynamic connection weight corresponding to the monitoring response window with the latest window end time within the previous construction phase. If no monitoring response window is formed in the previous construction phase, the dynamic connection weights of inherited connections are recalculated after the first monitoring response window is formed in the subsequent construction phase. The second monitoring response window is the second monitoring response window within the subsequent construction phase, arranged in ascending order of window end time. Starting from the second monitoring response window, the dynamic connection weights of inherited connections, new connections, and endpoint-related connections are recalculated based on the corresponding monitoring response window. The connection update type is written as deleted connection, new connection, endpoint-related connection, or inherited connection.

[0054] like Figure 5As shown, the dynamic connection weight distribution diagram is used to represent the weight distribution of each physical connection within the same construction phase as the monitoring response window changes. The vertical axis of the matrix represents the connection identifier, with each row corresponding to a physical connection in the construction phase; the horizontal axis represents the monitoring response window number, with each column corresponding to a monitoring response window, or the end time of the corresponding monitoring response window can be used as the horizontal axis marker; each cell in the matrix represents the dynamic connection weight of the corresponding physical connection under the corresponding monitoring response window.

[0055] The color of each matrix cell corresponds to the dynamic connection weight value, with a color scale ranging from 0 to 1. A dynamic connection weight closer to 1 indicates a higher proportion of the corresponding physical connection participating in the monitoring response propagation at that time step; a dynamic connection weight closer to 0 indicates a lower proportion of the corresponding physical connection participating in the monitoring response propagation at that time step. By comparing the weight changes of the same row under different monitoring response windows, we can determine how the dynamic connection weight of the same physical connection changes with the window; by comparing the color differences of different rows under the same monitoring response window, we can determine the differences in dynamic connection weights between different physical connections.

[0056] In this implementation plan, by comparing the connection identifiers of adjacent construction stages, physical connections are classified into deleted connections, newly added connections, endpoint-associated connections, and inherited connections. The weights are then cleared, recalculated, or reused between stages, respectively. This ensures that the dynamic connection weights during construction stage transitions correspond to the range of connection increases and decreases, avoiding the simultaneous reuse or recalculation of all connection weights. It also provides the corresponding dynamic connection weights for the physical consistency check and monitoring quantity over-limit risk assessment of the next construction stage.

[0057] Specifically, the calculation of the node balance error and the consistency error of the support axial force, and the generation of effective dynamic connection records include: The incremental force transfer of adjacent physical connections at the same monitoring point is converted into a force vector in the same coordinate system using the unit vector of the connection direction. The force vectors are then summed, and the average of the squares of the resultant force vector lengths is used as the node balance error. If no incremental force transfer is generated for any adjacent physical connection, no node balance error is generated for the corresponding monitoring point. The node balance error is calculated for monitoring points not marked as constraint boundary nodes in the design calculation model. The unit vector of the connection direction points from the first monitoring point to the second monitoring point. When the current monitoring point is the first monitoring point, the incremental force transfer is multiplied by the unit vector of the connection direction to obtain the stress vector; when the current monitoring point is the second monitoring point, the resulting force vector is inverted. When the monitoring point corresponds to the support action location, the incremental force transfer of the support connection is included as the incremental support reaction force in the summation of the force vectors. If there is an incremental external force applied to the monitoring point in the design calculation model or construction load record, the incremental external force is converted to the same coordinate system and included in the summation. When the constraint boundary reaction force cannot be obtained from the design calculation model or monitoring data, the corresponding monitoring point is written into the boundary node exclusion record, and the node balance error for that monitoring point is not calculated. For each monitoring point, the square of the resultant force vector length is calculated at each sampling time in the monitoring response window. The squares of the resultant force vector lengths at W sampling times are added together and divided by W to obtain the node balance error of the monitoring point in the corresponding monitoring response window. The unit of the node balance error is the square of the force. The node balance errors of different monitoring points are not averaged again.

[0058] The incremental force transmitted by the support connection is projected onto the axial direction of the support and subtracted from the measured axial force increment of the same support at the same sampling time. The average of the squared differences is taken as the support axial force consistency error. If the balance error of all nodes in the subsequent construction stage is not greater than the node balance error threshold, and the support axial force consistency error of all support connections is not greater than the support axial force consistency error threshold, the dynamic connection weight of the subsequent construction stage is written into the valid dynamic connection record. If any node balance error is greater than the node balance error threshold, any support axial force consistency error is greater than the support axial force consistency error threshold, or any physical connection involved in the inspection does not generate a connection force incrementality, the dynamic connection weight of the subsequent construction stage is not written, and a record of insufficient physical consistency is output. For each support connection, the squared difference between the predicted and measured axial force increments is calculated at W sampling times within the monitoring response window. The sum of the W squared differences is divided by W to obtain the support axial force consistency error of the corresponding support connection. The unit of the support axial force consistency error is the square of force; the errors of different support connections are not averaged again. This item takes a pair of adjacent construction stages in the aforementioned construction stage connection update record as the inspection object. The preceding construction stage is the construction stage with the earlier start time among the adjacent construction stages, and the following construction stage is the construction stage that immediately follows the preceding construction stage. The dynamic connection weight of the following construction stage refers to the dynamic connection weight written into the aforementioned construction stage connection update record and corresponding to the monitoring response window of the following construction stage.

[0059] The physical consistency insufficiency record includes the start time of the construction phase, the end time of the monitoring response window, the error type exceeding the threshold, the error value, and the corresponding error threshold. The node balance error and support axial force physical consistency error corresponding to the model validation samples are read, and the two types of errors are arranged in ascending order of value. The model training parameter record includes the error quantile number. The computer reads the node balance error value and support axial force physical consistency error value corresponding to the error quantile number, and records the reading results as the node balance error threshold and support axial force physical consistency error threshold, respectively. The effective dynamic connection record includes the start time of the construction phase, the end time of the monitoring response window, the two monitoring point numbers, the connection type, the predicted value, and the dynamic connection weight. The previous stable construction phase refers to the previous construction phase in which no measured monitoring value exceeded the engineering control limit, and each monitoring response window generated complete node balance error and support axial force consistency error, and the phase is adjacent to the next construction phase in time. When calculating the 95th percentile value, the corresponding errors are arranged in ascending order of value, the number of errors is recorded as N, and the error value corresponding to 0.95N (rounded up) is read. Records of insufficient physical consistency also include the corresponding monitoring point number, support number or connection identifier; records of exclusion of boundary nodes include the start time of the construction phase, the end time of the monitoring response window, the monitoring point number and the reason for exclusion.

[0060] In this implementation scheme, the incremental force of each physical connection, the incremental force of the support reaction, and the incremental force of the external force are uniformly transformed to the same coordinate system. The node balance error of the unconstrained boundary monitoring point and the consistency error of the support axial force of the support connection are calculated separately. This allows the dynamic connection weight to be subject to dual checks of the node force balance relationship and the measured support axial force. Only when both types of errors do not exceed the corresponding thresholds are valid dynamic connection records generated. This eliminates dynamic connection weights with missing connection force, excessive node resultant force deviation, or inconsistent support axial force prediction and measured results. This provides connection data that has undergone physical consistency checks for subsequent prediction limit judgment and abnormal response transmission path backtracking.

[0061] Specifically, the calculation of the predicted over-limit amplitude, determination of the earliest predicted over-limit time, and output of the over-limit risk warning and abnormal response transmission path for the monitored quantity include: When the physical consistency check result is a valid dynamic connection record, the computer reads the prediction results for the next H sampling times output by the phase connection prediction model and writes them into the prediction result record according to the construction phase start time, monitoring response window end time, predicted sampling time, monitoring point number, monitoring quantity type, and predicted value. The prediction result record is associated with the construction phase connection update record and the valid dynamic connection record through the construction phase start time and monitoring response window end time. The predicted value in the construction phase connection update record refers to the predicted value in the prediction result record read through the above association relationship. The computer receives the engineering control limit record, which includes the monitoring point number, monitoring quantity type, lower control limit, upper control limit, limit effective time, and limit invalidation time. The engineering control limit is taken from the special monitoring plan and uses the same physical dimension as the predicted value. For each predicted value, the computer reads the engineering control limit with the same monitoring point number and monitoring quantity type, and whose predicted sampling time is not earlier than the limit effective time and earlier than the limit invalidation time. If there is no corresponding engineering control limit, or the lower control limit is greater than the upper control limit, no limit violation judgment is performed on the predicted value. The engineering control limit is the monitoring quantity criterion for triggering the risk management of soil and rock instability in the special monitoring plan. The risk warning of exceeding the monitoring quantity limit indicates that the predicted monitoring quantity will exceed the corresponding criterion, but does not mean that the soil and rock mass has become unstable.

[0062] When the predicted value in the effective dynamic connection record is less than the lower control limit, the predicted over-limit amplitude is obtained by subtracting the predicted value from the lower control limit; when the predicted value is greater than the upper control limit, the predicted over-limit amplitude is obtained by subtracting the upper control limit from the predicted value; when the predicted value is not less than the lower control limit and not greater than the upper control limit, the predicted over-limit amplitude is written as zero; in ascending order of the predicted sampling time, the sampling time at which the first predicted over-limit amplitude is greater than zero is written as the earliest predicted over-limit time; when there are two or more over-limit results at the earliest predicted over-limit time, they are processed according to the monitoring quantity type; for over-limit results of the same monitoring quantity type, the over-limit result with the largest predicted over-limit amplitude is recorded, and when the predicted over-limit amplitudes are the same, the over-limit result with the smallest monitoring point number is recorded; for over-limit results of different monitoring quantity types, the corresponding over-limit results are retained respectively, and the numerical values ​​of the predicted over-limit amplitudes are not compared.

[0063] If the same monitoring point number and the same monitoring quantity type have out-of-limit results in all C adjacent monitoring response windows, an out-of-limit risk warning will be output. If no corresponding predicted out-of-limit record exists in any monitoring response window, or if the physical consistency check module outputs a record of insufficient physical consistency, the number of out-of-limit records for the corresponding monitoring point number and monitoring quantity type will be set to zero.

[0064] For a given monitoring response window, among the H predicted sampling times corresponding to that window and all monitoring points and monitoring quantity types with corresponding engineering control limits, the earliest predicted limit violation time is retrieved in ascending order of predicted sampling times. Each limit violation result corresponds to a predicted sampling time, a monitoring point number, and a monitoring quantity type, and includes the start time of the construction phase, the end time of the monitoring response window, the predicted value, the lower control limit, the upper control limit, and the predicted limit violation amplitude. When the earliest predicted limit violation time has limit violation results for different monitoring quantity types, the limit violation results are retained according to the monitoring quantity type; the comparison of the predicted limit violation amplitude is only performed between limit violation results of the same monitoring quantity type. The parameter C is a positive integer not less than 2; the monitoring scheme records the warning confirmation time and the monitoring response window sliding time interval, divides the warning confirmation time by the monitoring response window sliding time interval, rounds up the result, and adds 1 to obtain C. If the result is less than 2, C is set to 2. The C adjacent monitoring response windows refer to the C monitoring response windows within the same construction phase, arranged in ascending order of window end times and with consecutive sequence numbers.

[0065] The monitoring point corresponding to the earliest predicted out-of-limit result is taken as the end of the path. The connection propagation contribution value corresponding to the monitoring response window that generated the current prediction result is read. From the valid dynamic connection records, the physical connection connected to the current monitoring point with the largest connection propagation contribution value is selected sequentially. The monitoring point at the other end of the physical connection is written into the anomaly response transmission path. The connection propagation contribution record includes the feature source monitoring point number, the feature receiving monitoring point number, the connection identifier, and the connection propagation contribution value. During path backtracking, only the connection propagation contribution record where the feature receiving monitoring point number is equal to the current monitoring point number is selected, and the corresponding feature source monitoring point is taken as the predecessor monitoring point of the current monitoring point. Backtracking is not performed along records outside the opposite direction from the feature receiving monitoring point to the feature source monitoring point. After the same monitoring point or the same connection identifier is written into the anomaly response transmission path, it is recorded as an accessed object, and subsequent backtracking will not select records containing accessed predecessor monitoring points or accessed connection identifiers. Path backtracking ends when the number of physical connections in the anomaly response transmission path reaches the number of different monitoring points in the valid dynamic connection records minus one. The last monitoring point written at the end of the backtracking is recorded as the response source monitoring point. When there are earliest predicted out-of-limit results for different monitoring quantity types, the monitoring point corresponding to each earliest predicted out-of-limit result is used as the end point of the path to perform path backtracking.

[0066] The system continues to read connected physical connections using newly written monitoring points. Physical connections to monitoring points already written into the abnormal response propagation path are deleted, and the physical connection with the largest connection propagation contribution value is selected. If no remaining adjacent physical connections exist, or if the connection propagation contribution values ​​of the remaining physical connections are all equal to zero, the path backtracking ends. If the connection propagation contribution values ​​are the same, the physical connection with the smallest monitoring point number at the other end is selected. The backtracked monitoring points and physical connections are arranged in reverse order to form an abnormal response propagation path. The system outputs the monitoring quantity exceeding the limit risk warning result. The monitoring quantity exceeding the limit risk warning result includes the warning release result, the earliest predicted exceeding time, the monitoring point number, the monitoring quantity type, the predicted exceeding amplitude, the start time of the construction phase, the effective dynamic connection record, the support force consistency check result, and the abnormal response propagation path. The warning release result includes two values: release and no release. A release is written when the judgment conditions of C adjacent monitoring response windows are met; a no release is written when the judgment conditions are not met. The support force consistency check result includes the support number, the support axial force consistency error, the support axial force consistency error threshold, and the comparison result. Different types of monitoring data generate separate early warning results for risks of exceeding monitoring limits.

[0067] In this implementation plan, the predicted values ​​output by the stage connection prediction model are compared with the engineering control limits within the corresponding monitoring point, monitoring quantity type, and effective time range to determine the predicted limit exceedance amplitude and the earliest predicted limit exceedance time. Continuous limit exceedance results in C adjacent monitoring response windows are used as the conditions for issuing early warnings. Simultaneously, based on the characteristic propagation direction in the connection propagation contribution record, the preceding monitoring point is traced back, and visited monitoring points and physical connections are excluded to form an abnormal response transmission path from the response source monitoring point to the predicted limit exceedance monitoring point. This results in the output of a monitoring quantity limit exceedance risk early warning result that includes the limit exceedance location, limit exceedance time, limit exceedance amplitude, and response propagation path.

[0068] Embodiment 2 of this invention provides an early warning system for soil and rock instability based on a physical constraint neural network, comprising: a construction stage determination module, used to determine the support number and activation time of each support, form a construction stage transition timetable, divide the construction stage by adjacent times, and output a construction stage sequence. The construction stage determination module receives layered excavation completion records, a list of support design connection points, support connection point installation and acceptance records, and a support axial force monitoring sequence. It determines the support activation time based on the comparison results of B consecutive axial force sampling values ​​after the support installation completion time with the support activation threshold, and generates a list of excavated empty areas, remaining soil, and activated support numbers for the stage based on the layered excavation completion records before the stage start time.

[0069] The physical connection generation module is used to filter physical connections by combining the list of excavated voids and activated support numbers for each construction stage, and write the reference stiffness of each physical connection. The physical connection generation module receives the construction stage sequence, monitoring point layout records, and design calculation models, establishes retaining structure connections, candidate soil connections, candidate soil-retaining structure connections, and support connections, deletes candidate connections that intersect with the interior of the excavated voids of the stage, and connections with empty or non-zero reference stiffness, and generates a unit vector of connection direction based on the three-dimensional coordinates of the monitoring points at both ends of the connection, and outputs the physical connection sequence of the construction stage.

[0070] The monitoring response window construction module is used to construct the monitoring response window and generate standardized response increments and physical dimension response increments. The module extracts the common sampling times of all participating monitoring sequences within the same construction phase, selects the last W sampling times to form the monitoring response window, subtracts the corresponding monitoring value of the first sampling time from each monitoring value within the window to generate the physical dimension response increment, and reads the minimum resolution of the corresponding sensor according to the sensor number to generate the standardized response increment.

[0071] The dynamic connection weight update module employs a gated cyclic graph neural network training algorithm to establish a stage connection prediction model, calculate dynamic connection weights, compose connection identifiers, and form a construction stage connection update record. The module inputs the standardized response increment into the time-series coding layer to generate monitoring point time-series response features. It then inputs the monitoring point time-series response features at both ends of the connection, the one-hot vector of the connection type, and the stiffness ratio into the connection weight generation layer, outputting dynamic connection weights. The module also inputs the monitoring point time-series response features and physical connection propagation features into the response prediction layer, outputting the predicted response for the next H sampling times. Finally, it compares the connection identifiers of adjacent construction stages, classifying physical connections into deleted connections, newly added connections, endpoint-associated connections, and inherited connections, and writes them into the construction stage connection update record according to the corresponding update rules.

[0072] The risk warning output module calculates the node balance error and the support axial force consistency error, generates valid dynamic connection records, calculates the predicted limit exceedance amplitude, determines the earliest predicted limit exceedance time, and outputs a monitoring quantity limit exceedance risk warning and an abnormal response transmission path. The risk warning output module converts the connection transmission force increment into a force vector in the same coordinate system, calculates the node balance error of each non-boundary monitoring point within the monitoring response window, and compares the predicted support axial force increment with the measured support axial force increment to calculate the support axial force consistency error; valid dynamic connection records are generated when both errors do not exceed the corresponding thresholds. The risk warning output module compares the predicted value output by the stage connection prediction model with the engineering control limits of the corresponding monitoring point and monitoring quantity type, generates predicted limit exceedance results, and outputs a monitoring quantity limit exceedance risk warning when predicted limit exceedance results exist in C adjacent monitoring response windows for the same monitoring point number and the same monitoring quantity type; it traces back the physical connection according to the preceding direction in the connection propagation contribution record, using the earliest predicted limit exceedance monitoring point as the path end, and outputs the abnormal response transmission path.

[0073] In summary, this application generates a construction phase sequence through a construction phase determination module, a physical connection generation module generates a construction phase physical connection sequence based on the construction phase sequence, a monitoring response window construction module generates standardized response increments and physical dimension response increments, a dynamic connection weight update module generates dynamic connection weights, future monitoring responses, and construction phase connection update records, and a risk warning output module generates effective dynamic connection records, monitoring quantity over-limit risk warning results, and abnormal response transmission paths based on physical consistency check results and predicted over-limit results.

[0074] The terms "first," "second," etc., in the specification, claims, and drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the term "comprising," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, apparatus, product, or device that includes a series of steps or units is not limited to the listed steps or modules, but may optionally include steps or modules not listed, or may optionally include other step units inherent to these processes, methods, apparatuses, products, or devices.

[0075] In the embodiments of this application, the terms "module" or "unit" refer to a computer program or part of a computer program that has a predetermined function and works with other related parts to achieve a predetermined goal, and can be implemented wholly or partially using software, hardware (such as processing circuitry or memory), or a combination thereof. Similarly, a processor (or multiple processors or memory) can be used to implement one or more modules or units. Furthermore, each module or unit can be part of an overall module or unit that includes the functionality of that module or unit.

[0076] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementations should not be considered beyond the scope of this application.

[0077] The methods and related apparatuses provided in this application are described with reference to the method flowcharts and / or structural diagrams provided in this application. Specifically, each block of the method flowchart and / or structural diagram, as well as combinations of blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing device to create a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, generate instructions for implementing the process. Figure 1 A schematic diagram of one or more processes and / or structures. Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to operate in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1A schematic diagram of one or more processes and / or structures. Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 A process or multiple processes and / or structures illustrate the steps of the functions specified in one or more boxes.

[0078] The above-disclosed embodiments are merely preferred embodiments of this application and should not be construed as limiting the scope of this application. Therefore, any equivalent variations made in accordance with the claims of this application shall still fall within the scope of this application.

Claims

1. An early warning method for soil and rock instability based on physical constraint neural networks, characterized in that, include: S1, determine the support number and support activation time of each support, form a construction stage transition timetable, divide the construction stages by adjacent times, and output the construction stage sequence; S2, combine the list of excavated voids and activated support numbers for each construction stage to filter physical connections, and write the reference stiffness of each physical connection. S3, construct the monitoring response window and generate standardized response increments and physical dimension response increments; S4. A gated cyclic graph neural network training algorithm is used to establish a stage connection prediction model, and dynamic connection weights are obtained to form connection identifiers and create a construction stage connection update record. S5 calculates the node balance error and the consistency error of the support axial force, generates an effective dynamic connection record, calculates the predicted over-limit amplitude, determines the earliest predicted over-limit time, and outputs the monitoring quantity over-limit risk warning and abnormal response transmission path.

2. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 1, characterized in that, The determination of the support number and support activation time for each support includes: Receive the layered excavation completion record, the initial ground elevation of the project, the initial soil range of the project, the list of support design connection points, the support connection point installation and acceptance record, the support axial force monitoring sequence, the minimum resolution of the sensor, the number of axial force determination samples B, the support activation threshold, the sampling interval, and the axial force sensor range; compare the design connection point number in the support design connection point list with the connection point number in the installation and acceptance record for the same support line item by item. If there are installation and acceptance records with the same number for each design connection point, record the latest acceptance time in the installation and acceptance record as the support installation completion time; Read B adjacent axial force sampling values ​​after the end of support installation. When the B axial force sampling values ​​are not empty and are within the range of the axial force sensor, the difference between adjacent sampling times is not greater than the sampling interval, and the B axial force sampling values ​​are all greater than or equal to the support activation threshold, record the Bth axial force sampling time as the support activation time, and output the support number and support activation time.

3. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 2, characterized in that, The process of forming a construction phase transition timetable, dividing construction phases by adjacent times, and outputting a construction phase sequence includes: Read the excavation completion time and support activation time between the start and end times of this analysis, sort them in ascending order of time and delete duplicate times to form a construction stage transition time table. Divide the construction stage by adjacent times, with the previous time as the stage start time and the next time as the stage end time. For each construction stage, read the excavation completion records of the layers whose excavation completion time is no later than the stage start time. Record the minimum excavation bottom elevation in the layer excavation completion records as the stage excavation bottom elevation. Record the geometric union of the excavation space range in the layer excavation completion records of this layer as the stage excavated void. Subtract the stage excavated void from the initial soil range of the project to obtain the stage remaining soil. Write the support numbers whose support activation time is no later than the stage start time into the stage activated support number list to form a construction stage sequence.

4. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 3, characterized in that, The process of filtering physical connections by combining the list of excavated voids and activated support numbers from each construction stage, and writing the reference stiffness of each physical connection, includes: The system receives the construction phase sequence, monitoring point layout records, and design calculation model. It uses the monitoring point number as a graph node. When the retaining structure segment numbers corresponding to two monitoring points are both written into the adjacent records of the same retaining structure segment in the design calculation model, a retaining structure connection is established between the two monitoring points. When the soil element numbers corresponding to two monitoring points are both written into the common boundary record of the same soil element in the design calculation model, a candidate soil connection is established between the two monitoring points. When the soil element corresponding to one monitoring point and the retaining structure segment corresponding to another monitoring point both exist in the same soil-retaining structure contact surface record, a candidate soil-retaining structure connection is established. When two monitoring points correspond to two support connection points of the same support, and the support number is within the list of activated support numbers for the phase, a support connection is established. Spatially compare the internal regions of the two soil units corresponding to the two ends of the candidate soil connection and the corresponding contact surface of the candidate soil-retaining structure connection with the excavated void area of ​​the stage. If there is an intersection of internal regions, delete the connection; if only the boundary is connected, retain the connection. The established support connection is directly retained. Read the corresponding reference stiffness from the design calculation model according to the connection type and delete the connection with an empty reference stiffness or not greater than zero. Arrange the two monitoring points at both ends of the physical connection in ascending order of numerical value. With the first monitoring point as the starting point and the second monitoring point as the ending point, subtract the three-dimensional coordinates of the first monitoring point from the three-dimensional coordinates of the second monitoring point to obtain the coordinate difference vector. When the length of the coordinate difference vector is equal to zero, delete the corresponding physical connection. When the length of the coordinate difference vector is greater than zero, divide the coordinate difference vector by the length of the coordinate difference vector to obtain the unit vector of the connection direction, forming the physical connection sequence of the construction stage.

5. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 4, characterized in that, The construction of the monitoring response window and the generation of standardized response increments and physical dimension response increments include: For each construction stage, the common sampling times of all activated monitoring data from all activated monitoring points between the start and end times of the stage are extracted, and the last W sampling times are selected to construct a monitoring response window; if there are fewer than W common sampling times, no window is generated; the physical dimension response increment is obtained by subtracting the corresponding monitoring value of the first sampling time from each monitoring value in the window, and then divided by the minimum resolution of the sensor corresponding to the monitoring value to obtain the standardized response increment; and the physical dimension response increment and standardized response increment of the axial force of the activated support are calculated.

6. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 5, characterized in that, The process of using a gated cyclic graph neural network training algorithm to establish a stage connection prediction model and calculate dynamic connection weights includes: A stage connection prediction model is established using a gated cyclic graph neural network training algorithm. The model includes a temporal coding layer, a connection weight generation layer, a physical connection weighted propagation layer, and a response prediction layer. For each construction stage, the standardized response increments of W sampling times within the monitoring response window are input to the gated cyclic unit, outputting the temporal response features of the monitoring points. For each physical connection in a construction stage, the reference stiffness is divided by the maximum reference stiffness of the same connection type within the same construction stage to obtain the stiffness ratio. The connection weight generation layer uses the temporal response features of the monitoring points at both ends of the connection and the stiffness ratio as input, and generates dynamic connection weights for each physical connection within the current monitoring response window via a multilayer perceptron. For each monitoring point, the temporal response features of adjacent monitoring points are weighted and summed according to the dynamic connection weights and stiffness ratios to obtain the physical connection propagation features. The temporal response features of the monitoring points and the physical connection propagation features are concatenated and input to the response prediction layer to obtain the standardized response increments for the next H sampling times, which are then converted into physical dimension response increments based on the sensor's minimum resolution. The temporal response characteristics of adjacent monitoring points corresponding to each physical connection are multiplied by the dynamic connection weight and stiffness ratio to obtain the connection propagation vector. The absolute values ​​of each component of the connection propagation vector are summed to obtain the connection propagation contribution value. For physical connections where the unit vector of the connection direction is not empty, the physical dimension displacement increments at both ends of the connection are converted into displacement increment vectors. The projection of the difference between the displacement increment vectors at both ends onto the connection direction is calculated, and the projection result is multiplied by the reference stiffness and dynamic connection weight to obtain the connection transmission force increment. For support connections, the support axial unit vector is determined based on the three-dimensional coordinates of the two support connection points, and the connection direction unit vector of the support connection is set as the support axial unit vector. The connection transmission force increment calculated according to the support connection direction is directly recorded as the support axial force prediction increment. H measured sampling times within the same construction stage after the monitoring response window are selected as prediction targets, and the prediction error is calculated. The predicted increment of support axial force is subtracted from the physical dimension response increment of support axial force at the same sampling time for the same support. The difference is divided by the minimum resolution of the axial force sensor, and the average of the squares of the resulting values ​​is calculated to obtain the normalized support axial force training error. The average of the squares of the differences in the dynamic connection weights of adjacent monitoring response windows within the same construction stage is calculated to obtain the connection weight time series error. The prediction error, support axial force consistency error, and connection weight time series error are added together to obtain the training error. The Adam algorithm and backpropagation are used to update the network parameters until the change in training error is less than or equal to the training termination threshold, or the number of training rounds reaches the training round threshold. After training, the monitoring response window with the latest window end time and the corresponding physical connection record of the construction stage are input into the stage connection prediction model, and the dynamic connection weights of each physical connection are output.

7. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 6, characterized in that, The component connection identifiers, forming the connection update record during the construction phase, include: The two monitoring point numbers are arranged numerically and combined with the connection type to form a connection identifier. The connection identifiers of adjacent construction stages are compared. Connections that exist only in the previous construction stage are marked as deleted connections and their weights are set to zero. Connections that exist only in the subsequent construction stage are marked as new connections. Connections that exist in both construction stages and share monitoring points with deleted or new connections are marked as endpoint-associated connections. The new connections and endpoint-associated connections, along with the monitoring response window of the subsequent construction stage, are input into the stage connection prediction model. The temporal response features of the monitoring points at both ends are extracted by the temporal coding layer. Then, combined with the stiffness ratio, the dynamic connection weights are recalculated by the connection weight generation layer. The remaining connections that exist in both construction stages are marked as inherited connections. The dynamic connection weights at the end of the previous construction stage are used to form a construction stage connection update record.

8. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 7, characterized in that, The calculation node balance error and the support axial force consistency error generate effective dynamic connection records including: The incremental force transmission of adjacent physical connections at the same monitoring point is converted into a force vector in the same coordinate system according to the unit vector of the connection direction. The force vectors are summed, and the average of the squares of the resultant force vector lengths is used as the node balance error. The incremental force transmission of the support connection is projected onto the support axial direction and subtracted from the measured axial force increment of the same support at the same sampling time. The average of the squares of the difference is used as the support axial force consistency error. When the balance error of all nodes in the subsequent construction stage is not greater than the node balance error threshold, and the consistency error of the support axial force of all support connections is not greater than the support axial force consistency error threshold, the dynamic connection weight of the subsequent construction stage is written into the valid dynamic connection record. When any node balance error is greater than the node balance error threshold, any support axial force consistency error is greater than the support axial force consistency error threshold, or any physical connection involved in the inspection does not generate a incremental force transmission, the dynamic connection weight of the subsequent construction stage is not written, and a record of insufficient physical consistency is output.

9. The method for early warning of soil and rock instability based on a physically constrained neural network according to claim 8, characterized in that, The calculation and prediction of the over-limit amplitude, determination of the earliest predicted over-limit time, and output of the over-limit risk warning and abnormal response transmission path for the monitored quantity include: When the predicted value in the effective dynamic connection record is less than the lower control limit, the predicted over-limit amplitude is obtained by subtracting the predicted value from the lower control limit; when the predicted value is greater than the upper control limit, the predicted over-limit amplitude is obtained by subtracting the upper control limit from the predicted value; when the predicted value is not less than the lower control limit and not greater than the upper control limit, the predicted over-limit amplitude is written as zero; in ascending order of the predicted sampling time, the sampling time at which the first predicted over-limit amplitude is greater than zero is written as the earliest predicted over-limit time; when there are two or more over-limit results at the earliest predicted over-limit time, the over-limit result with the largest predicted over-limit amplitude is recorded; when the predicted over-limit amplitudes are the same, the over-limit result with the smallest monitoring point number is recorded; when there are over-limit results for the same monitoring point number and the same monitoring quantity type in C adjacent monitoring response windows, an over-limit risk warning for the monitoring quantity is output; The monitoring point corresponding to the earliest predicted out-of-limit result is taken as the end of the path. The connection propagation contribution value corresponding to the monitoring response window that generated the current prediction result is read. The physical connection with the largest connection propagation contribution value connected to the current monitoring point is selected from the effective dynamic connection record in sequence. The monitoring point at the other end of the physical connection is written into the abnormal response propagation path. The monitoring points and physical connections obtained by backtracking are arranged in reverse order to form the abnormal response propagation path.

10. An early warning system for soil and rock instability based on a physically constrained neural network, characterized in that, include: The construction phase determination module is used to determine the support number and support activation time of each support, form a construction phase transition timetable, divide the construction phases by adjacent times, and output the construction phase sequence. The physical connection generation module is used to filter physical connections by combining the list of excavated voids and activated support numbers at each construction stage, and to write the reference stiffness of each physical connection. The monitoring response window building module is used to build the monitoring response window and generate standardized response increments and physical dimension response increments. The dynamic connection weight update module is used to establish a stage connection prediction model by using a gated cyclic graph neural network training algorithm, and to obtain dynamic connection weights to form connection identifiers and create a construction stage connection update record. The risk warning output module is used to calculate the node balance error and the consistency error of the support axial force, generate effective dynamic connection records, calculate the predicted over-limit amplitude, determine the earliest predicted over-limit time, and output the risk warning and abnormal response transmission path for the monitoring quantity over-limit.

Images