A telescopic multi-stage wall protection sampling device suitable for deep overburden loose stratum and a sampling method thereof

Abstract

The present application relates to the technical field of geological exploration drilling sampling, and discloses a telescopic multi-stage wall protection sampling device suitable for deep overburden loose stratum and a sampling method thereof, which solves the problem that the prior art cannot reliably determine the actual wall protection demand corresponding to the uncovered hole section in the hole and lacks direct corresponding relationship between the risk judgment result and the wall protection execution action, leading to disconnection of control. The present application scheme comprises the following steps: constructing a hole working condition sequence and a wall protection extension state, generating an instability risk evidence quantity, a risk trigger flag and a wall protection coverage demand quantity through an evidence time series convolution network; when the risk is triggered, a rolling wall protection plan is generated through a rolling time domain constraint branch; a telescopic motor is driven to execute and the encoder is checked to the position, and the wall protection extension state is updated; and the updated state is taken as the condition quantity of the next control time, so that a closed loop is formed.

Description

Technical Field

[0001] This invention relates to the field of geological exploration drilling and sampling technology, specifically to a retractable multi-stage wall-supported sampling device and sampling method suitable for thick overburden loose strata. Background Technology

[0002] In engineering geological exploration, mineral sampling, and deep overburden exploration, drilling and sampling are often required in thick, loose overburden strata. These strata typically exhibit characteristics such as loose structure, uneven particle size distribution, high local water content, and poor borehole wall self-stabilization. As drilling depth increases, borehole disturbances accumulate significantly, leading to phenomena such as borehole wall spalling, diameter reduction, drill bit burial, and borehole collapse. Especially when continuous drilling and sampling are carried out simultaneously, if borehole wall protection is not implemented in a timely manner, the exposed section of the borehole wall increases, making deeper sections more prone to instability. This not only affects sampling continuity but also leads to sample disturbance, stratigraphic mixing, and sampling failure. Therefore, in such scenarios, coordinating borehole wall protection with the sampling process and adjusting the wall coverage area promptly according to changes in borehole conditions has always been a crucial issue in the design and control methods of related sampling equipment.

[0003] In existing technologies, borehole sampling in loose formations typically employs methods such as casing wall support, casing-following drilling, segmented lowering of the casing wall, and rotary propulsion sampling to maintain borehole wall stability. Some devices incorporate a retractable casing wall, which is extended via a hydraulic, motor, or rack and pinion transmission mechanism, and its position is detected by displacement sensors or encoders. For control, parameters such as drilling depth, torque, current, rotational speed, and casing wall displacement are typically collected during drilling. These parameters are then combined with preset thresholds, empirical rules, or simple state judgment logic to monitor changes in drilling resistance and the casing wall's follow-up status. When certain conditions are met, control commands are issued to extend or retract the casing wall, pause drilling, or continue sampling, thus achieving basic casing wall control during the sampling operation.

[0004] However, during sampling in deep, loosely packed formations, changes in operating conditions such as torque, current, and rotational speed are simultaneously affected by multiple factors, including variations in formation density, water disturbance, and insufficient wall coverage. Existing solutions often struggle to distinguish between "normal drilling load changes" and "restricted changes caused by insufficient wall coverage," making it difficult to reliably determine the actual wall coverage requirements for uncovered sections within the borehole. Furthermore, existing wall control methods often rely on continuous displacement settings or empirical calculations, failing to establish unified constraints with the remaining stroke, minimum step distance, and actual reach of multi-stage wall coverage. This results in a lack of direct correlation between risk assessment results and wall execution actions, leading to insufficient control accuracy and response timeliness. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a retractable multi-stage wall protection sampling device and sampling method suitable for deep overburden loose strata, which solves the problem that the existing technology is difficult to reliably determine the actual wall protection needs corresponding to the uncovered borehole section and that there is no direct correspondence between the risk assessment results and the wall protection execution actions, resulting in control disconnect.

[0006] The technical solution adopted by the present invention to solve the above-mentioned technical problems is as follows:

[0007] On one hand, the present invention provides a retractable multi-stage wall protection sampling device suitable for use in deep overburden loose strata, which includes: a sampling drive mechanism, a retractable multi-stage wall protection mechanism, a sampling mechanism, a sealing mechanism, and a control module;

[0008] The sampling drive mechanism is located at the top of the device and is connected to the first-stage protective wall tube in the retractable multi-stage protective wall mechanism, and forms a transmission connection with the sampling mechanism. The sampling drive mechanism includes a servo motor, a transmission connector and a torque sensor, which are used to provide sampling propulsion power and sampling rotation power to the sampling mechanism, and to collect sampling torque signals in real time.

[0009] The retractable multi-stage wall protection mechanism is sleeved around the sampling mechanism and includes a primary wall protection tube, a secondary wall protection tube, a tertiary wall protection tube, a telescopic drive assembly, and a wall protection encoder; it is used to perform graded wall protection and follow-up support on the borehole wall during drilling, and output wall extension amount and displacement feedback signals; the telescopic drive assembly includes a telescopic motor.

[0010] The sampling mechanism is fitted inside the retractable multi-stage retaining wall mechanism and is used to drill and collect soil samples in the thick overburden loose strata, and to provide feedback on the sampling load change signal.

[0011] The sealing mechanism is located at the bottom of the sampling mechanism and is used to maintain the bottom of the sampling mechanism in a sealed state during the sampling process;

[0012] The control module is electrically connected to the servo motor and torque sensor in the sampling drive mechanism, the telescopic motor and wall encoder in the telescopic multi-stage wall protection mechanism, and the sealing mechanism, respectively, and is used to receive the working condition data in the hole and the wall displacement data, and output the instability risk identification result and the wall telescopic control command.

[0013] Based on the structural design of the above sampling device, the collaborative operation of drilling sampling and multi-stage wall protection is realized. From the hardware architecture perspective, it provides structural support for accurately distinguishing drilling load types, identifying borehole wall instability risks in real time, and reliably executing wall extension and contraction actions.

[0014] On the other hand, the present invention also provides a sampling method applied to the above-mentioned sampling device, the method comprising:

[0015] S1. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, construct the hole working condition sequence according to the time sequence, and determine the wall extension state based on the collected data;

[0016] S2. Based on the upper limit of the stroke of the multi-stage retaining wall and the rack and pinion drive step distance, generate a set of discrete retaining wall actions that adapt to the execution capability of the mechanism;

[0017] S3. Input the working condition sequence inside the hole into the evidence temporal convolutional network, and after feature extraction and fusion calculation, output the amount of evidence of instability risk, risk triggering flags, and the amount of wall covering required;

[0018] S4. When the risk triggering flag meets the preset conditions, a rolling retaining wall plan is generated by rolling time-domain solution based on the retaining wall coverage demand, retaining wall extension status and discrete retaining wall action set.

[0019] S5. Generate wall extension control commands according to the rolling wall plan, drive the extension motor of the retractable multi-stage wall mechanism to perform wall extension action, and perform position verification through the wall encoder, and update the wall extension status according to the verification result;

[0020] S6. Use the updated wall extension state as the wall condition quantity for the next control moment.

[0021] In this scheme, by constructing a borehole working condition sequence according to a unified time sequence and determining the wall extension state, the multi-source working condition data and the wall state are time-aligned, eliminating the discrimination error caused by data asynchrony and providing a reliable input basis for risk identification. Based on the upper limit of the multi-level wall travel and the rack and pinion drive step distance, a discrete wall action set is generated, transforming continuous wall requirements into executable action units that match the mechanical structure, thus avoiding the problem of mismatch between control commands and execution capabilities from the source. By inputting the borehole working condition sequence into an evidence-based temporal convolutional network for processing, torque and current characteristics can be coupled and combined with wall state constraints to accurately distinguish between normal drilling load changes and constrained changes caused by insufficient wall coverage, improving the accuracy and targeting of instability risk identification. When the risk triggering conditions are met, a rolling time-domain solution is used to generate a rolling wall protection plan. Under the constraints of stroke, the action units can be screened, combined, and sorted to ensure that the wall protection action plan fits the real-time risk and coverage requirements in the borehole. The telescopic motor is driven to execute the wall protection action according to the rolling wall protection plan, and the encoder completes the positioning verification, which can ensure the accurate execution of the wall extension action. At the same time, the actual execution results are used to correct the wall protection state and eliminate execution deviations and cumulative errors. Finally, the updated wall extension state is fed back to the next control moment as a conditional quantity, forming a complete closed loop of working condition acquisition, risk identification, action planning, execution feedback and state update. This allows the wall protection to continuously adapt to the changes in working conditions during the drilling process and realize the dynamic coordination between sampling of loose formations in deep overburden and wall protection.

[0022] Furthermore, step S1 specifically includes:

[0023] S11. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, and align these parameters at a unified control time.

[0024] S12. Extract the sampled torque, servo motor current and servo motor speed within the most recent continuous control moment, arrange them in chronological order to form a sequence of working conditions inside the hole, and retain the timing position of each control moment corresponding to the displacement of the advance scale and the wall encoder.

[0025] S13. Determine the current total extension of the wall based on the wall encoder displacement, determine the difference between the current advance and the front edge of the wall based on the current advance and the current total extension of the wall, and limit the value boundary of the current advance with the target hole depth;

[0026] S14. The current total wall extension and the difference between the current advance and the leading edge of the wall are used to form the wall extension status.

[0027] Furthermore, step S2 specifically includes:

[0028] S21. Obtain the upper limit of the second-stage wall protection stroke, the upper limit of the third-stage wall protection stroke, and the rack and pinion drive step distance, and determine the rack and pinion drive step distance as the second-stage minimum action unit and the third-stage minimum action unit, respectively;

[0029] S22. Determine the current remaining stroke of the secondary wall encoder, the current remaining stroke of the tertiary wall encoder, and the total remaining stroke based on the wall encoder displacement, the upper limit of the secondary wall stroke, and the upper limit of the tertiary wall stroke;

[0030] S23. The second-level current remaining stroke, the third-level current remaining stroke, the second-level minimum action unit, and the third-level minimum action unit are discretized to form action units with corresponding telescopic motor execution steps. The combinable range of action units is limited according to the total remaining stroke and the current total extension of the wall, and a discrete wall action set is generated.

[0031] S24. Determine the current reachable coverage based on the discrete wall protection action set, and combine the second-level current remaining travel, the third-level current remaining travel, the second-level minimum action unit, the third-level minimum action unit, the total remaining travel, and the current reachable coverage into an action code.

[0032] Furthermore, in step S3, the evidence temporal convolutional network includes an evidence temporal convolutional network backbone and a rolling temporal constraint branch;

[0033] The evidence temporal convolutional network backbone includes a joint channel layer, multi-scale convolutional branches, a convolutional fusion layer, a two-level fully connected layer, an evidence output layer, and a cover output layer.

[0034] The combined channel layer combines the sampled torque and the servo motor current into a combined channel, while retaining the servo motor speed as an independent channel;

[0035] The multi-scale convolutional branch includes three one-dimensional convolutional branches with kernel lengths of 3, 5, and 7 respectively. Each branch has two convolutional layers, and the convolutional outputs of each branch are concatenated to form sequence features.

[0036] The convolutional fusion layer conditionally fuses the sequence features with the wall extension state to obtain fused features;

[0037] The fusion features are formed as evidence features after passing through two fully connected layers.

[0038] The evidence output layer is used to output the instability risk value and credibility.

[0039] The coverage output layer is used to output the required amount of wall protection coverage;

[0040] The rolling time-domain constraint branch includes an action level layer, an action selection layer, and a rolling time-domain planning layer;

[0041] The rolling time-domain constraint branch receives instability risk value, confidence level, retaining wall coverage requirement, second retaining wall extension status and 6-dimensional action code to form 11-dimensional decision input;

[0042] The 11-dimensional decision input is sequentially input into an action level layer containing 12 neurons and an action selection layer containing 6 neurons, and outputs the three-dimensional action unit displacement level corresponding to small displacement, medium displacement, and large displacement.

[0043] The rolling time-domain planning layer filters action units based on the displacement level of the three-dimensional action unit and the discrete retaining wall action set, and generates a rolling retaining wall plan under the upper limit constraint of the stroke, so that the retaining wall coverage requirement directly corresponds to the actual achievable extension of the secondary and tertiary retaining walls.

[0044] Furthermore, step S3 specifically includes:

[0045] S31. Input the working condition sequence inside the hole into the backbone of the evidence temporal convolutional network, and couple the sampling torque and servo motor current according to the temporal position through the joint channel layer to form a joint channel, and retain the servo motor speed as an independent channel;

[0046] S32. Input the joint channel and the independent channel into the multi-scale convolution branch for convolution extraction, and then concatenate to obtain the sequence features;

[0047] S33. Conditionally fuse the sequence features and the wall protrusion state in the convolutional fusion layer to obtain the fused features;

[0048] S34. The fused features are passed through two fully connected layers to generate evidence features, then the evidence output layer generates the amount of instability risk evidence, and the coverage output layer generates the amount of wall coverage demand.

[0049] S35. Compare the instability risk value and confidence level with the corresponding thresholds respectively, and determine the risk triggering flag when the double threshold condition is met;

[0050] S36. Calculate the required amount of retaining wall coverage based on the amount of evidence of instability risk, the retaining wall extension status, and the current advance constraints.

[0051] Furthermore, step S4 specifically includes:

[0052] S41. When the risk trigger flag meets the conditions, input the instability risk value, confidence level, wall cover requirement, wall extension status and discrete wall action set into the rolling time domain constraint branch, and convert the discrete wall action set into action codes for the corresponding wall extension amounts at each level.

[0053] S42. The decision input is sequentially sent to the action level layer and the action selection layer. Based on the instability risk value, confidence level and wall cover requirement, the action unit displacement level including small displacement, medium displacement and large displacement is generated.

[0054] S43. Pre-select the action units in the discrete wall action set according to the displacement level of the action unit, and combine the wall extension state and the upper limit of the stroke to complete the reachability determination and obtain the candidate action units;

[0055] S44. Based on the candidate action units, perform rolling time-domain solution, filter, combine and sort the action units, and directly generate an action unit sequence adapted to the telescopic motor.

[0056] S45. Based on the wall protection coverage requirement, perform cumulative coverage verification on the candidate action units, and determine the rolling wall protection plan under the travel limit constraint;

[0057] S46. Based on the amount of evidence of instability risk, the rolling retaining wall plan is classified into displacement levels, the instability risk value is used to determine the extension amplitude level, and the confidence level is used to limit the conditions for switching displacement levels.

[0058] S47. Determine the extension amplitude and action sequence of each action unit according to the displacement level of the final action unit to form an action unit sequence that can directly generate wall extension and retraction control commands.

[0059] Furthermore, step S5 specifically includes:

[0060] S51. Parse the sequence of action units in the rolling retaining wall plan into retaining wall extension and retraction control commands corresponding to each level of retaining wall, and write the extension amplitude of each action unit into the telescopic motor control quantity according to the action sequence.

[0061] S52. Drive the telescopic motor to execute the wall extension control command, so that the secondary wall and the tertiary wall extend sequentially according to the rolling wall plan, and the actual extension action is consistent with the action unit defined by the discrete wall action set;

[0062] S53. During the operation of the telescopic motor, the positioning signal fed back by the wall encoder is received, and the positioning signal is checked against the target action unit in the rolling wall plan to form a positioning check state;

[0063] S54. Update the current total wall extension based on the in-place verification status, recalculate and determine the difference between the current advance and the leading edge of the wall, and combine the updated current total wall extension and the difference between the current advance and the leading edge of the wall to form the wall extension status.

[0064] Furthermore, step S6 specifically includes:

[0065] S61. Receive the updated wall extension status after step S5, and extract the updated current total wall extension and the difference between the current advance and the wall leading edge;

[0066] S62. Integrate the updated current total wall extension with the current advance and the difference between the wall leading edge and the current advance into the wall condition quantity for the next control time, so that the wall condition quantity corresponds to the time sequence of the working conditions in the hole at the next control time.

[0067] S63. Input the wall protection condition quantity into the convolutional fusion layer of the evidence temporal convolutional network backbone, so that the next control time step generates the instability risk evidence quantity based on the wall protection condition quantity and the working condition data.

[0068] S64. The wall protection condition quantity is used as the state constraint input for the rolling time domain constraint branch, and combined with the discrete wall protection action set to limit the reachable extension and stroke limit of each level of wall protection.

[0069] The beneficial effects of this invention are:

[0070] (1) Effectively improves the accuracy and relevance of instability risk identification within the borehole, avoiding misjudgment:

[0071] This invention no longer treats sampling torque, servo motor current, and servo motor speed as independent operating parameters for simple threshold judgment. Instead, it uses an evidence-based temporal convolutional network to form a joint channel between sampling torque and servo motor current, and introduces the wall extension state ("current total wall extension + difference between current advance and wall leading edge") as a conditional quantity to participate in the generation of instability risk evidence. This design can accurately distinguish between normal drilling resistance fluctuations and hole segment constraints caused by insufficient wall follow-up during continuous advance in thick, loose formations. It breaks through the limitations of empirical rules or fixed threshold identification methods in existing technologies, upgrading the output results from a traditional single alarm signal to three correlated results: "instability risk value, confidence level, and wall coverage requirement," which can directly provide a clear basis for subsequent wall control.

[0072] Comparative testing of a classic loose sand-silt alternating overburden prototype showed that, under the same advance and rotation speed conditions, the invention can improve the lead time for identifying insufficient wall protection conditions by about 20%-30%, reduce the number of invalid triggers by about 25%, and control the deviation between the wall protection coverage requirement directly obtained from the identification results and the actual uncovered borehole section within 10%, providing stable and reliable pre-support for wall protection control.

[0073] (2) Solve the problem of disconnect between risk assessment and wall protection execution in the existing technology, and realize control closed loop:

[0074] This invention incorporates the upper limits of the stroke of the secondary and tertiary retaining walls, the rack and pinion drive step distance, the current remaining stroke, and the current achievable coverage into the discrete retaining wall action set. Through rolling time-domain constraint branches, it directly completes candidate action screening, displacement grading, and action sequencing at the discrete action unit level, completely avoiding the control disconnect phenomenon of "first generating continuous displacement, then splitting it again by the execution end" in existing technologies. The resulting rolling retaining wall plan is naturally adapted to the action units that the telescopic motor can execute. Combined with the position verification function of the retaining wall encoder and the feedback mechanism of the retaining wall state in the next control cycle, a complete closed-loop link of "operating condition acquisition—risk identification—action planning—execution feedback—state update" is formed.

[0075] This closed-loop control design ensures that the wall protection action is highly consistent with the mechanical boundary constraints, effectively reducing control fluctuations caused by overtravel, undertravel, and repeated corrections. Through continuous sampling and comparative testing of similar prototypes, under the same hole depth conditions, the one-time completion rate of the wall protection action can be increased by about 15%-20%, the average number of displacement corrections after the action is executed can be reduced by about 30%, the duration of the uncovered hole section can be shortened by about 25%, and the number of operation interruptions caused by hole collapse or significant diameter reduction can be reduced by about 20%. Attached Figure Description

[0076] Figure 1This is an overall flowchart of the wall protection sampling method in this invention.

[0077] Figure 2 This is a flowchart illustrating the construction of the working condition sequence and the wall extension state in the borehole in this invention.

[0078] Figure 3 This is a flowchart of the generation and encoding process for the discrete wall protection action set in this invention.

[0079] Figure 4 This is a flowchart of the evidence-based temporal convolutional network for risk identification and demand calculation in this invention.

[0080] Figure 5 This is a flowchart of the rolling retaining wall plan generation process in this invention.

[0081] Figure 6 This is a flowchart of the wall protection execution and status update process in this invention.

[0082] Figure 7 This is a flowchart of the wall protection condition quantity update process in this invention.

[0083] Figure 8 This is a comparison diagram of the combined working condition spatial distribution and confidence ellipse in an embodiment of the present invention.

[0084] Figure 9 This is a color-filled equivalent effect diagram of the required wall covering quantity in an embodiment of the present invention. Detailed Implementation

[0085] This invention aims to provide a retractable multi-stage wall protection sampling device and its sampling method suitable for deep overburden loose formations. It addresses the problems of existing technologies, such as the difficulty in reliably identifying the actual wall protection requirements corresponding to uncovered sections within the borehole and the lack of a direct correspondence between risk assessment results and wall protection actions, leading to control disconnect. The core idea is that this invention directly solves the core problems of insufficient wall protection follow-up, instability identification bias, and control-execution disconnect in deep overburden loose formation sampling through a closed-loop technical path of in-bore condition evidence extraction, wall protection requirement identification, discrete action planning, and execution feedback updates. This achieves precise adaptation between sampling operations and wall protection control.

[0086] (1) An innovative algorithm for instability identification and wall protection demand linkage is proposed. This algorithm abandons the independent threshold judgment mode for working conditions such as sampling torque, servo motor current, and speed in the existing technology. Through the evidence-based temporal convolutional network backbone, the sampling torque and servo motor current are coupled together. At the same time, the current total wall extension and the difference between the current advance and the wall front edge are introduced as conditional quantities and integrated into the convolutional fusion layer. This allows the interpretation of changes in working conditions in the borehole to be simultaneously constrained by both the load signal and the wall coverage state. This enables the accurate distinction between "drilling load change" and "limited change caused by insufficient wall follow-up". The algorithm directly outputs the instability risk value, credibility and wall coverage demand, solving the problem that the wall protection demand is difficult to reliably determine in deep overburden loose strata.

[0087] (2) A novel rolling time-domain constraint branch construction method is constructed, which directly encodes the upper limit of the multi-level retaining wall stroke, the rack and pinion drive step distance, the current remaining stroke and the current achievable coverage as a discrete retaining wall action set. On this discrete set, the action level determination, candidate action screening and rolling retaining wall plan generation are completed. From the decision input stage, it fits the mechanical structure execution capability, so that the generated retaining wall plan is naturally adapted to the action unit sequence that the telescopic motor can execute. At the same time, the action amplitude is constrained by the instability risk value, the displacement level switching is limited by the credibility, and the retaining wall coverage requirement is cumulatively checked to form a continuous parameter link of risk evidence, coverage requirement and retaining wall action, so as to solve the problem of the disconnect between the identification result and the execution action.

[0088] (3) Establish a closed-loop control system for the wall protection for sampling of loose strata in deep overburden. Through the complete technical process of “construction of working condition sequence in borehole - generation of instability risk evidence - discrete action planning - encoder positioning verification - wall protection status update”, risk identification, action planning and execution feedback are incorporated into the same control loop. Not only is the positioning verification performed using the wall protection encoder, but the updated wall extension status is also fed back to the next control moment as the common input of the evidence temporal convolutional network backbone and the rolling temporal constraint branch. This allows the wall protection status to continuously participate in instability judgment and action constraints, maintain the dynamic correspondence between the wall protection front and the uncovered borehole section, realize the sustainable operation of the wall protection follow-up closed-loop control, and ultimately improve the stability and reliability of sampling of loose strata in deep overburden.

[0089] To facilitate understanding of the technical solution of this invention, some technical terms involved in this invention are explained below:

[0090] Evidence-based temporal convolutional networks: a neural network structure for processing time-series operational data. It extracts short-term mutation and continuous change features through convolutional kernels of different lengths and outputs instability risk values ​​and their credibility, rather than just providing a single risk judgment.

[0091] Wall extension status: refers to the actual follow-up status of the current wall, which is composed of the total wall extension and the difference between the current advance and the leading edge of the wall. It is used to reflect whether the borehole section is covered or not, and is an important condition quantity for risk identification and wall control.

[0092] Discrete retaining wall action set: refers to a set of executable action units pre-generated based on the upper limit of the stroke of the secondary and tertiary retaining walls and the rack and pinion drive step distance. Its function is to transform continuous displacement requirements into discrete expansion and contraction combinations that the equipment can actually complete.

[0093] Rolling time-domain constraint branch: This refers to the control module that generates the wall protection plan in time periods based on the current risk outcome, wall protection status, and action codes. It dynamically filters and sorts executable action units under constraints such as travel limit, remaining travel, and coverage requirements.

[0094] In specific implementation, the present invention first provides a retractable multi-stage wall protection sampling device suitable for deep overburden loose strata, which includes: a sampling drive mechanism, a retractable multi-stage wall protection mechanism, a sampling mechanism, a sealing mechanism and a control module;

[0095] The sampling drive mechanism is located at the top of the device and is connected to the first-stage protective wall tube in the retractable multi-stage protective wall mechanism, and forms a transmission connection with the sampling mechanism. The sampling drive mechanism includes a servo motor, a transmission connector and a torque sensor, which are used to provide sampling propulsion power and sampling rotation power to the sampling mechanism, and to collect sampling torque signals in real time.

[0096] The retractable multi-stage wall protection mechanism is sleeved around the sampling mechanism and includes a primary wall protection tube, a secondary wall protection tube, a tertiary wall protection tube, a telescopic drive assembly, and a wall protection encoder; it is used to perform graded wall protection and follow-up support on the borehole wall during drilling, and output wall extension amount and displacement feedback signals; the telescopic drive assembly includes a telescopic motor.

[0097] The sampling mechanism is fitted inside the retractable multi-stage retaining wall mechanism and is used to drill and collect soil samples in the thick overburden loose strata, and to provide feedback on the sampling load change signal.

[0098] The sealing mechanism is located at the bottom of the sampling mechanism and is used to maintain the bottom of the sampling mechanism in a sealed state during the sampling process;

[0099] The control module is electrically connected to the servo motor and torque sensor in the sampling drive mechanism, the telescopic motor and wall encoder in the telescopic multi-stage wall protection mechanism, and the sealing mechanism, respectively. It is used to receive in-hole working condition data and wall displacement data, and output instability risk identification results and wall extension control commands. More specifically, the control module executes the control program in the memory through the processor. The control program is configured to construct an in-hole working condition sequence, wall extension state, and discrete wall action set based on the target hole depth, current advance, sampling torque, servo motor current, servo motor speed, and wall encoder displacement. It outputs instability risk value, confidence level, and wall coverage requirement through the evidence temporal convolutional network backbone, and generates a rolling wall protection plan and wall extension control commands through a rolling temporal constraint branch to drive the secondary and tertiary walls to perform extension control according to the action unit sequence and update the wall extension state.

[0100] The functions of each component in the above-mentioned sampling device and their arrangement are shown in Table 1.

[0101] Table 1. Module Structure of the Sampling Device

[0102]

[0103] Based on the above sampling device, the sampling method provided by the present invention is as follows: Figure 1 As shown, it includes the following implementation steps:

[0104] S1. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, construct the hole working condition sequence according to the time sequence, and determine the wall extension state based on the collected data;

[0105] In one exemplary implementation, the process for constructing the borehole working condition sequence and the wall extension state in this step is as follows: Figure 2 As shown, it includes the following sub-steps:

[0106] S11. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, and match the current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement with a unified control time, so that a set of hole working condition sampling values ​​are formed at the same control time.

[0107] S12. Extract the sampled torque, servo motor current and servo motor speed within the most recent continuous control moments, arrange them in chronological order to form the working condition sequence inside the hole, and retain the timing position of each control moment corresponding to the forward advance and the wall encoder displacement, so that the working condition sequence inside the hole and the wall extension state correspond one-to-one in the subsequent convolutional fusion layer.

[0108] S13. Determine the current total extension of the wall based on the wall encoder displacement, determine the difference between the current advance and the front edge of the wall based on the current advance and the current total extension of the wall, and limit the value boundary of the current advance with the target hole depth;

[0109] S14. The current total wall extension and the difference between the current advance and the wall leading edge are used to form the wall extension state. The wall extension state is used as the condition input for step S3 and as the state basis for the upper limit constraint of the stroke and the reachability determination of the discrete wall action set in step S4.

[0110] In this step, time alignment is achieved by relying on multi-source operating parameters collected in real time on site. The working condition characterization and wall protection status quantification can be completed using only the inherent monitoring signals of drilling operations, without the need for additional detection equipment. Through the time-location association design, the data correlation between the working condition sequence in the borehole and the wall protection extension status is ensured, providing a standardized and unified preliminary data foundation for subsequent network feature fusion, accurate identification of instability risks, and determination of wall protection action constraints.

[0111] S2. Based on the upper limit of the stroke of the multi-stage retaining wall and the rack and pinion drive step distance, generate a set of discrete retaining wall actions that adapt to the execution capability of the mechanism;

[0112] In one exemplary implementation, the process for generating and encoding the discrete wall protection action set in this step is described in [reference needed]. Figure 3 This includes the following sub-steps:

[0113] S21. Obtain the upper limit of the second-stage wall protection stroke, the upper limit of the third-stage wall protection stroke, and the rack and pinion drive step distance, and determine the rack and pinion drive step distance as the second-stage minimum action unit and the third-stage minimum action unit, respectively;

[0114] S22. Determine the current remaining stroke of the secondary wall encoder, the current remaining stroke of the tertiary wall encoder, and the total remaining stroke based on the wall encoder displacement, the upper limit of the secondary wall stroke, and the upper limit of the tertiary wall stroke;

[0115] S23. The current remaining stroke of the second level, the current remaining stroke of the third level, the minimum action unit of the second level, and the minimum action unit of the third level are discretized in a hierarchical manner to form action units with corresponding telescopic motor execution step lengths. The combinable range of the action units is limited according to the total remaining stroke and the current total extension of the guard wall, and a discrete guard wall action set is generated so that the achievable extension of the second-level and third-level guard walls is limited from the time the discrete guard wall action set is generated.

[0116] S24. Determine the current achievable coverage based on the discrete wall action set, and combine the second-level current remaining stroke, the third-level current remaining stroke, the second-level minimum action unit, the third-level minimum action unit, the total remaining stroke, and the current achievable coverage into an action code, so that the action code and the wall extension state together serve as the decision input for step S4.

[0117] S3. Input the working condition sequence inside the hole into the evidence temporal convolutional network, and after feature extraction and fusion calculation, output the amount of evidence of instability risk, risk triggering flags, and the amount of wall covering required;

[0118] The evidence-based temporal convolutional network provided by this invention comprises two main parts: an evidence-based temporal convolutional network backbone and a rolling temporal constraint branch. The evidence-based temporal convolutional network backbone is responsible for extracting instability signs from the in-hole working condition sequence and the wall extension state, directly outputting the instability risk value, confidence level, and wall coverage requirement. The rolling temporal constraint branch receives the instability risk value, confidence level, wall coverage requirement, wall extension state, and discrete wall action set, further generating action unit displacement levels, and then forming a rolling wall plan based on these displacement levels. The event-triggered rolling temporal multi-level wall extension / retraction decision is not independently attached to the evidence-based temporal convolutional network, but rather embedded as a rolling temporal constraint branch after the evidence output, enabling instability identification and multi-level wall actions to be transmitted within the same parameter link.

[0119] The input to the evidence-based temporal convolutional network backbone consists of two parts: the first part is the sequence of working conditions within the borehole, taking the most recent 24 control moments, with an input dimension of 24*3, and three features: sampling torque, servo motor current, and servo motor speed. The second part is the wall extension state, with an input dimension of 2, and two features: the current total wall extension and the difference between the current advance and the wall leading edge. The sampling torque and servo motor current together reflect changes in sampling resistance, the servo motor speed reflects changes in propulsion release, and the difference between the current advance and the wall leading edge reflects the length of the uncovered borehole segment. The rolling temporal constraint branch then receives the action codes of the discrete wall action set, with an action code dimension of 6, and six features: the second-level current remaining stroke, the third-level current remaining stroke, the second-level minimum action unit, the third-level minimum action unit, the total remaining stroke, and the current achievable coverage.

[0120] The backbone of the temporal convolutional network for evidence first sets up a joint channel layer, forming a 24*2 joint channel for the sampled torque and servo motor current, while retaining the servo motor speed as a 24*1 independent channel. Then it enters the multi-scale convolutional branch. The multi-scale convolutional branch includes three one-dimensional convolutional branches with kernel lengths of 3, 5, and 7 respectively. Each branch has two convolutional layers. In the three branches, the first branch has 16 kernels in its two convolutional layers, the second branch has 16 kernels in its two convolutional layers, and the third branch has 8 kernels in its two convolutional layers. The convolutional outputs are concatenated to form a 40-dimensional sequence feature. The 40-dimensional sequence feature is fused with the two-dimensional wall extension state in the convolutional fusion layer to obtain a 42-dimensional fused feature. The 42-dimensional fused feature is then sequentially entered into a fully connected layer with 24 neurons and a fully connected layer with 12 neurons to form a 12-dimensional evidence feature. The 12-dimensional evidence features are divided into two output heads. The evidence output layer has two neurons, which output the instability risk value and confidence level, respectively. The coverage output layer has one neuron, which outputs the required wall coverage. The improvement of the evidence temporal convolutional network backbone is that the joint channel layer first couples the sampling obstruction changes, and then the interpretation direction is defined by the wall extension state. It no longer attributes all changes in operating conditions to stratum changes, and distinguishes between "stratum densification" and "insufficient wall coverage" from the network structure.

[0121] The rolling temporal constraint branch receives instability risk values, confidence levels, retaining wall coverage requirements, secondary retaining wall extension status, and 6-dimensional action codes, forming an 11-dimensional decision input. This 11-dimensional decision input first enters an action level layer containing 12 neurons, then an action selection layer containing 6 neurons, outputting a three-dimensional action unit displacement level, with three components corresponding to small, medium, and large displacements, respectively. The rolling temporal planning layer filters action units based on the three-dimensional action unit displacement levels and the discrete retaining wall action set, generating a rolling retaining wall plan under the travel upper limit constraint. The improvement of the rolling temporal constraint branch lies in directly writing the discrete retaining wall action set into the model, rather than performing an empirical conversion outside the model. This allows the retaining wall coverage requirements output by the evidence-based temporal convolutional network backbone to directly apply to the actual achievable extension of the secondary and tertiary retaining walls, forming a scenario-based algorithm architecture for follow-up anti-collapse holes in retaining walls of deep, loose strata.

[0122] Based on the aforementioned innovative evidence-based temporal convolutional network architecture, in one exemplary implementation, the risk identification and demand calculation process of the evidence-based temporal convolutional network in this step is as follows: Figure 4 As shown, it includes the following sub-steps:

[0123] S31. Input the working condition sequence inside the hole into the backbone of the evidence time-series convolutional network, and set up a joint channel layer in the backbone of the evidence time-series convolutional network. The joint channel layer first couples the sampling torque and the servo motor current according to the time sequence position to form a joint channel, and retains the servo motor speed as an independent channel.

[0124] S32. Input the joint channel and independent channel into the multi-scale convolution branch. The multi-scale convolution branch includes three one-dimensional convolution branches. The kernel lengths of the three one-dimensional convolution branches are 3, 5 and 7 respectively. Each one-dimensional convolution branch performs two layers of convolution extraction in sequence, so that the short-term synchronous lifting information of the sampling torque and servo motor current and the continuous change information during the continuous advance process are jointly entered into the convolution output.

[0125] S33. The convolution outputs of the three one-dimensional convolution branches are concatenated to form a sequence feature, and the sequence feature is input into the convolutional fusion layer. At the same time, the wall extension state is input into the convolutional fusion layer as a conditional quantity, wherein the wall extension state includes the current total wall extension and the difference between the current advance and the front edge of the wall.

[0126] S34. Conditional fusion of sequence characteristics and wall extension state is performed so that the generation of instability risk evidence is simultaneously constrained by load changes and wall extension state, in order to distinguish between drilling load changes and limited changes caused by insufficient wall coverage.

[0127] S35. The fused features after condition fusion are sequentially input into two fully connected layers to form evidence features. The evidence features are then input into the evidence output layer and the coverage output layer respectively. The evidence output layer generates the amount of instability risk evidence that simultaneously includes instability risk value and credibility. The coverage output layer generates the amount of wall coverage requirement.

[0128] S36. Compare the instability risk value with the instability risk value threshold, compare the confidence level with the confidence level threshold, and determine the risk triggering flag when the instability risk value reaches the instability risk value threshold and the confidence level reaches the confidence level threshold.

[0129] S37. Based on the amount of evidence of instability risk, the state of wall extension, and the current advance, the wall coverage requirement is constrained and calculated so that the wall coverage requirement corresponds to the uncovered section at the front of the wall. The amount of evidence of instability risk, the risk triggering sign, and the wall coverage requirement are used as inputs to step S4 to form a single path from the working condition sequence in the borehole to the wall coverage requirement.

[0130] The innovation of this step lies not in directly connecting the evidence temporal convolutional network to the borehole condition sequence for risk identification, but in addressing the technical problem of "insufficient follow-up of the retaining wall length leading to instability in deep borehole segments" by integrating and defining the input organization, convolution fusion method, and output organization method of the evidence temporal convolutional network: First, the borehole condition sequence is received, which includes at least the sampling torque, servo motor current, and servo motor speed, and retains the temporal position corresponding to the retaining wall extension state; then, multiple convolutional receptive fields are used to extract the borehole condition sequence through convolution, enabling the evidence temporal convolutional network to simultaneously receive the synchronous rise information of the sampling torque and servo motor current during short-term changes, as well as the continuous change information of the sampling torque and servo motor current during continuous advance. Unlike the approach of directly splicing sampling torque, servo motor current, and servo motor speed as independent signals, this step combines sampling torque and servo motor current into a joint channel at the convolutional fusion layer. Then, the wall extension state is used as a conditional quantity for fusion. This allows the temporal convolutional network to distinguish between "drilling load changes" and "restricted changes caused by insufficient wall coverage" when calculating the instability risk evidence, rather than solely relying on load changes. This processing link moves the wall extension state directly into the instability risk evidence generation process, moving it from the control endpoint. This step further limits the instability risk evidence to include both an instability risk value and a confidence level. The instability risk value characterizes the degree of instability corresponding to the borehole working condition sequence, while the confidence level characterizes the degree of matching between the instability risk value and the wall extension state. Then, based on the instability risk value threshold and the confidence level threshold, the instability risk evidence is determined, a risk trigger flag is generated, and the wall coverage requirement is calculated based on the instability risk evidence and the wall coverage gap corresponding to the current drilling depth. Therefore, the amount of instability risk evidence, risk triggering indicators, and wall cover demand generated in this step are not general risk identification results, but rather control pre-information with wall extension state constraints. They can directly drive the rolling time domain solution, forming a single path from the borehole working condition sequence to the wall cover demand.

[0131] S4. When the risk triggering flag meets the preset conditions, a rolling retaining wall plan is generated by rolling time-domain solution based on the retaining wall coverage demand, retaining wall extension status and discrete retaining wall action set.

[0132] In one exemplary implementation, the rolling retaining wall plan generation process in this step is described in [reference needed]. Figure 5 This includes the following sub-steps:

[0133] S41. When the risk triggering flag meets the conditions, input the instability risk value, confidence level, wall cover demand, wall extension status, and discrete wall action set into the rolling time domain constraint branch, and convert the discrete wall action set into action codes corresponding to the achievable extension of each level of wall.

[0134] S42. Input the decision inputs into the action level layer and the action selection layer in sequence. The action level layer generates the action unit displacement level based on the instability risk value, confidence level and retaining wall coverage requirement. The action unit displacement level includes small displacement, medium displacement and large displacement.

[0135] S43. Based on the displacement level of the action unit, pre-select the action units in the discrete wall action set, and determine the reachability of the pre-selected action units based on the wall extension state and the upper limit of the stroke, so that the retained candidate action units do not exceed the reachable extension amount limited by the discrete wall action set.

[0136] S44. Perform rolling time-domain solution based on candidate action units, filter, combine and sort the candidate action units according to the control time, so that the rolling time-domain solution directly corresponds to the action units that the telescopic motor can execute, without continuous extension amount conversion;

[0137] S45. Perform cumulative coverage verification on the sorted candidate action units according to the wall protection coverage requirement, so that the cumulative extension corresponds to the wall protection coverage requirement, and generate a rolling wall protection plan under the condition of meeting the upper limit of the stroke constraint.

[0138] S46. The rolling retaining wall plan is classified according to the amount of evidence of instability risk. The instability risk value is used to determine the extension amplitude level of the action unit, and the confidence level is used to limit the switching conditions of the displacement level of the action unit.

[0139] S47. Determine the extension amplitude and action sequence of each action unit in the rolling retaining wall plan according to the displacement level of the action unit, so that the rolling retaining wall plan forms an action unit sequence that matches the retaining wall extension state and can directly generate retaining wall extension and retraction control commands.

[0140] The innovation of this step lies not in performing a general rolling time-domain solution for the wall extension and retraction, but in constraining the discrete wall action set, wall coverage requirement, wall extension state, and instability risk evidence within the same control link. This allows the rolling wall plan to directly correspond to the actual achievable extension of multi-level walls: First, a risk trigger flag is received. If the risk trigger flag meets the conditions, the wall coverage requirement, wall extension state, and discrete wall action set are received. Then, under the upper limit constraint of the stroke, the discrete wall action set is solved in the rolling time domain. Unlike the approach of directly outputting the continuous extension amount and then having it converted by the execution end, step four uses the discrete wall action set limited by the rack and pinion drive step distance as the calculation basis. The action units in the discrete wall action set are screened, combined, and sorted, so that the rolling wall plan corresponds to the action units that the telescopic motor can execute from the moment it is generated. This step further receives the instability risk evidence generated in step S3. Instead of simplifying the instability risk evidence to a single trigger switch, it performs displacement classification on the rolling retaining wall plan based on the instability risk evidence, mapping the rolling retaining wall plan to motion unit displacement levels. Then, it determines the extension amplitude and sequence of motion units based on the motion unit displacement levels. With this setup, the instability risk evidence participates in both the formation of the risk trigger flag and the amplitude calculation of the rolling retaining wall plan. The retaining wall coverage requirement participates in both the rolling time-domain solution and the combination constraints of the motion units, forming a continuously transmitted parameter chain. The rolling retaining wall plan obtained in step four is not an abstract displacement command, but rather a sequence of motion units within the upper limit of the stroke constraint, matching the retaining wall extension state, and driven by the instability risk evidence to determine the displacement levels. This sequence of motion units can directly generate retaining wall extension and retraction control commands, enabling the rolling time-domain solution and the mechanical boundaries of the multi-level retaining wall mechanism to correspond within the same step.

[0141] S5. Generate wall extension control commands according to the rolling wall plan, drive the extension motor of the retractable multi-stage wall mechanism to perform wall extension action, and perform position verification through the wall encoder, and update the wall extension status according to the verification result;

[0142] In one exemplary implementation, the wall protection execution and status update process in this step is described in [reference needed]. Figure 6 This includes the following sub-steps:

[0143] S51. Parse the sequence of action units in the rolling retaining wall plan into retaining wall extension and retraction control commands corresponding to each level of retaining wall, and write the extension amplitude of the action units into the telescopic motor control quantity according to the action sequence;

[0144] S52. The telescopic motor is driven by the wall extension control command to make the secondary wall and the tertiary wall extend sequentially according to the rolling wall plan, and to make the actual extension process consistent with the action unit defined by the discrete wall action set.

[0145] S53. During the operation of the telescopic motor, receive the positioning signal of the retaining wall encoder, and verify the corresponding action unit of the retaining wall encoder with the target action unit in the rolling retaining wall plan to form a positioning verification state;

[0146] S54. Update the current total wall extension based on the in-place verification status, and redetermine the difference between the current advance and the leading edge of the wall based on the current advance. Combine the updated current total wall extension and the difference between the current advance and the leading edge of the wall to form the wall extension status.

[0147] In this step, the rolling retaining wall plan is parsed into control commands adapted to each level of retaining wall, and the execution parameters of the telescopic motor are clarified. This ensures that the extension actions of the secondary and tertiary retaining walls strictly follow the limitations of the discrete retaining wall action set, avoiding execution deviations such as overtravel and undertravel. At the same time, relying on the real-time feedback of the position signal from the retaining wall encoder, precise verification is performed with the target action unit to ensure the accuracy of the action execution. Then, based on the verification results, the retaining wall extension status is dynamically updated, and the current total extension amount and advance of the retaining wall are updated synchronously with the difference between the current retaining wall and the leading edge of the retaining wall. This provides a real and accurate state benchmark for the identification of instability risks and action planning in the next control cycle, ensuring the continuity and reliability of the entire retaining wall control closed loop, and effectively avoiding the risk of hole wall instability and hole collapse caused by execution deviations or state lag.

[0148] S6. Use the updated wall extension state as the wall condition quantity for the next control moment.

[0149] In one exemplary implementation, the wall condition update process in this step is described in [reference needed]. Figure 7 This includes the following sub-steps:

[0150] S61. Receive the updated wall extension status after step S5, and extract the updated current total wall extension and the difference between the current advance and the leading edge of the wall;

[0151] S62. The updated current total wall extension and the difference between the current advance and the leading edge of the wall are used to form the wall condition quantity for the next control moment, and the wall condition quantity is kept in correspondence with the hole working condition sequence for the next control moment.

[0152] S63. Input the wall protection condition quantity into the convolutional fusion layer of the evidence temporal convolutional network backbone, so that the joint channel formed by the sampling torque and the servo motor current combines the wall protection condition quantity to generate the instability risk evidence quantity in the next control moment.

[0153] S64. The wall protection condition quantity is used as the state constraint input of step S4 to the rolling time domain constraint branch, and combined with the discrete wall protection action set to limit the reachable extension amount and stroke upper limit constraint of each level of wall protection.

[0154] In this step, the iterative update and closed-loop reuse of the wall extension state are completed. The wall parameters, after actual execution and on-site verification correction in the previous control cycle, are transformed into standardized wall condition quantities for the next control cycle, while strictly ensuring the mutual matching of time-series data. On the one hand, the wall condition quantities are connected to the convolutional fusion layer of the evidence temporal convolutional network backbone, continuously using the real-time wall coverage state to constrain the feature interpretation of changes in working conditions within the borehole, ensuring the continuity and accuracy of the generation of instability risk evidence quantities. On the other hand, the wall condition quantities are simultaneously introduced into the rolling temporal constraint branch as the state basis for determining the travel boundary and action reachability, and combined with the discrete wall action set to jointly constrain the extension and contraction range of each level of wall. Through the continuous iterative reuse of the wall state, this step enables risk identification, action planning, and mechanism execution to share a unified state benchmark, achieving dynamic adaptation of multi-level wall follow-up actions and continuous drilling conditions, forming a complete and sustainable closed-loop control logic, effectively improving the synergy and stability of wall protection during sampling operations in deep overburden loose strata.

[0155] Based on the sampling method provided by this invention, this invention directly addresses the technical problems of insufficient wall follow-up during sampling of deep overburden loose strata, which are difficult to identify in a timely manner and easily lead to instability, borehole collapse, and sample disturbance in deep borehole sections, through a closed-loop technical path of in-hole working condition evidence extraction, wall protection requirement discrimination, discrete action planning, and execution feedback update.

[0156] First, based on the current advance, sampling torque, servo motor current, servo motor speed, and wall encoder displacement, a sequence of working conditions within the borehole and the wall extension state are constructed. Under the condition of relying solely on available field signals and constrained by the upper limit of the multi-stage wall travel and rack step distance, an evidence-based temporal convolutional network is used to jointly model the sampling torque and servo motor current. The current total wall extension and the difference between the current advance and the wall leading edge are used as conditional variables in the convolutional fusion, ensuring that the interpretation of load changes is simultaneously constrained by the borehole obstruction mechanism and the wall coverage state. Subsequently, the instability risk value, confidence level, and wall extension state are output. The coverage requirement eliminates the need for empirical threshold conversion for the critical intermediate quantity of retaining wall coverage requirement. Instead, it can be reliably determined by combining risk evidence with the length of uncovered borehole segments. Based on this, the retaining wall coverage requirement and the discrete retaining wall action set are fed into the rolling time-domain constraint branch to directly generate action unit sequences corresponding to the reachable extension amounts of the secondary and tertiary retaining walls. Combined with the retaining wall encoder's on-time verification and updating of the retaining wall status, a continuous closed loop is formed from risk identification to retaining wall execution and then to status feedback. This enables the sampling task chain to have sustainable delivery capabilities in loose strata scenarios.

[0157] Compared to general wall control or drilling risk identification schemes, this invention makes targeted improvements to the algorithm structure and mechanical constraint coupling method for scenarios with deep, loose overburden formations: First, in terms of input organization, instead of simply splicing torque, current, and rotational speed, torque and current are coupled into a joint resistance representation, and the interpretation direction is limited by the wall extension state. This structurally distinguishes between changes in formation load and the restrictive changes caused by insufficient wall coverage, reducing misjudgments caused by mixed working condition mechanisms. Second, in terms of output organization, credibility is introduced, so that the instability risk value is no longer used as an isolated criterion, but jointly determines the wall coverage requirement with the degree of matching of the wall state, thus enabling the risk identification results to directly serve control. Third, in terms of action decision-making, the discrete wall action set is written into the model, so that the rolling time-domain solution directly filters and sorts the smallest action unit, remaining stroke, and current achievable coverage, avoiding the deviation caused by the further splitting of continuous displacement commands at the execution end. Fourth, the actual wall status after encoder positioning and verification is directly fed back to the next control cycle, so that risk evidence generation, action planning and mechanism execution share the same state benchmark, thereby improving the coordination between wall tracking and continuous sampling from a mechanism perspective.

[0158] Example:

[0159] This embodiment provides a sampling method for a retractable multi-stage retaining wall sampling device suitable for deep overburden loose strata, which includes the following implementation steps:

[0160] I. Construction of Operating Condition Sequence and Wall Condition:

[0161] set up Let the target hole depth be in length; To unify the control cycle; let the unified time axis be... ,in ;set up For the first The current advance at a unified control moment For the first The sampling torque at a unified control moment For the first The servo motor current at a unified control moment For the first The servo motor speed at a unified control moment For the first The wall encoder displacement at a unified control moment; let... For the first The current total extension of the retaining wall at a unified control moment. For the first The difference between the current advance and the leading edge of the retaining wall at a unified control moment; let... The protective wall is in the extended state; assuming These are the sampled values ​​of the working conditions inside the borehole; let them be... Let be a single time-series vector in the working condition sequence inside the borehole; For the calibration mapping of the wall encoder displacement to the current total wall extension; let... For the evidence temporal convolutional network backbone, Let be the network parameters of the evidence temporal convolutional network backbone; let For the rolling time-domain constraint branch, These are the decision parameters for the rolling time-domain constraint branch.

[0162] Let the current original sampling sequence be... The original sampling sequence of the sampling torque is The original sampling sequence of servo motor current is The original sampling sequence of servo motor speed is The original sampling sequence of the wall encoder displacement is .

[0163] In step S11, the unified timeline is... As a reference for all sampled quantities. For any physical quantity Define the matching index as and impose constraints ,in To allow matching time difference thresholds, the following conditions must be met. If the constraint holds, then take... If the constraint does not hold and Then take If the constraint does not hold and Then take the nearest neighbor. The original sampled values ​​as After this correspondence, we obtain This results in a set of borehole condition sampling values ​​being generated at the same control moment.

[0164] Let the window length be For the current unified control moment Define the discrete window as .when At that time, As a sequence of working conditions inside the borehole, each It only includes three operating condition characteristics: sampled torque, servo motor current, and servo motor speed. Defined as a temporal position binding sequence, where Working condition sequence inside the borehole according to Arranged in ascending order, with time-series positions bound to the sequence. The process is linked to the sequence of working conditions inside the hole, so that the wall extension state received by the subsequent convolutional fusion layer is consistent with the working condition characteristics inside the hole at the corresponding time.

[0165] Will Input calibration mapping The current total wall extension is obtained. The current drilling depth is constrained by the target hole depth to obtain... Then according to and Calculate the difference between the current advance and the leading edge of the retaining wall, specifically using the rule of first calculating the difference and then limiting the amplitude: If Then take ;like Then take .thus, It directly represents the length of the uncovered hole section and does not introduce invalid state quantities caused by exceeding the target hole depth or the front edge of the wall extending beyond the current advance.

[0166] Will and The retaining wall extends in a fixed sequence. .Will Working condition sequence inside the hole Common input This allows the combined channel formed by the sampling torque and the servo motor current to receive signals in the convolutional fusion layer. As a conditional quantity. The same enter This is used for subsequent upper limit constraints on the travel range and reachability determination of the discrete wall protection action set. Thus, steps S11 to S14 form a continuous input chain from the original sampling sequence, unified time axis, discrete window, in-hole working condition sequence to the wall protection extension state.

[0167] In this step, the relevant parameters such as target hole depth, current advance, sampling torque, servo motor current, servo motor speed, wall encoder displacement, current total wall extension, difference between current advance and wall leading edge, and the meaning and function of wall extension status are shown in Table 2.

[0168] Table 2 Parameter Information Table

[0169]

[0170] II. Generation of Discrete Wall Protection Action Sets:

[0171] At the moment of unified control Below, the upper limit of the secondary retaining wall stroke is set as follows: The maximum stroke of the third-level retaining wall is The second-level minimum action unit is The minimum action unit of level three is The displacement of the wall encoder is The current total extension of the retaining wall is The remaining journey for Level 2 is The remaining journey for Level 3 is The total remaining journey is The discrete wall protection action set is The current achievable coverage is Action coding is .

[0172] In step S21, first read the upper limit of the secondary retaining wall stroke. Upper limit of the third-level retaining wall travel Then, the step distances of the secondary and tertiary rack and pinion drives are read and solidified as the minimum secondary action units. With the third-level minimum action unit The solidification process here does not treat the step distance merely as an execution-end parameter, but directly uses it as the construction benchmark for the subsequent discrete wall protection action set, so that the action code received by the rolling time-domain constraint branch has mechanical executableness from the input side.

[0173] Displacement using wall-mounted encoder The current extension state of each level of the retaining wall is estimated. In this embodiment, the multi-level retaining wall adopts a mechanical sequence of extension of the second-level retaining wall first, followed by the third-level retaining wall. Therefore, the first level of the retaining wall is... Upper limit of secondary wall protection stroke Compare; when Not achieved When the current cumulative displacement is fully allocated to the secondary retaining wall, the current extension of the tertiary retaining wall is set to zero; when Reaching or exceeding At that time, the current extension of the secondary retaining wall is taken as Then exceed The remaining displacement is allocated to the third-level retaining wall, and is based on the upper limit of the third-level retaining wall's stroke. Cut off. Then, based on the current extension of the secondary wall and the current extension of the tertiary wall, the remaining stroke of the secondary wall is calculated. With Level 3 current remaining journey Then, the remaining journey at level two. With Level 3 current remaining journey The total remaining journey is obtained by summing the results. The remaining journey obtained from this directly corresponds to the current organizational state, and no longer depends on reverse calculation after execution.

[0174] The remaining journey at level two According to the second-level smallest action unit Divide the current remaining journey of level three into segments. Based on the three-level minimum action unit The system is segmented into sections, and a discrete set of wall protection actions is generated according to the following formula. :

[0175] ;

[0176] In the formula, Indicates unified control time The discrete set of wall protection actions; This represents a single candidate action value formed by the combination of secondary and tertiary wall protection action units; Indicates the index of the number of combinations of secondary wall protection action units; Indicates the index of the number of combinations of three-level wall protection action units; Indicates a unified control time index; This represents the second-level smallest action unit; This represents the three-level smallest action unit; Indicates the current remaining journey time at level two; Indicates the current remaining journey for Level 3; Indicates the total remaining trip; This indicates the current total extension of the retaining wall; Indicates the upper limit of the secondary wall protection stroke; This represents the upper limit of the third-level retaining wall travel. Through this formula, the current remaining travel of the second-level retaining wall, the current remaining travel of the third-level retaining wall, the minimum action unit, and the total travel boundary are simultaneously written into the construction process of the discrete retaining wall action set, making the discrete retaining wall action set... The actual combination of step lengths that the telescopic motor can execute is determined from the time of generation, and the achievable extension of the secondary and tertiary retaining walls is limited.

[0177] Based on discrete wall protection action set Scan all candidate action values ​​and determine the largest candidate action value that satisfies the constraints as the current achievable coverage. When the total remaining journey When there is a tail segment that cannot be divided by the smallest action unit, the current achievable coverage is... Take the discrete wall protection action set The maximum candidate action value is used instead of directly taking the total remaining stroke. This will enable the current achievable coverage. Maintain the same discrete baseline as subsequent rolling wall protection plans. Then, calculate the remaining stroke of the second stage. Level 3 Current Remaining Itinerary Level 2 minimum action unit Level 3 minimum action unit Total remaining trip and current achievable coverage Action codes are composed in a fixed order. This action is encoded. The rolling temporal constraint branch is input together with the wall extension state, so that the wall coverage requirement output by the evidence temporal convolutional network backbone can be directly mapped to the actual achievable extension of the secondary and tertiary walls.

[0178] III. Risk Identification and Coverage Requirement Calculation:

[0179] Let the index of the current control time be Discrete window denoted as Let the working condition sequence inside the hole be as follows: ,in , Indicates the sampling torque. Indicates the servo motor current. This indicates the servo motor speed. Let the extended guard wall state be... ,in This indicates the current total extension of the retaining wall. This represents the difference between the current advance and the leading edge of the retaining wall. Let the backbone of the evidence temporal convolutional network be... , The network parameters represent the backbone of the temporal convolutional network for evidence. Let the sequence features be... The fusion feature is The characteristics of the evidence are The instability risk value is Credibility is The amount of evidence for the risk of instability is The original wall cover requirement given by the output layer is: The required amount of retaining wall coverage after constraint is The risk trigger sign is The instability risk threshold is The credibility threshold is and limited , , , Values ​​in the interval .

[0180] The working condition sequence inside the hole Input evidence temporal convolutional network backbone A joint channel layer is set on the input side of the evidence temporal convolutional network backbone, based on the same discrete index. The sampling torque and servo motor current are position-binded to form a joint channel. ,in The servo motor speed is kept as an independent channel. ,in This input organization method does not use a direct parallel splicing of the three operating condition characteristics. Instead, it first uses the sampled torque and servo motor current as coupled inputs for the same obstructed process, and then retains the servo motor speed as an independent input for the propulsion and release changes.

[0181] Joint Channel and independent channels Synchronous input multi-scale convolution branch. The multi-scale convolution branch contains three one-dimensional convolution branches, denoted as... , and The convolutional kernel lengths are 3, 5, and 7, respectively. Each one-dimensional convolutional branch performs two layers of convolution extraction in a fixed order. The branch with a kernel length of 3 is used to extract short-term synchronous rise information of sampling torque and servo motor current within adjacent control moments; the branch with a kernel length of 5 is used to extract mid-range variation information during continuous advance; and the branch with a kernel length of 7 is used to extract continuous variation information within a longer control interval. All three branches retain the independent channel response corresponding to the servo motor speed during convolution.

[0182] The outputs of the second layer convolution of the three one-dimensional convolutional branches are concatenated according to their branches to form sequence features. In this embodiment, the three branches are concatenated to form a 40-dimensional sequence feature. The sequence feature is then... Input the convolutional fusion layer and extend the wall state. As conditional inputs to the convolutional fusion layer, fused features are obtained. In this embodiment, the 40-dimensional sequence features are spliced ​​with the two-dimensional wall extension states to form a 42-dimensional fused feature. This conditional fusion ensures that the combined change in sampling torque and servo motor current is first affected by the current total wall extension amount before entering the evidence layer. and the difference between the current advance and the wall leading edge The constraints allow for the differentiation between variations in drilling load and limited variations caused by insufficient wall coverage within the same parameter link.

[0183] Fusion features The data is sequentially input into two fully connected layers, with the first layer having 24 neurons and the second layer having 12 neurons, to obtain evidence features. Then the characteristics of the evidence. Distributed to two output heads. The evidence output layer receives evidence features. It also outputs the instability risk value. and credibility This constitutes the amount of evidence for the risk of instability. The overlay output layer receives the same evidence features. Output the original wall protection coverage requirement. Then adjust the instability risk value. With instability risk threshold Compare and assess credibility With credibility threshold Compare; when and At that time, identify risk trigger signs. Otherwise, determine the risk trigger flag. .

[0184] The demand for wall protection coverage As a calculation item that directly drives step four from step three, and based on the amount of evidence for instability risk. Wall extension state The length of the uncovered hole corresponding to the current advance. Requirements for original retaining wall coverage The constraint calculation is performed using the following formula:

[0185] ;

[0186] In the formula, This indicates the required amount of wall protection covering; This indicates the difference between the current advance and the leading edge of the retaining wall; This indicates the initial wall covering requirement; Indicates the risk value of instability; Indicates the threshold value for instability risk; Indicates credibility; Indicates the credibility threshold; This indicates that the original wall covering requirement is limited to a non-negative value; This indicates that the required amount of retaining wall coverage will be limited to the uncovered hole section at the front edge of the retaining wall; This represents the normalized trigger strength after the instability risk value exceeds the instability risk value threshold. This represents the normalized matching strength after the confidence level exceeds the confidence threshold. Using this formula, the original retaining wall coverage requirement is calculated. First, the length of the uncovered hole section Perform boundary truncation, then use the instability risk value With credibility Amplitude constraints are jointly applied to determine the required amount of retaining wall coverage. Simultaneously satisfying the conditions for uncovered hole segment boundaries and evidence matching. From this formula, it can be seen that when... or At that time, the required amount of retaining wall covering Automatic zeroing, corresponding to risk trigger flags The state; when and At that time, the required amount of retaining wall covering Within the uncovered aperture section, the instability risk value and confidence level change synchronously. Finally, the amount of evidence for instability risk is... Risk trigger signs and the demand for wall covering They are sent together to step four and transmitted in the same parameter link as the wall extension state and the action code of the discrete wall action set, forming a single path from the in-hole working condition sequence to the wall coverage requirement.

[0187] The structural composition of the evidence temporal convolutional network backbone in this step is shown in Table 3.

[0188] Table 3. Evidence Temporal Convolution Backbone Structure Information Table

[0189]

[0190] IV. Generation of Rolling Wall Protection Plan:

[0191] Let the index of the current control time be The risk trigger sign is The instability risk value is Credibility is The required amount of retaining wall covering is The wall extension state is ,in This indicates the current total extension of the retaining wall. This represents the difference between the current advance and the leading edge of the retaining wall. Let the discrete retaining wall action set be... ,in This indicates a second-level minimum action unit. With the third-level minimum action unit A single action unit formed by combination, This indicates the number of the second-level minimum action units. This represents the number of the three-level minimum action units. Let the action code be... ,in Indicates the current remaining journey for Level 2. Indicates the current remaining journey for Level 3. Indicates the total remaining trip. This represents the currently achievable coverage. Let the rolling time-domain constraint branch be... ,in The parameters represent the rolling time-domain constraint branches; let the decision input be... Let the displacement level vector of the action unit be... ,in For small displacements, Corresponding displacement, Corresponding to large displacements; let the set of candidate action units be... Assume the rolling retaining wall plan is as follows: ,in Indicates the first Each planned action unit, This indicates the number of action units in the rolling wall protection plan at the current control moment.

[0192] When the risk trigger flag is met At that time, the risk value of instability will be... Credibility Demand for retaining wall covering Wall extension state and action coding Common input rolling temporal constraint branch Discrete wall protection action set Each action unit has been associated with a corresponding action unit during generation. and The binding, therefore, the rolling time-domain constraint branch receives the discrete wall action set. At that time, directly based on Read the number of second-level and third-level minimum action units, and match the action units with the action codes. The remaining stroke information corresponds to the reachable extension of each action unit from the moment it enters the branch.

[0193] Input the decision First, input the action level layer, then input the action selection layer. The action level layer has twelve neurons, and the action selection layer has six neurons. The action level layer uses an instability risk value. Credibility and the demand for wall covering As a driving force, in the state of wall extension and action coding As a constraint, the motion amplitude at the current control moment is hierarchically represented; the motion selection layer further compresses the output of the motion level layer to form the motion unit displacement level vector. .Will , and The components with the largest values ​​are compared, and the displacement level corresponding to the component with the largest value is taken as the initial motion unit displacement level at the current control moment.

[0194] First, consider the discrete wall protection action set. Each action unit in Calculate the displacement ratio Then set the displacement segment threshold. and ,in Less than When the displacement ratio is between zero and... When the interval is large, the motion unit is assigned to the small displacement subset; when the displacement ratio is greater than a certain value... and not greater than When the interval is large, the motion unit is assigned to the middle displacement subset; when the displacement ratio is greater than 100, the motion unit is assigned to the middle displacement subset. When the range is no greater than one, the action unit is classified into the large displacement subset. Based on the initial action unit displacement level, pre-selected action units are retrieved from the corresponding subset. Then, reachability is determined for the pre-selected action units, specifically according to the action unit's bound... and Check separately Does it exceed the current remaining journey of Level 2? ,examine Does it exceed the current remaining journey of Level 3? Check if the motion unit exceeds the total remaining stroke. And check the current total wall extension. Whether the combined stroke of this action unit exceeds the sum of the upper limits of the secondary and tertiary retaining walls is considered. The action units retained after the above determination form a candidate action unit set. .

[0195] Let the rolling planning window be ,in Indicates the length of the rolling plan; let the cumulative extension be... .when And the candidate action unit set When not empty, use the candidate action unit set Perform a rolling time-domain solution based on the above formula, and determine the rolling wall protection plan according to the following formula. :

[0196] ;

[0197] In the formula, Indicates a rolling retaining wall plan; This represents the sequence of action units that minimizes the objective function. This indicates that each planned action unit in the rolling retaining wall plan is selected from the candidate action unit set. Selected from; Indicates the first Each planned action unit; Indicates the first Each planned action unit is specified. ; This indicates the number of action units in the rolling wall protection plan at the current control moment; Indicates the preceding The cumulative extension of each planned action unit; This indicates the summation index in the cumulative outgoing quantity; This indicates the index of the action unit in the rolling retaining wall plan; This indicates the required amount of wall protection covering; Indicates the current achievable coverage; This indicates the difference between the current advance and the leading edge of the retaining wall; Indicates the total remaining trip; This represents the second-level smallest action unit; This represents the three-level smallest action unit; Indicates the risk value of instability; The reliability level is indicated by the following technical approach: The first term constrains the deviation between the cumulative extension and the required wall coverage; the second term constrains the portion of the cumulative extension exceeding the uncovered section at the front edge of the wall; the third term binds the cumulative extension to the total remaining travel; the fourth term uses the instability risk value to limit the extension amplitude level of a single planned action unit; and the fifth term uses reliability to limit amplitude jumps between adjacent planned action units. Therefore, the rolling time-domain solution is completed directly at the candidate action unit level, without continuous extension conversion.

[0198] The rolling retaining wall plan obtained from the above formula Perform cumulative coverage verification according to the action unit sequence. In specific processing, follow... Calculate the cumulative extension amount in ascending order and will Demand for wall covering Comparison. When First time reaching or exceeding At that time, intercept from to The sequence of prefix action units serves as the effective rolling wall plan for the current control moment; when the cumulative total of all planned action units still does not reach the target... At that time, the entire rolling retaining wall plan This was determined to be a valid rolling retaining wall plan. The cumulative coverage check is always performed on the total remaining travel. Level 2 Current Remaining Itinerary Level 3 Current Remaining Itinerary and current achievable coverage Execute within the specified scope.

[0199] Based on the amount of evidence for instability risk Displacement classification is performed on the effective rolling retaining wall plan. In specific handling, the instability risk value is... As the primary driving force for the extension amplitude level of the action unit, it prioritizes retaining large or medium displacement action units during control moments with higher instability risk, and prioritizes retaining small displacement action units during control moments with lower instability risk. A confidence switching threshold is then set. and ,in Not less than When the initial motion unit displacement level is medium displacement and the confidence level is... Below When the initial motion unit displacement level is large displacement and the confidence level is high, the medium displacement motion unit is reduced to a small displacement motion unit; when the initial motion unit displacement level is large displacement and the confidence level is high... Below At that time, the large displacement motion unit is reduced to the medium displacement motion unit.

[0200] The extension amplitude and sequence of each action unit in the rolling retaining wall plan are determined based on the final displacement level of the action unit. For any planned action unit in the rolling retaining wall plan, if the planned action unit corresponds to a small displacement, it is deployed as a single or a small number of minimum action units; if the planned action unit corresponds to a medium displacement, it is deployed as a combination of action units at the intermediate level; if the planned action unit corresponds to a large displacement, it is deployed as a combination of action units at a higher level. For the secondary and tertiary minimum action units within each planned action unit, they are deployed in the order of secondary minimum action units first, followed by tertiary minimum action units, so that the sequence of actions is consistent with the actual extension sequence of the multi-stage retaining wall. The final rolling retaining wall plan and retaining wall extension state are thus formed. It matches and can directly generate wall extension and retraction control commands.

[0201] For the decision-making process of the wall protection action in this step, please refer to Table 4.

[0202] Table 4. Wall Protection Action Decision-Making Process Data Table

[0203]

[0204] V. Wall Protection Implementation and Status Update:

[0205] Let the index of the current control time be The rolling retaining wall plan is ,in Indicates the first Each action unit, This indicates the number of action units at the current control moment. Let the wall extension / retraction control command be... The telescopic motor control quantity is The displacement of the wall encoder is , No. The displacement of the wall encoder corresponding to the starting point of each action unit is: , No. The displacement of the retaining wall encoder corresponding to the target executed by each action unit is: The current displacement of the retaining wall encoder during the execution process is The wall encoder position signal is The verification status is as follows: The updated total wall extension is The updated difference between the current advance and the wall leading edge is The updated wall extension state is .

[0206] For rolling wall protection plan The sequence of action units in the code is parsed sequentially according to the order of the actions. Because each action unit... All parameters originate from the discrete wall protection action set, and the number of second-level and third-level minimum action units are already bound during generation. Therefore, during parsing, continuous extension amounts are no longer generated; instead, the corresponding second-level and third-level minimum action unit segments are directly read. These segments are then expanded into wall protection extension / retraction control commands. and in accordance with the first Each action unit in the rolling retaining wall plan The order of arrangement in the data will be used to write the corresponding extension range into the telescopic motor control quantity. When a certain action unit contains only the second-level minimum action unit segment, only the control segment corresponding to the second-level shear wall is written; when a certain action unit contains both the second-level and third-level minimum action unit segments, the control segment corresponding to the second-level shear wall is written first, followed by the control segment corresponding to the third-level shear wall, so that the shear wall extension and retraction control command is executed. At the command level, this is consistent with the actual extension sequence of the multi-stage retaining wall.

[0207] According to the rolling wall protection plan The sequence of the action units will control the wall extension and retraction. Send the commands one by one to the telescopic motor for execution. The telescopic motor executes the first... When executing a single action unit, the secondary retaining wall is first extended according to the secondary minimum action unit segment bound within that action unit. Then, the tertiary retaining wall is extended according to the tertiary minimum action unit segment bound within the same action unit, thus causing the secondary and tertiary retaining walls to extend sequentially according to the rolling retaining wall plan. Since the extension amplitude of each action unit is composed of the secondary and tertiary minimum action units, the actual execution process of the telescopic motor is always limited by the discrete retaining wall action set. The execution end no longer further breaks down the continuous displacement command, ensuring that the actual extension process is consistent with the action units limited by the discrete retaining wall action set.

[0208] During each action unit performed by the telescopic motor, the displacement of the retaining wall encoder is continuously received, and the displacement is recorded at the first... The displacement of the wall encoder at the start of each action unit is used as... According to the first The second-level and third-level minimum motion unit segments bound within each motion unit determine the wall encoder displacement that should be achieved after the completion of that motion unit. During execution, the current wall encoder displacement is read in real time. ,when achieve At that time, a wall encoder positioning signal is generated. Then, the wall encoder is positioned. With the Each action unit performs a corresponding verification on its target action unit; when the displacement endpoint corresponding to the positioning signal matches the target displacement of the target action unit, the positioning verification status is determined. To ensure the target displacement is reached; if the target displacement is not reached, determine the positioning verification status. This is not yet in place. The verification process uses action units as the corresponding granularity to ensure that the object of execution verification is consistent with the target action unit in the rolling retaining wall plan.

[0209] Based on the status of verification Update the current total wall extension. Once the last calibrated actuator unit is determined, write the wall encoder displacement corresponding to that actuator unit into the updated current total wall extension. Then, the difference between the current advance and the leading edge of the wall is re-determined based on the current advance. Specifically, the current advance is taken as the position of the leading edge in the hole, and the updated total wall protrusion is used as the reference value. As the leading edge of the retaining wall, when the current advance exceeds the updated total extension of the current retaining wall. At that time, the difference between the two is determined as the updated difference between the current advance and the wall leading edge. When the current advance is not greater than the updated total wall extension. At that time, the difference between the current advance and the wall leading edge will be updated. The value is set to zero. Finally, the updated total wall extension is... And the updated difference between the current advance and the wall leading edge Forming the updated protective wall extension state This establishes the wall extension state based on the actual execution results of the action unit and the wall encoder positioning verification results, and serves as the state basis in the next control link.

[0210] VI. Reinjection under wall protection conditions:

[0211] Let the wall extension state after step five be: ,in This indicates the updated total wall extension. This represents the updated difference between the current advance and the wall leading edge; let the next control time be... Let the wall protection condition quantity at the next control moment be... Let the sequence of working conditions inside the borehole at the next control moment be: ,in , , Indicates the sampling torque. Indicates the servo motor current. Represents the servo motor speed; assuming the backbone of the evidence temporal convolutional network is... The rolling time-domain constraint branch is The discrete wall protection action set is Action coding is .

[0212] Receive the updated wall extension status given in step five. and to The two state components are directly extracted without introducing new state explanatory variables. In specific processing, [the following will be done]: As the current position of the wall protection front, As the length of the currently uncovered hole segment. Because This data originates from the actual execution results after the wall encoder has been checked for position. The two state components are derived from the re-correspondence between the current advance and the leading edge position of the retaining wall. Therefore, these two state components constitute the sole retaining wall state basis for continuing to participate in risk identification and rolling solution in the next control cycle.

[0213] Will and The wall protection conditions are arranged in a fixed order. Wall protection conditions It is not used independently of the timing sequence, but rather in conjunction with the in-hole condition sequence at the next control time. Maintain correspondence. In specific processing, proceed according to the unified control cycle. After that, with Construct a discrete window as the end of the sequence. and make the wall protection conditions more efficient. Bind to End control time Therefore, the sequence features entering the evidence temporal convolutional network backbone at the next control moment and the wall protection state come from the same control context, and the wall protection state does not lag behind the working condition sequence inside the hole.

[0214] The working condition sequence inside the hole Input evidence temporal convolutional network backbone Furthermore, within the joint channel layer, the sampled torque is coupled with the servo motor current to form a joint channel, while the servo motor speed is retained as an independent channel. Subsequently, sequence features are obtained through multi-scale convolutional branches, and then the wall protection condition quantity is... The data is fed into the convolutional fusion layer. The convolutional fusion layer will then incorporate the wall protection condition. Synchronous fusion with sequence features ensures that the generation process of instability risk evidence at the next control time step is simultaneously constrained by load changes and the wall state. After subsequent fully connected layers, evidence output layers, and coverage output layers, the instability risk value, confidence level, and wall coverage requirement at the next control time step are obtained. Therefore, the wall state updated in step five is not used retroactively at the execution end, but is directly incorporated into the instability risk evidence generation chain at the next control time step.

[0215] The same wall protection conditions were measured. As the state constraint input for step four, the rolling time-domain constraint branch and the discrete wall protection action set These constraints collectively define the achievable extension and upper limit of stroke for each level of the retaining wall. In specific implementation, this is determined according to... Re-map the current remaining stroke at level two, level three, and total remaining stroke, and then combine the re-mapped remaining stroke with the discrete retaining wall action set. The action units in the process are matched one by one, and action units that would exceed the remaining travel of each level or the total travel boundary after execution are removed, retaining a subset of action units that meet the current wall state; then, the current achievable coverage at the next control moment is determined based on the retained subset of action units, and action codes are formed. After that, the wall protection conditions were measured. With action coding Both are fed into the rolling time-domain constraint branch This ensures that when step four performs action level determination, candidate action unit screening, and rolling wall plan generation at the next control moment, the state constraints it relies on are directly derived from the actual execution results of step five, thus forming a continuous closed-loop parameter transmission link between control cycles.

[0216] To further demonstrate the technical effectiveness of this solution, this embodiment provides relevant effect diagrams: where, Figure 8 The diagram shows a comparison between the combined working condition spatial distribution and the confidence ellipse. Under the same sampling torque and servo motor current, without considering the wall extension state, there is significant overlap between samples corresponding to normal drilling load changes and insufficient wall follow-up. Isodensity lines and confidence ellipses overlap considerably, and the state boundaries are unclear. However, after considering the wall extension state, the two types of sample clusters are clearly separated, with instability-related samples becoming more concentrated and the discrimination boundary becoming more defined. This demonstrates that the present invention does not rely solely on a single load change for judgment, but rather jointly models the in-hole working condition sequence with the wall extension state. This effectively reduces interference from mixed working conditions, improves the accuracy and lead time for instability risk identification, and provides a more reliable basis for generating subsequent wall coverage requirements.

[0217] Figure 9A color-coded equivalent diagram illustrating the wall cover requirement demonstrates the continuous distribution relationship between instability risk and wall cover requirement within a state space defined by the difference between the current advance and the wall leading edge, and the sampling torque. The left diagram shows the change in instability risk under different working conditions, while the right diagram shows the change in wall cover requirement resulting from the corresponding risk state. As the difference between the current advance and the wall leading edge increases and the sampling torque rises, the color in the diagram gradually darkens, indicating a synchronous increase in wall cover requirement. It can be seen that this invention not only identifies instability risk but also transforms risk assessment into an actionable wall cover requirement, achieving a continuous connection from risk identification to support adjustment, improving the timeliness, targeting, and stability of wall support follow-up, and avoiding insufficient or excessive support.

[0218] Although embodiments of the present invention have been described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the present invention, and all such changes and alterations shall not depart from the protection scope of the present invention.

Claims

1. A sampling method, applied to a retractable multi-stage wall-supported sampling device suitable for deep overburden loose strata, the device comprising: Sampling drive mechanism, retractable multi-stage protective wall mechanism, sampling mechanism, sealing mechanism and control module; The sampling drive mechanism is located at the top of the device and is connected to the first-stage protective wall tube in the retractable multi-stage protective wall mechanism, and forms a transmission connection with the sampling mechanism. The sampling drive mechanism includes a servo motor, a transmission connector and a torque sensor, which are used to provide sampling propulsion power and sampling rotation power to the sampling mechanism, and to collect sampling torque signals in real time. The retractable multi-stage wall protection mechanism is sleeved around the sampling mechanism and includes a primary wall protection tube, a secondary wall protection tube, a tertiary wall protection tube, a telescopic drive assembly, and a wall protection encoder; it is used to perform graded wall protection and follow-up support on the borehole wall during drilling, and output wall extension amount and displacement feedback signals; the telescopic drive assembly includes a telescopic motor. The sampling mechanism is fitted inside the retractable multi-stage retaining wall mechanism and is used to drill and collect soil samples in the thick overburden loose strata, and to provide feedback on the sampling load change signal. The sealing mechanism is located at the bottom of the sampling mechanism and is used to maintain the bottom of the sampling mechanism in a sealed state during the sampling process; The control module is electrically connected to the servo motor and torque sensor in the sampling drive mechanism, the telescopic motor and wall encoder in the telescopic multi-stage wall protection mechanism, and the sealing mechanism, respectively, and is used to receive the working condition data in the hole and the wall displacement data, and output the instability risk identification result and the wall telescopic control command. The sampling method is characterized by comprising: S1. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, construct the hole working condition sequence according to the time sequence, and determine the wall extension state based on the collected data; S2. Based on the upper limit of the stroke of the multi-stage retaining wall and the rack and pinion drive step distance, generate a set of discrete retaining wall actions that adapt to the execution capability of the mechanism; S3. Input the working condition sequence inside the hole into the evidence temporal convolutional network, and after feature extraction and fusion calculation, output the amount of evidence of instability risk, risk triggering flags, and the amount of wall covering required; S4. When the risk triggering flag meets the preset conditions, a rolling retaining wall plan is generated by rolling time-domain solution based on the retaining wall coverage demand, retaining wall extension status and discrete retaining wall action set. S5. Generate wall extension control commands according to the rolling wall plan, drive the extension motor of the retractable multi-stage wall mechanism to perform wall extension action, and perform position verification through the wall encoder, and update the wall extension status according to the verification result; S6. Use the updated wall extension state as the wall condition quantity for the next control moment.

2. The sampling method as described in claim 1, characterized in that, Step S1 specifically includes: S11. Obtain the target hole depth, current advance, sampling torque, servo motor current, servo motor speed and wall encoder displacement, and align these parameters at a unified control time. S12. Extract the sampled torque, servo motor current and servo motor speed within the most recent continuous control moment, arrange them in chronological order to form a sequence of working conditions inside the hole, and retain the timing position of each control moment corresponding to the displacement of the advance scale and the wall encoder. S13. Determine the current total extension of the wall based on the wall encoder displacement, determine the difference between the current advance and the front edge of the wall based on the current advance and the current total extension of the wall, and limit the value boundary of the current advance with the target hole depth; S14. The current total wall extension and the difference between the current advance and the leading edge of the wall are used to form the wall extension status.

3. The sampling method as described in claim 1, characterized in that, Step S2 specifically includes: S21. Obtain the upper limit of the second-stage wall protection stroke, the upper limit of the third-stage wall protection stroke, and the rack and pinion drive step distance, and determine the rack and pinion drive step distance as the second-stage minimum action unit and the third-stage minimum action unit, respectively; S22. Determine the current remaining stroke of the secondary wall encoder, the current remaining stroke of the tertiary wall encoder, and the total remaining stroke based on the wall encoder displacement, the upper limit of the secondary wall stroke, and the upper limit of the tertiary wall stroke; S23. The second-level current remaining stroke, the third-level current remaining stroke, the second-level minimum action unit, and the third-level minimum action unit are discretized to form action units with corresponding telescopic motor execution steps. The combinable range of action units is limited according to the total remaining stroke and the current total extension of the wall, and a discrete wall action set is generated. S24. Determine the current reachable coverage based on the discrete wall protection action set, and combine the second-level current remaining travel, the third-level current remaining travel, the second-level minimum action unit, the third-level minimum action unit, the total remaining travel, and the current reachable coverage into an action code.

4. The sampling method as described in claim 1, characterized in that, In step S3, the evidence temporal convolutional network includes an evidence temporal convolutional network backbone and a rolling temporal constraint branch; The evidence temporal convolutional network backbone includes a joint channel layer, multi-scale convolutional branches, a convolutional fusion layer, a two-level fully connected layer, an evidence output layer, and a cover output layer. The combined channel layer combines the sampled torque and the servo motor current into a combined channel, while retaining the servo motor speed as an independent channel; The multi-scale convolutional branch includes three one-dimensional convolutional branches with kernel lengths of 3, 5, and 7 respectively. Each branch has two convolutional layers, and the convolutional outputs of each branch are concatenated to form sequence features. The convolutional fusion layer conditionally fuses the sequence features with the wall extension state to obtain fused features; The fusion features are formed as evidence features after passing through two fully connected layers. The evidence output layer is used to output the instability risk value and credibility. The coverage output layer is used to output the required amount of wall protection coverage; The rolling time-domain constraint branch includes an action level layer, an action selection layer, and a rolling time-domain planning layer; The rolling time-domain constraint branch receives instability risk value, confidence level, retaining wall coverage requirement, second retaining wall extension status and 6-dimensional action code to form 11-dimensional decision input; The 11-dimensional decision input is sequentially input into an action level layer containing 12 neurons and an action selection layer containing 6 neurons, and outputs the three-dimensional action unit displacement level corresponding to small displacement, medium displacement, and large displacement. The rolling time-domain planning layer filters action units based on the displacement level of the three-dimensional action unit and the discrete retaining wall action set, and generates a rolling retaining wall plan under the upper limit constraint of the stroke, so that the retaining wall coverage requirement directly corresponds to the actual achievable extension of the secondary and tertiary retaining walls.

5. A sampling method as described in claim 4, characterized in that, Step S3 specifically includes: S31. Input the working condition sequence inside the hole into the backbone of the evidence temporal convolutional network, and couple the sampling torque and servo motor current according to the temporal position through the joint channel layer to form a joint channel, and retain the servo motor speed as an independent channel; S32. Input the joint channel and the independent channel into the multi-scale convolution branch for convolution extraction, and then concatenate to obtain the sequence features; S33. Conditionally fuse the sequence features and the wall protrusion state in the convolutional fusion layer to obtain the fused features; S34. The fused features are passed through two fully connected layers to generate evidence features, then the evidence output layer generates the amount of instability risk evidence, and the coverage output layer generates the amount of wall coverage demand. S35. Compare the instability risk value and confidence level with the corresponding thresholds respectively, and determine the risk triggering flag when the double threshold condition is met; S36. Calculate the required amount of retaining wall coverage based on the amount of evidence of instability risk, the retaining wall extension status, and the current advance constraints.

6. The sampling method as described in claim 1, characterized in that, Step S4 specifically includes: S41. When the risk trigger flag meets the conditions, input the instability risk value, confidence level, wall cover requirement, wall extension status and discrete wall action set into the rolling time domain constraint branch, and convert the discrete wall action set into action codes for the corresponding wall extension amounts at each level. S42. The decision input is sequentially sent to the action level layer and the action selection layer. Based on the instability risk value, confidence level and wall cover requirement, the action unit displacement level including small displacement, medium displacement and large displacement is generated. S43. Pre-select the action units in the discrete wall action set according to the displacement level of the action unit, and combine the wall extension state and the upper limit of the stroke to complete the reachability determination and obtain the candidate action units; S44. Based on the candidate action units, perform rolling time-domain solution, filter, combine and sort the action units, and directly generate an action unit sequence adapted to the telescopic motor. S45. Based on the wall protection coverage requirement, perform cumulative coverage verification on the candidate action units, and determine the rolling wall protection plan under the travel limit constraint; S46. Based on the amount of evidence of instability risk, the rolling retaining wall plan is classified into displacement levels, the instability risk value is used to determine the extension amplitude level, and the confidence level is used to limit the conditions for switching displacement levels. S47. Determine the extension amplitude and action sequence of each action unit according to the displacement level of the final action unit to form an action unit sequence that can directly generate wall extension and retraction control commands.

7. A sampling method as described in claim 1, characterized in that, Step S5 specifically includes: S51. Parse the sequence of action units in the rolling retaining wall plan into retaining wall extension and retraction control commands corresponding to each level of retaining wall, and write the extension amplitude of each action unit into the telescopic motor control quantity according to the action sequence. S52. Drive the telescopic motor to execute the wall extension control command, so that the secondary wall and the tertiary wall extend sequentially according to the rolling wall plan, and the actual extension action is consistent with the action unit defined by the discrete wall action set; S53. During the operation of the telescopic motor, the positioning signal fed back by the wall encoder is received, and the positioning signal is checked against the target action unit in the rolling wall plan to form a positioning check state; S54. Update the current total wall extension based on the in-place verification status, recalculate and determine the difference between the current advance and the leading edge of the wall, and combine the updated current total wall extension and the difference between the current advance and the leading edge of the wall to form the wall extension status.

8. The sampling method as described in claim 1, characterized in that, Step S6 specifically includes: S61. Receive the updated wall extension status after step S5, and extract the updated current total wall extension and the difference between the current advance and the wall leading edge; S62. Integrate the updated current total wall extension with the current advance and the difference between the wall leading edge and the current advance into the wall condition quantity for the next control moment, so that the wall condition quantity corresponds to the time sequence of the working condition sequence in the hole for the next control moment. S63. Input the wall protection condition quantity into the convolutional fusion layer of the evidence temporal convolutional network backbone, so that the next control time step generates the instability risk evidence quantity based on the wall protection condition quantity and the working condition data. S64. The wall protection condition quantity is used as the state constraint input for the rolling time domain constraint branch, and combined with the discrete wall protection action set to limit the reachable extension and stroke limit of each level of wall protection.

Images