A roadway surrounding rock intelligent perception and dynamic support decision system and method

Abstract

This invention relates to the field of intelligent mining and rock strata control technology, and discloses an intelligent sensing and dynamic support decision-making system and method for roadway surrounding rock. The system includes an underground sensing and interactive execution layer, an edge decision server, and a cloud-based training and evolution platform. The system collects multimodal time-series data packets from the cutting process, and generates a three-dimensional surrounding rock strength field using an edge-side surrounding rock strength sensing model and D-H kinematic reconstruction technology. Combined with a support decision-making knowledge graph inference engine, it generates dynamic support parameter adjustment schemes based on geological characteristics and multiple constraints. The cloud platform parses feedback data chains containing records of human intervention and iteratively updates the model and knowledge base. This invention solves the problems of lagging surrounding rock detection and static support schemes in existing systems, achieving real-time accurate detection and dynamic closed-loop control of support parameters based on multi-source information fusion, thus improving the adaptability of roadway support.

Description

Technical Field

[0001] This invention relates to the field of intelligent mining and rock strata control technology, specifically to an intelligent sensing and dynamic support decision-making system and method for roadway surrounding rock. Background Technology

[0002] Tunnel support is a core component of ensuring safe mine production, and the rationality and timeliness of its plan formulation directly depend on the accurate assessment of the strength and stability of the surrounding rock. In current mine tunneling operations, the acquisition of geological information about the surrounding rock mainly relies on prior geological borehole coring and laboratory uniaxial compressive strength tests. Data obtained through this traditional method suffers from time lag, as test results are typically obtained after tunneling operations, failing to guide real-time decision-making at the current working face. Spatially, borehole coring only provides geological information from sparse locations, making it difficult to capture continuous changes in lithology ahead of tunnel excavation, and its ability to identify local geological abrupt changes such as weak interlayers and fracture zones is insufficient.

[0003] Due to the lack of real-time, full-section geological information, current tunnel support schemes are typically determined based on static geological data, generally adopting a fixed approach of one policy per section or even one policy per entire tunnel. This static decision-making mechanism results in support parameters failing to respond to the dynamic changes in surrounding rock conditions during tunneling and exposure. When encountering geologically anomalous zones, insufficient support strength can easily lead to roof collapse accidents, while in areas with stable surrounding rock, excessive support often results in material waste and reduced tunneling efficiency.

[0004] While methods exist for estimating lithology using fluctuations in single operating parameters such as the cutting motor current and cylinder pressure of tunneling machines, these signals are susceptible to interference from equipment conditions and operational factors, including the wear of the cutting teeth, fluctuations in hydraulic system efficiency, unstable power grid voltage, and operator habits. This leads to a high misjudgment rate and makes it difficult to establish standardized and quantifiable rock strength classifications. Furthermore, simple models built based on historical data from specific working faces or equipment have limited versatility and often fail after changing working faces or equipment due to differences in data distribution, facing the challenge of model adaptation due to a lack of samples under new operating conditions. How to utilize multi-source data generated by the interaction between tunneling equipment and surrounding rock to accurately detect rock strength in the absence of a large number of labeled samples, and how to instantly transform the detection results into dynamic support schemes that comply with safety regulations, achieving real-time linkage between detection and parameter adjustment, are urgent technical problems to be solved in the current intelligent construction of mines. Summary of the Invention

[0005] To address the problems in existing technologies where roadway excavation support scheme formulation relies on human experience, resulting in strong subjectivity, lagging surrounding rock perception, and difficulty in adapting to dynamic changes in geological conditions, leading to mismatches between support parameters and actual working conditions, this invention provides a roadway surrounding rock intelligent perception and dynamic support decision-making system and method. By constructing a cloud-edge-device collaborative architecture and utilizing multimodal signal fusion perception and knowledge graph reasoning technology, it achieves real-time and accurate identification of surrounding rock strength and dynamic optimization decision-making of support parameters.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of the present invention provides an intelligent sensing and dynamic support decision-making system for roadway surrounding rock, the system comprising an underground sensing and interactive execution layer, an edge decision server, and a cloud training and evolution platform.

[0008] The downhole sensing and interactive execution layer is used to trigger the acquisition of multimodal time-series data packets based on the effective cut-off state, and to receive and execute dynamic support parameter adjustment schemes. The edge decision server includes a surrounding rock strength sensing model and a support decision knowledge graph inference engine; the edge decision server receives multimodal time-series data packets, uses the surrounding rock strength sensing model to infer and generate surrounding rock strength distribution data, and uses the support decision knowledge graph inference engine to generate dynamic support parameter adjustment schemes based on this data, while simultaneously transmitting the complete feedback data chain to the cloud. The cloud training and evolution platform includes a model pre-training and optimization module and a knowledge base management and evolution module; the cloud training and evolution platform receives the complete feedback data chain, performs iterative optimization of the model and knowledge base, generates evolved model parameters and a knowledge base version, and distributes them to the edge decision server to complete the system update.

[0009] Preferably, the downhole sensing and interactive execution layer includes a multi-source heterogeneous sensor array, which is used to acquire multi-source physical signals to encapsulate and generate multimodal time-series data packets. Specific signal acquisition methods include: using a current sensor installed on the power supply side of the cutting motor to acquire high-frequency current signals reflecting the cutting load characteristics; using a vibration sensor installed at the root of the cutting arm or the gearbox to acquire triaxial vibration acceleration signals during the cutting process; using an acoustic emission sensor installed on the cutting section's protective plate or the machine body to acquire high-frequency acoustic emission signals generated by rock fracturing; and using a vision sensor installed at the front of the machine body to acquire images of falling rock and video streams of the cutting area. A data acquisition card is used to aggregate the above signals and encapsulate them into multimodal time-series data packets.

[0010] Preferably, the downhole sensing and interactive execution layer is used to trigger the generation of multimodal timing data packets based on the working conditions of the tunneling machine. The edge decision server, acting as the master clock node, establishes a microsecond-level synchronous clock reference with the data acquisition card, which acts as the slave clock node, using the IEEE 1588 clock synchronization protocol. The downhole sensing and interactive execution layer reads the working condition data from the tunneling machine's PLC controller and identifies the effective pure cutting stage based on the microsecond-level synchronous clock reference and preset logic threshold criteria, thereby triggering the acquisition of multi-source signals. The logic threshold criteria include: a preset no-load current threshold, set to 1.1 to 1.2 times the rated no-load current value of the cutting motor; a preset static pressure fluctuation threshold, set to 1.5 times the peak pressure fluctuation of the hydraulic system itself when the cutting arm is in a static hovering state; and a propulsion speed threshold, set to greater than 0 m / min.

[0011] Preferably, the surrounding rock strength perception model is trained using a model-independent meta-learning strategy, and backpropagation and gradient descent updates are performed on the task-specific layer at the edge side using a minimal sample calibration set to adapt to the lithological characteristics of the current working face. Simultaneously, the model utilizes an attention mechanism module to dynamically allocate weights based on the signal-to-noise ratio of each modal signal in the multimodal time-series data packet, performing weighted fusion of multimodal features for inference to generate instantaneous surrounding rock strength inference values.

[0012] Preferably, the edge decision server is used to perform the strength field reconstruction step to generate surrounding rock strength distribution data. Specific steps include: real-time reading of the joint motion variables of the cutting arm; establishing a multi-link kinematic model of the cutting arm using the DH parameter method; calculating the three-dimensional spatial coordinates of the cutting head tip; subsequently mapping the inference results of the surrounding rock strength sensing model to these three-dimensional spatial coordinates; and filling data gaps using ordinary kriging spatial interpolation, thereby generating continuous and complete surrounding rock strength distribution data across the entire cross-section.

[0013] Preferably, the support decision knowledge graph reasoning engine internally stores a mine support domain ontology model using an entity relation triplet topology structure. This engine performs reasoning based on surrounding rock strength distribution data: first, based on the geological feature areas identified in the surrounding rock strength distribution data, it performs rule matching to retrieve a set of candidate support components in the knowledge base; then, it introduces the tunnel cross-sectional dimensions and ground stress conditions as boundary constraints to filter the set of candidate support components, and uses the support cost entity as the optimization objective function, ultimately generating a dynamic support parameter adjustment scheme that includes anchor bolt length, diameter, spacing, and preload values.

[0014] Preferably, the downhole sensing and interactive execution layer also includes an explosion-proof human-machine interface terminal. This terminal receives and displays surrounding rock strength distribution data and dynamic support parameter adjustment schemes, and provides three interactive operation options: confirm execution, parameter fine-tuning, and rejection with manual input. The terminal associates and packages the surrounding rock strength distribution data, the initial dynamic support parameter adjustment scheme, the manual operation log, and the support effect data to construct a complete feedback data chain and feeds it back to the edge decision server.

[0015] Preferably, the edge decision server is used to perform data processing before transmitting the complete feedback data chain: parsing the complete feedback data chain to identify abnormal intervention records that include parameter fine-tuning operations, rejection operations, or top plate delamination exceeding a safety threshold; marking the samples corresponding to the abnormal intervention records as high-value difficult case samples; and performing encryption processing on the high-value difficult case samples so that the complete feedback data chain can be transmitted to the cloud training and evolution platform. The safety threshold is set to 50mm.

[0016] Preferably, the knowledge base management and evolution module is used to perform knowledge updates using a complete feedback data chain: traversing the complete feedback data chain, calculating the Euclidean distance between the system-recommended support parameters and the final manual execution parameters to identify decision conflict events; for decision conflict events, using an association rule mining algorithm to analyze the implicit association between geological features and manually corrected parameters, generating new candidate rules; performing logical conflict verification on the new candidate rules, and updating the new candidate rules to the ontology layer of the knowledge graph after the verification passes, generating the evolved model parameters and knowledge base version.

[0017] This invention provides an intelligent sensing and dynamic support decision-making system and method for roadway surrounding rock. It has the following beneficial effects: 1. This invention effectively solves the problems of lagging and insufficient accuracy in surrounding rock perception by constructing a cloud-edge-device collaborative architecture and integrating multimodal perception technology. The system uses a multi-source heterogeneous sensor array in the well to synchronously collect cutting current, vibration and visual images. With the help of the model-independent element learning strategy on the edge side, it can quickly adapt to the lithological characteristics of the new working face under very few sample conditions. This approach not only overcomes the information island effect of a single monitoring method, but also avoids the dependence of traditional deep learning models on massive labeled data, and realizes real-time and accurate identification of the strength of the surrounding rock in the tunneling working face.

[0018] 2. This invention utilizes a support decision knowledge graph reasoning engine and kinematic reconstruction technology to achieve dynamic real-time optimization of support schemes. By establishing a cutting arm motion model through the DH parameter method to reconstruct the three-dimensional surrounding rock strength field, and performing rule matching and multi-constraint solution based on a knowledge base of entity relationship entity topology, the system can generate a quantitative execution scheme including anchor bolt parameters in real time according to the currently exposed surrounding rock conditions. This mechanism shortens the support decision response time from hours to minutes, realizing the transformation from static experience design to dynamic and accurate decision-making, and ensuring a high degree of matching between support parameters and actual geological conditions.

[0019] 3. This invention establishes a closed-loop knowledge evolution mechanism based on human-machine collaboration, which solves the problem that the system is difficult to adapt to complex and ever-changing geological conditions. Through an explosion-proof human-machine interaction terminal, the system collects manual fine-tuning or rejection operations on the support scheme. The system automatically identifies and encrypts and transmits high-value and difficult case samples to the cloud. The cloud platform uses these feedback data to mine the implicit rules behind decision conflicts, iteratively updates the perception model and knowledge base, and distributes them to the edge side, thereby enabling the system to have continuous self-learning capabilities and continuously improve the robustness and reliability of decision-making as the application time increases. Attached Figure Description

[0020] Figure 1 This is a schematic diagram of the overall architecture of a roadway surrounding rock intelligent sensing and dynamic support decision-making system according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the core workflow and data closed loop of a roadway surrounding rock intelligent sensing and dynamic support decision-making method according to an embodiment of the present invention. Figure 3 A schematic diagram illustrating the principle of synchronous acquisition of multi-source signals from underground and intelligent slicing based on PLC operating conditions, according to one embodiment of the invention; Figure 4 This is a schematic diagram of the edge-side real-time perception and knowledge graph-based decision reasoning process according to an embodiment of the present invention. Figure 5 This is a simulation interface diagram of a downhole explosion-proof human-machine interface terminal displaying a rock strength cloud map and a support suggestion scheme, according to one embodiment.

[0021] Among them, 100, downhole perception and interactive execution layer; 101, current sensor; 102, vibration sensor; 103, acoustic emission sensor; 104, vision sensor; 105, explosion-proof human-machine interaction terminal; 200, edge decision server; 201, surrounding rock strength perception model; 202, support decision knowledge graph reasoning engine; 300, cloud training and evolution platform; 301, model pre-training and optimization module; 302, knowledge base management and evolution module. Detailed Implementation

[0022] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0023] See attached document Figure 1 The intelligent sensing and dynamic support decision-making system for roadway surrounding rock provided by this invention adopts a physically distributed and logically collaborative "cloud-edge-device" three-layer architecture. The system mainly includes an underground sensing and interaction execution layer 100, an edge decision server 200, and a cloud training and evolution platform 300. The underground sensing and interaction execution layer 100, the edge decision server 200, and the cloud training and evolution platform 300 are connected and interact with data through a mining industrial network, forming a closed-loop control system.

[0024] The underground sensing and interactive execution layer 100 is deployed on the roadheader or roadheader-anchor machine at the tunneling face, configured to collect multi-source physical signals and interact with and execute the final decision-making scheme. The underground sensing and interactive execution layer 100 internally includes a multi-source heterogeneous sensor array and human-machine interface devices. The multi-source heterogeneous sensor array includes four sensor modules: a current sensor 101, a vibration sensor 102, an acoustic emission sensor 103, and a vision sensor 104.

[0025] A current sensor 101 is installed on the power supply side of the cutting motor and is configured to acquire high-frequency current signals reflecting the characteristics of the cutting load. A vibration sensor 102 is installed at the root of the cutting arm or at the gearbox and is configured to acquire triaxial vibration acceleration signals during the cutting process. An acoustic emission sensor 103 is installed on the cutting section guard plate or the machine body and is configured to acquire high-frequency acoustic emission signals generated by rock fracturing. A vision sensor 104 is installed at the front of the machine body and is configured to acquire images of falling rocks and video streams of the cutting area. The current sensor 101, vibration sensor 102, acoustic emission sensor 103, and vision sensor 104 are all communicatively connected to the edge decision server 200 for transmitting multi-source synchronization signals. In addition, the downhole sensing and interactive execution layer 100 also includes an explosion-proof human-machine interface terminal 105, which is deployed in the tunneling machine operator's cab and establishes a two-way data connection with the edge decision server 200. The explosion-proof human-machine interface terminal 105 is configured to receive "surrounding rock strength: manual confirmation" requests and "support suggestions: parameter adjustment" schemes from the edge decision server 200, and to send confirmation instructions or modified parameters back to the edge decision server 200.

[0026] The edge decision server 200 is deployed at the power distribution point or refuge chamber within the roadway, and adopts an intrinsically safe or explosion-proof design for mining. The edge decision server 200 contains two core functional modules: a surrounding rock strength perception model 201 and a support decision knowledge graph inference engine 202. The surrounding rock strength perception model 201 is configured to receive multi-source synchronous signals from the underground sensing and interactive execution layer 100, and infers and outputs the surrounding rock strength level through deep learning algorithms. The support decision knowledge graph inference engine 202 internally stores structured rock mechanics knowledge, support specifications, and expert rules. It is configured to receive surrounding rock strength level data, combine it with roadway geometric parameters to perform logical reasoning and multi-constraint solving, and generate specific dynamic adjustment schemes for support parameters. There is a two-way interaction between the edge decision server 200 and the cloud training and evolution platform 300: the edge decision server 200 sends "new sample data", "model update request" and "effect feedback" to the cloud training and evolution platform 300; the cloud training and evolution platform 300 pushes "pre-trained model" and "knowledge base version push" to the edge decision server 200.

[0027] The cloud-based training and evolution platform 300 is deployed in a ground-based data center or a remote cloud server cluster, configured to perform global data aggregation and model iteration tasks. Internally, the cloud-based training and evolution platform 300 includes a model pre-training and optimization module 301 and a knowledge base management and evolution module 302. The model pre-training and optimization module 301 utilizes aggregated historical data to perform pre-training, meta-learning training, and hyperparameter optimization of the general perception model. The knowledge base management and evolution module 302 is responsible for version control, rule mining, and update maintenance of the domain knowledge graph.

[0028] In terms of inter-layer communication and coordination, the current sensor 101, vibration sensor 102, acoustic emission sensor 103, and vision sensor 104 within the underground sensing and interactive execution layer 100 are connected to the data acquisition card via shielded cables. The data acquisition card communicates with the edge decision server 200 through an underground industrial gigabit ring network or a 5G private network, employing the IEEE 1588 precision clock synchronization protocol to ensure that the timestamp synchronization accuracy of multi-source signals reaches the microsecond level. The edge decision server 200 establishes an encrypted VPN channel with the cloud training and evolution platform 300 through the mine backbone network and the surface exit gateway, utilizing a message queue protocol for asynchronous data transmission.

[0029] See attached document Figure 2 This invention provides a method for intelligent sensing and dynamic support decision-making of surrounding rock in roadways, comprising the following steps: S1, Signal Synchronous Acquisition and Intelligent Slicing: The system reads the working condition data of the tunneling machine PLC, identifies the effective pure cutting stage, and triggers the current sensor 101, vibration sensor 102, acoustic emission sensor 103 and vision sensor 104 to perform synchronous acquisition. The acquired multi-source signals are sliced ​​and packaged according to the cutting cycle to generate multi-modal time sequence data packets. S2, Real-time edge perception: The edge decision server 200 receives the multimodal time-series data packets output by S1, uses the built-in surrounding rock strength perception model 201 to extract and infer the features of the signal, and at the same time updates the parameters of the surrounding rock strength perception model 201 by combining the incremental fine-tuning data fed back from the previous cycle S4, and finally outputs the surrounding rock strength distribution data of the current exposed section. S3, Strength Field Reconstruction and Decision Reasoning: The support decision knowledge graph reasoning engine 202 receives the surrounding rock strength distribution data output by S2, reconstructs the surrounding rock strength field cloud map by combining the spatial location information of the cutting head, and performs rule matching and multi-constraint solution in the knowledge base based on the geological feature areas identified by the surrounding rock strength field cloud map to generate a dynamic support parameter adjustment scheme. S4, Human-machine collaborative execution and feedback: The explosion-proof human-machine interaction terminal 105 receives and displays the dynamic support parameter adjustment scheme and surrounding rock strength field cloud map output by S3, which are used by the operator for confirmation or parameter fine-tuning. The system records the final executed support parameters and the subsequently monitored mine pressure data to form a complete feedback data chain, and feeds back the parameter correction information in the complete feedback data chain as incremental fine-tuning data to S2. S5, Data Closed Loop and Cloud Evolution: The cloud training and evolution platform 300 receives the complete feedback data chain uploaded by S4, uses the new sample data in the complete feedback data chain to train the perception model and optimize the rules of the knowledge base, generates the evolved model parameters and knowledge base version, and sends them to the edge decision server 200 where S1 and S2 are located for signal acquisition and perception reasoning in the next cycle.

[0030] The steps described above will be explained in detail below.

[0031] See attached document Figure 3 In the intelligent sensing and dynamic support decision-making method for roadway surrounding rock provided by this invention, the high-precision synchronous acquisition and intelligent slicing mechanism of multi-source heterogeneous signals in step S1 is specifically implemented through the following sub-steps: S101, intelligent identification of effective pure cutting stage based on tunneling machine PLC operating data.

[0032] The edge decision server 200 reads the register status data of the tunneling machine PLC controller in real time at millisecond intervals via the mine industrial ring network or fieldbus interface. The data includes the real-time three-phase current value of the cutting motor, the hydraulic pressure values ​​of the cutting arm lifting cylinder and the slewing cylinder, and the propulsion speed value of the walking track.

[0033] To extract valid data containing geological information from a continuous operation process, the edge decision server 200 identifies valid pure truncation stages based on preset logical threshold criteria. The specific setting method for the logical threshold criteria is as follows: Preset no-load current threshold: Set to 1.1 to 1.2 times the rated no-load current value of the cutting motor, used to distinguish whether the motor is in no-load state or load cutting state; The preset static pressure fluctuation threshold is set to 1.5 times the peak pressure fluctuation of the hydraulic system when the cutting arm is in a static, hovering state. This is used to identify whether the cutting arm has undergone substantial movement. The cutting condition logic threshold is set to the track braking state or the propulsion speed being lower than the micro-motion threshold (e.g., 0.5 m / min). The judgment logic is as follows: When the edge decision server 200 detects that the real-time current value of the cutting motor is continuously higher than the preset no-load current threshold, and the pressure change rate of the cutting arm cylinder exceeds the preset static pressure fluctuation threshold, and the track braking or micro-speed propulsion conditions are met, it is determined that the tunneling machine cutting head has begun to make substantial contact with and break the rock mass, and this moment is marked as the start point of the time window. Subsequently, when the current value of the cutting motor is detected to fall below the preset no-load current threshold and remain below it for a set duration (e.g., 5 seconds), or when the propulsion speed of the traveling unit is detected to return to zero, it is determined that the current cutting behavior has ended, and this moment is marked as the end point of the time window.

[0034] like Figure 3 As shown in the diagram, the timing relationship between "PLC condition identification" and "multi-source signal acquisition" is illustrated within "one tunneling cycle (timeline: 0-8 minutes)". The system divides this cycle into three stages: During "Phase I" (corresponding to time axis "0-1 minutes" and operating condition "idling / preparation"), although the cutting head is rotating, the "current signal" shows a "stable waveform", the "vibration signal" shows "low amplitude", the "acoustic emission" shows "low event rate", and the "visual sequence" shows "cutting head idling". The edge decision server 200 determines it to be in the idling or preparation phase and does not execute data slicing.

[0035] During "Phase II" (corresponding to "1-5 minutes" on the timeline and "effective pure cutting stroke"), due to a sudden increase in cutting load, the "current signal" exhibits "violent fluctuations," the "vibration signal" displays "high-frequency components," the "acoustic emission" reaches a "peak event rate," and the "visual sequence" captures "rock collapse." The edge decision server 200 locks this period as the effective pure cutting stroke and activates the "smart slicing window" shown below the diagram.

[0036] During "Phase III" (corresponding to "5-8 minutes" on the timeline and "reversal / stop" operating conditions), the "current signal" shows a "decreasing waveform," the "vibration signal" returns to "low amplitude," the "acoustic emission" recovers to "low event rate," and the "visual sequence" shows "cutting head reversal." The system determines that this phase does not constitute effective cutting and stops sampling.

[0037] S102 is a downhole multi-source heterogeneous signal synchronous acquisition system based on the IEEE 1588 precision clock synchronization protocol, enabling microsecond-level synchronous acquisition.

[0038] To address the challenge of data temporal alignment caused by the large differences in sampling frequencies and dispersed physical locations of the sensors in the downhole sensing and interactive execution layer 100, this system adopts a synchronous acquisition strategy based on the IEEE 1588 precision clock synchronization protocol. Specifically, the edge decision server 200 acts as the PTP master clock node, and the data acquisition cards connected to the current sensor 101, vibration sensor 102, acoustic emission sensor 103, and vision sensor 104 act as PTP slave clock nodes.

[0039] The synchronization acquisition process is as follows: The edge decision server 200 periodically broadcasts synchronization messages containing the transmission time to the network. Each slave clock node receives the message and records the reception time, then sends a delay request message to the edge decision server 200. The edge decision server 200 records the arrival time of the request message and feeds it back to the slave clock nodes. Each slave clock node uses the four timestamp data recorded by both the master and slave nodes to calculate the network transmission delay and the local clock deviation, and then dynamically adjusts the local crystal oscillator count to eliminate clock drift.

[0040] Through the above mechanism, the edge decision server 200 ensures that the local time base deviation of all sensor nodes is kept within 1 microsecond. On this basis, the current waveform collected by the current sensor 101, the vibration signal collected by the vibration sensor 102, the acoustic emission signal collected by the acoustic emission sensor 103, and the video stream collected by the vision sensor 104 are all marked with a unified PTP absolute timestamp at the source end.

[0041] S103 is used for cleaning, slicing, and vectorizing multimodal temporal data within a truncation loop.

[0042] The edge decision server 200 processes the original multi-source data streams synchronously collected in S102 based on the time window of the effective pure truncation stage determined in S101.

[0043] First, the edge decision server 200 performs data cleaning: for the data from the current sensor 101, a digital bandpass filter is used to filter out power frequency interference; for the data from the vibration sensor 102 and the acoustic emission sensor 103, a wavelet soft thresholding denoising algorithm is used to remove background noise. These filtering and denoising algorithms are well-known technologies in the field and will not be elaborated upon here.

[0044] Subsequently, the edge decision server 200 performs alignment and encapsulation of multimodal data. To address the issue of inconsistent sampling rates, the edge decision server 200 uses a preset model input frequency as a benchmark to perform linear interpolation upsampling on low-frequency signals and sliding window statistical feature extraction (calculating root mean square value and peak factor) on high-frequency signals, thereby aligning the data of each channel in the time dimension.

[0045] Finally, the edge decision server 200 stacks and combines the time-aligned current sequence, vibration sequence, acoustic emission feature sequence, and visual image frame sequence according to the time step, encapsulating them into a standardized multimodal time-series data packet. This multimodal time-series data packet serves as the input data for the surrounding rock strength sensing model 201 in the subsequent step S2.

[0046] See attached document Figure 4 In the intelligent sensing and dynamic support decision-making method for roadway surrounding rock provided by this invention, the edge-side small-sample adaptive and real-time surrounding rock strength sensing in step S2 is specifically implemented through the following sub-steps: S201, Acquisition and calibration of a very small number of true value samples in the field based on a portable point load instrument.

[0047] When the tunneling face is initially deployed or encounters significant geological structural changes, the system enters the initialization calibration mode. On-site geological engineers use a portable point load tester to conduct on-site strength tests on newly exposed rockfalls or rock walls at the current working face. In specific operations, engineers select rock samples from no fewer than five different spatial locations on the tunnel section, determine their point load strength index, and convert them into uniaxial compressive strength (UCS) values ​​according to well-known rock mechanics formulas.

[0048] While conducting point load tests, the system records the absolute timestamp when the tunneling machine's cutting head acts on the test point. The explosion-proof human-machine interface 105 provides a calibration data input interface for engineers to input the true value of the uniaxial compressive strength obtained from the test. The edge decision server 200 uses this absolute timestamp as an index key to retrieve and extract the signal segment corresponding to the time period from the locally cached historical multimodal time-series data packets generated by S1.

[0049] The edge decision server 200 uses the extracted "multimodal signal fragments" as feature input (SupportInput) and the "manually entered true values ​​of uniaxial compressive strength" as labels to construct a "minimal sample calibration set" containing 5 to 10 sets of data. This calibration set accurately reflects the mapping relationship between physical signals and rock strength under specific geological environments (such as sandstone and mudstone interbedded layers) and specific equipment operating conditions (such as the wear state of cutting teeth).

[0050] S202, Fast edge-side transfer and parameter fine-tuning of a general perception model based on meta-learning strategy.

[0051] The edge decision server 200 has a pre-downloaded rock strength sensing model 201 downloaded from the cloud training and evolution platform 300. The rock strength sensing model 201 is not a traditional fixed-parameter model, but a general initialization model trained using a model-independent meta-learning (MAML) strategy. Its core technical feature is that the initial parameter positions of the rock strength sensing model 201 are optimized to a manifold space that is highly sensitive to gradient updates, allowing the model to quickly converge to the optimal solution for a new task with only a very small number of gradient descent steps.

[0052] After obtaining the "minimal sample calibration set" of S201, the edge decision server 200 initiates a parameter fine-tuning procedure: First, the edge decision server 200 freezes the weight parameters of the feature extraction layer in the surrounding rock strength perception model 201. The feature extraction layer includes a convolutional neural network layer for extracting texture features from the image of the visual sensor 104, and a long short-term memory network layer for extracting temporal features from the current sensor 101, vibration sensor 102, and acoustic emission sensor 103. The freezing operation ensures that the model's ability to extract general physical features (such as current waveform envelope and vibration spectrum structure) is not compromised.

[0053] Simultaneously, the edge decision server 200 unfreezes the task-specific layer (i.e., the fully connected regression layer) of the model. The edge decision server 200 calculates the prediction loss using a "minimal sample calibration set" and updates the task-specific layer using backpropagation and gradient descent based on this loss. To meet the real-time requirements of the edge side, the system limits the number of iterations to less than 10. Through this process, the general model is quickly updated to a specialized model adapted to the lithological characteristics of the current working face, effectively solving the technical challenge of traditional deep learning models being unable to work in new geological environments when large-scale labeled data is lacking.

[0054] S203, Multimodal signal feature extraction and real-time inference of surrounding rock strength level at discrete sampling points.

[0055] During the normalized tunneling process after model parameter fine-tuning, the edge decision server 200 continuously receives real-time multimodal time-series data packets (i.e., query sets) output from step S1. For example... Figure 4 As shown in the upper-level "perception process", after the "multimodal sample data package" enters the model, it first undergoes "feature extraction and fusion" processing.

[0056] The edge decision server 200 utilizes the attention mechanism module in the model to perform weighted fusion of multimodal features. The attention mechanism module dynamically allocates weights based on the signal-to-noise ratio of each modal signal: when the contrast of the image acquired by the visual sensor 104 is reduced due to dust occlusion, the weight of the visual features is automatically reduced; when the vibration sensor 102 detects a strong truncation response, the weight of the vibration features is automatically increased.

[0057] The weighted and fused high-dimensional feature vectors are input into the task-specific layer of the fine-tuned "small-sample fast-adaptation perception model" in S202. The "small-sample fast-adaptation perception model" first outputs the predicted continuous value of uniaxial compressive strength (UCS), and then maps it to the "discrete strength classification result" corresponding to the current cut instant based on a preset threshold. This result contains three key dimensions: Grid coordinates (X,Y): Marks the planar coordinates of the cutting head on the roadway cross-section corresponding to this inference result (these coordinates are calculated and provided synchronously by the S301 spatial position calculation module); Strength grade: The mechanical strength category of the surrounding rock at the current location (specifically divided into: soft rock <20MPa, medium-hard rock 20-60MPa, hard rock >60MPa); Confidence level: The model's probability prediction of the classification result (range 0-1).

[0058] The edge decision server 200 caches the above discrete inference results until a complete truncation loop ends, forming a set of discrete sampling points covering the current section, which serves as the basis for generating surrounding rock strength distribution data and strength field reconstruction.

[0059] See attached document Figure 4 In the intelligent sensing and dynamic support decision-making method for roadway surrounding rock provided by this invention, the three-dimensional reconstruction of the surrounding rock strength field and the knowledge graph-driven decision reasoning in step S3 are specifically implemented through the following sub-steps: S301, Spatial position calculation and coordinate mapping based on the kinematic parameters of the tunneling machine cutting arm.

[0060] When the edge decision server 200 receives the discrete strength values ​​from the surrounding rock strength distribution data output in step S2, it simultaneously performs spatial position calculation of the cutting head. The edge decision server 200 reads the joint motion variables of the cutting arm in real time through the tunneling machine's body communication interface. These variables include: the horizontal rotation angle corresponding to the stroke of the slewing cylinder, the vertical lifting angle corresponding to the stroke of the lifting cylinder, and the telescopic length corresponding to the telescopic mechanism.

[0061] To achieve precise spatial positioning, the edge decision server 200 constructs a local three-dimensional coordinate system with the geometric center of the tunnel boring machine as the origin. The system uses the Denavit-Hartenberg (DH) parameter method to establish a multi-link kinematic model of the cutting arm. The read joint motion variables are substituted into a homogeneous transformation matrix for chain multiplication to calculate the three-dimensional coordinates (X, Y, Z) of the cutting head tip relative to the tunnel cross-section. This type of forward kinematics algorithm based on DH parameters is common knowledge in the field of robotics and will not be elaborated upon here.

[0062] Subsequently, the edge decision server 200 performs spatiotemporal alignment and binding of the calculated three-dimensional coordinates with the "intensity level" and "confidence" obtained from step S2 reasoning, generates discrete intensity sampling points with clear spatial attributes, and maps these sampling points to a pre-constructed two-dimensional rasterized mesh model of the tunnel cross-section.

[0063] S302, Reconstruction of the whole-section surrounding rock strength field cloud map and identification of characteristic regions based on Kriging interpolation.

[0064] Because the trajectory of the cutting head within a single cutting cycle is discrete, the mesh model generated by S301 contains "data holes" that are not reached by the cutting head. For example... Figure 4 As shown in the “Intensity Field Reconstruction” process, the edge decision server 200 uses the “spatial interpolation algorithm (Kriging method)” to reconstruct the intensity estimate of these void regions.

[0065] Specifically, the edge decision server 200 first calculates the variance-distance relationship between discrete sampling point data pairs, fitting a semi-variogram model to quantify the spatial correlation of surrounding rock strength attenuation with distance. Based on this model, the system calculates the weighting coefficient between each grid point with unknown strength and its surrounding known sampling points. The calculation of this weighting coefficient follows the principles of "unbiased estimation" and "minimization of estimation variance".

[0066] The edge decision server 200 uses the intensity values ​​of known sampling points and their corresponding weighting coefficients to perform linear weighted summation, thereby calculating the intensity estimates of all blank grid points across the entire cross-section and generating a continuous and complete "pseudo-color grading cloud map of surrounding rock intensity". Based on this, the edge decision server 200 performs feature region identification on the "pseudo-color grading cloud map of surrounding rock intensity". Specifically, it uses an image connected component analysis algorithm to extract continuous pixel regions of the same intensity level in the cloud map, calculates the geometric center coordinates and coverage area of ​​each connected region, and marks soft or hard rock regions with an area exceeding a preset threshold (e.g., 0.5 square meters) as "geological feature regions" with semantic information.

[0067] S303, Ontology construction, rule definition and constraint coding for knowledge graphs in the support decision-making domain.

[0068] The support decision-making knowledge graph reasoning engine 202 internally stores a pre-built ontology model of the mine support domain. For example... Figure 4 As shown in the lower layer "Knowledge Graph Decision Reasoning", this graph uses an entity relationship triplet topology structure to store expert knowledge.

[0069] The edge decision server 200 is pre-configured with the following map elements: Ontology Construction: Defines the core entity concepts involved in support decision-making, including "surrounding rock strength grade", "ground stress conditions", "tunnel cross-sectional dimensions", "anchor bolt parameters" (further subdivided into "length", "diameter", "spacing", and "preload") and "support cost".

[0070] Rule definition: Defines the logical influence relationships between entities, specifically including: The "surrounding rock strength grade" points to the "anchor bolt parameters" through the "requirement" relationship, indicating the direct determining effect of strength on support parameters; The “surrounding rock strength grade” points to the “geostress condition” through the “recommendation” relationship, representing an empirical rule for inferring geostress from lithology; "Anchor bolt parameters" are linked to "support costs" through the "optimization" relationship, indicating the impact of parameter adjustments on economic efficiency.

[0071] Constraint coding: This involves converting existing coal mine roadway support technical specifications into attribute constraints on the edges of a graphical relationship. For example, in a "constraint" relationship (such as...) Figure 4 The lines connecting "surrounding rock strength grade" and "tunnel cross-sectional dimensions" and "anchor bolt parameters" and "tunnel cross-sectional dimensions" are shown in the figure, which encode geometric hard constraints.

[0072] S304, dynamic support scheme inference generation based on graph path search and multi-objective optimization.

[0073] The support decision knowledge graph reasoning engine 202 receives the "surrounding rock strength pseudo-color classification cloud map" and geological feature area data output by S302, and instantiates them as the starting node in the graph through the "real-time mapping" module.

[0074] Support decision knowledge graph reasoning engine 202 starts multi-constraint solution process: First, the engine starts with the current "surrounding rock strength level" entity and searches the graph path along the "requirement" and "recommendation" relationship to retrieve all candidate support components in the knowledge base that are suitable for this geological condition.

[0075] Secondly, the engine introduces "tunnel cross-sectional dimensions" and "geostress conditions" as boundary constraints, activating the logical rules in the "constraint" relationship edges. The system automatically filters out solutions in the candidate set that do not meet the geometric or mechanical constraints.

[0076] Finally, for the remaining set of feasible solutions, the engine uses the "support cost" entity as the optimization objective function, and combines the "optimization" relationship between "anchor bolt parameters" and "support cost". For example... Figure 4 As shown, the information flow of "tunnel cross-section dimensions", "anchor bolt parameters", and "support cost" ultimately converges to the "multi-constraint solution engine". This engine calculates the comprehensive performance score of each scheme, selects the combination with the highest comprehensive score that meets all safety specifications, and generates a "support parameter adjustment scheme" that includes specific values ​​for "length", "diameter", "spacing", and "preload".

[0077] The solution was then sent to the explosion-proof human-machine interface terminal 105 for display, and after receiving manual confirmation, it was fed back to the downhole sensing and interaction execution layer 100 for execution.

[0078] See attached document Figure 5 In the intelligent perception and dynamic support decision-making method for roadway surrounding rock provided by the present invention, the visualized human-computer collaborative interaction and closed-loop feedback execution in step S4 is specifically implemented through the following sub-steps: S401, Visualization of the pseudo-color grading cloud map of surrounding rock strength and the dynamic support suggestion scheme.

[0079] The explosion-proof human-machine interface terminal 105 receives the surrounding rock intensity field cloud map data packet output from the edge decision server 200 via an industrial Ethernet interface. The built-in graphics rendering engine of the explosion-proof human-machine interface terminal 105 parses and visualizes the above data.

[0080] In the main display area (70%), the explosion-proof human-computer interaction terminal 105 maps the discrete grid strength values ​​in the surrounding rock strength field cloud map to a pseudo-color spectrum based on the rock mechanics uniaxial compressive strength standard. The specific mapping logic is as follows: Mesh with a uniaxial compressive strength of less than 20 MPa is rendered in red to represent weak and dangerous areas. The 20MPa to 60MPa mesh is rendered in yellow tones to represent the medium-hard transition zone; Mesh with a strength greater than 60 MPa is rendered in green to represent stable hard rock zones.

[0081] This visualization method allows the tunneling machine operator to observe the lithological distribution of the current cross-section, especially to identify... Figure 5 The red soft rock area in the upper right corner is shown as a geological anomaly. Simultaneously, the system overlays 70% of the main display area with the perceived metadata, including the current cycle advance: 1.0m, cutting time: 4 minutes 15 seconds, the confidence level of the perception model inferenced by the surrounding rock strength perception model 201: 94%, and data quality: excellent.

[0082] In the support suggestion section (30% of the control area on the right), the explosion-proof human-machine interface terminal 105 converts structured decision data into text for display. The interface displays the title "Recommended Solution A: Safety First," and lists the key risk areas and parameters identified by the system. Identified area: Upper right soft rock area; Area dimensions: approximately 2.5m long, approximately 3m² in area. 2 For this area, the interface suggests the following adjustments: Anchor bolt spacing: 1000×1000mm → 800×800mm; Anchor bolt length: 2.4m → 2.8m; Preload: 60kN→80kN.

[0083] In addition, the system calculates and displays the expected impact of the solution: safety factor +25%, cost +5%.

[0084] S402 is a human-computer interaction confirmation, parameter correction, and rejection mechanism based on the experience of on-site experts.

[0085] The explosion-proof human-machine interface terminal 105 is equipped with a touch interface, providing a human-machine collaborative decision-making mechanism. The system has three response logics set in the 30% control area on the right side, allowing operators to process the dynamic support parameter adjustment schemes generated by S3. It also provides feedback on the previous cycle's effect (roof delamination <5mm) and historical matching cases (3 similar working conditions as decision-making references). Confirm Execution: When the operator determines that the intensity cloud map displayed by the system matches the on-site visual observation and agrees that the recommended plan A prioritizes safety, the operator can trigger the command by clicking the √ Confirm Execution button. The explosion-proof human-machine interface terminal 105 locks the set of support parameters and converts them into execution commands, which are then sent to the tunneling machine's anchor drilling rig control system or guided through voice broadcast.

[0086] Parameter Fine-tuning: When the operator approves the system's identification of soft rock areas but determines that the recommended parameters are not entirely applicable, they can trigger the edit mode by clicking the parameter fine-tuning button. The system allows the operator to adjust specific values ​​within a preset safety range. This safety range is pre-set by the edge decision server 200 according to roadway support specifications (e.g., the adjustment range of anchor bolt preload is limited to ±20kN). The adjusted parameters must pass the built-in verification before taking effect.

[0087] Reject and Manual Input: When the operator determines that the perception model has misjudged (e.g., misclassifying sensor noise as soft rock), they can switch to manual mode by clicking the "Reject and Manual Input" button. The operator sets the support parameters based on field experience and selects the reason for rejection (such as "sensor failure" or "geological misjudgment") from the drop-down menu. This rejection operation will be recorded by the system as a negative sample label for subsequent model error correction training.

[0088] S403, the construction of a full-process feedback data chain.

[0089] To enable the continuous evolution of the perception model and decision map, the explosion-proof human-computer interaction terminal 105 and the edge decision server 200 work together to build a complete feedback data chain containing causal logic.

[0090] The complete feedback data chain uses a unique "tunneling loop ID" (as attached). Figure 5 Using the cycle number (023) in the status bar as the primary key, the time series data from the following four dimensions are associated and packaged: Sensing status data: The original multimodal signal characteristics and inferred surrounding rock intensity distribution data output in step S2, as well as the spatiotemporal information such as working face: E5201 and time: 2024-05-16 14:30 displayed in the status bar; Decision recommendation data: Initial dynamic support parameter adjustment scheme generated in step S3; Interactive execution data: In step S402, the operator finally confirms the actual support parameters to be executed, as well as the operator's interactive behavior log (confirmation, fine-tuning, or rejection). At this time, the status bar displays the communication status: normal and edge server: online, ensuring that the data interaction is correct. Support effect data: Within a preset time window after the support is completed, the amount of roof delamination, the monitoring value of anchor bolt force, and the amount of roadway surface displacement are collected by the mine pressure monitoring station in the roadway.

[0091] The edge decision server 200 encapsulates the data from the four dimensions mentioned above into a standard sample package. If the sample package contains records of "parameter fine-tuning" or "rejection" operations, or if the "support effect data" shows that the roof delamination exceeds the safety threshold (e.g., 50mm), the edge decision server 200 marks the sample package as a "high-value difficult example sample" and prioritizes sending it to the cloud training and evolution platform 300 via the transmission channel. This mechanism uses parameter correction information as incremental fine-tuning data to ensure that the cloud model can be retrained in a targeted manner based on real field feedback.

[0092] See attached document Figure 2 In the intelligent sensing and dynamic support decision-making method for roadway surrounding rock provided by this invention, the data closure and cloud evolution in step S5 are specifically implemented through the following sub-steps: S501 enables encrypted transmission and decryption of edge-side data and the construction of a global sample pool in the cloud.

[0093] The edge decision server 200 serializes and encapsulates the complete feedback data chain formed in step S4. The complete feedback data chain includes surrounding rock strength distribution data, dynamic support parameter adjustment schemes, manual operation logs, and mine pressure monitoring feedback data. To ensure data security during public network transmission, the edge decision server 200 uses its built-in encryption module and employs the AES-256 industrial-grade encryption algorithm to encrypt the encapsulated data packets, and then transmits them to the cloud training and evolution platform 300 via an IPSec VPN secure tunnel.

[0094] The Cloud Training and Evolution Platform 300 is equipped with a high-concurrency data receiving gateway, which performs decryption and integrity verification upon receiving encrypted data packets. Data that passes verification is parsed and stored in a distributed database, constructing a "cloud-based global sample pool." The Cloud Training and Evolution Platform 300 manages samples hierarchically based on the label information in the data packets. Basic sample library: Samples whose perception results are consistent with manual confirmation and whose support effect data indicate that the roadway is in a stable state (such as the amount of roof delamination is within a safe range) are stored in the "basic sample library" to maintain the statistical distribution characteristics of the model. High-Value Difficulty Case Library: Samples that have undergone manual "parameter fine-tuning" or "rejection" operations, or whose roof delamination exceeds the preset safety threshold after support, are stored in the "High-Value Difficulty Case Library." This library is the core data source for subsequent model iterations and is mainly used to solve the problem of perception inaccuracy under long-tail conditions.

[0095] S502, cloud-based incremental training, performance evaluation, and hyperparameter optimization of perceptual models.

[0096] The model pre-training and optimization module 301 within the cloud-based training and evolution platform 300 periodically (e.g., weekly or when the number of samples in the high-value difficult example library increases by 1000) triggers a retraining task. The model pre-training and optimization module 301 extracts samples from the "high-value difficult example library" and mixes them with data from the "basic sample library" at a preset ratio (e.g., 1:5) to construct an incremental training dataset.

[0097] During training, the model pre-training and optimization module 301 employs a transfer learning strategy, freezing the weights of the first few layers of the basic feature extraction network in the pre-trained model and focusing on gradient updates for the weights of the high-level semantic understanding network and classifier. This approach can quickly adapt to new geological conditions while retaining the original general feature extraction capabilities and preventing catastrophic forgetting in the model.

[0098] After training, the model pre-training and optimization module 301 performs automated performance evaluation of the new model on an independent validation set. Evaluation metrics include mean average precision (mAP), recall, and F1 score. If the overall performance of the new model is better than the old model currently running online (e.g., an improvement in mAP of more than 1%), the model pre-training and optimization module 301 will initiate a hyperparameter optimization process, using a Bayesian optimization algorithm to fine-tune hyperparameters such as learning rate, regularization coefficient, and batch size to further explore the model's potential, ultimately generating an "evolved perceptual model version" to be released.

[0099] S503 enables automatic mining, conflict detection, and version iteration of knowledge graph rules based on feedback data.

[0100] The knowledge base management and evolution module 302 within the cloud-based training and evolution platform 300 is responsible for self-improvement of the support decision-making logic. This process aims to address the discrepancy between preset expert rules and actual on-site conditions, specifically including three stages: conflict detection, rule mining, and version iteration.

[0101] In the conflict detection phase, the knowledge base management and evolution module 302 traverses the complete feedback data chain to calculate the Euclidean distance between the "system recommended support parameters" vector and the "human final execution parameters" vector. If the calculated Euclidean distance exceeds a preset tolerance threshold, the system determines that the operation is a "decision conflict event".

[0102] In the rule mining stage, for accumulated "decision conflict events," the knowledge base management and evolution module 302 uses association rule mining algorithms (such as the FP-Growth algorithm) to analyze the implicit association between geological features and manually corrected parameters. The system filters effective rules by calculating "support" and "confidence." For example, if the system statistically discovers the phenomenon that "when the visual sensor 104 detects obvious rock wall water seepage characteristics and the intensity is less than 30MPa, the preload of the anchor bolt is always increased from 60kN to 80kN," and if the support of this phenomenon is greater than 0.05 and the confidence is greater than 0.8, the knowledge base management and evolution module 302 will automatically generate a new candidate rule: "If water seepage exists and the rock is soft, then a preload of 80kN is recommended."

[0103] During the version iteration phase, newly generated candidate rules, after being verified by the background simulation system, are written into the ontology layer of the knowledge graph, updating the original reasoning logic. The knowledge base management and evolution module 302 packages the updated knowledge graph into a new version file. Finally, the cloud training and evolution platform 300 pushes the "evolved perception model version" and the "new version of the knowledge graph" to the underground edge decision server 200 through the cloud-edge collaboration channel, completing the process. Figure 2 The closed-loop update of the "evolved model and knowledge base" shown enables the system to have more accurate perception capabilities and more on-site decision-making capabilities in the next stage of operation.

[0104] To verify the practical engineering application effects of the system and method proposed in this invention, a detailed description is provided below with reference to specific engineering embodiments. This embodiment selects a large-section coal roadway excavation face in a deep mining area of ​​a large coal mine as the verification scenario.

[0105] Example environment deployment and initialization configuration: Before the system officially goes into operation, hardware selection and environmental deployment are completed first. The downhole sensing and interaction execution layer 100 is deployed on an EBZ260 cantilevered longitudinal shaft tunneling machine. To monitor the multi-dimensional physical field of the cutting process, the multi-source heterogeneous sensor array is configured as follows: The current sensor 101 is an open-loop Hall current sensor with a range of 0-600A and a sampling frequency of 2kHz. It is installed on the U-phase power cable on the output side of the cutting motor inverter and is configured to collect the stator current signal that reflects the transient change of the cutting resistance.

[0106] The vibration sensor 102 is an intrinsically safe MEMS triaxial accelerometer with a range of ±50g and a sampling frequency of 10kHz. It is fixed to the root of the cutting arm rotary table by a strong magnetic base and epoxy resin, and is configured to pick up the broadband vibration response when the cutting head hits the rock.

[0107] The acoustic emission sensor 103 is a wideband piezoelectric ceramic sensor with a frequency response range of 20kHz-200kHz, which is coupled and installed on the inner side of the cutting section guard plate and configured to capture high-frequency stress wave signals generated by micro-fractures in the rock.

[0108] The visual sensor 104 is a mining explosion-proof low-light camera with a resolution of 1920×1080. It is installed above the hydraulic pump station on the top of the machine body. The field of view covers a range of 3 meters in front of the cutting head. It is configured to collect texture images of the cutting section and video streams of falling rock blocks.

[0109] The edge decision server 200 uses a mining edge computing terminal with an explosion-proof and intrinsically safe design, integrating a high-performance AI computing module (such as the NVIDIA Jetson series), and is deployed in a mobile substation chamber about 200 meters away from the tunneling face. The edge decision server 200 is connected to the underground sensing and interaction execution layer 100 via fiber optic Ethernet, and establishes communication with the cloud training and evolution platform 300 through the mine backbone network.

[0110] During the initialization phase, to address the adaptability of the pre-trained model to specific geological environments, geological engineers used a portable point load tester to conduct strength tests on the rock at eight randomly selected feature points on the initial section of the tunneling face, obtaining measured values ​​of uniaxial compressive strength. The edge decision server 200 simultaneously recorded signal segments as the tunneling machine passed these points. The system used these eight sets of "multi-source signal features - measured strength values" data pairs as a minimal sample calibration set (SupportSet), and fine-tuned the weights of the fully connected classification layer of the surrounding rock strength perception model 201 using the Model Agnostic Meta-Learning Algorithm (MAML) (e.g., performing a 5-step gradient descent update) to quickly adapt it to the geological characteristics of the current working face.

[0111] Validation of the entire "perception-decision-execution" process within a single tunneling cycle: After system initialization, it enters normalized tunneling operation mode. Taking the 623rd tunneling cycle as an example, the process from S1 to S5 is executed as follows: S1, synchronous signal acquisition and intelligent slicing.

[0112] The system reads the tunneling machine's PLC operating status data via the OPCUA protocol. When the system detects that the "cutting motor is on," "cutting arm is extended," and "track brake" signals are all valid simultaneously, it identifies this as a "valid pure cutting phase" and triggers current sensor 101, vibration sensor 102, acoustic emission sensor 103, and vision sensor 104 to perform microsecond-level synchronous data acquisition. When the PLC signal indicates that cutting has ended, the system slices and packages the acquired multi-source signals according to the cutting cycle, generating a multimodal time-series data packet. This data packet eliminates idle and standby segments, containing only valid cutting operation data.

[0113] S2, real-time edge-side perception.

[0114] Edge decision server 200 receives the multimodal time-series data packets output by S1 and uses the built-in surrounding rock strength perception model 201 to extract features and infer from the signals. The surrounding rock strength perception model 201 discretizes the current cut section into 50 grid cells and outputs the strength category probability for each cell. Simultaneously, edge decision server 200 updates the parameters of surrounding rock strength perception model 201 based on incremental fine-tuning data fed back from the previous loop S4. Finally, edge decision server 200 outputs the surrounding rock strength distribution data of the currently exposed section. In this embodiment, the inference results show that most of the section is "Class II medium-hard rock (strength 25-45 MPa)," but there is a concentrated "Class III soft rock (strength 5-15 MPa)" area in the upper right corner of the section.

[0115] S3, Intensity Field Reconstruction and Decision Reasoning.

[0116] The support decision knowledge graph reasoning engine 202 receives the surrounding rock strength distribution data output by S2 and reconstructs the surrounding rock strength field cloud map by combining it with the spatial location information of the cutting head. Based on the geological feature areas identified in the surrounding rock strength field cloud map, the support decision knowledge graph reasoning engine 202 performs rule matching in the knowledge base. In this example, the engine matching rule is: "If there is a continuous soft rock area exceeding 0.5 square meters in the corner of the roof with a strength lower than 15 MPa, it belongs to a high-risk area prone to shear failure." Accordingly, the support decision knowledge graph reasoning engine 202 performs multi-constraint solving, generating a dynamic support parameter adjustment scheme under the premise of meeting safety specifications and cost optimization objectives. The specific content of this scheme is: "For the upper right soft rock area, the spacing between anchor bolts is increased from 1000mm×1000mm to 800mm×800mm; the anchor bolt length is increased from 2.4m to 2.8m; and the preload is increased from 60kN to 80kN."

[0117] S4, Human-Machine Collaborative Execution and Feedback.

[0118] The explosion-proof human-machine interface terminal 105 receives and displays the dynamic support parameter adjustment plan and surrounding rock strength field cloud map output by S3, allowing operators to confirm or fine-tune the parameters. The tunneling machine operator, combining on-site visual observation, determines that the perception is accurate and the plan is reasonable, and clicks "Confirm Execution." The system records the final executed support parameters and subsequent mine pressure data monitored by mine pressure monitoring equipment (such as a roof delamination meter) (showing a roof delamination of 3mm, meeting the requirements). Figure 5 The “<5mm” safety standard shown forms a complete feedback data chain. The explosion-proof human-machine interface terminal 105 feeds back the geological feature label information, which has been manually confirmed or corrected, from the complete feedback data chain as incremental fine-tuning data to S2 for online calibration of the surrounding rock strength sensing model 201.

[0119] S5: Data Closed Loop and Cloud Evolution.

[0120] The cloud-based training and evolution platform 300 receives the complete feedback data chain uploaded by S4. The model pre-training and optimization module 301 uses the new sample data in the complete feedback data chain to incrementally train the perceptual model, identifying it as a high-confidence sample and adding it to the global sample pool. Simultaneously, the knowledge base management and evolution module 302 uses the feedback data to optimize the rules and adjust the weights of the knowledge base. Finally, the cloud-based training and evolution platform 300 generates the evolved model parameters and knowledge base version, and distributes them to the edge decision server 200, where S1 and S2 reside, for signal acquisition and perceptual inference in the next cycle, thereby achieving continuous iteration and performance improvement of the system.

Claims

1. A roadway surrounding rock intelligent sensing and dynamic support decision-making system, characterized in that, include: The downhole sensing and interactive execution layer (100) is used to trigger the acquisition of multimodal time-series data packets based on the effective cut-off state, and to receive and execute dynamic support parameter adjustment schemes; An edge decision server (200) contains a surrounding rock strength perception model (201) and a support decision knowledge graph inference engine (202). The edge decision server (200) is used to receive the multimodal time-series data packets, use the surrounding rock strength perception model (201) to infer and generate surrounding rock strength distribution data, use the support decision knowledge graph inference engine (202) to generate the dynamic support parameter adjustment scheme based on the surrounding rock strength distribution data, and transmit the complete feedback data chain to the cloud. The cloud training and evolution platform (300) includes a model pre-training and optimization module (301) and a knowledge base management and evolution module (302). The cloud training and evolution platform (300) is used to receive the complete feedback data chain, perform iterative optimization of the model and knowledge base, generate the evolved model parameters and knowledge base version, and send them to the edge decision server (200) to complete the update.

2. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 1, characterized in that, The downhole sensing and interactive execution layer (100) contains a multi-source heterogeneous sensor array and a data acquisition card. The multi-source heterogeneous sensor array is used to acquire multi-source physical signals for encapsulation to generate the multi-modal time-series data packet. The multi-source heterogeneous sensing array specifically includes: A current sensor (101) installed on the power supply side of the cutting motor is used to collect high-frequency current signals that reflect the characteristics of the cutting load. The vibration sensor (102) installed at the root of the cutting arm or at the gearbox is used to collect triaxial vibration acceleration signals during the cutting process. The acoustic emission sensor (103) installed on the cutter guard plate or the machine body is used to collect high-frequency acoustic emission signals generated by rock fracturing. A vision sensor (104) is installed at the front of the fuselage. The vision sensor (104) is used to acquire images of the rockfall and video streams of the cut area. The data acquisition card is used to aggregate the high-frequency current signal, the triaxial vibration acceleration signal, the high-frequency acoustic emission signal, and the rockfall image and cut area video stream, and encapsulate the high-frequency current signal, the triaxial vibration acceleration signal, the high-frequency acoustic emission signal, and the rockfall image and cut area video stream into the multimodal time-series data packet.

3. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 2, characterized in that, The downhole sensing and interaction execution layer (100) is used to generate the multimodal time-series data packet based on the tunneling machine's operating conditions. The specific steps include: Using the data acquisition card as a slave clock node, a microsecond-level synchronous clock reference is established with the edge decision server (200) as the master clock node using the IEEE 1588 clock synchronization protocol; Read the operating data of the tunneling machine PLC controller, and identify the effective pure truncation stage based on the microsecond-level synchronous clock reference and the preset logic threshold criterion, trigger the acquisition of multi-source signals and encapsulate them to generate the multi-modal timing data packet; The logic threshold criteria include: a preset no-load current threshold, set to 1.1 to 1.2 times the rated no-load current value of the cutting motor; a preset static pressure fluctuation threshold, set to 1.5 times the peak value of the hydraulic system pressure fluctuation when the cutting arm is in a static hovering state; and a propulsion speed threshold, set to greater than 0 m / min.

4. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 1, characterized in that, The surrounding rock strength sensing model (201) is configured to receive and process the multimodal time-series data packets in the following manner: The training is based on a model-independent meta-learning strategy, and backpropagation and gradient descent updates are performed on the task-specific layer at the edge using a minimal sample calibration set to adapt to the lithological characteristics of the current working face. The built-in attention mechanism module dynamically allocates weights based on the signal-to-noise ratio of each modal signal in the multimodal time-series data packet, and performs weighted fusion of multimodal features to generate instantaneous surrounding rock strength inference values.

5. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 4, characterized in that, The edge decision server (200) is used to perform the following intensity field reconstruction steps to generate the surrounding rock intensity distribution data: The joint motion variables of the cutting arm are read in real time, and the multi-link kinematic model of the cutting arm is established using the DH parameter method to calculate the three-dimensional spatial coordinates of the cutting head tip. The instantaneous surrounding rock strength inference value is mapped to the three-dimensional spatial coordinates, and the ordinary Kriging spatial interpolation algorithm is used to fill the data voids to generate continuous and complete surrounding rock strength distribution data for the entire cross section.

6. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 5, characterized in that, The support decision knowledge graph reasoning engine (202) internally stores a mine support domain ontology model with an entity relation triplet topology structure, and is used to perform the following reasoning steps based on the surrounding rock strength distribution data: Based on the geological feature areas identified in the surrounding rock strength distribution data, a set of candidate support components is retrieved by rule matching in the knowledge base; The candidate support component set is filtered by introducing the tunnel cross-sectional dimensions and ground stress conditions as boundary constraints, and the dynamic support parameter adjustment scheme, which includes the anchor bolt length, diameter, spacing and preload value, is generated with the support cost entity as the optimization objective function.

7. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 1, characterized in that, The downhole sensing and interactive execution layer (100) also includes an explosion-proof human-machine interface terminal (105), which is used to perform the following human-machine collaborative interaction steps based on the dynamic support parameter adjustment scheme: Receive and display the surrounding rock strength distribution data and the dynamic support parameter adjustment scheme, and provide three interactive response logics: confirm execution, parameter fine-tuning, and rejection with manual input. The surrounding rock strength distribution data, the initial dynamic support parameter adjustment scheme, the manual operation log, and the support effect data are associated and packaged to construct the complete feedback data chain and fed back to the edge decision server (200).

8. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 7, characterized in that, The edge decision server (200) is used to perform the following data processing steps before transmitting the complete feedback data chain: Analyze the complete feedback data chain to identify abnormal intervention records that include parameter fine-tuning operations, rejection operations, or roof delamination exceeding the safety threshold; The samples corresponding to the abnormal intervention records are marked as high-value difficult case samples; Encryption is performed on the high-value difficult sample in order to transmit the complete feedback data chain to the cloud training and evolution platform (300). The safety threshold is 50mm.

9. The intelligent sensing and dynamic support decision-making system for roadway surrounding rock according to claim 1, characterized in that, The knowledge base management and evolution module (302) is used to perform the following knowledge evolution steps using the complete feedback data chain: By traversing the complete feedback data chain, the Euclidean distance between the system-recommended support parameters and the final manual execution parameters is calculated to identify decision conflict events. For the aforementioned decision conflict events, an association rule mining algorithm is used to analyze the implicit association between geological features and manually corrected parameters to generate new candidate rules; The new candidate rules are logically conflict-checked. If the check passes, the new candidate rules are updated to the ontology layer of the knowledge graph, and the evolved model parameters and knowledge base version are generated.

10. A method for intelligent sensing and dynamic support decision-making of roadway surrounding rock, characterized in that, The intelligent sensing and dynamic support decision-making system for roadway surrounding rock, as described in any one of claims 1-9, comprises the following steps: The underground sensing and interactive execution layer (100) reads the working condition data of the tunneling machine PLC, identifies the effective pure cutting stage, and triggers the multi-source heterogeneous sensor array to collect data synchronously, generating multi-modal time-series data packets. The edge decision server (200) receives the multimodal time-series data packet, uses the built-in surrounding rock strength perception model (201) to extract and infer the signal features, generates instantaneous surrounding rock strength inference value, and combines it with the three-dimensional spatial coordinate mapping of the cutting head, and outputs the surrounding rock strength distribution data of the current exposed section through spatial interpolation. The support decision knowledge graph reasoning engine (202) receives the surrounding rock strength distribution data, reconstructs the surrounding rock strength field cloud map by combining the spatial location information of the cutting head, and performs rule matching and multi-constraint solution in the knowledge base based on the geological feature areas identified by the surrounding rock strength field cloud map to generate a dynamic support parameter adjustment scheme. The explosion-proof human-machine interaction terminal (105) receives and displays the dynamic support parameter adjustment scheme and the surrounding rock strength field cloud map for the operator to confirm or fine-tune the parameters. The system records the final support parameters and the subsequent monitored mine pressure data to form a complete feedback data chain and feeds it back to the edge decision server (200). The edge decision server (200) transmits the complete feedback data chain to the cloud training and evolution platform (300). The cloud training and evolution platform (300) uses the new sample data in the complete feedback data chain to train the perception model and optimize the rules of the knowledge base, generate the evolved model parameters and knowledge base version, and send them to the edge decision server (200).

Images