A multi-dimensional perception and self-adaptive tool changing method and system for a shield machine under low-speed pressure maintaining

Abstract

The application discloses a multi-dimensional perception and self-adaptive tool changing method and system for a shield machine under low-speed pressure keeping, the tool changing system comprises a perception layer, a decision layer and a control layer; the tool changing method comprises the following steps: obtaining three-dimensional comprehensive data through a sensor; the three-dimensional comprehensive data is sent into a multi-modal space-time graph neural network MST-GNN through a sensor data fusion module, and a tool changing priority, a tool changing type label, a tool changing number and a tool changing number are output; the target tool is instantaneously positioned through the tool changing type label; angle error calculation is performed, a zero relative window is confirmed, and tool changing is realized. The application proposes the following method to cope with the long process that the traditional method must be completely stopped and depressurized, that is, under the micro-stop working condition of the low-speed cutter head and the pressure keeping of the soil bin, three-dimensional perception of defects-life-geology is fused, millisecond-level decision is made, RFID instantaneous positioning is performed, five times spline closed-loop tool changing is performed, and the tool changing time is reduced from hours to minutes.

Description

Technical Field

[0001] This invention relates to the field of shield tunneling cutterhead replacement technology, specifically to a multi-dimensional sensing and adaptive cutterhead replacement method and system for shield tunneling machines under low-speed pressure maintenance. Background Technology

[0002] Generally speaking, shield tunnels need to pass through complex strata such as uneven hardness, highly abrasive boulders, soft upper layers and hard lower layers, and well-developed karst caves.

[0003] Due to the complex geological conditions within the tunnel, tunnel boring machines (TBMs) require cutter replacements for continued excavation. Existing TBM cutter replacement technologies suffer from the following problems:

[0004] 1) Traditional cutterhead replacement technology relies on the complete shutdown of the cutterhead and the depressurization of the earth chamber, which leads to tunneling interruptions for several hours, forcing the construction window to be extended, increasing labor and equipment costs exponentially, and the depressurization operation in high gas and water-rich environments is extremely risky.

[0005] 2) After shutdown, manual visual inspection or disassembly inspection is generally used, which cannot capture multi-source defects such as wear, cracks, corrosion, chipping and abrupt changes in rock strata in real time;

[0006] 3) The lack of a real-time fusion model of defect-life-geological three-dimensional data means that the timing of tool replacement is judged solely by experience. Too early a tool replacement results in tool waste, while too late a tool replacement leads to tool breakage, tool head jamming, or even ground subsidence.

[0007] 4) The positioning process requires an additional rotating ring at the same speed or a large mechanical compensation mechanism, which is complex in structure, heavy in weight, has many failure points, is time-consuming and expensive to maintain, and has strict requirements on the space of the cutter head, which is not conducive to compact design. Summary of the Invention

[0008] The purpose of this invention is to address the lack of specific and targeted methods for micro-stop cutter replacement in tunnel boring machines (TBMs) under complex terrain conditions. This invention provides a multi-dimensional sensing and adaptive cutter replacement method and system for TBMs under low-speed pressure maintenance, which can offer a unified, accurate, and efficient cutter replacement method for TBMs.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: a multi-dimensional sensing and adaptive cutterhead replacement method for tunnel boring machines under low-speed pressure maintenance, the cutterhead replacement method comprising the following steps:

[0010] S1. In the perception layer, the three-dimensional comprehensive data of the tool is acquired in real time through sensors;

[0011] S2. In the decision-making layer, the three-dimensional integrated data is fused into the multimodal spatiotemporal graph neural network MST-GNN through the sensor data fusion module in the perception layer, and outputs the tool change priority, tool change type label, tool change quantity and tool change number.

[0012] S3. In the control layer, the target tool is instantaneously positioned by means of the tool change type label and in conjunction with the low-speed rotation of the tool turret;

[0013] S4. In the control layer, after the tool robot synchronizes with the target tool's angular velocity, the angular error is calculated to confirm the zero relative window, so that the tool robot can complete the tool removal, tool insertion and locking, and realize the tool change.

[0014] As a further embodiment of the present invention: in the perception layer, the three-dimensional integrated data includes wear defect data A1, corrosion defect data A2 and fracture defect data A3 in defect dimension A, rock layer hardness data B1 in geological hardness dimension B, and life data C1 in life dimension C.

[0015] Among them, the wear defect data A1 is obtained by using an electromagnetic inductance wear sensor to obtain the radial wear amount ΔW of the cutter ring;

[0016] Corrosion defect data A2, the percentage of corrosion area ΔC in the bearing sealing region was obtained by an electrochemical impedance spectroscopy corrosion sensor;

[0017] Fracture defect data A3 is obtained by using an acoustic emission crack sensor to obtain the tool shaft crack length ΔL or the cumulative acoustic emission energy AE.

[0018] Rock hardness data B1, the uniaxial compressive strength UCS of the rock strata was obtained by using an advanced ground-penetrating radar antenna;

[0019] Lifetime data C1 is used to obtain the equivalent load cycle count Neq through a strain gauge load cycle counter;

[0020] The three-dimensional integrated data is processed by the sensor data fusion module, and an N×5×T spatiotemporal tensor is output for the decision-making layer to make decisions.

[0021] As a further aspect of the present invention: In the sensing layer, during a micro-stop condition where the cutterhead rotates at a low speed of ≤2 rpm and the soil chamber is pressurized, the necessity of tool replacement is determined based on defect dimension A, geological hardness dimension B, and lifespan dimension C, including:

[0022] In defect dimension A, tool defects are divided into:

[0023] Wear defect data A1 is based on the radial wear amount ΔW of the tool ring. When ΔW ≥ 25mm, A1 is determined to have reached the tool replacement threshold.

[0024] Corrosion defect data A2 is defined by the percentage of corrosion area ΔC in the bearing seal region. When ΔC ≥ 30%, A2 is determined to have reached the tool replacement threshold.

[0025] Fracture defect data A3. Taking the cutter shaft crack length ΔL or the cumulative value of acoustic emission energy AE as the index, when ΔL≥15mm or AE≥100mV·ms, it is determined that the A3 defect reaches the tool change threshold;

[0026] Among them, if any tool defect reaches its threshold, the tool change instruction for defect dimension A is triggered;

[0027] In the geological hardness dimension B, the geological hardness grade is defined as follows:

[0028] Soft rock: UCS≤30Mpa, replace with a scraper type tool;

[0029] Medium hard rock: 30MPa<UCS≤90MPa, replace with a hob type tool;

[0030] Extremely hard rock: UCS>90MPa, heavy hob type tool;

[0031] When the real-time value of UCS obtained by the advanced geological radar antenna or the inversion of cutterhead vibration crosses from a low grade to a high grade, or when the continuous tunneling length in the extremely hard rock grade is ≥10m, the tool change instruction for the geological hardness dimension B is triggered, and the tool type is replaced according to the geological hardness grade;

[0032] In the life dimension C, the defined service life of the tool is the equivalent load cycle number Neq. When Neq≥the designed rated life Nr, the tool change instruction for the life dimension C is triggered.

[0033] As a further solution of the present invention: In the sensing layer, in order to obtain the three-dimensional comprehensive data of the defect dimension A, the geological hardness dimension B and the life dimension C in real time, a sensor system is arranged on the cutterhead and tools of the shield machine, and is arranged and connected in the following manner:

[0034] In the defect dimension A:

[0035] Wear monitoring module. Inside the cutter ring matrix of each replaceable hob on the front of the cutterhead, two groups of electromagnetic mutual inductance wear sensors are buried along the circumference at the same radius and at equal angular intervals, for real-time output of the radial wear amount ΔW of the cutter ring;

[0036] Corrosion monitoring module: A group of electrochemical impedance spectroscopy corrosion sensors are attached to the inner surface of the bearing seal end cover of each hob, for real-time output of the corrosion area percentage ΔC of the bearing seal area;

[0037] Fracture monitoring module: A group of acoustic emission crack sensors are pasted on the transition fillet area of the cutter shaft of each hob, for real-time capture of the cumulative value of acoustic emission energy AE of crack initiation and propagation;

[0038] In geological hardness dimension B: At least three sets of ground-penetrating radar antennas are equally spaced along the circumferential direction on the outer periphery of the cutterhead. These antennas are used to transmit electromagnetic waves in front of the tunnel face before the tunnel boring machine advances and to receive reflected signals in order to obtain the uniaxial compressive strength (UCS) of the rock strata.

[0039] In the life dimension C: On the cutter shaft surface of each cutter, at least one pair of strain-type load cycle counters are symmetrically attached along the axial centerline to collect the torque and thrust load spectrum of the cutter in real time during the tunneling process and calculate the equivalent load cycle number Neq.

[0040] Sensor data fusion module: Located in the explosion-proof electrical control box at the center rotary joint of the cutter head, it includes a multi-channel synchronous sampling unit, a time synchronization unit, and an edge computing unit;

[0041] The multi-channel synchronous sampling unit is connected to A1, A2, A3, B1 and C1 respectively via shielded cables, with a sampling frequency ≥1kHz;

[0042] The time synchronization unit receives a 1PPS pulse signal from the main control PLC of the tunnel boring machine, which is used to timestamp the three-dimensional integrated data, with a time synchronization error of ≤1ms.

[0043] As a further aspect of the present invention: In the perception layer, a multi-dimensional state matrix is ​​constructed based on the aforementioned three-dimensional integrated data, and the "isolated state of a single tool" is upgraded to a globally coupled relationship network of "tool head-tool-geology-time" through graph structure modeling. This step specifically includes:

[0044] Using N replaceable cutting tools on the tool turret as nodes, a five-dimensional state vector is generated in real time for each tool i:

[0045] S it =[Wear, corrosion, fracture, geological hardness, life] T

[0046] =[ΔW it ΔC it ΔL it / AE it UCS it Neq it ] T

[0047] Wherein, ΔW it ΔC it ΔL it / AE it UCS it and Neq it Provided directly by A1, A2, A3, B1 and C1 respectively, where i represents the tool number and t represents a specific moment of monitoring;

[0048] 2) Stack the five-dimensional state vectors of the N tools in order of tool number to form an N×5-dimensional global state matrix:

[0049]

[0050]

[0051] 3) Using the hobs as nodes, with node features being the aforementioned five-dimensional state vector, and the mechanical coupling relationship between the hobs and the rotational dynamics of the cutter head as edges, construct an undirected graph G = (V, E), where the edge weight formula in the undirected graph G is as follows:

[0052]

[0053] Edge weight Q ij The physical distance d between hobs i and j on the cutter head ij The speed of the cutter head, ω, and σ and α are determined together, and are preset hyperparameters.

[0054] 4) Introduce a time dimension into the node features of the undirected graph G, and transform the state matrix M over T consecutive sampling periods. t Stacked as N×5×T spacetime tensors.

[0055] As a further aspect of the present invention: in the decision-making layer, the N×5×T spatiotemporal tensor is processed by a multimodal spatiotemporal graph neural network (MST-GNN) to obtain the tool-changing decision result, specifically including:

[0056] The N×5×T spatiotemporal tensor is modally decomposed into five-channel tensors: wear channel, corrosion channel, fracture channel, geological hardness channel, and lifetime channel. LayerNorm is applied independently to each channel to obtain the normalized five-channel feature tensor.

[0057] X c ∈R N×T c∈{1,…,5};

[0058] Spatial Graph Convolutional Network (S-TCN) uses the adjacency matrix of an undirected graph G as a mask to perform graph convolution at each time point τ∈{1,…,T}:

[0059]

[0060] Where l represents the layer index and D is the degree matrix. These are trainable spatial weights; after passing through the Ls layer, the spatial feature tensor is obtained. ds represents the spatial embedding dimension;

[0061] The temporal causal convolutional layer T-TCN performs dilated causal convolution along the temporal dimension for each node i:

[0062] Z i =CausalConv1D(H s [i,:,:];W t ,k t ,d t )

[0063] Where, k t d is the kernel size. t As the expansion factor, via L s The time feature tensor is obtained after the layer. T′ is the time step after downsampling;

[0064] Multimodal attention fusion: After linearly mapping the five-channel feature tensor, calculate the cross-modal attention weights.

[0065]

[0066] Among them, Q c K c V c They are query, key-value mapping, and output fused features respectively.

[0067] The results are output through the task output header. The output results include:

[0068] Tool change priority P∈[0,1] N ;

[0069] Tool replacement type label T∈{wear type, corrosion type, fracture type, geological adaptability type, end-of-life type} N ;

[0070] Tool change count K = Round! (∑) i σ(P i ≥θ)), the threshold θ is adjusted in real time by engineering constraints;

[0071] Tool change number set ID = {i | P i ≥θ}.

[0072] As a further aspect of the present invention: in the control layer, instantaneous positioning of the target tool includes the following steps:

[0073] An absolute encoder is coaxially mounted on the end face of the tool turret spindle, and the output signal is transmitted to the tool robot controller in real time via an EtherCAT bus.

[0074] Each replaceable hob has a passive UHF-RFID tag embedded inside its base. The passive UHF-RFID tag contains a UID and a fixed offset Δθ. i , where Δθ iThis is the mechanical angle difference between the tool centerline and the encoder zero position;

[0075] After the tool robot controller is powered on, it performs a one-time scan: by slowly rotating the tool head, the RFID read / write antenna array sequentially reads the UID and Δθ of each tool. i A non-power-loss index array Index[N] = {Δθ1, Δθ2, ..., Δθ} is created in the memory of the tool robot controller. N};

[0076] At any time t0, the real-time angle of the tool head is read by the absolute encoder. Then the instantaneous target angular coordinates of the Nth tool are:

[0077]

[0078] Dynamic compensation and closed-loop control specifically include: the tool robot controller will adjust θ... target (t0) is input into the quintic spline trajectory planner to generate the angular displacement-time curve of the end effector.

[0079] As a further aspect of the present invention: in the control layer, the instantaneous target angular coordinate θ is obtained by instantaneously positioning the target tool. target (t), under the micro-stop tool change condition, the tool robot performs trajectory pre-planning, speed chasing, and error closed-loop control. This step includes:

[0080] Within the joint space of the cutting tool robot, with the current joint angle vector q curr Let q be the starting and target joint angle vectors. target Assuming the endpoint is t, a fifth-order polynomial trajectory q(t) is generated. A fifth-order spline trajectory pre-planning is then performed, and its expression is:

[0081] q(t) = a0 + a1t + a2t 2 +a3t 3 +a4t 4 +a5t 5 ,(0≤t≤Δt pred )

[0082] The boundary conditions that the fifth-order polynomial trajectory q(t) must satisfy at the starting and ending points include:

[0083] Starting from the robot's "current actual joint angles," avoid position jumps:

[0084] q(0)=q curr

[0085] At the specified time Δt pred The inner diameter accurately reaches the target tool angle, achieving "zero relative" alignment:

[0086] q(Δt pred )=q target

[0087] Starting speed and acceleration are zero to ensure a smooth start.

[0088]

[0089] The end-effector velocity must be matched with the angular velocity ω of the cutter head to keep the robot end-effector relatively stationary with respect to the cutter.

[0090]

[0091] The final acceleration should be zero to avoid mechanical impact or residual vibration.

[0092]

[0093] Where ω is the instantaneous angular velocity of the cutter head, which is calculated in real time by the differential of the encoder:

[0094]

[0095] The spline solver is based on the maximum joint velocity v max Maximum acceleration a max The theoretical arrival time Δt is automatically given. pred ;

[0096] To perform velocity feedforward and angular synchronization, the controller uses -ω as the end-effector angular velocity feedforward command to ensure that at t = Δt pred When the relative angular velocity between the robot end effector and the target tool is Δω≈0, the system switches to angular synchronization mode after reaching the target angle.

[0097] 4) Complete velocity feedforward and angle synchronization, implement error closure and reentry mechanisms, and the real-time angle error is:

[0098] Δθ(t)=θ target (t)-θ end-effector (t)

[0099] When |Δθ(t)|≤0.5° and the holding time>50ms, the zero relative window is determined to be established, and the tool changing mechanism is allowed to operate;

[0100] If |Δθ(t)|>0.5° or the joint torque mutation>15N·m, immediately trigger reentry: the tool robot decelerates to zero and records q. reentry Recalculate θ target (t).

[0101] As a further aspect of the present invention: In the control layer, once the zero relative window is established and confirmed, the tool changing mechanism is allowed to operate, and the tool robot controller immediately initiates the following tool changing sequence, the process of which specifically includes:

[0102] After the zero relative window is confirmed, the locking state is detected. The normal force Fn of the tool locking surface is monitored in real time by a six-dimensional force sensor. If Fn≤2N and lasts for more than 50ms, the locking force is determined to be released and the tool can be pulled out.

[0103] After the electromagnetic / pneumatic lock is released, the Z-axis electric slide table pushes outward by 50mm at a speed of 100mm / s to complete the tool removal.

[0104] The photoelectric switch detects the position at the end of the travel; if the position is not reached within 3 seconds, the system will interrupt and trigger an alarm.

[0105] The tool robot picks up a new tool, and the Z-axis slide inserts it into the tool holder in the opposite direction at 80mm / s.

[0106] After insertion, apply a locking torque of 180 N·m, with a torque closed-loop tolerance of ±5 N·m;

[0107] After locking is completed, the six-dimensional force sensor confirms that the normal force Fn' ≥ 800N, ensuring locking;

[0108] The end-effector UHF-RFID reader reads the new tool's UID and compares it with the tool robot controller's array. If the verification is successful, the lifespan counter is reset to zero. If the UID verification fails, and two rereads still fail, an anomaly is marked.

[0109] A cutter changing system based on a multi-dimensional sensing and adaptive cutter changing method for tunnel boring machines under low-speed pressure maintenance, the cutter changing system comprising:

[0110] The sensing layer senses and collects data on the tool's condition, including electromagnetic inductance wear sensors, electrochemical impedance spectroscopy corrosion sensors, acoustic emission crack sensors, advanced ground-penetrating radar antennas, strain gauge load cycle counters, and sensor data fusion modules.

[0111] The decision layer is used to intelligently judge the tool status based on the three-dimensional comprehensive data obtained from the perception layer, and generates instructions through the multimodal spatiotemporal graph neural network MST-GNN, outputting tool change priority, tool change type label, tool change quantity and tool change number, and sending them to the control layer through the EtherCAT bus.

[0112] The control layer performs instantaneous positioning of the tool to be replaced and performs trajectory closed-loop control of the tool robot based on its position to complete the tool changing action. It includes the tool robot, absolute encoder, passive UHF-RFID tag, tool robot controller and six-dimensional force sensor.

[0113] The beneficial effects of this invention are:

[0114] 1) Through the micro-stop design of low-speed rotation of the cutterhead and pressure maintenance of the soil chamber, the traditional lengthy process of having to completely stop and depressurize is completely eliminated, and the interruption time of a single cutter change is reduced from hours to minutes, significantly shortening the tunnel excavation cycle.

[0115] 2) The defect-life-geology three-dimensional perception system replaces manual visual inspection. The data of electromagnetic mutual inductance, electrochemical impedance, acoustic emission, ground radar and strain gauge are integrated in real time, making wear, corrosion, cracks and changes in rock hardness clear at a glance. The accuracy of tool replacement decision is increased by more than 100%, avoiding economic losses and safety risks caused by premature scrapping or sudden chipping.

[0116] 3) The absolute rotary encoder and passive UHF-RFID tag work together for positioning, and with the joint space five-fold spline trajectory and angular synchronization closed loop, a millisecond-level relative static window is achieved. The robot can accurately insert and remove tools without the need for a rotating ring at the same speed. The overall structure is lighter, more reliable, and requires less maintenance.

[0117] 4) By dynamically adjusting the threshold through the multimodal spatiotemporal graph neural network MST-GNN, scrapers, roller cutters or heavy roller cutters can be automatically matched according to different strata, improving the tool utilization rate by more than 30%, while reducing spare parts inventory and transportation costs, providing an efficient, safe and economical integrated solution for long-distance tunnel construction in complex geology.

[0118] 5) To address the lengthy process of traditional methods that require complete shutdown and depressurization, this invention proposes a micro-shutdown operation with low cutter head speed and soil chamber pressure maintenance. It integrates three-dimensional perception of defects, lifespan, and geology, enables millisecond-level decision-making, instantaneous RFID positioning, and five-strip closed-loop tool changing, eliminating the need for shutdown and depressurization or the need for a rotating ring at the same speed. The tool changing time is reduced from hours to minutes. Attached Figure Description

[0119] Figure 1 This is a schematic diagram of the overall process structure of the present invention;

[0120] Figure 2 This is a flowchart illustrating the signal transmission process of the present invention.

[0121] Figure 3 This is an analysis diagram of the tool changing types, sensor selection, and position distribution of the present invention;

[0122] Figure 4 This is a flowchart and framework diagram of the overall system of the decision-making layer of this invention;

[0123] Figure 5 This is a flowchart and framework diagram of the overall control layer system of the present invention. Detailed Implementation

[0124] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0125] Embodiment 1, as Figure 1 and Figure 2 shown, this embodiment provides a multi-dimensional perception and adaptive tool changing method under low-speed pressure maintenance of a shield machine. This tool changing method divides the entire process into a perception layer, a decision-making layer, and a control layer based on the micro-stop condition of the cutter head rotating (turning) at a low speed and the soil chamber maintaining pressure. The specific steps are as follows:

[0126] First: In the perception layer, in the micro-stop working condition where the cutter head rotates at a low speed of ≤2 rpm and the soil chamber maintains pressure, judge the necessity of tool changing based on the defect dimension A, the geological hardness dimension B, and the life dimension C.

[0127] Judge the necessity of tool changing through the above three dimensions and execute according to the following method, as Figure 3 shown:

[0128] 1) In the defect dimension A, divide the tool defects into:

[0129] 11) Wear defect data A1, with the radial wear amount ΔW of the cutter ring as the index. When ΔW≥25 mm, it is determined that A1 reaches the tool changing threshold;

[0130] 12) Corrosion defect data A2, with the percentage of the corrosion area ΔC in the bearing seal area as the index. When ΔC≥30%, it is determined that A2 reaches the tool changing threshold;

[0131] 13) Fracture defect data A3, with the crack length ΔL of the tool shaft or the cumulative value AE of the acoustic emission energy as the index. When ΔL≥15 mm or AE≥100 mV·ms, it is determined that the A3 defect reaches the tool changing threshold;

[0132] Among them, if any tool defect reaches its threshold, the tool changing instruction for the defect dimension A is triggered.

[0133] 2) In the geological hardness dimension B, define the geological hardness grade:

[0134] 21) Soft rock: UCS≤30 Mpa, replace with a scraper type tool;

[0135] 22) Medium hard rock: 30 MPa < UCS≤90 MPa, replace with a hob type tool;

[0136] 23) Extremely hard rock: UCS>90 MPa, heavy hob type tool;

[0137] When the real-time or cutterhead vibration inversion UCS value obtained by the advanced ground-penetrating radar antenna (or cutterhead vibration inversion) crosses from a lower level to a higher level, or when the continuous tunneling length in the extremely hard rock level is ≥10m, the cutter change command of geological hardness dimension B is triggered, and the cutter type is changed according to the geological hardness level.

[0138] 3) In the life dimension C, the tool's service life is defined as the equivalent load cycle number Neq. When Neq ≥ the design rated life Nr, the tool change command in the life dimension C is triggered.

[0139] It should be noted that ΔW, ΔC, ΔL, AE, UCS, and Neq are all collected in real time by a sensor system deployed on the tool turret and the tool, and processed synchronously by the sensor data fusion module.

[0140] Second: In the sensing layer, in order to acquire three-dimensional comprehensive data of defect dimension, geological hardness dimension, and life dimension in real time, a sensor system is deployed on the shield machine cutterhead and cutters. This sensor system includes:

[0141] 1) Obtain wear defect data, corrosion defect data, and crack defect data in defect dimension A.

[0142] 11) Wear monitoring module: At least two sets of electromagnetic inductance wear sensors are embedded in the cutter ring base of each replaceable hob on the front of the cutter head, with the same radius and equal angle intervals along the circumference. The at least two sets of electromagnetic inductance wear sensors are arranged vertically and vertically staggered in the thickness direction of the cutter ring, with a stagger distance of 30%-40% of the thickness of the cutter ring, to form a differential measurement structure for real-time output of the radial wear amount ΔW of the cutter ring.

[0143] 12) Corrosion monitoring module: At least one set of electrochemical impedance spectroscopy corrosion sensors are attached to the inner surface of the bearing sealing end cover of each hob. The probe of the electrochemical impedance spectroscopy corrosion sensor maintains an electrolyte gap of 0.1mm-0.3mm with the end cover metal surface and is isolated from the external mud by a sealing ring. It is used to output the percentage of corrosion area ΔC of the bearing sealing area in real time.

[0144] 13) Fracture monitoring module: At least one set of acoustic emission crack sensors are attached to the transition fillet area of ​​the cutter shaft of each hob; the center frequency of the acoustic emission crack sensor is 150kHz±10%, and it is tightly attached to the surface of the cutter shaft through a high-temperature resistant coupling agent, which is used to capture the cumulative acoustic emission energy AE of crack initiation and propagation in real time.

[0145] 2) Obtain rock layer hardness data in geological hardness dimension B.

[0146] Three sets of advanced ground-penetrating radar antennas are evenly spaced along the circumference of the cutterhead. These antennas are shielded dipole antennas with their radiating surfaces flush with the front of the cutterhead and isolated from the external rock and soil by a wear-resistant ceramic cover. The three sets of antennas operate in a time-division manner, forming a 120° sector scan. They are used to transmit electromagnetic waves with a center frequency of 100MHz in the range of 0–30m in front of the tunnel face before the tunnel boring machine advances and to receive reflected signals in order to obtain the uniaxial compressive strength (UCS) of the rock strata ahead.

[0147] 3) Obtain lifespan data in lifespan dimension C.

[0148] At least one pair of strain gauge load cycle counters are symmetrically attached to the cutter shaft surface of each cutter along the axial centerline. The strain gauge load cycle counter is a full-bridge strain gauge group, and the measurement direction is consistent with the direction of the maximum principal stress of the cutter shaft. It is protected by a waterproof adhesive layer and a metal foil shielding layer. It is used to collect the torque and thrust load spectrum of the cutter in real time during the tunneling process and calculate the equivalent load cycle number Neq.

[0149] 4) Sensor data fusion module: Located in the explosion-proof electrical control box at the center rotary joint of the cutter head, it includes a multi-channel synchronous sampling unit, a time synchronization unit, and an edge computing unit.

[0150] Among them, the multi-channel synchronous sampling unit is connected to A1, A2, A3, B1 and C1 respectively through shielded cables, and the sampling frequency is ≥1kHz;

[0151] The time synchronization unit receives a 1PPS pulse signal from the main control PLC of the tunnel boring machine, which is used to timestamp all three-dimensional integrated data. The time synchronization error is ≤1ms.

[0152] The edge computing unit performs real-time calculations and caches ΔW, ΔC, AE, UCS, and Neq, and uploads the three-dimensional integrated data to the tool robot controller via the EtherCAT bus.

[0153] Third: In the perception layer, a multi-dimensional state matrix is ​​constructed based on real-time acquired 3D integrated data, and the "isolated state of a single tool" is upgraded to a globally coupled relationship network of "tool head-tool-geology-time" through graph structure modeling. This step specifically includes:

[0154] 1) Using the N replaceable tools on the tool turret as nodes, a five-dimensional state vector is generated in real time for each tool i:

[0155] S it =[Wear, corrosion, fracture, geological hardness, life] T

[0156] =[ΔW it ΔC it ΔLit / AE it UCS it New it ] T

[0157] Wherein, ΔW it ΔC it ΔL it / AE it UCS it and Neq it Provided directly by A1, A2, A3, B1 and C1 respectively, where i represents the tool number and t represents a specific moment of monitoring;

[0158] 2) Stack the five-dimensional state vectors of the N tools in order of tool number to form an N×5-dimensional global state matrix:

[0159]

[0160] 3) Using the hobs as nodes, with node features being the aforementioned five-dimensional state vector, and the mechanical coupling relationship between the hobs and the rotational dynamics of the cutter head as edges, construct an undirected graph G = (V, E), where the edge weight formula in the undirected graph G is as follows:

[0161]

[0162] Edge weight Q ij The physical distance d between hobs i and j on the cutter head ij The speed of the cutter head, ω, and σ and α are determined together, and are preset hyperparameters.

[0163] 4) Introduce a time dimension into the node features of the undirected graph G, and transform the state matrix M over T consecutive sampling periods. t Stacked into N×5×T spatiotemporal tensors, providing raw data with "zero latency" for subsequent decisions.

[0164] Fourth: In the decision-making layer, the N×5×T spatiotemporal tensor is processed by a multimodal spatiotemporal graph neural network (MST-GNN) to obtain the tool-changing decision results. Spatial graph convolution (S-TCN) captures the mechanical coupling between tools, and temporal causal convolution (T-TCN) captures the evolutionary trend of defects, geology, and lifespan, such as... Figure 4 As shown, this step specifically includes:

[0165] 1) In the input adaptation, the N×5×T spatiotemporal tensor is modally decomposed into five-channel tensors: wear channel, corrosion channel, fracture channel, geological hardness channel, and lifespan channel. LayerNorm is applied independently to each channel to obtain the normalized five-channel feature tensor:

[0166] Xc ∈R N×T c∈{1,…,5};

[0167] 2) Spatial Graph Convolution (S-TCN) uses the adjacency matrix of the undirected graph G as a mask to perform graph convolution at each time point τ∈{1,…,T}:

[0168]

[0169] Where l represents the layer index and D is the degree matrix. These are trainable spatial weights; after passing through the Ls layer, the spatial feature tensor is obtained. ds represents the spatial embedding dimension;

[0170] 3) The Temporal Causal Convolutional Layer (T-TCN) performs dilated causal convolution along the temporal dimension for each node i:

[0171] Z i =CausalConv1D(H s [i,:,:];W t ,k t ,d t )

[0172] Where, k t d is the kernel size. t As the expansion factor, via L s The time feature tensor is obtained after the layer. T′ is the time step after downsampling;

[0173] 4) Multimodal attention fusion: After linear mapping of the five-channel feature tensor, cross-modal attention weights are calculated.

[0174]

[0175] Among them, Q c K c V c They are query, key-value mapping, and output fused features respectively.

[0176] 5) Output the results through the task output header. The output results include:

[0177] Tool change priority P∈[0,1] N ;

[0178] Tool replacement type label T∈{wear type, corrosion type, fracture type, geological adaptability type, end-of-life type} N ;

[0179] Tool change count K = Round! (∑) i σ(P i≥θ)), the threshold θ is adjusted in real time by engineering constraints;

[0180] Tool change number set ID = {i | P i ≥θ}.

[0181] Among them, the optimal weight model for training the multimodal spatiotemporal graph neural network MST-GNN is deployed in the local edge computing unit of the tunnel boring machine, with a single forward inference latency of ≤50ms; the inference results are sent to the control layer via the EtherCAT bus with an update cycle of 1kHz.

[0182] Fifth: In the control layer, based on the tool change type label output by the multimodal spatiotemporal graph neural network MST-GNN, and in conjunction with the low-speed rotation state of the tool head, the required tool (target tool) is instantaneously positioned. For example... Figure 5 As shown, the process specifically includes:

[0183] 1) A 25-bit absolute encoder with a resolution of 0.0055° / tick is coaxially mounted on the end face of the tool turret spindle. The output signal is transmitted to the tool robot controller in real time via EtherCAT bus at a frequency of 4kHz.

[0184] 2) Each replaceable hob has a passive UHF-RFID tag embedded inside its base. The passive UHF-RFID tag contains a 64-bit UID and a fixed offset Δθ. i , where Δθ i This is the mechanical angular difference between the tool centerline and the encoder zero position, with an accuracy of 0.001°;

[0185] 3) After the tool robot controller is powered on, it performs a one-time scan: the tool head rotates slowly for one revolution, and the RFID read / write antenna array sequentially reads the UID and Δθ of each tool. i A non-power-loss index array Index[N] = {Δθ1, Δθ2, ..., Δθ} is created in the memory of the tool robot controller. N};

[0186] 4) At any time t0, the real-time angle of the tool turret is read by the absolute encoder. Then the instantaneous target angular coordinates of the Nth tool are:

[0187]

[0188] This formula is refreshed by the tool robot controller every 0.25ms to ensure that the positioning error is ≤0.01°;

[0189] 5) Dynamic compensation and closed-loop control, specifically including: the tool robot controller will adjust θ target (t0) is input into the quintic spline trajectory planner to generate the angular displacement-time curve of the end effector.

[0190] It is important to note that during the movement of the cutter head, feedback is provided via 4kHz EtherCAT. The trajectory is corrected in real time to achieve angular synchronization closed loop, ensuring that the relative angular velocity between the end effector and the target tool is ≤0.02° / s; if RFID reading fails, the absolute encoder and mechanical zero point dual-channel verification are enabled, and the positioning error is still kept ≤0.02°.

[0191] Sixth: In the control layer, receive the instantaneous target angular coordinates θ. target After (t), the tool robot's trajectory is pre-planned, its speed is caught up, and its error is controlled in a closed-loop manner under the micro-stop tool change condition. For example... Figure 5 As shown, this step includes:

[0192] 1) Within the joint space of the tool robot, using the current joint angle vector q curr Let q be the starting and target joint angle vectors. target Assuming the endpoint is t, a fifth-order polynomial trajectory q(t) is generated. A fifth-order spline trajectory pre-planning is then performed, and its expression is:

[0193] q(t) = a0 + a1t + a2t 2 +a3t 3 +a4t 4 +a5t 5 ,(0≤t≤Δt pred )

[0194] 2) The boundary conditions that the fifth-order polynomial trajectory q(t) must satisfy at the starting and ending points include:

[0195] 21) Avoid position jumps by starting from the robot's "current actual joint angle":

[0196] q(0)=q curr

[0197] 22) At the specified time Δt pred The inner diameter accurately reaches the target tool angle, achieving "zero relative" alignment:

[0198] q(Δt pred )=q target

[0199] The starting speed and acceleration are zero to ensure a smooth start (without shaking):

[0200]

[0201] The end-effector velocity must be matched with the angular velocity ω of the cutter head to keep the robot end-effector relatively stationary with respect to the cutter.

[0202]

[0203] The final acceleration should be zero to avoid mechanical impact or residual vibration.

[0204]

[0205] Where ω is the instantaneous angular velocity of the cutter head, which is calculated in real time by the differential of the encoder:

[0206]

[0207] The spline solver is based on the maximum joint velocity v max Maximum acceleration a max The theoretical arrival time Δt is automatically given. pred Typical range: 0.15–0.30 s;

[0208] 3) Perform velocity feedforward and angular synchronization. The controller uses -ω as the end-point angular velocity feedforward command to ensure that at t = Δt pred When the relative angular velocity between the robot end effector and the target tool is Δω≈0, after reaching the target angle, the system switches to angular synchronization mode with a control cycle of 250μs, position loop gain Kp≥800Hz, and velocity feedforward coefficient Kv=1.

[0209] 4) Complete velocity feedforward and angle synchronization, implement error closure and reentry mechanisms, and the real-time angle error is:

[0210] Δθ(t)=θ target (t)-θ end-effector (t)

[0211] When |Δθ(t)|≤0.5° and the holding time>50ms, the zero relative window is determined to be established, and the tool changing mechanism is allowed to operate;

[0212] If |Δθ(t)|>0.5° or the joint torque mutation>15N·m, immediately trigger reentry: the robot decelerates to zero and records q. reentry Recalculate θ target (t); with q reentry Five splines are generated again starting from the current point, with a maximum of three re-entries allowed, and the total time is less than 1.5 seconds. If three consecutive re-entries fail, the tool change is abandoned and the next opportunity is waited for.

[0213] It is important to note that once the zero relative window is established and confirmed, the tool change mechanism is allowed to operate, and the tool robot controller immediately initiates the following tool change sequence, the process of which includes:

[0214] After the zero relative window is confirmed, the locking state is detected. The normal force Fn of the tool locking surface is monitored in real time by a six-dimensional force sensor. If and only if Fn≤2N and lasts for more than 50ms, the locking force is determined to be released and the tool removal step is allowed.

[0215] After the electromagnetic / pneumatic lock is released, the Z-axis electric slide table pushes outward by 50mm at a speed of 100mm / s to complete the tool removal.

[0216] The photoelectric switch detects the position at the end of the travel; if the position is not reached within 3 seconds, the system will interrupt and trigger an alarm.

[0217] The tool robot picks up a new tool, and the Z-axis slide inserts it into the tool holder in the opposite direction at 80mm / s.

[0218] After insertion, apply a locking torque of 180 N·m, with a torque closed-loop tolerance of ±5 N·m;

[0219] After locking is completed, the six-dimensional force sensor confirms that the normal force Fn' ≥ 800N, ensuring reliable locking.

[0220] The end-of-line UHF-RFID reader reads the new tool's UID within 0.2s and compares it with the tool robot controller array. If the verification is successful, the lifespan counter is reset to zero. If the UID verification fails, it is marked as abnormal after two rereads. The entire unplug-plug-lock-verify process takes ≤1.2s. If the timeout or deviation exceeds the limit, it immediately returns to the safe position and retryes once. If it fails twice in a row, the tool change is abandoned.

[0221] Example 2: This example provides a multi-dimensional sensing and adaptive cutterhead changing system for tunnel boring machines under low-speed pressure maintenance. This cutterhead changing system can implement the cutterhead changing method in Example 1. The cutterhead changing system includes:

[0222] 1) Sensing layer: This layer senses and collects data on the tool's status, including:

[0223] 11) Electromagnetic mutual inductance wear sensor: embedded inside the cutter ring of each hob, it measures the radial wear amount ΔW of the cutter ring in real time to determine the "wear-type" tool replacement requirement;

[0224] 12) Electrochemical impedance spectroscopy corrosion sensor: attached to the inner surface of the bearing sealing end cap, it outputs the percentage of corrosion area ΔC in the bearing sealing area in real time, which is used to determine the "corrosion-type" tool replacement requirement;

[0225] 13) Acoustic emission crack sensor: fixed to the transition fillet of the cutter shaft, it captures the crack length ΔL or the cumulative acoustic emission energy AE of the cutter shaft in real time, and is used to determine the "fracture type" tool replacement requirement;

[0226] 14) Advanced ground-penetrating radar antenna: Three sets are evenly distributed around the outer perimeter of the cutterhead, transmitting 100MHz electromagnetic waves forward from 0–30m to invert the uniaxial compressive strength (UCS) of the rock strata, which is used to determine the "geologically adaptable" cutter replacement requirements.

[0227] 15) Strain-type load cycle counter: pasted on the surface of the tool shaft, it records the torque-thrust load spectrum in real time and calculates the equivalent load cycle number Neq, which is used to determine the tool replacement requirement of "life end type".

[0228] 16) Sensor data fusion module: Located in the explosion-proof electrical control box at the center of the cutter head, it synchronously samples (≥1kHz) the data from the above five types of sensors and timestamps them, and outputs an N×5×T spatiotemporal tensor for the decision-making layer to make decisions;

[0229] 2) Decision layer, used to intelligently judge the tool status of the three-dimensional comprehensive data obtained by the perception layer, and generate instructions through multimodal spatiotemporal graph neural network MST-GNN, including the shield machine local edge computing unit: running multimodal spatiotemporal graph neural network MST-GNN, completing a forward inference within 50ms, outputting tool change priority, tool change type label, tool change quantity and tool change number, and sending it to the control layer at a frequency of 1kHz through EtherCAT bus;

[0230] 3) Control layer: This layer performs instantaneous positioning of the tool to be replaced and establishes a closed-loop trajectory for the tool robot based on its position to complete the tool change action. This includes:

[0231] 31) Tool robot: Under the micro-stop condition of low-speed rotation of the tool head, it automatically completes the whole set of actions of "positioning - trajectory planning - angle synchronization - tool removal - tool insertion - locking - identity confirmation" according to the "tool change number + target angle coordinates" command issued by the decision-making level;

[0232] 32) 25-bit absolute encoder: mounted on the end face of the cutter head spindle, with a resolution of 0.0055° / tick, measures the cutter head rotation angle φ(t) in real time, providing a reference for "millisecond-level instantaneous positioning";

[0233] 33) Passive UHF-RFID tag: Embedded in the base of each tool, storing a 64-bit UID and a fixed angular offset Δθ from the encoder zero position. i Used to calculate the instantaneous target angular coordinate θ of the tool. target (t);

[0234] 34) Tool robot controller: Receives encoder and RFID data, generates five-dimensional spline trajectory in real time, realizes robot joint space trajectory planning and angular synchronization closed loop, with a control cycle of 250μs;

[0235] 35) Six-dimensional force sensor: Installed at the end of the robot, it detects the normal force and torque during the tool removal / insertion process in real time, and is used for locking / unlocking status judgment and safety protection.

[0236] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0237] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A multi-dimensional sensing and adaptive cutterhead changing method for tunnel boring machines under low-speed pressure maintenance, characterized in that, The tool changing method includes the following steps: S1. In the perception layer, three-dimensional comprehensive data of the tool is obtained in real time through sensors. S2. In the decision layer, the three-dimensional comprehensive data is fused through the sensor data fusion module in the perception layer and sent into the multi-modal spatio-temporal graph neural network MST-GNN, and the tool changing priority, tool changing type label, tool changing quantity and tool changing number are output. S3. In the control layer, instantaneous positioning of the target tool is carried out through the tool changing type label and in cooperation with the low-speed rotation state of the cutter head. S4. In the control layer, after the angular velocity of the tool robot is synchronized with that of the target tool, angular error calculation is carried out to confirm the zero relative window, so that the tool robot completes tool extraction, tool insertion and locking, and tool changing is realized.

2. The tool changing method according to claim 1, characterized in that: In the perception layer, the three-dimensional comprehensive data includes wear defect data A1, corrosion defect data A2 and fracture defect data A3 in the defect dimension A, rock layer hardness data B1 in the geological hardness dimension B, and life data C1 in the life dimension C. Among them, the wear defect data A1 obtains the radial wear amount ΔW of the cutter ring through an electromagnetic mutual inductance type wear sensor. The corrosion defect data A2 obtains the corrosion area percentage ΔC of the bearing seal area through an electrochemical impedance spectroscopy corrosion sensor. The fracture defect data A3 obtains the crack length ΔL of the tool shaft or the accumulated acoustic emission energy AE through an acoustic emission crack sensor. The rock layer hardness data B1 obtains the uniaxial compressive strength UCS of the rock layer through an advanced geological radar antenna. The life data C1 obtains the equivalent load cycle number Neq through a strain type load cycle counter. The three-dimensional comprehensive data is processed by the sensor data fusion module and outputs an N×5×T spatio-temporal tensor for decision-making in the decision layer.

3. The tool changing method according to claim 1, characterized in that: In the perception layer, in the micro-stop working condition where the cutter head rotates at a low speed of ≤2 rpm and the soil bin is kept pressurized, the necessity of tool changing is judged based on the defect dimension A, the geological hardness dimension B and the life dimension C, including: In the defect dimension A, the tool defects are classified as: For the wear defect data A1, with the radial wear amount ΔW of the cutter ring as the index, when ΔW≥25 mm, it is determined that A1 reaches the tool changing threshold. For the corrosion defect data A2, with the corrosion area percentage ΔC of the bearing seal area as the index, when ΔC≥30%, it is determined that A2 reaches the tool changing threshold. For the fracture defect data A3, with the crack length ΔL of the tool shaft or the accumulated acoustic emission energy AE as the index, when ΔL≥15 mm or AE≥100 mV·ms, it is determined that the A3 defect reaches the tool changing threshold. Among them, if any tool defect reaches its threshold, the tool changing instruction for the defect dimension A is triggered. In the geological hardness dimension B, the geological hardness grade is defined as: Soft rock: UCS≤30 Mpa, replace with a scraper type tool. Medium hard rock: 30 MPa < UCS≤90 MPa, replace with a hob type tool. Extremely hard rock: UCS>90 MPa, heavy hob type tool. When the real-time UCS value obtained by the advanced geological radar antenna or the tool disc vibration inversion jumps from a low grade to a high grade, or when the continuous tunneling length in the extremely hard rock grade is ≥10 m, the tool changing instruction for the geological hardness dimension B is triggered, and the tool type is replaced according to the geological hardness grade. In the life dimension C, the tool's service life is defined as the equivalent load cycle number Neq. When Neq ≥ the design rated life Nr, the tool change command in the life dimension C is triggered.

4. The tool changing method according to claim 2, characterized in that: In the perception layer, to acquire real-time comprehensive three-dimensional data of defect dimension A, geological hardness dimension B, and lifespan dimension C, a sensor system is deployed on the tunnel boring machine cutterhead and cutters, and arranged and connected as follows: In defect dimension A: The wear monitoring module has two sets of electromagnetic mutual inductance wear sensors embedded in the cutter ring base of each replaceable hob on the front of the cutter head, with the same radius and equal angle intervals along the circumference, to output the radial wear amount ΔW of the cutter ring in real time. Corrosion monitoring module: A set of electrochemical impedance spectroscopy corrosion sensors are attached to the inner surface of the bearing sealing end cap of each hob to output the percentage of corrosion area ΔC in the bearing sealing area in real time; Fracture monitoring module: A set of acoustic emission crack sensors are attached to the transition fillet area of ​​the cutter shaft of each hob to capture the cumulative acoustic emission energy AE of crack initiation and propagation in real time; In geological hardness dimension B: At least three sets of ground-penetrating radar antennas are equally spaced along the circumferential direction on the outer periphery of the cutterhead. These antennas are used to transmit electromagnetic waves in front of the tunnel face before the tunnel boring machine advances and to receive reflected signals in order to obtain the uniaxial compressive strength (UCS) of the rock strata. In the life dimension C: On the cutter shaft surface of each cutter, at least one pair of strain-type load cycle counters are symmetrically attached along the axial centerline to collect the torque and thrust load spectrum of the cutter in real time during the tunneling process and calculate the equivalent load cycle number Neq. Sensor data fusion module: Located in the explosion-proof electrical control box at the center rotary joint of the cutter head, it includes a multi-channel synchronous sampling unit, a time synchronization unit, and an edge computing unit; The multi-channel synchronous sampling unit is connected to A1, A2, A3, B1 and C1 respectively via shielded cables, with a sampling frequency ≥1kHz; The time synchronization unit receives a 1PPS pulse signal from the main control PLC of the tunnel boring machine, which is used to timestamp the three-dimensional integrated data, with a time synchronization error of ≤1ms.

5. The tool changing method according to claim 2, characterized in that: In the perception layer, a multi-dimensional state matrix is ​​constructed based on the aforementioned three-dimensional integrated data, and the "isolated state of a single tool" is upgraded to a globally coupled relationship network of "tool head-tool-geology-time" through graph structure modeling. This step specifically includes: Using N replaceable cutting tools on the tool turret as nodes, a five-dimensional state vector is generated in real time for each tool i: Sᵢ t = ; = ； in, , , , and Provided directly by A1, A2, A3, B1 and C1 respectively, where i represents the tool number and t represents a specific moment of monitoring; 2) Stack the five-dimensional state vectors of the N tools in order of tool number to form an N×5-dimensional global state matrix: M= ； = ； 3) Using the hobs as nodes, with node features being the aforementioned five-dimensional state vector, and the mechanical coupling relationship between the hobs and the rotational dynamics of the cutter head as edges, construct an undirected graph G=(V, E), where the edge weight formula in the undirected graph G is as follows: ； The edge weight Qᵢⱼ is determined by the physical distance dᵢⱼ between hobs i and j on the cutter head and the rotational speed of the cutter head. Joint decision, , These are preset hyperparameters; 4) Introduce a time dimension into the node features of the undirected graph G, and transform the state matrix M over T consecutive sampling periods. t Stacked as N×5×T spacetime tensors.

6. The tool changing method according to claim 5, characterized in that: In the decision-making layer, the N×5×T spatiotemporal tensor is processed by a multimodal spatiotemporal graph neural network (MST-GNN) to obtain the tool-changing decision result, specifically including: The N×5×T spatiotemporal tensor is modally decomposed into five-channel tensors: wear channel, corrosion channel, fracture channel, geological hardness channel, and lifetime channel. LayerNorm is applied independently to each channel to obtain the normalized five-channel feature tensor. ； Spatial Graph Convolutional Network (S-TCN) uses the adjacency matrix of an undirected graph G as a mask to perform graph convolution at each time point τ∈{1,…,T}: ； ； Where l represents the layer index and D is the degree matrix. These are trainable spatial weights; after passing through the Ls layer, the spatial feature tensor is obtained. ds represents the spatial embedding dimension; The temporal causal convolutional layer T-TCN performs dilated causal convolution along the temporal dimension for each node i: ； in, The kernel size is [size]. As the expansion factor, after The time feature tensor is obtained after the layer. T′ is the time step after downsampling; Multimodal attention fusion: After linearly mapping the five-channel feature tensor, calculate the cross-modal attention weights. ； ； in, , , They are query, key-value mapping, and output fused features respectively. ; The results are output through the task output header. The output results include: Tool change priority ; Tool Change Type Tag Number of tool changes The threshold θ is adjusted in real time by engineering constraints; Tool change number set .

7. The tool changing method according to claim 6, characterized in that: In the control layer, instantaneous positioning of the target tool includes the following steps: An absolute encoder is coaxially mounted on the end face of the tool turret spindle, and the output signal is transmitted to the tool robot controller in real time via an EtherCAT bus. Each replaceable hob has a passive UHF-RFID tag embedded inside its base. The passive UHF-RFID tag contains a fixed UID and a fixed offset. ,in, This is the mechanical angle difference between the tool centerline and the encoder zero position; After the tool robot controller is powered on, it performs a one-time scan: the UID of each tool is read sequentially by the RFID read / write antenna array through the slow rotation of the tool head. Establish a non-power-loss index array in the memory of the tool robot controller. ; In any At any given moment, the real-time angle of the tool turret is read from the absolute encoder. Then the instantaneous target angular coordinates of the Nth tool are: ； Dynamic compensation and closed-loop control specifically include: the tool robot controller will... Input the data into the quintic spline trajectory planner to generate the angular displacement-time curve of the end effector.

8. The tool changing method according to claim 7, characterized in that: In the control layer, the instantaneous target angular coordinates are obtained by instantaneously positioning the target tool. In the case of micro-stop tool changing, the robot performs trajectory pre-planning, speed chasing, and error closed-loop control. This step includes: Within the joint space of the cutting tool robot, using the current joint angle vector The starting and target joint angle vectors As the endpoint, generate a fifth-order polynomial trajectory. Five-order spline trajectory pre-planning is performed, and its expression is: ； Fifth-order polynomial locus The boundary conditions that must be satisfied at the start and end points include: Starting from the robot's "current actual joint angles," avoid position jumps: ； Within the specified time The inner diameter accurately reaches the target tool angle, achieving "zero relative" alignment: ； Starting speed and acceleration are zero to ensure a smooth start. ； The end-effector velocity must be matched with the angular velocity ω of the cutter head to keep the robot end-effector relatively stationary with respect to the cutter. ； The final acceleration should be zero to avoid mechanical impact or residual vibration. ； Where ω is the instantaneous angular velocity of the cutter head, which is calculated in real time by the differential of the encoder: ； The spline solver is based on the maximum joint velocity. Maximum acceleration Automatically provides theoretical arrival time ; To perform velocity feedforward and angular synchronization, the controller uses −ω as the end-point angular velocity feedforward command to ensure that... When the relative angular velocity between the robot end effector and the target tool is Δω≈0, the system switches to angular synchronization mode after reaching the target angle. 4) Complete velocity feedforward and angle synchronization, implement error closure and reentry mechanisms, and the real-time angle error is: ； When |Δθ(t)|≤0.5° and the holding time>50ms, the zero relative window is determined to be established, and the tool changing mechanism is allowed to operate; If |Δθ(t)|>0.5° or the joint torque mutation>15N·m, immediately trigger reentry: the tool robot decelerates to zero and records. Recalculate .

9. The tool changing method according to claim 8, characterized in that, In the control layer, once the zero relative window is established and confirmed, the tool changer is allowed to move. The tool robot controller immediately initiates the following tool change sequence, the process of which includes: After the zero relative window is confirmed, the locking state is detected. The normal force Fn of the tool locking surface is monitored in real time by a six-dimensional force sensor. If Fn≤2N and lasts for more than 50ms, the locking force is determined to be released and the tool can be pulled out. After the electromagnetic / pneumatic lock is released, the Z-axis electric slide table pushes outward by 50mm at a speed of 100mm / s to complete the tool removal. The photoelectric switch detects the position at the end of the travel; if the position is not reached within 3 seconds, the system will interrupt and trigger an alarm. The tool robot picks up a new tool, and the Z-axis slide inserts it into the tool holder in the opposite direction at 80mm / s. After insertion, apply a locking torque of 180 N·m, with a torque closed-loop tolerance of ±5 N·m; After locking is completed, the six-dimensional force sensor confirms that the normal force Fn' ≥ 800N, ensuring locking; The end-effector UHF-RFID reader reads the new tool's UID and compares it with the tool robot controller's array. If the verification is successful, the lifespan counter is reset to zero. If the UID verification fails, and two rereads still fail, an anomaly is marked.

10. A cutter changing system for implementing the multi-dimensional sensing and adaptive cutter changing method for low-speed pressure maintenance of a tunnel boring machine as described in any one of claims 1 to 9, characterized in that, The tool changing system includes: The sensing layer senses and collects data on the tool's condition, including electromagnetic inductance wear sensors, electrochemical impedance spectroscopy corrosion sensors, acoustic emission crack sensors, advanced ground-penetrating radar antennas, strain gauge load cycle counters, and sensor data fusion modules. The decision layer is used to intelligently judge the tool status based on the three-dimensional comprehensive data obtained from the perception layer, and generates instructions through the multimodal spatiotemporal graph neural network MST-GNN, outputting tool change priority, tool change type label, tool change quantity and tool change number, and sending them to the control layer through the EtherCAT bus. The control layer performs instantaneous positioning of the tool to be replaced and performs trajectory closed-loop control of the tool robot based on its position to complete the tool changing action. It includes the tool robot, absolute encoder, passive UHF-RFID tag, tool robot controller and six-dimensional force sensor.

Images