Artificial Intelligence-Based Rapid Construction Method for Collaborative Initial Support of Tunnel Equipment

Abstract

This invention relates to the field of intelligent construction control technology, specifically to a rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence. The method includes the following steps: acquiring parameter ratios to construct a coupled stiffness gradient; constructing a basic reward matrix based on a Markov decision process and calculating a negative penalty scalar; quantifying collision avoidance constraint factors using Euclidean distance; generating a collaboratively updated reward matrix and filtering and outputting target travel increment commands; and executing underlying bus conflict detection to establish a collaborative operation sequence. In this invention, a gradient system constructed by calculating multi-source physical parameters overcomes the lag in experience-dependent perception, deeply maps and predicts the deformation trend of surrounding rock and transient pressure changes, quantifies interference risks through spatial coordinate calculations, deeply integrates collision avoidance factors with the control sequence, and executes underlying time-domain translation scheduling to achieve interlocking of actions and obstacle avoidance coordination among multiple devices in confined working environments, completely eliminating the risk of trajectory collisions in parallel operations.

Description

Technical Field

[0001] This invention relates to the field of intelligent construction control technology, and in particular to a rapid construction method for initial support of tunnel equipment based on artificial intelligence. Background Technology

[0002] The field of intelligent construction control technology involves core aspects of building and civil engineering construction, such as automated equipment scheduling, on-site work environment status monitoring, joint operation of multiple construction machines, and issuance of construction procedure execution instructions. It mainly collects physical parameters of equipment operation by arranging displacement and tension / compression sensors on the robotic arms and chassis of various types of construction machinery. The geographical coordinates and travel parameters of the machinery are aggregated to the monitoring station using communication cables laid on site. The operator at the control station issues mechanical action instructions such as forward, backward, or rotation to multiple machines of different types on site based on paper construction drawings and predetermined schedule plans, thereby completing the operation arrangement of multiple machines in the entire construction work area.

[0003] The traditional tunnel equipment collaborative initial support rapid construction method refers to the temporary load-bearing structure erection operation carried out after the underground excavation section of the tunnel exposes the rock mass. The construction workers guide the multi-arm drilling rig to the working face through hand flags or voice intercom devices and manually fine-tune the inclination angle of the drill arm to carry out drilling operations. Then, the workers manually operate the hydraulic crank of the wet spraying rig to spray cement slurry layer by layer to cover the excavation face according to the spraying layer thickness requirements marked on the drawings. During the spraying intervals, the steel arch frame is manually lifted and bolted together for installation. The various construction tools are operated alternately in strict accordance with the fixed assembly line action steps of fixed-point drilling, anchor bolt insertion, pressure grouting, steel frame erection, steel mesh suspension, and shotcrete.

[0004] Traditional construction methods rely excessively on personnel to guide machinery and make fine-tuning adjustments using hand flags and walkie-talkies. Under this heavily manual intervention mechanism, there is a significant time and space lag in information transmission and coordination between independent processes. Faced with the stress environment of underground rock strata, operators cannot cope with the dynamic changes in surrounding rock pressure based solely on sensory experience. When multiple heavy machines are working alternately on a narrow working face, the lack of spatial constraints and motion-level overall coordination and scheduling can easily lead to execution conflicts and trajectory overlap interference between equipment, resulting in project delays and serious risks of collapse and equipment damage. Summary of the Invention

[0005] To address the problem that traditional construction methods rely excessively on personnel using hand flags and walkie-talkies for mechanical guidance and posture fine-tuning, resulting in significant time and space delays in information transmission and coordination between independent processes, operators struggle to cope with dynamic changes in surrounding rock pressure based solely on sensory experience when facing underground rock stress environments. Furthermore, the lack of spatial constraints and coordinated scheduling when multiple heavy machines operate alternately at narrow tunnel faces easily leads to execution conflicts and trajectory overlaps, causing project delays and serious risks of collapse and equipment damage. This invention provides a rapid construction method for initial support of tunnel equipment based on artificial intelligence.

[0006] To achieve the above objectives, this invention employs an artificial intelligence-based method for rapid initial support construction using collaborative tunnel equipment, comprising the following steps:

[0007] S1: Obtain the scalar of tension force of the construction equipment robotic arm and the scalar of displacement of the tunnel opening, perform first-order difference operation to obtain the tension force increment and displacement increment and calculate the proportional value, and construct the real-time coupled stiffness gradient.

[0008] S2: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm, input them into the Markov decision process execution state mapping, construct the basic action reward matrix, and calculate the negative penalty scalar when the real-time coupled stiffness gradient is less than the soft rock yield gradient.

[0009] S3: Call the negative penalty scalar, calculate the Euclidean distance between the robotic arms through spatial coordinates, calculate the collision avoidance limit factor when it is less than the safety threshold, and calculate the negative penalty scalar and the collision avoidance limit factor based on preset weights to construct a collaboratively updated reward matrix;

[0010] S4: Analyze the collaborative update reward matrix, use a greedy algorithm to filter directional control commands and candidate hydraulic valve opening duration parameters, calculate the product of candidate hydraulic valve opening duration parameters and flow constant, and generate target stroke increment commands;

[0011] S5: Obtain the underlying control bus address of the construction equipment, associate the target stroke increment command and direction control command to construct a feature mapping, perform conflict detection, and construct a collaborative operation sequence.

[0012] As a further aspect of the present invention, the real-time coupled stiffness gradient includes the tangent stiffness coefficient, the slope of the surrounding rock characteristic curve, and the elastoplastic hardening modulus; the basic action reward matrix includes the state transition probability, the action value function, and the policy gradient increment; the collaborative update reward matrix includes the Euclidean distance penalty term, the obstacle avoidance safety margin, and the multi-agent consistency weight; the target stroke increment instruction includes the hydraulic cylinder piston displacement, the flow control rated value, and the actuator end displacement vector; and the collaborative operation sequence includes the bus object dictionary, the instruction priority weight, and the logical interlock criterion.

[0013] As a further aspect of the present invention, the specific steps of S1 are as follows:

[0014] S101: The monitoring pressure detection branch acquires the scalar of the mechanical arm tension force of the construction equipment, monitors the feedback frequency to collect the scalar of the tunnel opening displacement, performs signal filtering and noise reduction and zero-point offset calibration, extracts the feature values ​​of the original sensing voltage waveform, maps the sampling channel and the physical parameter response weights, and establishes a set of working condition monitoring scalars.

[0015] S102: Call the set of working condition monitoring scalars, perform a first-order difference operation on the scalar of the robot arm tension force and the scalar of the tunnel opening displacement, calculate the difference in tension load variation and the difference in tunnel surrounding rock displacement, map the sensor dynamic resolution and signal acquisition frequency matching attributes, and generate tension force increment and displacement increment.

[0016] S103: Based on the tension force increment and displacement increment, calculate the ratio of tension force increment to displacement increment, map the face pressure distribution vector, aggregate the stress and displacement response variation amplitude, associate the elastic recovery capability of the shield support system, and generate a real-time coupled stiffness gradient.

[0017] As a further aspect of the present invention, the specific steps of S2 are as follows:

[0018] S201: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm and the joint angle sequence of the end effector, perform joint feature extraction on the real-time coupled stiffness gradient and pose data, map the pose data and stress response distribution, aggregate displacement deviation and load change to calculate the transfer probability distribution, and establish an environmental state mapping vector.

[0019] S202: Based on the environmental state mapping vector, evaluate the effectiveness of the real-time coupled stiffness gradient execution strategy, calculate the total expected amount of node step reward, combine motion smoothness weight and energy loss allocation coefficient to aggregate the weight of multi-objective reward function, and establish a basic action reward matrix.

[0020] S203: Call the basic action reward matrix. If the real-time coupling stiffness gradient is less than the soft rock yield gradient, collect the instantaneous fluctuation frequency and rock settlement rate. Based on the deviation coefficient between the deformation trend and the instability threshold, combine the defense compensation coefficient and the safety penalty factor to perform normalization and generate a negative penalty scalar.

[0021] As a further aspect of the present invention, the instability threshold is determined by extracting the spatial stress extremum through the analytical environmental state mapping vector, calculating the absolute difference between the soft rock yield gradient and the real-time coupled stiffness gradient as the basic deformation tolerance, and then weighting and combining the basic deformation tolerance, the spatial stress extremum, and the energy loss allocation coefficient based on the parameter influence weight.

[0022] As a further aspect of the present invention, the specific steps of S3 are as follows:

[0023] S301: Call the negative penalty scalar, collect the spatial coordinates of the robotic arm, perform Euclidean operation, calculate the relative displacement span between the centroids of the end effectors and joint axes of adjacent robotic arms, perform normalization mapping for the dynamic working radius and motion envelope, quantify the risk overlap, and obtain the spatial distance value between robotic arms.

[0024] S302: For the spatial distance value between the robotic arms, when it is less than the safety threshold, retrieve the preset collision avoidance gain weight factor, perform exponential decay mapping, quantify the urgency of spatial conflict and the negative correlation strength of trajectory overlap constraint, associate the dynamic safety margin correction coefficient, and obtain the collision avoidance limit factor.

[0025] S303: Based on the collision avoidance constraint factor and the negative penalty scalar, perform scalar linear weighted product operation, map the operation result to the state-action pair score node of the Markov decision process, associate the multi-machine cooperative joint action probability distribution, perform reward value discretization reconstruction and dimension tensor merging, and establish a collaboratively updated reward matrix.

[0026] As a further aspect of the present invention, the specific steps of S4 are as follows:

[0027] S401: For the collaborative update reward matrix, extract node values ​​and perform comparison, filter the state parameter vector corresponding to the node with the largest score, obtain spatial displacement features, retrieve the hydraulic valve control rule mapping table, associate the robot arm joint posture parameters to perform vector dot product, quantify the motion trend, and generate direction control commands.

[0028] S402: Based on the direction control command, call the collaborative update reward matrix, extract the time axis component, read the operation response delay parameter and perform summation, associate time boundary threshold to remove outlier data, match the hydraulic mechanism response time constant and perform proportional scaling to obtain the candidate hydraulic valve opening duration parameter;

[0029] S403: Based on the opening duration parameter of the candidate hydraulic valve, obtain the valve port flow constant, multiply the opening duration parameter of the candidate hydraulic valve by the flow constant, calculate the cylinder volume parameter, obtain the piston cross-sectional area and perform division operation, and use the result as the modulus and the direction control command vector for product operation to generate the target stroke increment command.

[0030] As a further aspect of the present invention, the time boundary threshold is determined by extracting the response time distribution characteristics of historical operating cycles within the collaborative update reward matrix.

[0031] As a further aspect of the present invention, the specific steps of S5 are as follows:

[0032] S501: Call the target travel increment instruction and direction control instruction, obtain the underlying control bus address, parse the data frame to extract the register index, establish the instruction port mapping relationship, perform drive voltage amplitude quantization processing through the mapping relationship and the hardware response delay, and establish the underlying drive control mapping table.

[0033] S502: Based on the underlying drive control mapping table, monitor the overlapping period of duty cycle of multiple commands, extract the load component and peak inrush current coefficient, perform vector summation on the load component in the overlapping period, compare with the bus power supply reference value, identify the concurrent execution logic mutual exclusion state, and establish a command conflict detection matrix.

[0034] S503: Based on the instruction conflict detection matrix, call the underlying driver control mapping table to retrieve the safety instruction packet corresponding to the zero element matrix position, perform time-domain phase shift scheduling for non-zero element conflict items, extract the actuator cooperative response priority, arrange instructions according to logical order and pose constraints, and establish a cooperative operation sequence.

[0035] As a further aspect of the present invention, the step of performing time-domain phase shift scheduling for non-zero element conflict items refers to obtaining the overlapping time span contained in the non-zero element conflict items, calculating the absolute difference between the preset safety response interval and the overlapping time span to generate a target phase offset, and superimposing the target phase offset onto the initial execution time node associated with the non-zero element conflict item to complete the shift.

[0036] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0037] In this invention, geological parameters and mechanical status are sensed in real time, and the proportional relationship between tension force and displacement increment is calculated to construct a coupled stiffness gradient. This effectively overcomes the lag of purely empirical judgment, accurately maps the deformation trend and pressure change characteristics of the surrounding rock, and performs Markov decision-making and environmental assessment based on this. By combining spatial coordinates and Euclidean calculations, the safe overlap range of mechanical operation is quantified, collision avoidance limit factors and negative penalty scalars are calculated and integrated into the multi-machine cooperative control matrix, and the temporal phase translation scheduling of the underlying commands is executed synchronously. This enables the interlocking of actions and dynamic obstacle avoidance coordination of multiple devices in a confined space, completely eliminating the risk of trajectory interference in alternating operations. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the accompanying drawings without creative effort.

[0039] Figure 1 This is a schematic diagram of the steps of the present invention;

[0040] Figure 2 This is a detailed schematic diagram of S1 of the present invention;

[0041] Figure 3 This is a detailed schematic diagram of S2 of the present invention;

[0042] Figure 4 This is a detailed schematic diagram of S3 of the present invention;

[0043] Figure 5 This is a detailed schematic diagram of S4 of the present invention;

[0044] Figure 6 This is a detailed schematic diagram of S5 of the present invention. Detailed Implementation

[0045] The technical solution of the present invention will now be described with reference to the accompanying drawings.

[0046] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0047] Please see Figure 1 This invention provides a rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence, comprising the following steps:

[0048] S1: Obtain the scalar of tension force of the construction equipment robotic arm and the scalar of displacement of the tunnel opening, perform first-order difference operation to obtain the tension force increment and displacement increment and calculate the proportional value, and construct the real-time coupled stiffness gradient.

[0049] S2: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm, input them into the Markov decision process execution state mapping, construct the basic action reward matrix, and calculate the negative penalty scalar when the real-time coupled stiffness gradient is less than the soft rock yield gradient.

[0050] S3: Call the negative penalty scalar, calculate the Euclidean distance between the robotic arms through spatial coordinates, calculate the collision avoidance limit factor when it is less than the safety threshold, and calculate the negative penalty scalar and the collision avoidance limit factor based on preset weights to construct a collaboratively updated reward matrix;

[0051] S4: Analyze the collaborative update reward matrix, use a greedy algorithm to filter directional control commands and candidate hydraulic valve opening duration parameters, calculate the product of candidate hydraulic valve opening duration parameters and flow constant, and generate target stroke increment commands;

[0052] S5: Obtain the underlying control bus address of the construction equipment, associate the target travel increment command and direction control command to construct a feature map, perform conflict detection, and construct a collaborative operation sequence.

[0053] The real-time coupled stiffness gradient includes the tangent stiffness coefficient, the slope of the surrounding rock characteristic curve, and the elastoplastic hardening modulus. The basic action reward matrix includes the state transition probability, the action value function, and the policy gradient increment. The collaborative update reward matrix includes the Euclidean distance penalty term, the obstacle avoidance safety margin, and the multi-agent consistency weight. The target stroke increment command includes the hydraulic cylinder piston displacement, the flow control rating, and the actuator end displacement vector. The collaborative operation sequence includes the bus object dictionary, the command priority weight, and the logical interlock criterion.

[0054] Please see Figure 2 The specific steps of S1 are as follows:

[0055] S101: The monitoring pressure detection branch acquires the scalar of the mechanical arm tension force of the construction equipment, monitors the feedback frequency to collect the scalar of the tunnel opening displacement, performs signal filtering and noise reduction and zero-point offset calibration, extracts the feature values ​​of the original sensing voltage waveform, maps the sampling channel and the physical parameter response weights, and establishes a set of working condition monitoring scalars.

[0056] A thin-film pressure sensor array positioned at the bottom of the hydraulic cylinder of the construction equipment is used to collect analog voltage signals in real time during the process of the robotic arm supporting the rock surface at a sampling frequency of 500 Hz. Analog-to-digital conversion is performed on the analog voltage signals to obtain the scalar value of the robotic arm tension force. Simultaneously, a fiber optic displacement gauge installed 10 meters behind the tunnel face acquires the scalar value of the tunnel opening displacement at the same sampling frequency. To address high-frequency noise interference in the collected robotic arm tension force and tunnel opening displacement values, a mean filtering mechanism is invoked. A sliding window size of 15 sampling points is set, and the arithmetic mean of the data within the window is calculated point by point to replace the center point value, thus completing the signal filtering and noise reduction process. Next, the baseline values ​​of tension force and displacement are extracted for 3 seconds while the robotic arm is in an unloaded hovering state. The corresponding baseline values ​​are subtracted from the current filtered robotic arm tension force and tunnel opening displacement values, respectively, to perform zero-point offset calibration and eliminate static errors caused by sensor temperature drift. Subsequently, the root mean square (RMS) and peak-to-peak values ​​of the time-series waveform composed of 50 consecutive calibrated sampling points were extracted as the original sensing voltage waveform feature values. Channel mapping configuration rules were then obtained. The RMS value of tension force extracted from the first sampling channel was multiplied by the physical response weight of tension force, and the RMS value of displacement extracted from the second sampling channel was multiplied by the physical response weight of displacement. The feature values ​​after weight adjustment were then merged. The advantage of this operational logic is that by performing noise reduction and benchmark calibration on multi-source sensor data within a time window, combined with weighted mapping of physical features, electromagnetic interference from the construction environment and static load deviations caused by equipment weight are effectively eliminated. The settings for the physical response weights of tension force and displacement were determined through statistical analysis of historical construction conditions, specifically setting the physical response weight of tension force to 0.6 and the physical response weight of displacement to 0.4. For example, when the root mean square value of the tension force extracted by the first sampling channel is 120 millivolts, multiplying it by a weight of 0.6 yields an adjusted tension force characteristic value of 72, and the root mean square value of the displacement extracted by the second sampling channel is 45 millivolts, multiplying it by a weight of 0.4 yields an adjusted displacement characteristic value of 18, thus establishing a set of working condition monitoring scalars containing elements 72 and 18.

[0057] Table 1: Initial Acquisition Parameters of Sensor

[0058]

[0059] Table 1 shows the test data for the extraction process of the scalar set of operating condition monitoring, and displays the acquisition parameters of different channels.

[0060] S102: Call the set of scalars for working condition monitoring, perform a first-order difference operation on the scalar of the robot arm tension force and the scalar of the tunnel opening displacement, calculate the difference in tension load variation and the difference in tunnel surrounding rock displacement, map the sensor dynamic resolution and signal acquisition frequency matching attributes, and generate tension force increment and displacement increment.

[0061] The system retrieves the time series of tension force and displacement characteristic values ​​from the scalar set of working condition monitoring, and performs a first-order difference operation on data nodes of two adjacent sampling periods. Specifically, it subtracts the previous moment's robotic arm tension force scalar from the current moment's scalar to calculate the tension load variation difference. Simultaneously, it subtracts the previous moment's tunnel opening displacement scalar from the current moment's scalar to calculate the tunnel surrounding rock displacement variation difference. Then, it obtains the analog-to-digital converter dynamic resolution parameters and the aforementioned 500 Hz signal acquisition frequency from the sensor datasheet, divides the tension load variation difference by the dynamic resolution parameters of the tension force sensor, and multiplies it by the conversion factor between the sampling frequency and time unit to generate a standardized tension force increment. Similarly, it divides the tunnel surrounding rock displacement variation difference by the dynamic resolution parameters of the displacement sensor and multiplies it by the corresponding frequency matching attribute to generate a standardized displacement increment. The advantage of this operational logic is that by performing a first-order difference distribution resolution standardization operation on adjacent sampling points, absolute physical quantities are transformed into relative change rate indicators with uniform dimensions. The dynamic resolution parameter of the aforementioned tension sensor is set to 0.05 MPa per digit, and the dynamic resolution parameter of the displacement sensor is set to 0.02 mm per digit. The frequency matching attribute is fixed at 1. In the actual calculation, assuming the current tension scalar of the robotic arm is 15.2 MPa and the previous one was 14.8 MPa, the difference in tension load variation is 0.4 MPa. Dividing 0.4 by 0.05 and multiplying by 1, the calculated tension increment is 8. The current displacement scalar is 6.5 mm and the previous one was 6.2 mm. Dividing the current displacement scalar by 0.3 mm, the calculated displacement variation is 0.3 mm. Dividing 0.3 by 0.02 and multiplying by 1, the calculated displacement increment is 15.

[0062] S103: Based on the increment of tension force and the increment of displacement, calculate the ratio of the increment of tension force to the increment of displacement, map the pressure distribution vector at the tunnel face, aggregate the magnitude of stress and displacement response changes, associate the elastic recovery capability of the shield support system, and generate a real-time coupled stiffness gradient.

[0063] The generated tension and displacement increments are obtained. Using the tension increment as the numerator and the displacement increment as the denominator, a division operation is performed to calculate the ratio of tension to displacement increments. Subsequently, the initial prestress values ​​of five different monitoring zones on the tunnel face cross-section are extracted to construct a face pressure distribution vector. The calculated ratio of tension to displacement increments is then multiplied by each element in the face pressure distribution vector to aggregate the displacement response variation and stress state, generating dynamic compressive strength characteristic values ​​for multiple zones. This further yields the elastic recovery coefficient of the shield support system. This elastic recovery coefficient is calibrated through compression and rebound tests of the shield machine's hydraulic cylinders. Specifically, the dynamic compressive strength characteristic values ​​of multiple zones are multiplied by the elastic recovery coefficient to generate a real-time coupled stiffness gradient representing the local rock mass stiffness characteristics. The advantage of this calculation logic is that the ratio of tension to displacement increments directly reflects the dynamic constitutive relationship of the surrounding rock. The elastic recovery coefficient mentioned above is set based on the Young's modulus and thickness parameters of the support material, with a preset value of 0.85. For example, substituting the calculated tension increment of 8 and displacement increment of 15 into a division operation yields a tension increment to displacement increment ratio of 0.533. If the initial prestress value of a certain zone in the face pressure distribution vector is 20 MPa, multiplication yields a dynamic compressive strength characteristic value of 10.66. This value is then multiplied by the elastic recovery coefficient of 0.85, ultimately generating a real-time coupled stiffness gradient of 9.061 for that zone.

[0064] Please see Figure 3 The specific steps of S2 are as follows:

[0065] S201: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm and the joint angle sequence of the end effector, perform joint feature extraction on the real-time coupled stiffness gradient and pose data, map the pose data and stress response distribution, aggregate displacement deviation and load change to calculate the transfer probability distribution, and establish an environmental state mapping vector.

[0066] The generated real-time coupled stiffness gradient is used as an input for environmental features. Absolute encoders located at multiple joints of the robotic arm collect real-time coordinate data sequences of the robotic arm in 3D space and the joint angle sequences of the three rotational degrees of freedom of the end effector. Feature extraction is performed on the continuous real-time coupled stiffness gradient. Specifically, local peak points where the gradient change rate exceeds a specific fluctuation are selected. A mapping relationship is established between the spatial coordinate data of the robotic arm's end effector and the extracted local peak points at the corresponding time to reflect the stress response distribution law of the surrounding rock under a specific pose. Then, the spatial Euclidean distance between the current pose and the pose of the previous action cycle is calculated as the displacement deviation. The difference in real-time coupled stiffness gradient between the two cycles is obtained as the load change. The displacement deviation and load change are weighted and summed to construct a state transition index. The state transition index is input into the exponential distribution function to perform probability calculation, resulting in the probability distribution of transitions to multiple adjacent work nodes. The spatial coordinates, joint angles, real-time coupled stiffness gradient, and transition probability distribution are concatenated into a one-dimensional array in a fixed order. The advantage of this computational logic lies in achieving a deep quantitative representation of the working environment state by fusing the changes in the robotic arm's pose with the environmental stiffness characteristics. Assuming a displacement deviation weight of 0.4, a load change weight of 0.6, a current cycle displacement deviation of 50 mm, and a load change of 12, the calculated state transition exponent is 27.2. After substituting this into the exponential decay function, a transition probability of 0.76 is generated. By concatenating the coordinates 120 and 85 and 90, the angles 15 and 30 and 45, the real-time stiffness gradient of 9.061, and the transition probability of 0.76, an environmental state mapping vector of length 8 elements can be generated.

[0067] S202: Based on the environmental state mapping vector, evaluate the effectiveness of the real-time coupled stiffness gradient execution strategy, calculate the total expected amount of node step reward, combine motion smoothness weight and energy loss allocation coefficient to aggregate the weight of multi-objective reward function, and establish the basic action reward matrix.

[0068] For the established environment state mapping vector, the real-time coupling stiffness gradient component and transition probability distribution are extracted to evaluate the effectiveness of the execution strategy. Specifically, this includes a pre-defined evaluation network model containing an input layer, two fully connected hidden layers, and an output layer. The number of neurons in the input layer is aligned with the 8-dimensional dimension of the environment state mapping vector. The first hidden layer contains 64 neurons and uses a linear rectified activation function to filter negative features. The second hidden layer contains 32 neurons and also uses a linear rectified activation function. Full connections are used between layers. The output layer is a single neuron with no activation function, directly outputting a real value. The environmental state mapping vector is input into the network, and the expected total step reward for each selectable action node is calculated through forward propagation. Then, the trajectory curvature integral during the robotic arm's action is extracted as a motion smoothness variable, and the hydraulic pump output power integral is extracted as an energy loss variable. The expected total step reward is multiplied by the corresponding base gain parameter, the motion smoothness variable is multiplied by the motion smoothness weight, and the energy loss variable is multiplied by the energy loss allocation coefficient. A linear summation operation is performed on these three products to achieve the aggregate calculation of the multi-objective reward function weights. A basic action reward matrix is ​​established based on the aggregate calculation results of the action nodes. The advantage of this operational logic is that, through a multi-dimensional penalty and reward superposition mechanism, it guides the subsequent decision-making process towards a smooth and low-energy-consumption evolution. The baseline value for the motion smoothness weight is set to -0.3, and the energy loss allocation coefficient is set to -0.5. For example, the expected total step reward of a certain action node output by the network model is 85, the basic gain parameter is 1, the motion smoothness variable of the action process is calculated to be 15, and the energy loss variable is 20. 85×1+(-0.3×15)+(-0.5×20), that is, 85-4.5-10, finally obtains the aggregation result of the node as 70.5, which is then filled into the corresponding position of the basic action reward matrix.

[0069] Table 2: Action Reward Weight Allocation Table

[0070]

[0071] As shown in Table 2, a detailed calculation example of node step reward and multi-class weight aggregation is presented, clearly reflecting the derivation logic of the final reward value.

[0072] S203: Call the basic action reward matrix. If the real-time coupling stiffness gradient is less than the soft rock yield gradient, collect the instantaneous fluctuation frequency and rock settlement rate. Based on the deviation coefficient between the deformation trend and the instability threshold, combine the defense compensation coefficient and the safety penalty factor to perform normalization and generate a negative penalty scalar.

[0073] The system iterates through the established basic action reward matrix, extracts the corresponding real-time coupling stiffness gradient value, and compares it with the pre-set soft rock yield gradient benchmark value. The soft rock yield gradient benchmark value is set to 8.5 based on the uniaxial compressive strength of the rock mass provided in the geological exploration report. If the comparison shows that the real-time coupling stiffness gradient of 9.061 is not less than 8.5, the process continues to the next step. If it is assumed that the real-time coupling stiffness gradient is 8.2, which meets the condition of being less than the soft rock yield gradient of 8.5, then the high-frequency vibration sensor is activated to collect the instantaneous fluctuation frequency of the rock mass at the working face. Simultaneously, the rock strata subsidence rate is collected within a 5-second time window using a laser displacement meter. The ratio of the rock strata subsidence rate to the fluctuation frequency is calculated as the deformation trend. The stress components in the aforementioned environmental state mapping vector are analyzed to obtain the spatial stress extremum. The deviation coefficient is obtained by subtracting the instability threshold from the deformation trend. The yield gradient of soft rock (8.5) is subtracted from the actual real-time coupled stiffness gradient (8.2), and the absolute value of the calculated result (0.3) is taken as the basic deformation tolerance. Then, the basic deformation tolerance (0.3), the spatial stress extremum, and the energy loss distribution coefficient are multiplied together. After obtaining the product, this result is added to the defense compensation coefficient and then divided by the safety penalty factor to achieve normalization, finally generating a negative penalty scalar. The advantage of this operation logic is that it can instantly trigger a highly sensitive penalty mechanism when encountering soft and fractured strata. The defense compensation coefficient is set to 2.0, and the safety penalty factor is set to 10.0. For example, the basic deformation tolerance is 0.3, the spatial stress extreme value is 45, and the aforementioned energy loss distribution coefficient is -0.5. Multiplying these values ​​together, we get 0.3 × 45 × (-0.5) = -6.75. Adding the defense compensation coefficient of 2.0 to -6.75 gives us -4.75. Dividing this by the safety penalty factor of 10.0 gives us a negative penalty scalar of -0.475.

[0074] Please see Figure 4 The specific steps of S3 are as follows:

[0075] S301: Call the negative penalty scalar, collect the spatial coordinates of the robotic arm, perform Euclidean calculation, calculate the relative displacement span between the centroids of the end effectors and joint axes of adjacent robotic arms, perform normalization mapping for the dynamic working radius and motion envelope, quantify the risk overlap, and obtain the spatial distance value between robotic arms.

[0076] The generated negative penalty scalar is used as a pre-parameter for system state control. Spatial coordinate arrays of multiple collaborative robotic arms within the current working area are synchronously acquired via a vision capture system and joint encoder. For any two adjacent robotic arms, the coordinates of their end effector center point and joint axis centroid are extracted. In a three-dimensional Cartesian coordinate system, Euclidean algorithm (sum of squared differences and square roots of corresponding coordinate components) is used to calculate multiple sets of relative displacement spans between the end effectors and joint axis centroids of adjacent robotic arms. The current extension / retraction amount of the hydraulic cylinder is multiplied by the maximum stroke parameter to obtain the dynamic working radius. A motion envelope surface for the robotic arms is constructed in space using the dynamic working radius as the center radius. Multiple sets of relative displacement spans are divided by the interference diameter of the motion envelope surface, and a normalization mapping operation is performed. The mapped ratio is used as a risk overlap index. The minimum value is selected from all risk overlap indices, and the absolute distance associated with this minimum value is defined as the spatial distance between the robotic arms. The advantage of this algorithm is that it transforms abstract spatial coordinates into relative distance parameters that intuitively reflect collision risk. For example, the center-of-mass coordinates of robotic arm 1 are 200, 300, and 150, and the center-of-mass coordinates of robotic arm 2 are 200, 350, and 150. The difference between the corresponding components is 0 and -50, and the sum of the squares is 2500. The square root of the sum of the squares gives a relative displacement span of 50 mm. The dynamic working radius is obtained, and the diameter of the motion envelope is 100 mm. Dividing 50 by 100 gives a normalized risk overlap of 0.5. The spatial distance between the robotic arms after comprehensive calculation is extracted to be 50 mm.

[0077] S302: For the spatial distance between robotic arms, when it is less than the safety threshold, retrieve the preset collision avoidance gain weight factor, perform exponential decay mapping, quantify the urgency of spatial conflict and the negative correlation strength of trajectory overlap constraint, and associate the dynamic safety margin correction coefficient to obtain the collision avoidance limit factor.

[0078] The calculated spatial distance between robotic arms is conditionally assessed and compared with a preset safety threshold, which is set to 80 mm based on the equipment's maximum braking distance and response delay. When the spatial distance of 50 mm is confirmed to be less than the safety threshold of 80 mm, the corresponding preset collision avoidance gain weight factor is retrieved from the system controller's memory database. This preset collision avoidance gain weight factor is set to 1.5, with the natural constant as the base. The difference between the safety threshold and the spatial distance is divided by the safety threshold, and the resulting value is used as the exponent, with a negative sign appended. Exponential decay mapping calculation is then performed to obtain the basic collision avoidance decay value, thereby quantifying the urgency of spatial conflict caused by the shortened distance. Simultaneously, the negative correlation strength of trajectory overlap constraints is extracted. The basic collision avoidance decay value is multiplied by the negative correlation strength of trajectory overlap constraints, and then added to the dynamic safety margin correction coefficient to finally calculate the collision avoidance limiting factor. The advantage of this calculation logic is that the exponential decay model can non-linearly amplify the collision avoidance control force in such close-range crisis scenarios. Here, the negative correlation strength of the trajectory overlap constraint is set to 0.8, and the dynamic safety margin correction coefficient is set to 0.2. In the specific example, the safety threshold of 80 minus the spatial distance value of 50 equals 30. 30 divided by 80 equals 0.375. Calculating the natural constant to the power of -0.375 gives an approximate basic collision avoidance attenuation value of 0.687. Multiplying 0.687 by the preset collision avoidance gain weight factor of 1.5 gives 1.0305. Multiplying this by the negative correlation strength of the trajectory overlap constraint of 0.8 gives 0.8244. Finally, adding the dynamic safety margin correction coefficient of 0.2 gives a collision avoidance limiting factor of 1.0244.

[0079] S303: Based on the collision avoidance constraint factor and negative penalty scalar, perform scalar linear weighted product operation, map the operation result to the state-action pair score node of the Markov decision process, associate the probability distribution of multi-machine collaborative joint action, perform reward value discretization reconstruction and dimension tensor merging, and establish a collaboratively updated reward matrix;

[0080] The calculated collision avoidance constraint factor and the negative penalty scalar containing environmental negative feedback attributes are extracted. The collision avoidance constraint factor is multiplied by a first superposition coefficient, and the negative penalty scalar is multiplied by a second superposition coefficient. A scalar linear weighted product operation is performed on the two products to obtain a comprehensive risk calibration value. This comprehensive risk calibration value is mapped and associated with the state-action pair score nodes in the previously constructed multi-machine cooperative Markov decision process model. The soft maximization probability of the candidate joint actions is calculated for the Markov decision process model, and the probability distribution of the multi-machine cooperative joint actions is output. The original reward value is then adjusted using the comprehensive risk calibration value. Specifically, the original reward value is added to the product of the comprehensive risk calibration value and a preset grading threshold. A reward value discretization reconstruction operation is performed, causing the continuous reward distribution to converge to multiple preset discrete action levels. Finally, the reconstructed 1D reward vectors are merged into dimensional tensors according to the robot arm numbering order. The advantage of this operation logic is that it effectively integrates the collision avoidance constraints of multiple devices and the geological condition penalties of a single machine into the reward framework of reinforcement learning. The first superposition coefficient is set to 0.7, and the second superposition coefficient is set to 0.3. The collision avoidance constraint factor of 1.0244 is multiplied by 0.7 to get 0.717, and the negative penalty scalar of -0.475 is multiplied by 0.3 to get -0.1425. The two are added together to get the comprehensive risk calibration value of 0.5745. If the original reward value of a certain state action pair is 70.5, after adding 0.5745 and multiplying by 10, it is discretized, that is, 70.5 plus 5.745 equals 76.245. Discretize it to the nearest integer level of 76, and then merge it with the reward data of the other robotic arms to generate a collaboratively updated reward matrix.

[0081] Please see Figure 5 The specific steps of S4 are as follows:

[0082] S401: For the collaborative update reward matrix, extract node values ​​and perform comparison, filter the state parameter vector corresponding to the node with the highest score, obtain spatial displacement features, retrieve the hydraulic valve control rule mapping table, associate the robot arm joint posture parameters to perform vector dot product, quantify the motion trend, and generate directional control commands.

[0083] The evaluation value nodes corresponding to the joint actions stored in the aforementioned collaborative update reward matrix are extracted. A pairwise comparison and sorting operation is performed on all node values ​​to select the node with the highest value. The state parameter vector mapped to this node with the highest score is retrieved, and the relative coordinate difference expected to be reached by the robotic arm end effector is parsed from this vector as a spatial displacement feature. This spatial displacement feature is input into a pre-configured hydraulic valve control rule mapping table for retrieval, and the corresponding hydraulic control reference vector is obtained. Subsequently, the actual angle of the current robotic arm joint is read by the encoder and converted into joint posture parameters. The hydraulic control reference vector and the joint posture parameters are multiplied in the corresponding three-dimensional spatial coordinate system to calculate a scalar result to quantify the actual motion trend amplitude on the action axis. Control channels are allocated based on the sign and absolute value of the multiplication result. The advantage of this operation logic is that by using the posture parameters of the node with the highest reward for projection, the high-dimensional motion strategy is quickly reduced to a one-dimensional direction command that can be recognized by the underlying hydraulic hardware. For example, in the collaborative update reward matrix, the node with the highest score, 76, is selected. The spatial displacement feature is obtained as 50 mm laterally. The reference vectors are 1, 0.5, and 0, obtained by searching the hydraulic valve control rule mapping table. The vectors corresponding to the current joint posture parameters are 20, 10, and 0. The vector dot product is executed: 1×20+0.5×10+0=25. The positive value 25 indicates that the motion trend extends in the positive direction. This generates a directional control command that requires the main control valve to be opened in the positive direction.

[0084] S402: Based on the direction control command, call the collaborative update reward matrix, extract the time axis components, read the operation response delay parameter and perform summation, associate the time boundary threshold to remove outlier data, match the hydraulic mechanism response time constant and perform proportional scaling to obtain the candidate hydraulic valve opening duration parameter;

[0085] After generating the aforementioned direction control command, the action time axis component attached to the collaborative update reward matrix is ​​reread. The hardware operation response delay parameter is read from the underlying controller register. The action time axis component and the operation response delay parameter are summed to obtain the basic action cycle prediction value. Subsequently, the response time distribution characteristics of the past 20 historical operating cycles within the collaborative update reward matrix are traversed. The average and standard deviation of the actual response time of these 20 cycles are calculated. The average plus twice the standard deviation is used as the upper time boundary threshold, and the average minus twice the standard deviation is used as the lower time boundary threshold. The basic action cycle prediction value is compared with these upper and lower limits, and outlier time series data exceeding the boundaries are eliminated, retaining valid data. The hydraulic mechanism response time constant inherent in the hydraulic pipeline manual is retrieved. The retained valid basic action cycle prediction value is divided by this response time constant, and a scaling operation is performed to finally obtain the candidate hydraulic valve opening duration parameter. The advantage of this operation logic is that it dynamically integrates historical statistical features to eliminate abnormal delays, ensuring the smoothness and accuracy of control duration allocation. The hydraulic mechanism response time constant is calibrated to 0.05 seconds. For example, if the extracted time axis component is 0.3 seconds and the operation response delay parameter is 0.02 seconds, the sum is 0.32 seconds. If the average value calculated from the historical 20 cycles is 0.31 seconds and the standard deviation is 0.01 seconds, the upper limit is calculated to be 0.33 seconds and the lower limit is 0.29 seconds. The 0.32 seconds is not eliminated within the range. Then, the 0.32 seconds is divided by the time constant of 0.05 seconds to obtain a ratio of 6.4. After converting the units, the candidate hydraulic valve opening duration parameter is determined to be 640 milliseconds.

[0086] S403: Based on the candidate hydraulic valve opening duration parameter, obtain the valve port flow constant, multiply the candidate hydraulic valve opening duration parameter by the flow constant, calculate the cylinder volume parameter, obtain the piston cross-sectional area and perform a division operation, and use the result as the module length and the direction control command vector for product operation to generate the target stroke increment command;

[0087] The system obtains the opening duration parameter of the generated candidate hydraulic valve, accesses the system configuration library to extract the valve orifice flow constant of the main control valve under a specific pipe diameter, and directly multiplies the previously determined candidate hydraulic valve opening duration parameter with the valve orifice flow constant to calculate the expected in-cylinder volume parameter flowing into the actuator cylinder during this opening time. Then, it reads the effective cross-sectional area of ​​the piston in the actuator from the equipment nameplate or assembly parameter table, uses the calculated in-cylinder volume parameter as the dividend and the piston cross-sectional area as the divisor, and performs a division operation to obtain the theoretical stroke displacement scalar. This theoretical stroke displacement scalar is assigned a directional attribute and is vector-superimposed with the directional control command vector generated in the previous steps, fusing the directional vector specified by the directional control command with the displacement scalar. The advantage of this calculation logic is that, based on a rigorous fluid dynamics volume formula, it transforms abstract time parameters into precise physical space displacement commands, significantly reducing the overshoot error at the actuator end. The valve orifice flow constant is set to 15 cubic centimeters per millisecond, and the piston cross-sectional area is set to 120 square centimeters. For example, by substituting the candidate hydraulic valve opening duration parameter of 640 milliseconds and multiplying it by the valve orifice flow constant of 15, the expected cylinder volume parameter is calculated to be 9600 cubic centimeters. Dividing this by the piston cross-sectional area of ​​120, the theoretical stroke displacement scalar is calculated to be 80 centimeters. Combined with the aforementioned positive direction control command, the command content is generated as a target stroke increment command of 80 centimeters forward advancement.

[0088] Table 3: Target Travel Increment Parameter Table

[0089]

[0090] Table 3 shows the parameters and calculation results in the above volume and stroke conversion logic, and provides stroke extrapolation under different opening durations.

[0091] Please see Figure 6 The specific steps of S5 are as follows:

[0092] S501: Call the target travel increment instruction and direction control instruction, obtain the underlying control bus address, parse the data frame to extract the register index, establish the instruction port mapping relationship, perform drive voltage amplitude quantization processing through the mapping relationship and the hardware response delay in a direct ratio, and establish the underlying drive control mapping table.

[0093] The generated target travel increment command and direction control command are extracted. The underlying device node description file of the system is accessed to obtain the underlying control bus address of the communication controller. Bit operations are performed on the sent data frame content to extract the register index number used to store the drive voltage configuration. A corresponding command port mapping relationship is established between the underlying control bus address and the register index. The hardware response delay between the command issuance and the output voltage signal of the digital-to-analog converter is measured. The hardware response delay ratio is calculated. The base drive voltage is multiplied by this ratio and a corresponding gain coefficient is added. Drive voltage amplitude quantization processing is performed to determine the analog control voltage output to the proportional valve. The port address, register index, and quantized drive voltage amplitude are bound in the form of key-value pairs to establish an underlying drive control mapping table that records the control topology of the entire system. The advantage of this operation logic is that it compensates for the timing deviation caused by hardware delay during control bus transmission, ensuring accurate application of the drive voltage. The measured hardware response delay is 12 milliseconds, the reference period is 10 milliseconds, the calculated hardware response delay ratio is 1.2, and the gain coefficient is fixed at 1.5. For example, the base drive voltage is set to 5 volts according to the direction control command. Multiplying 5 volts by the proportional ratio 1.2 equals 6 volts. Then multiplying by the gain coefficient 1.5, after quantization, the drive voltage amplitude is 9 volts. Subsequently, the bus address 192, register index 4 and voltage value 9 are combined and stored in the underlying drive control mapping table.

[0094] S502: Based on the underlying drive control mapping table, it monitors the overlapping periods of duty cycles of multiple commands, extracts load components and peak inrush current coefficients, performs vector summation on the load components within the overlapping periods, compares with the bus power supply reference value, identifies the concurrent execution logic mutual exclusion state, and establishes a command conflict detection matrix.

[0095] The underlying driver control mapping table is invoked to scan all planned output control commands in chronological order. The trigger timestamps and pulse widths of the commands are compared to extract the overlapping duty cycle periods of multiple commands. Within these overlapping periods, the device status table is queried to obtain the load components allocated to the actuators. The peak inrush current coefficient of the current circuit state is read from the power management module. The load components of all parallel actuators within the overlapping periods are vector-summed to obtain the total load demand. This total load demand is multiplied by the peak inrush current coefficient to estimate the transient maximum power, obtaining the system bus power supply reference value. The estimated transient maximum power is compared with the bus power supply reference value. If the transient maximum power is greater than the bus power supply reference value, a concurrent execution logic mutual exclusion state is identified for the relevant command group. The mutually exclusive command channels are recorded in a row-column interleaved format, generating a command conflict detection matrix where 1 represents a conflict and 0 represents no conflict. The advantage of this operation logic is that it performs power supply overload prediction before the hardware-level commands are issued, avoiding bus voltage drops or even power-off restart faults caused by excessively high current peaks. Assume the bus power supply baseline is 5000 watts and the peak inrush current factor is 1.8. For example, if a scan finds that actions 1 and 2 overlap between 200 and 300 milliseconds, the load component of action 1 is extracted as 1500 watts and the load component of action 2 is extracted as 2000 watts. The total load demand is 3500 watts. Multiplying this by the peak inrush current factor of 1.8 gives a transient maximum power of 6300 watts. A comparison shows that 6300 is greater than the baseline value of 5000, triggering a logic mutual exclusion state. A 1 is entered in the first row and second column of the instruction conflict detection matrix to mark it as a conflict item.

[0096] S503: Based on the instruction conflict detection matrix, call the underlying driver control mapping table to retrieve the safety instruction packet corresponding to the zero element matrix position, perform time-domain phase shift scheduling for non-zero element conflict items, extract the actuator cooperative response priority, arrange instructions according to logical order and pose constraints, and establish a cooperative operation sequence.

[0097] The generated instruction conflict detection matrix is ​​read, and its elements are traversed. For zero elements (0), the corresponding conflict-free instruction parameter combination is retrieved from the underlying driver control mapping table and packaged into a safety instruction package for direct issuance. For non-zero conflict elements (1), the overlap time span of the two concurrent instructions involved in the conflict is obtained. The safety response interval parameter pre-stored in the safety control strategy library is read, and the absolute difference between the safety response interval and the overlap time span is calculated. A compensation margin is added to this difference to generate a target phase offset. This target phase offset is then summed with the initial execution time node of the secondary priority actuator to complete the translation. Based on the equipment process requirement configuration table, the actuator collaborative response priority is extracted. The execution time axis of the instructions is rearranged according to the priority from high to low and the pose constraints of non-interference, constructing a conflict-free collaborative operation sequence. The advantage of this operation logic is that by implementing dynamic phase shift processing at the time domain level for conflicting instructions, the power supply contention is resolved without changing the spatial movement trajectory, maximizing the equipment collaborative efficiency. The safety response interval is set to 150 milliseconds, with an additional compensation margin of 20 milliseconds. For example, if the overlap time span between actions 1 and 2 is 100 milliseconds, the absolute difference of 50 milliseconds is obtained by subtracting the overlap time span of 100 milliseconds from the safety response interval of 150 milliseconds. Adding the additional compensation margin of 20 milliseconds, the target phase offset is generated as 70 milliseconds. If action 2 has a lower priority and its initial execution time node is 200 milliseconds, then 70 milliseconds are added to shift it to start at 270 milliseconds. Finally, the collaborative operation sequence is generated according to this time sequence.

[0098] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of protection of the described technical solutions.

Claims

1. A rapid construction method for initial support of tunnel equipment based on artificial intelligence, characterized in that, Includes the following steps: S1: Obtain the scalar of tension force of the construction equipment robotic arm and the scalar of displacement of the tunnel opening, perform first-order difference operation to obtain the tension force increment and displacement increment and calculate the proportional value, and construct the real-time coupled stiffness gradient. S2: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm, input them into the Markov decision process execution state mapping, construct the basic action reward matrix, and calculate the negative penalty scalar when the real-time coupled stiffness gradient is less than the soft rock yield gradient. S3: Call the negative penalty scalar, calculate the Euclidean distance between the robotic arms through spatial coordinates, calculate the collision avoidance limit factor when it is less than the safety threshold, and calculate the negative penalty scalar and the collision avoidance limit factor based on preset weights to construct a collaboratively updated reward matrix; S4: Analyze the collaborative update reward matrix, use a greedy algorithm to filter directional control commands and candidate hydraulic valve opening duration parameters, calculate the product of candidate hydraulic valve opening duration parameters and flow constant, and generate target stroke increment commands; S5: Obtain the underlying control bus address of the construction equipment, associate the target stroke increment command and direction control command to construct a feature mapping, perform conflict detection, and construct a collaborative operation sequence.

2. The rapid construction method for initial support of tunnel equipment based on artificial intelligence according to claim 1, characterized in that, The real-time coupled stiffness gradient includes the tangent stiffness coefficient, the slope of the surrounding rock characteristic curve, and the elastoplastic hardening modulus. The basic action reward matrix includes the state transition probability, the action value function, and the policy gradient increment. The collaborative update reward matrix includes the Euclidean distance penalty term, the obstacle avoidance safety margin, and the multi-agent consistency weight. The target stroke increment command includes the hydraulic cylinder piston displacement, the flow control rated value, and the actuator end displacement vector. The collaborative operation sequence includes the bus object dictionary, the command priority weight, and the logical interlock criterion.

3. The rapid construction method for initial support of tunnel equipment based on artificial intelligence according to claim 1, characterized in that, The specific steps of S1 are as follows: S101: The monitoring pressure detection branch acquires the scalar of the mechanical arm tension force of the construction equipment, monitors the feedback frequency to collect the scalar of the tunnel opening displacement, performs signal filtering and noise reduction and zero-point offset calibration, extracts the feature values ​​of the original sensing voltage waveform, maps the sampling channel and the physical parameter response weights, and establishes a set of working condition monitoring scalars. S102: Call the set of working condition monitoring scalars, perform a first-order difference operation on the scalar of the robot arm tension force and the scalar of the tunnel opening displacement, calculate the difference in tension load variation and the difference in tunnel surrounding rock displacement, map the sensor dynamic resolution and signal acquisition frequency matching attributes, and generate tension force increment and displacement increment. S103: Based on the tension force increment and displacement increment, calculate the ratio of tension force increment to displacement increment, map the face pressure distribution vector, aggregate the stress and displacement response variation amplitude, associate the elastic recovery capability of the shield support system, and generate a real-time coupled stiffness gradient.

4. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 3, characterized in that, The specific steps of S2 are as follows: S201: Call the real-time coupled stiffness gradient, collect the spatial coordinates of the robotic arm and the joint angle sequence of the end effector, perform joint feature extraction on the real-time coupled stiffness gradient and pose data, map the pose data and stress response distribution, aggregate displacement deviation and load change to calculate the transfer probability distribution, and establish an environmental state mapping vector. S202: Based on the environmental state mapping vector, evaluate the effectiveness of the real-time coupled stiffness gradient execution strategy, calculate the total expected amount of node step reward, combine motion smoothness weight and energy loss allocation coefficient to aggregate the weight of multi-objective reward function, and establish a basic action reward matrix. S203: Call the basic action reward matrix. If the real-time coupling stiffness gradient is less than the soft rock yield gradient, collect the instantaneous fluctuation frequency and rock settlement rate. Based on the deviation coefficient between the deformation trend and the instability threshold, combine the defense compensation coefficient and the safety penalty factor to perform normalization and generate a negative penalty scalar.

5. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 4, characterized in that, The instability threshold is determined by extracting the spatial stress extremum from the analytical environmental state mapping vector, calculating the absolute difference between the soft rock yield gradient and the real-time coupled stiffness gradient as the basic deformation tolerance, and then weighting and combining the basic deformation tolerance, the spatial stress extremum, and the energy loss allocation coefficient based on the parameter influence weight.

6. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 5, characterized in that, The specific steps for S3 are as follows: S301: Call the negative penalty scalar, collect the spatial coordinates of the robotic arm, perform Euclidean operation, calculate the relative displacement span between the centroids of the end effectors and joint axes of adjacent robotic arms, perform normalization mapping for the dynamic working radius and motion envelope, quantify the risk overlap, and obtain the spatial distance value between robotic arms. S302: For the spatial distance value between the robotic arms, when it is less than the safety threshold, retrieve the preset collision avoidance gain weight factor, perform exponential decay mapping, quantify the urgency of spatial conflict and the negative correlation strength of trajectory overlap constraint, associate the dynamic safety margin correction coefficient, and obtain the collision avoidance limit factor. S303: Based on the collision avoidance constraint factor and the negative penalty scalar, perform scalar linear weighted product operation, map the operation result to the state-action pair score node of the Markov decision process, associate the multi-machine cooperative joint action probability distribution, perform reward value discretization reconstruction and dimension tensor merging, and establish a collaboratively updated reward matrix.

7. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 6, characterized in that, The specific steps of S4 are as follows: S401: For the collaborative update reward matrix, extract node values ​​and perform comparison, filter the state parameter vector corresponding to the node with the largest score, obtain spatial displacement features, retrieve the hydraulic valve control rule mapping table, associate the robot arm joint posture parameters to perform vector dot product, quantify the motion trend, and generate direction control commands. S402: Based on the direction control command, call the collaborative update reward matrix, extract the time axis component, read the operation response delay parameter and perform summation, associate time boundary threshold to remove outlier data, match the hydraulic mechanism response time constant and perform proportional scaling to obtain the candidate hydraulic valve opening duration parameter; S403: Based on the opening duration parameter of the candidate hydraulic valve, obtain the valve port flow constant, multiply the opening duration parameter of the candidate hydraulic valve by the flow constant, calculate the cylinder volume parameter, obtain the piston cross-sectional area and perform division operation, and use the result as the modulus and the direction control command vector for product operation to generate the target stroke increment command.

8. The rapid construction method for initial support of tunnel equipment based on artificial intelligence according to claim 7, characterized in that, The time boundary threshold is determined by extracting the response time distribution characteristics of historical operating cycles within the collaborative update reward matrix.

9. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 7, characterized in that, The specific steps of S5 are as follows: S501: Call the target travel increment instruction and direction control instruction, obtain the underlying control bus address, parse the data frame to extract the register index, establish the instruction port mapping relationship, perform drive voltage amplitude quantization processing through the mapping relationship and the hardware response delay, and establish the underlying drive control mapping table. S502: Based on the underlying drive control mapping table, monitor the overlapping period of duty cycle of multiple commands, extract the load component and peak inrush current coefficient, perform vector summation on the load component in the overlapping period, compare with the bus power supply reference value, identify the concurrent execution logic mutual exclusion state, and establish a command conflict detection matrix. S503: Based on the instruction conflict detection matrix, call the underlying driver control mapping table to retrieve the safety instruction packet corresponding to the zero element matrix position, perform time-domain phase shift scheduling for non-zero element conflict items, extract the actuator cooperative response priority, arrange instructions according to logical order and pose constraints, and establish a cooperative operation sequence.

10. The rapid construction method for collaborative initial support of tunnel equipment based on artificial intelligence according to claim 9, characterized in that, The step of performing time-domain phase shift scheduling for non-zero element conflict items refers to obtaining the overlapping time span contained in the non-zero element conflict items, calculating the absolute difference between the preset safety response interval and the overlapping time span to generate the target phase offset, and superimposing the target phase offset onto the initial execution time node associated with the non-zero element conflict item to complete the shift.

Images