A PID fast excavation equipment adaptive control method based on GBO algorithm optimized RBFNN

Abstract

This invention belongs to the field of intelligent control of rapid tunneling machinery in coal mines. Specifically, it designs a trajectory control method for a horizontal-axis roadheader based on a PID controller optimized by a gradient optimization algorithm (GBO) using an RBFNN. This method takes the offset between the cutting teeth of the horizontal-axis rapid tunneling equipment and the preset sidewall position as the control objective. Addressing the time-varying characteristics of the damping and inertia of its power system, a variable time lag coefficient is introduced to establish a time-varying hydraulic transfer function. An RBF neural network PID adaptive controller is constructed with as input. The GBO algorithm is used to collaboratively optimize the initial parameters of the RBF network through gradient search rules (GSR) and local escape operators (LEO). Global optimization is performed using the output position error of the control system as the evaluation criterion. The optimal parameters are then loaded onto the controller to drive the hydraulic system and achieve real-time trajectory correction of the roadheader in the tunnel. This invention effectively solves the influence of the time-varying characteristics of the hydraulic system on control accuracy and improves the trajectory tracking performance of the roadheader in complex tunnel environments.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for engineering machinery, specifically the trajectory correction and high-precision tracking control of the hydraulic drive system of a tunneling and anchoring machine. It designs an adaptive control method for PID fast tunneling equipment based on GBO algorithm-optimized RBFNN. Background Technology

[0002] The horizontal axis rapid tunneling machine is a core piece of equipment in coal mines that simultaneously performs tunnel excavation and bolt support operations. Precise trajectory control and automatic edge-fitting of its traveling mechanism are crucial for ensuring tunnel formation quality and operational safety. The traveling drive system of the tunneling and bolting machine uses a valve-controlled hydraulic motor as its core actuator, exhibiting inherent characteristics such as high inertia, strong nonlinearity, and variable time delay, making high-precision motion trajectory control a significant engineering challenge. Currently, scholars both domestically and internationally have conducted fruitful research in areas such as adaptive trajectory tracking, neural network adaptive control, and hysteresis-compensated intelligent control. In the field of dedicated control for tunneling equipment, methods such as digital twin correction, attitude model identification, and BP neural network trajectory tracking have also emerged. However, the above solutions still have significant shortcomings: existing hydraulic system modeling generally uses a fixed time constant and does not consider the time-varying characteristics of clearance leakage caused by friction pair wear; the control architecture mostly uses a single feedback channel and lacks dedicated feedforward compensation for the large dead zone of hydraulic systems; offline optimization of BP neural network PID parameters mostly relies on traditional swarm intelligence algorithms such as particle swarm optimization (PSO) and genetic algorithm (GA), which are prone to getting trapped in local extrema in high-dimensional parameter spaces and have low convergence efficiency; the integral element in PID feedback control is prone to integral saturation when large deviations accumulate continuously, and there is a lack of effective integral separation and amplitude limiting protection.

[0003] While conventional PID controllers are simple, efficient, and dominant in industrial applications, their fixed parameters cannot adapt to the complex and ever-changing operating conditions of roadheader-and-anchor-machine (BAM). They often result in slow start-up response, severe mechanical overshoot, and persistent system oscillations when faced with the inherent time-delay characteristics of hydraulic systems. Introducing radial basis function (RBF) neural networks for real-time tuning of PID parameters can overcome the limitations of fixed-parameter PID controllers to some extent. However, the control performance of RBF neural networks is highly sensitive to their initial parameters; blind parameter initialization can lead to system divergence or severe mechanical shocks in the initial control phase. Therefore, there is an urgent need for a roadheader-and-anchor-machine trajectory control method that can accurately characterize the time-varying dynamic characteristics of the hydraulic system, effectively overcome hydraulic dead zones and integral saturation, and ensure the quality of the initial parameters of the RBF neural network using an efficient global optimization algorithm, thereby simultaneously improving the system's response speed, steady-state accuracy, and robustness. Summary of the Invention

[0004] To address the shortcomings of existing roadheader trajectory control technologies, such as the lack of consideration for time-varying leakage characteristics caused by friction pair wear in hydraulic system modeling, the absence of hydraulic dead zone feedforward compensation in single feedback channels, the reliance of RBF neural networks on PSO / GA initial parameters leading to local extrema and low convergence efficiency, and the tendency for PID integral elements to saturate under large deviations, this invention provides a roadheader trajectory control method based on gradient basis optimization algorithm to improve radial basis neural network PID, thereby simultaneously enhancing the system's response speed, steady-state accuracy, and robustness.

[0005] Implementation of the present invention: A method for trajectory control of a roadheader / anchor machine based on an improved radial basis function neural network PID algorithm using gradient basis optimization (GBO) is characterized by the following steps: The first step is to establish a mathematical model of the valve-controlled hydraulic motor. Based on the transfer function between the output flow and input voltage of the electro-hydraulic proportional directional valve, and combined with the flow continuity equation and torque balance equation of the hydraulic motor, a time-varying time coefficient is introduced to construct a complete transfer function of the valve-controlled hydraulic motor from the control voltage to the motor output speed. The second step is to establish a reference preset voltage signal-closed-loop feedback control architecture. Based on the mechanical dead zone compensation of the fast excavator, a reference preset voltage is generated to cross the dead zone of the hydraulic system. The feedback channel adopts a position PID controller with integral separation and amplitude limiting. The deviation between the actual coordinates of the cutting tooth and the preset sidewall position is used as the closed-loop feedback control quantity to synthesize the absolute voltage command of the final drive. The third step is to construct the RBF neural network architecture, using trajectory position offset, offset change rate and offset angle as network input vectors. The hidden layer uses Gaussian radial basis function as activation function. The three outputs of the output layer correspond to the real-time tuning values ​​of proportional coefficient, integral coefficient and derivative coefficient, respectively, which are used to synthesize feedback control quantity. The network parameters of each layer are updated online through gradient descent method. The fourth step employs the Gradient Basis Optimization (GBO) algorithm to perform global offline optimization of the hidden layer node centers, basis width parameters, and output layer weights of the RBF neural network. The fitness function is based on the integral of time multiplied by the absolute position error (ITAE), and the GBO rule is based on population optimum. with the worst individual vector The difference guides the global search direction, and the Local Escape Operator (LEO) applies random perturbations with a trigger probability to escape local extrema.

[0006] The method for constructing the time-varying time coefficients mentioned in the first step is as follows: select the motor output angular velocity. For state variables, input flow As a control input As an external disturbance, the pressure difference is eliminated by simultaneously solving the continuous flow equation of the hydraulic motor and the shaft end torque balance equation, resulting in the standard first-order linear time-varying state equation.

[0007] in For time-varying system matrices; Input matrix for time-varying system; Here is the perturbation matrix. Time-varying internal leakage coefficient Based on the laminar leakage coefficient of the concentric annular gap in the friction pair, a nonlinear relationship between internal leakage and wear is established. A first-order Taylor expansion of the cubic nonlinear gap term approximates it as a linear expression. Ignoring the product of the leakage and damping terms relative to the square of the displacement, a linear time-varying time coefficient is obtained.

[0008] The transfer function of the control system after introducing a variable time lag coefficient:

[0009] in Using the Laplace operator, this system is a third-order complex time-varying system. Its large inertia originates from the huge equivalent rotational inertia under heavy load conditions, resulting in a mechanically sluggish dynamic response. Its nonlinearity is due to the exponential effect of oil temperature changes and the triple nonlinear accumulation of wear between friction pairs. It also exhibits a first-order time-varying hysteresis influenced by the dynamic coupling of wear and temperature. .

[0010] The control input signal mentioned in the second step Includes reference voltage and feedback control voltage, with preset reference voltage output. By using a preset driving force, the system's operating point is shifted to the hydraulically sensitive zone, detaching the steady-state load and overcoming the mechanical dead zone; feedback channel Responsible for precise adjustment of dynamic deviation, the calculation method is as follows:

[0011] in This is the term for accumulating integrals. This represents the real-time trajectory deviation of the cutting teeth.

[0012] The positional PID controller employs an integral separation threshold and an integral accumulation term limit. When the absolute value of the deviation is greater than the integral separation threshold, the integral action is canceled; when it is less than the threshold, the integral action is engaged and the integral accumulation term is limited.

[0013] The third step involves establishing the three-layer network architecture of the RBFNN. Taking a hidden layer node as an example, the desired target at discrete time is the edge position, and the actual output feedback is the real-time position of the cutting tooth. The core error of the system is defined. with deflection angle As the learning target of the network; the final output is the PID parameters, and the hidden layer is... The center vector of each node and the input vector The Euclidean distance is used to obtain the hidden layer output through a Gaussian activation function. The output layer obtains three PID parameters by weighted summation of the hidden layer weight matrices; the calculation method is as follows:

[0014] in It is the output layer weight matrix. For the hidden layer output, in the base The parameters are updated to their optimal values ​​in real time.

[0015] The RBF network parameters are updated online using gradient descent. The update process includes updating the weight term w, the base width b, and the center vector c. The update method using an inertia coefficient and a small learning rate is as follows:

[0016] Similarly, the updates to the neuron's center vector and basis width are as follows:

[0017] in Based on the mapping relationship between the tunneling and anchoring machine body offset and the control voltage signal, the following information can be obtained for the system's Jacobian information:

[0018] in The radius of the drive wheel, For amplifier gain, For valve flow gain, Motor displacement Define the center width; define the loss function. ,when < At this point, the network parameters stop updating, and the optimal value is obtained. Value, output the current optimal control. This enables closed-loop control.

[0019] In the fourth step, the optimization of the GBO algorithm first encodes the RBF neural network parameters to be optimized as the position matrix X of the individuals in the GBO population, with the following dimensions:

[0020] The position update of the GBO algorithm consists of two stages. The first stage is the gradient search rule (GSR): it uses the vector difference between the best and worst individuals in the population to guide the global search direction.

[0021] in , These are the optimal and worst positions in the population, respectively. For a standard normally distributed random vector, It is a very small constant.

[0022] The second phase is the Local Escape Operator (LEO): based on the trigger probability... Apply random perturbations to the updated positions to help the population escape local optima:

[0023] in , Different individuals are randomly selected from the population; , δ is a uniformly distributed random vector; δ is a fixed disturbance amplitude coefficient.

[0024] A dynamic linear decreasing strategy is introduced for the step size coefficient and the direction control coefficient. The output position error of the control system is used as the evaluation criterion. The integral of time-in-absolute error (ITAE) is used as the fitness function. The minimum value of ITAE is calculated iteratively, and then the optimal particle position matrix is ​​output.

[0025] The overall implementation idea of ​​the above technical solutions of the present invention is as follows: firstly, introduce a time-varying time coefficient. Instead of the traditional fixed time constant, a forward simulation model is established by combining the mathematical model of the valve-controlled hydraulic motor and the dynamic model of the roadheader track. The difference between the real-time coordinates of the cutting teeth and the preset sidewall position is used as the tracking target deviation, and the target deviation and the forward simulation model form a closed-loop feedback loop. In the feedback control channel, the RBF neural network PID controller uses the trajectory deviation, deviation change rate, and deflection angle after feedforward dead zone compensation as inputs to tune the three gain coefficients of the PID in real time online. The gradient basis optimization algorithm (GBO) optimizes the initial parameters of the RBFNN. This includes GBO simulating numerical gradient information through the gradient search rule (GSR) to guide the population to converge toward the optimal region, the local escape operator (LEO) triggering the population to escape local extrema with a trigger probability, and using ITAE as the fitness function to output the optimal initial parameter matrix. The optimized initial values ​​are assigned to the RBF neural network PID controller, enabling the controller to obtain the optimal initial parameters, thereby achieving adaptive online real-time tuning of parameters under varying working conditions of the roadheader. Attached Figure Description

[0026] Figure 1Overall architecture diagram of the track control system for a roadheader and anchor machine based on GBO algorithm and RBFNN-PID; Figure 2 Flowchart of the controller parameter optimization algorithm based on GBO-LEO; Figure 3 Gradient basis optimization parameter values ​​and optimal fitness function values.

[0027] Figure 4 Example of correction trajectory simulation: Offset results of three control algorithms under the same preset edge and tooth coordinates. Detailed Implementation

[0028] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0029] like Figure 1 As shown, the track control system for a roadheader provided by this invention includes a pose detection device, a parameter adaptive negative feedback controller, an RBF neural network tuning module, a GBO offline optimization module, a valve-controlled hydraulic motor actuator, and a cutting tooth coordinate calculation module. The data flow and control logic between the modules are shown in the figure.

[0030] In this example, the valve-controlled hydraulic motor actuator of the roadheader's traveling mechanism uses an electro-hydraulic proportional directional valve as the core speed-regulating element. The overall transmission path is "electrical signal → valve core displacement → flow rate → motor speed → track speed". Within its rated operating range, the electro-hydraulic proportional directional valve can be approximated as a second-order oscillating element, and its control voltage... With output flow Transfer function between It can be represented as:

[0031] in These are the proportional amplifier gain and the valve flow gain, respectively. , These are the valve damping ratio and the valve natural frequency, respectively. The flow continuity equation and torque balance equation for a hydraulic motor are as follows:

[0032] Select motor output angular velocity For state variables, input flow As a control input As an external disturbance, the pressure difference is eliminated by simultaneously solving the continuous flow equation of the hydraulic motor and the shaft end torque balance equation, resulting in the standard first-order linear time-varying state equation.

[0033] in For time-varying system matrices; Input matrix for time-varying system; Here is the perturbation matrix. Time-varying internal leakage coefficient Based on the laminar leakage coefficient of the concentric annular gap in the friction pair, a nonlinear relationship between internal leakage and wear is established. A first-order Taylor expansion of the cubic nonlinear gap term approximates it as a linear expression. Ignoring the product of the leakage and damping terms relative to the square of the displacement, a linear time-varying coefficient is obtained.

[0034] The overall transfer function of the control system after introducing a variable time lag coefficient is:

[0035] To address the challenges of large dead zones, strong load impacts, and nonlinear hysteresis in the hydraulic system during downhole correction of roadheader machines, this invention designs a parameter adaptive negative feedback control architecture, whose overall control command... From dead zone reference voltage With feedback regulation voltage Synthesis, expressed as follows: in The positional deviation between the cutting teeth and the sidewall is due to... To cross the dead zone reference voltage signal value, the controller is subjected to integral separation threshold and integral accumulation term limiting. When the absolute value of the deviation is greater than the integral separation threshold, the integral action is canceled; when it is less than the threshold, the integral action is activated and the integral accumulation term is limited.

[0036] In this example, the RBFN network has 3 input layer nodes, 8 hidden layer nodes, and 3 output layer nodes. The input vector is... The loss function is defined as follows: The forward propagation Euclidean distance and the hidden layer output are as follows:

[0037] in , Let be the Euclidean distance between the node center vector and the input vector. The hidden layer output is activated using a Gaussian radial function. For the first Each base width parameter.

[0038] The three outputs of the output layer are the three parameters of the PID controller that are adjusted in real time, which are used to synthesize the feedback control quantity.

[0039] in It is the output layer weight matrix.

[0040] The output layer weights, hidden layer basis width parameters, and center vectors of the RBF neural network are updated online using the gradient descent method. The update method, which adjusts the inertia coefficient and the small learning rate, is as follows:

[0041] Similarly, the updates to the neuron's center vector and basis width are as follows:

[0042] in Based on the mapping relationship between the tunneling and anchoring machine body offset and the control voltage signal, the following information can be obtained for the system's Jacobian information:

[0043] in The radius of the drive wheel, For amplifier gain, For valve flow gain, Motor displacement Center width; when < At this point, the network parameters stop updating, and the optimal value is obtained. Value, output the current optimal control. This enables closed-loop control.

[0044] In this example, the learning rate is taken. Inertia weighting coefficient , Boundary constraints are applied to the updated network parameters. The base width parameter is constrained within the range of [0.1, 5.0], and the center vector is constrained within the coverage of the input vector space to prevent parameter drift from causing network degradation.

[0045] Iterative judgment, let Determine trajectory deviation Is it less than If the value is not less than 1, continue iterating and updating the parameters and calculating. If the value is true, output the optimal PID parameters.

[0046] like Figure 2 As shown, the Gradient Basis Optimization (GBO) algorithm is used to perform global offline optimization of the initial parameters of the RBF neural network. GBO guides the population to converge to the optimal region through the Gradient Search Rule (GSR) and escapes local optima through the Local Escape Operator (LEO), achieving efficient global search in a 56-dimensional high-dimensional parameter space. The specific process is as follows: Step 1: Convert the hidden layer node center vectors of the RBF neural network ( ), base width parameter Weight Encoded as a GBO population individual location vector:

[0047] The position vector dimension is 56, the population size is set to 100, and each dimension is initialized randomly within the physical region corresponding to the parameter.

[0048] Step 2: The core search mechanism of GBO is based on gradient search rules, guiding individuals in the population towards the optimal solution region by simulating numerical gradients. For the th individual in the population... individual Its gradient search rule is:

[0049] in , These are the optimal and worst positions in the population, respectively. For a standard normally distributed random vector, It is a very small constant. The step size coefficient of GSR is adopted using a dynamic linear decreasing strategy.

[0050] The direction control coefficient is based on the dynamic gradient factor. Sure. , A random number within (0, 1), dynamic gradient factor By randomly selecting 4 different individuals from the population , , , Solving for the position difference norm:

[0051] And Constrained within the interval [0.5, 2.0], adaptive adjustment of the search stride is achieved. The individual position update formula is:

[0052] Step 3: To prevent the population from getting trapped in local optima during the search process, GBO introduces a Local Escape Operator (LEO). This is based on the trigger probability... Apply a perturbation to the candidate solution:

[0053] in, Individuals randomly selected from the population; , Two different individuals were randomly selected; For the disturbance amplitude coefficient, a cubic nonlinear function attenuation strategy is adopted:

[0054] in, Depend on decay to The cubic decay strategy ensures that a strong perturbation is maintained in the early stage of the search to escape local extrema, while the perturbation is rapidly reduced in the later stage of the search to protect the convergence accuracy.

[0055] When the escape trigger condition is met (i.e., the random number is less than 1), Replace the current individual position with the candidate solution generated by LEO; otherwise, retain the position updated by GSR. This example takes...

[0056] Step 4: Use the integral of time multiplied by absolute position deviation (ITAE) as the fitness function:

[0057] in, This marks the simulation termination time. The ITAE metric penalizes both steady-state error and settling time, reflecting the overall dynamic quality of the system better than the IAE and ISE metrics. A smaller fitness value indicates better initial parameters for the RBF neural network corresponding to the individual. Increment the iteration count by 1; if the fitness value decreases, continue updating the optimal position. Continue until the maximum number of iterations is reached.

[0058] After decoding each individual into RBF neural network parameters, the entire control model is run in a simulation environment, and the ITAE index is calculated as the fitness value for that individual. Iterative optimization is then performed until the maximum number of iterations is reached. Then, the globally optimal position vector is output, and the optimal initial parameter matrix is ​​decoded. In this embodiment, the reference input voltage signals for both the left and right tracks are set to 100, the initial deviation is set to 150mm, the length of the excavator is 5000mm, and the width is 3000mm. The real-time coordinates of the cutting teeth are solved based on the Euler rotation matrix. The distance between these coordinates and the preset sidewall position is the monitoring quantity for the closed-loop feedback control of the entire vehicle. The simulation duration is set to... s, population size is The maximum number of iterations is .like Figure 3 As shown, the fitness value converges to the minimum value after 65 iterations. This indicates that the GBO algorithm has found the global optimum. By decomposing the optimal position vector and loading it into the RBF neural network, the GBO-RBF-PID controller can be obtained.

[0059] Simulations were performed on three algorithms—PID, RBF-PID, and GBO-RBF-PID—under the same positional deviation. The simulation time was set to 10 seconds, and the system response output curves were obtained as follows: Figure 4As shown, the RBF-PID control algorithm based on gradient basis optimization has the smallest overshoot and a smooth convergence process with no large oscillations. Compared with the traditional PID, the RBF-PID control reduces the overshoot by approximately 44% and 33%.

Claims

1. A method for improving the trajectory control of a tunnel anchor machine based on a radial basis neural network PID using a gradient-based optimization algorithm, characterized in that, Includes the following steps: Step S1: Establish a mathematical model of the valve-controlled hydraulic motor, based on the output flow rate of the electro-hydraulic proportional directional valve. With input voltage The transfer function between them, combined with the hydraulic motor flow continuity equation and the shaft end torque balance equation, is used to determine the time-varying clearance leakage coefficient caused by friction pair wear. Perform a Taylor expansion and introduce a linear time-varying time constant. Constructing a system from control voltage to motor output angular velocity The linear time-varying transfer function; Step S2: Establish a reference preset voltage signal-closed-loop feedback control architecture. Based on the mechanical dead zone compensation of the fast excavator, a reference preset voltage is generated to cross the dead zone of the hydraulic system. The feedback channel adopts a position PID controller with integral separation and amplitude limiting. The deviation between the actual coordinates of the cutting tooth and the preset sidewall position is used as the closed-loop feedback control quantity to synthesize the absolute voltage command of the final drive. Step S3: Construct an RBF neural network architecture, using trajectory position offset, offset change rate, and offset angle as network input vectors. The hidden layer uses Gaussian radial basis function as activation function. The three outputs of the output layer correspond to the real-time tuning values ​​of proportional coefficient, integral coefficient, and differential coefficient, respectively, which are used to synthesize feedback control quantity. The network parameters of each layer are updated online through gradient descent. Step S4: The gradient basis optimization algorithm (GBO) is used to optimize the control system of the horizontal axis rapid mining system. Global offline optimization is performed on the hidden layer node centers, basis width parameters, and output layer weights of the RBF neural network. The fitness function is the integral of time multiplied by the absolute position error and the deflection angle (ITAE). The gradient search rule (GBO) is based on population optimality. with the worst individual vector The difference guides the global search direction, and the Local Escape Operator (LEO) applies random perturbations with a trigger probability to escape local extrema.

2. The tunneling and anchoring machine trajectory control method according to claim 1, characterized in that: The method for constructing the time-varying time coefficient in step S1 is to select the motor output angular velocity. For state variables, input flow As a control input As an external disturbance, the pressure difference is eliminated by simultaneously solving the continuous flow equation of the hydraulic motor and the shaft end torque balance equation, resulting in the standard first-order linear time-varying state equation: in For time-varying system matrices; Input matrix for time-varying system; Here is the perturbation matrix. Time-varying internal leakage coefficient Based on the laminar leakage coefficient of the concentric annular gap in the friction pair, a nonlinear relationship between internal leakage and wear is established. A first-order Taylor expansion of the cubic nonlinear gap term approximates it as a linear expression. Ignoring the product of the leakage and damping terms relative to the square of the displacement, a linear time-varying coefficient is obtained. The transfer function of the control system after introducing a variable time lag coefficient: in Using the Laplace operator, this system is a third-order complex time-varying system. Its large inertia originates from the huge equivalent rotational inertia under heavy load conditions, resulting in a mechanically sluggish dynamic response. Its nonlinearity is due to the exponential effect of oil temperature changes and the triple nonlinear accumulation of wear between friction pairs. It also exhibits a first-order time-varying hysteresis influenced by the dynamic coupling of wear and temperature. .

3. The tunneling and anchoring machine trajectory control method according to claim 1, characterized in that: The control input signal in step S2 Includes reference voltage and feedback control voltage, with preset reference voltage output. By using a preset driving force, the system's operating point is shifted to the hydraulically sensitive zone, detaching the steady-state load and overcoming the mechanical dead zone; feedback channel Responsible for precise adjustment of dynamic deviation, the calculation method is as follows: in This is the term for accumulating integrals. This refers to the real-time trajectory deviation of the cutting teeth. The positional PID controller employs an integral separation threshold and an integral accumulation term limit. When the absolute value of the deviation is greater than the integral separation threshold, the integral action is canceled; when it is less than the threshold, the integral action is engaged and the integral accumulation term is limited.

4. The tunneling and anchoring machine trajectory control method according to claim 1, characterized in that: In step S3, the network topology architecture of RBFNN is established, and the number of hidden layer nodes is set to... The system core error is defined by taking the desired target at discrete time as the edge position and the actual output feedback as the real-time position of the cutting tooth. As a learning objective in the network; The input vector that constitutes the RBF neural network is, where is The fuselage deflection angle, the final output PID parameters, and the hidden layer... The center vector of each node and the input vector The Euclidean distance is used to obtain the hidden layer output through a Gaussian activation function. The output layer obtains three PID parameters by weighted summation of the hidden layer weight matrices; the calculation method is as follows: in It is the output layer weight matrix. For the hidden layer output, in the base The parameters are updated to their optimal values ​​in real time. The RBF network parameters are updated online using gradient descent. The update process includes updating the weight term w, the base width b, and the center vector c, and the update method uses an inertia coefficient and a small learning rate as follows: Similarly, the updates to the neuron's center vector and basis width are as follows: in Based on the mapping relationship between the tunneling and anchoring machine body offset and the control voltage signal, the following information can be obtained for the system's Jacobian information: in The radius of the drive wheel, For amplifier gain, For valve flow gain, Motor displacement Define the center width; define the loss function. ,when < At this point, the network parameters stop updating, and the optimal value is obtained. Value, output the current optimal control. This enables closed-loop control.

5. The tunneling and anchoring machine trajectory control method according to claim 1, characterized in that: In step S4, the optimization method of the GBO algorithm sets the fitness function as position deviation based on the target requirements of the tunneling and anchoring machine. angle with fuselage ITAE function: The ultimate optimization goal is to minimize the deviation and angle of deviation. The parameters of the RBF neural network to be optimized are encoded as the position matrix X of the individuals in the GBO population, with dimensions of 8m, as shown in the following expression: The position update of the GBO algorithm consists of two stages. The first stage is the gradient search rule (GSR): it uses the vector difference between the best and worst individuals in the population to guide the global search direction. in , These are the optimal and worst positions in the population, respectively. For a standard normally distributed random vector, It is a minimal constant. The second stage is the Local Escape Operator (LEO): with a trigger probability... Apply random perturbations to the updated positions to help the population escape local optima: in , Different individuals are randomly selected from the population; , Let be a uniformly distributed random vector; δ be a fixed disturbance amplitude coefficient. A dynamic linear decreasing strategy is introduced for the step size coefficient and the direction control coefficient. Using the output position error and deflection angle of the control system as evaluation criteria, the iteration cycle is completed, and the minimum value of ITAE is calculated. Then, the optimal particle position matrix is ​​output. The initial parameters assigned to RBFNN are decomposed, thus completing the GBO algorithm optimization.

6. A roadheader trajectory control system based on an improved radial basis function neural network PID algorithm using gradient basis optimization, characterized in that, The system includes: This system comprises five modules forming a three-layer closed loop: "offline-online-physical". The model building module establishes the time-varying transfer function of the valve-controlled hydraulic motor and serves as the shared physical interface for offline optimization and online control. The offline optimization module performs global optimization of the RBF network parameters under a multi-condition weighted criterion, unidirectionally injecting the optimal parameters into the online network. The RBF module takes deviation, deviation rate of change, and deflection angle as inputs and outputs the three parameter setpoints in real time. The reference preset voltage and closed-loop feedback control module implements position-based PID control on the reference input and actual deviation, superimposing the reference voltage to form the actuator absolute voltage command. The closed-loop feedback module completes the deviation detection between the actual position and the preset sidewall through the tooth coordinate solution, feeding back to the PID and RBF modules to form closed-loop feedback, thereby achieving high-precision trajectory tracking of the roadheader under complex working conditions. The model building module is used to establish the mathematical model of the valve-controlled hydraulic motor and construct the system transfer function under time-varying time coefficients. Reference preset voltage-closed-loop feedback control module: includes a reference preset signal and a position-type PID feedback module with integral separation, used to synthesize the absolute voltage command of the driver; RBF neural network tuning module: Input vector, output the tuned values ​​of proportional coefficient, integral coefficient and derivative coefficient in real time; GBO offline optimization module: used to perform global offline optimization of node center, base width and output layer weights of RBF neural network, and output the optimal initial parameter matrix to RBF neural network tuning module; Closed-loop feedback module: The deviation between the actual coordinates of the cutting tooth and the preset edge position is used as the feedback quantity to form a closed-loop negative feedback control.

Images