Robot driving control system based on particle swarm optimization

Abstract

The invention relates to the technical field of robot control and intelligent optimization control, in particular to a robot driving control system based on particle swarm optimization, which comprises a capability observation module used for receiving a theoretical trajectory instruction to output a system capability tensor; the manifold construction module is used for reading physical parameters of the robot body and initializing a knowledge manifold matrix containing an initial damping coefficient based on the physical parameters of the robot body so as to generate a target manifold matrix; the manifold optimization module is used for initializing a particle swarm on the surface of the target manifold matrix so as to output a target control parameter of the current control period; the driving closed loop module is used for converting a target control parameter into a driving waveform to control the driving motor, and performing numerical adjustment on an initial damping coefficient of the knowledge manifold matrix based on a track residual error; it is ensured that the robot has high real-time performance and low-overshoot safe driving capacity under the complex load and contact environment.

Description

Technical Field

[0001] This invention relates to the field of robot control and intelligent optimization control technology, specifically to a robot drive control system based on particle swarm optimization algorithm. Background Technology

[0002] As the requirements for robot motion accuracy and real-time performance continue to increase in scenarios such as refining and chemical maintenance, pipeline inspection, and industrial assembly, the stability control problem of robot drive control systems under complex loads, sudden friction changes, and multiple working conditions is becoming increasingly prominent. How to suppress overshoot, oscillation, and trajectory deviation while ensuring rapid response has become an important problem that needs to be solved in the field of robot control. Traditional robot drive control currently relies mainly on the following methods: servo control based on fixed parameters, parameter adjustment based on experience tuning, and offline parameter search based on conventional optimization algorithms; However, fixed parameter control, empirical tuning, and conventional offline optimization all have certain drawbacks. For example, fixed parameter control is difficult to adapt to real-time changes in load inertia and friction state, which can easily lead to overshoot or increased tracking error near the target position. Empirical tuning is highly dependent on operators and preset working conditions, making it difficult to adapt to dynamic switching in complex environments. Conventional optimization algorithms have a large number of iterations, making it difficult to meet the real-time requirements of millisecond-level control cycles, and they do not provide sufficient constraints on the physical boundaries of the robot body and the actual execution residuals, which can easily affect drive stability. Summary of the Invention

[0003] The purpose of this invention is to provide a robot drive control system based on particle swarm optimization algorithm, and to solve the following technical problems: This avoids the problems of traditional parameter optimization deviating from physical boundaries and excessive computation, thereby enabling rapid positioning and low overshoot control of robots under complex loads, complex frictions and high real-time requirements, and making them more adaptable to sudden changes in the physical environment and suppressing drive oscillations.

[0004] The objective of this invention can be achieved through the following technical solutions: A robot drive control system based on particle swarm optimization algorithm, the system comprising: The capability observation module is used to receive theoretical trajectory instructions from the host computer through the communication interface, collect the current data and position data of the robot's drive motor in real time through the configured current sensor and position encoder, and calculate the mechanical impedance characteristics based on the current data and position data to output a system capability tensor that includes physical dimensions such as friction force characteristics and load inertia characteristics. The manifold construction module is used to read the robot's physical parameters and initialize a knowledge manifold matrix based on the robot's physical parameters. This knowledge manifold matrix contains an initial damping coefficient pre-calibrated based on the ratio of the rated torque of the drive motor and the maximum allowable overload current. The module also uses the system capability tensor as an update weight to numerically update the local curvature of the knowledge manifold matrix to generate a target manifold matrix. The manifold optimization module is used to initialize the particle swarm on the surface of the target manifold matrix, extract the manifold gradient matrix by taking the derivative of the target manifold matrix, and perform single-step iterative optimization of the particle swarm within the current control cycle based on the manifold gradient matrix and the system capability tensor to output the target control parameters for the current control cycle. The drive closed-loop module is used to convert the target control parameters into drive waveforms to control the drive motor, collect the actual running trajectory of the drive motor, extract the trajectory residual between the actual running trajectory and the theoretical trajectory corresponding to the theoretical trajectory command, and numerically adjust the initial damping coefficient of the knowledge manifold matrix based on the trajectory residual.

[0005] Preferably, the capability observation module includes: A data synchronization unit is used to synchronously collect the current data and the position data at a preset frequency; The differential calculation unit is used to perform differential processing on the position data to obtain velocity differential data, and to perform online parameter identification based on the current data and the velocity differential data to calculate the friction force data and load inertia data. Tensor generation unit is used to encapsulate the friction force data and the load inertia data into the system capability tensor.

[0006] Preferably, the manifold building block includes: The parameter reading unit is used to read the reduction ratio parameter and rated torque parameter of the drive motor as the physical parameters of the robot body; A matrix initialization unit is used to construct a three-dimensional manifold surface based on the reduction ratio parameter and the rated torque parameter, and to convert the three-dimensional manifold surface into the knowledge manifold matrix; The target definition unit is used to map the control state, which represents the deviation rate between the theoretical trajectory and the actual trajectory at the target position, to a minimum point in the knowledge manifold matrix.

[0007] Preferably, the manifold construction module further includes a curvature update unit, which is used for: If the load inertia data is higher than the preset inertia threshold, the target location region is determined according to the theoretical trajectory corresponding to the theoretical trajectory instruction, and the curvature value of the matrix element corresponding to the knowledge manifold matrix in the target location region is increased according to the preset update step size. If the load inertia data is lower than a preset inertia threshold, the curvature value of the knowledge manifold matrix is ​​reduced according to a preset update step size; When the load inertia data is equal to a preset inertia threshold, the current curvature value of the knowledge manifold matrix remains unchanged; The preset inertia threshold is determined by superimposing the calibrated inertia reference value of the robot in an unloaded state with the dynamic margin of the current working condition.

[0008] Preferably, the manifold optimization module includes: The velocity update unit is used to initialize the current update velocity vector of the particle swarm, perform matrix multiplication of the current update velocity vector with the manifold gradient matrix, and perform element-wise weighted calculation of the product result with the system capability tensor to obtain the target update velocity. The parameter output unit is used to determine the spatial position of the particle swarm in the current control cycle based on the target update rate, and decode the spatial position into the target control parameters of the current control cycle, wherein the target control parameters include proportional coefficient, integral coefficient, derivative coefficient and feedforward coefficient.

[0009] Preferably, the drive closed-loop module includes: A waveform conversion unit is used to write the target control parameters into the controller register of the drive motor and convert them into pulse width modulation waveforms. A motor drive unit is used to drive the drive motor using the pulse width modulation waveform, and to make the robot's joints perform actions through the transmission components of the drive motor; The residual extraction unit is used to collect the actual running trajectory corresponding to the joint's execution action, and compare the actual running trajectory with the theoretical trajectory corresponding to the theoretical trajectory command to output the trajectory residual.

[0010] Preferably, the drive closed-loop module further includes a damping fine-tuning unit, which is used for: If the trajectory residual is higher than a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​increased by a preset adjustment step size; If the trajectory residual is lower than a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​reduced by a preset adjustment step size; When the trajectory residual is equal to a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix remains unchanged.

[0011] Preferably, the system is applied to industrial robotic arm control scenarios, wherein the physical parameters of the robot body include the arm span length parameter and joint mass parameter of the industrial robotic arm, and the knowledge manifold matrix is ​​used to constrain the overshoot of the end effector of the industrial robotic arm during the grasping process.

[0012] Preferably, the system is applied to pipeline robot driving scenarios, wherein the system capability tensor includes the real-time friction coefficient change of the pipeline inner wall, and the target manifold matrix is ​​used to adapt to sudden changes in pipeline load and suppress driving oscillations.

[0013] The beneficial effects of this invention are: 1. This invention reconstructs the traditional parameter optimization process into a single-step optimization within the manifold constrained by the system's physical capabilities; by eliminating time-consuming multi-round offline searches, it achieves rapid parameter updates with low computational load within millisecond-level control cycles, ensuring that the robot has a safe driving capability with high real-time performance and low overshoot in complex load and contact environments. 2. This invention introduces a data synchronization and online differential calculation mechanism, which effectively avoids non-physical fluctuations in mechanical impedance characteristics caused by inconsistent sensor sampling frequencies; by accurately extracting friction force and load inertia data and encapsulating them into a capacity tensor, it provides a stable and structured real-time physical state representation for the control system. 3. This invention explicitly embeds the physical parameters such as reduction ratio and rated torque into the knowledge manifold matrix and maps the overshoot-free target to the minimum point. This method brings the inherent hardware boundary forward before optimization, avoiding the algorithm from leading the parameters to an unsuitable aggressive region and eliminating high-frequency jitter of motor drive from the physical space level. 4. This invention designs a dynamic update strategy for curvature based on load inertia and the target location region; this mechanism enables the manifold to evolve synchronously with the real-time load, maintaining a smooth curve when moving away from the target to accelerate the approach, and increasing the curvature to form a potential barrier when approaching the target, thus achieving a dynamic balance between rapid approach at the far end and smooth braking at the near end. 5. This invention constructs a complete drive closed loop based on the residual of the actual running trajectory, and fine-tunes the initial damping coefficient of the manifold according to the residual; this mechanism directly converts the execution deviation into the correction amount of the smoothness of the manifold motion, enabling the controller to adapt to individual differences such as mechanical wear and temperature rise drift for a long time, and effectively suppress the cumulative error in long-term operation. 6. In the context of robotic arms, this invention introduces the arm span length and joint mass into the constraint system; it performs fine-grained potential energy adjustment to address the differences in inertia amplification caused by different postures, thus overcoming the shortcomings of traditional control in underestimating the differences in behavior between long and short arms, and achieving end-effector overshoot suppression and stable contact grasping of the robotic arm under complex extended postures. 7. In the pipeline robot scenario, this invention incorporates the real-time change in the friction coefficient of the pipeline inner wall into the capability tensor. This enables the controller to have extremely high environmental recognition, accurately distinguish between the actual increase in load and the deterioration of pipe wall friction, thereby quickly adapting to sudden load changes in complex pipelines and effectively suppressing drive oscillations when passing through scaled sections or bends. Attached Figure Description

[0014] The invention will now be further described with reference to the accompanying drawings.

[0015] Figure 1 This is a schematic diagram of a robot drive control system based on particle swarm optimization provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Please see Figure 1 A robot drive control system based on particle swarm optimization algorithm includes: a capability observation module, which is used to receive theoretical trajectory instructions issued by the host computer through a communication interface, collect current data and position data of the robot's drive motor in real time through configured current sensors and position encoders, and calculate mechanical impedance characteristics based on the current data and position data to output a system capability tensor containing physical dimensions such as friction force characteristics and load inertia characteristics. The manifold construction module is used to read the robot's physical parameters and initialize a knowledge manifold matrix based on the robot's physical parameters. This knowledge manifold matrix contains an initial damping coefficient pre-calibrated based on the ratio of the rated torque of the drive motor and the maximum allowable overload current. The module also uses the system capability tensor as an update weight to numerically update the local curvature of the knowledge manifold matrix to generate a target manifold matrix. The manifold optimization module is used to initialize the particle swarm on the surface of the target manifold matrix, extract the manifold gradient matrix by taking the derivative of the target manifold matrix, and perform single-step iterative optimization of the particle swarm within the current control cycle based on the manifold gradient matrix and the system capability tensor to output the target control parameters for the current control cycle. The drive closed-loop module is used to convert the target control parameters into drive waveforms to control the drive motor, collect the actual running trajectory of the drive motor, extract the trajectory residual between the actual running trajectory and the theoretical trajectory corresponding to the theoretical trajectory command, and numerically adjust the initial damping coefficient of the knowledge manifold matrix based on the trajectory residual.

[0018] This embodiment provides a robot drive control mechanism based on particle swarm optimization algorithm. Specifically, the mechanism is deployed in the underlying drive controller in the shutdown and maintenance scenario of a refining and chemical plant. The controller is connected to a current sensor, an absolute encoder, and a heterogeneous DSP / FPGA computing board. It is used to complete continuous control of capability observation, manifold update, parameter optimization, and drive closed loop within a millisecond-level control cycle when the robot performs high-precision trajectory actions. Specifically, the system receives theoretical trajectory instructions from the host computer; these instructions can be represented as a sequence of positions arranged in time. For example, in a simplified example, the theoretical positions of a joint at three consecutive control moments are 10°, 14°, and 17°. The capability observation module collects the corresponding current and position data at the same time, for example, the current sequence is 2.0A, 2.3A, and 2.5A, and the position feedback is 9.8°, 13.5°, and 16.6°. Based on these two types of data, the system calculates the mechanical impedance characteristics online. The mechanical impedance characteristics can be characterized by friction, inertia and equivalent damping. The equivalent damping is calculated by extracting the root mean square of the fluctuation of the current data within a preset time window and comparing it with the velocity decay rate corresponding to the position data. In a simplified example, it can be encapsulated as a two-dimensional capability vector [inertia level, friction level], such as [0.8, 0.3]; in practical applications, it can be extended to a multi-dimensional tensor, where each dimension corresponds to different joints or different physical capability components; the controller outputs the system capability tensor, which is used to characterize the dynamic capability boundary of the robot in the current posture, current load and current contact environment. Next, the manifold construction module reads the robot's physical parameters, such as reduction ratio, rated torque, joint allowable acceleration range, arm span or track traction capacity, and constructs a knowledge manifold matrix; this matrix can be understood as a potential energy distribution table after discretizing the control parameter space. For ease of explanation, assume that only the proportional coefficient and feedforward coefficient are constructed into a simplified 3×3 manifold matrix, whose initial values ​​can be represented as: First row: 9, 6, 4; Second row: 6, 3, 2; Third row: 5, 2, 1; where the smaller values ​​represent regions closer to the preset control target, such as regions with no overshoot and lower tracking error; The system writes the initial damping coefficient into this matrix to limit the particle's movement step on the manifold surface, so that parameter updates do not have overly sharp jumps. The initial damping coefficient is a constant pre-calibrated based on the ratio of the rated torque and the maximum allowable overload current of the drive motor, or a fixed empirical value identified and set by the system during the offline trial operation phase; The system capability tensor output by the capability observation module is used as update weights to participate in local curvature correction; for example, when the inertia increases, the matrix elements in the region near the target position increase, forming a higher local gradient; in this discrete knowledge manifold matrix, the local curvature is actually specifically defined as the rate of change of the potential energy difference between adjacent matrix elements, that is, the magnitude of the local gradient. When friction increases and torque redundancy is still sufficient, the mid-to-long-range region can be moderately flattened to allow particles to approach the high-gain region; after correction, the target manifold matrix is ​​generated. During the manifold optimization phase, the controller does not directly execute multiple rounds of particle swarm iteration in the traditional Euclidean parameter space, but initializes the particle swarm on the surface of the target manifold matrix and performs only a single-step iteration within the current control cycle. Taking the 3×3 simplified matrix as an example, if the initial landing points of the three particles correspond to matrix coordinates 1,1, 2,2, and 3,2 respectively, the manifold gradient matrix can be obtained after discretizing the target manifold matrix. Furthermore, by multiplying the particle's update velocity vector by the manifold gradient matrix and then performing a shrinkage operation with the system capability tensor, the target update velocity for the current operating condition can be obtained. The controller updates the particle's spatial position according to this target update velocity and decodes the position into the target control parameters for the current control cycle, such as a quadruple of proportional coefficient, integral coefficient, derivative coefficient, and feedforward coefficient. To avoid inconsistencies in the interpretation of symbols later in this paper, Kp, Ki, Kd, ​​and Kf will be used to represent proportional coefficient, integral coefficient, differential coefficient, and feedforward coefficient, respectively. These four letters will be used with this single meaning throughout the paper and will not be used for any other purpose. As a result, the parameter search is restricted to a safe space constrained by physical capabilities. During the closed-loop drive phase, the decoded target control parameters are written into the servo driver register and converted into the corresponding drive waveform, such as a PWM waveform, to drive the joint motor or the pipeline robot's propulsion motor. During execution, the system collects the actual running trajectory in real time and compares it with the theoretical trajectory to obtain the trajectory residual. For example, if the theoretical position is 17° and the actual position is 16.6°, then the position residual at this moment is 0.4°. The system can aggregate the residuals of multiple consecutive moments into the residual index of the current period, such as the mean square error or the maximum deviation value, and make numerical adjustments to the initial damping coefficient in the knowledge manifold matrix based on the residual, so that the manifold of the next control period is more in line with the current physical behavior. Furthermore, to avoid ambiguity in the dimensions of the manifold gradient matrix system capability tensor and particle velocity vector, all three are defined using the same control parameter axis in this embodiment; When the target control parameters include four terms, Kp, Ki, Kd and Kf, the particle position is represented by four-dimensional discrete coordinates, and the particle velocity is also represented by four components to show the update trend on the four parameter axes respectively; the knowledge manifold matrix can be a four-dimensional discrete table, or it can be degenerated into a two-dimensional or three-dimensional slice for the sake of illustration in the description, and the manifold gradient matrix represents the gradient components of each discrete node along each parameter axis. In single-joint scenarios, the system capability tensor is converted into capability weight vectors corresponding one-to-one with the parameter axes. In multi-joint scenarios, it is first calculated for each joint, and then independently invoked by the local control loops of the corresponding joints. In other words, in single-joint four-parameter control, the capability tensor can be organized as: constraint weights for Kp, constraint weights for Ki, constraint weights for Kd, and constraint weights for Kf, thus ensuring that the subsequent contraction operation is actually a numerical processing weighted according to the parameter axes, rather than an abstract operation with unknown dimensions. Furthermore, in this embodiment, the differentiation of the target manifold matrix is ​​preferably implemented using discrete difference, rather than requiring a continuous analytical surface. That is, for each parameter axis, the potential energy difference between the current node and the adjacent nodes is compared to obtain the local increase or decrease trend on that axis. If there are no adjacent nodes, boundary one-sided difference is used. After this processing, each item of the manifold gradient matrix can be directly interpreted as whether the system potential energy increases or decreases when a certain control parameter continues to increase or decrease, and the magnitude of the increase or decrease; accordingly, the so-called element-wise weighted calculation in this embodiment is preferably expressed as: the update amount of each parameter axis after the manifold gradient modulation is compressed, maintained or amplified axis by axis according to the current capability weight; For example, when the current load inertia is too large, the capability weights in the Kp and Kf directions can be smaller, so that the update steps in these two directions are compressed; when the friction increases but the torque margin is still sufficient, the capability weights in the Kf direction can be appropriately increased to allow the feedforward compensation to get closer to the adaptation region faster; thus, the calculation actions in the entire control chain can be mapped to specific discrete nodes, parameter axes and step size changes. It should be noted that if a certain type of input anomaly occurs during the control cycle, the system will enter the constraint processing mode; for example, when current data is missing but position data is available, the capability observation module will use the average current value of the previous cycle as a substitute and mark the capability tensor as low confidence. When position data suddenly jumps, such as the encoder suddenly jumping from 16.6° to 40° and not matching the current change, the point is judged as an abnormal sample and discarded directly, and the mechanical impedance characteristics of the previous cycle are used. When the particle falls outside the manifold boundary after the update, the particle is projected back to the nearest point on the boundary to prevent the output from exceeding the parameter range allowed by the driver. When the trajectory residual of multiple consecutive cycles exceeds the safety threshold, the system freezes the integral term and increases the damping coefficient to prioritize ensuring the stable stop of the actuator. For example, in the maintenance site of a refinery, the maintenance robot needs to first have the debugging robotic arm on the workshop station grab the calibration fixture, and then enter the process pipeline with bends to perform internal inspection and light-load operation. During the stage of the robotic arm grabbing the fixture, the theoretical trajectory requires the end to quickly approach the fixture without overshooting or collision. During operation within the pipeline section, the friction coefficient abruptly changes due to deposits on the pipe wall, causing the current to rise from 2.0A to 2.8A while the position advance decreases. Based on this, the system calculates the increase in friction and load inertia, and adjusts the knowledge manifold matrix in real time, so that the particles automatically move to a control parameter combination that is more suitable for the state in the next cycle. In this way, even if the physical environment changes rapidly in different maintenance stages, the controller can still maintain stable drive. The purpose of this step is to reconstruct the parameter optimization process, which was originally detached from the physical boundary, into an in-manifold optimization process constrained by the system capability, so as to achieve rapid positioning and low overshoot control of the robot under complex loads, complex frictions and high real-time requirements. Furthermore, to avoid the simplified example above and the tensor generation order below being interpreted as two different encoding rules, this manual makes a unified restriction on the component arrangement order of the system capability tensor when entering structured encapsulation and subsequent calculations: in a single-joint scenario containing only friction and inertia components, the formal encoding order is unified as friction force data, load inertia data; The content presented in the preceding text using inertia level and friction level is only used to illustrate that mechanical impedance characteristics can be jointly characterized by multiple physical capability components, and is not intended as the formal arrangement order for subsequent tensor encapsulation, matrix weighting, and element-wise weighted calculations. If other components such as torque margin and contact stiffness are introduced, they will be added sequentially according to the pre-registered fixed parameter axis order and remain unchanged within the same controller configuration cycle, thereby ensuring that the system capability tensor in the entire text has a unique meaning in terms of data structure and calling direction.

[0019] In a preferred embodiment of the present invention, the capability observation module includes: a data synchronization unit, used to synchronously collect the current data and the position data at a preset frequency; a differential calculation unit, used to perform differential processing on the position data to obtain velocity differential data, and to perform online parameter identification based on the current data and the velocity differential data to calculate friction force data and load inertia data; and a tensor generation unit, used to encapsulate the friction force data and the load inertia data into the system capability tensor. This embodiment provides a refinement mechanism for the capability observation module; specifically, during the continuous movement of the aforementioned maintenance robot, if only the current value and encoder position value are directly read without processing the sampling time deviation, the mechanical impedance characteristics are prone to non-physical fluctuations. For example, when the current is sampled at a frequency of 10kHz and the position is sampled at a frequency of 8kHz, the two are not strictly aligned within the same control cycle, which causes jitter in inertia identification and misleads subsequent manifold updates. Therefore, this embodiment introduces a data synchronization unit, a differential calculation unit, and a tensor generation unit to refine the underlying observation link. Specifically, the data synchronization unit synchronously collects current data and position data at a preset frequency; the preset frequency can be consistent with the main control cycle, for example, under a 1kHz control cycle, a synchronous sampling packet is formed every 1ms; if the original sampling frequency is higher than the control cycle, the median, average or last value can be taken within a 1ms window to form the representative value of that cycle; Assuming that the original current samples read in a certain control cycle are [2.1A, 2.2A, 2.1A, 2.3A] and the original position samples are [10.1°, 10.4°, 10.6°, 10.7°], the system can respectively obtain the representative current 2.175A and the representative position 10.45°, thus forming a data pair with a unified timestamp; The differential calculation unit performs differential processing on the position data to obtain velocity differential data. For ease of explanation, it is assumed that the positions of two adjacent control cycles are 10.45° and 10.85°, and the cycle duration is 1ms. Then the velocity differential value is approximately 400° / s. Combining the velocity change in adjacent cycles, the acceleration change trend can be further obtained. The system performs online parameter identification based on current, velocity differential, and acceleration trend. Under the simplified model, the current can be considered to mainly reflect the output torque, the velocity difference can be considered to mainly reflect the effect of friction, and the acceleration change is more related to the equivalent inertia. For example, if the current rise exceeds the set comparison threshold while the velocity increment decreases, it is more likely that friction is increasing; if the current rises but the acceleration increase is still limited, it is more likely that the load inertia is increasing. Therefore, the differential calculation unit outputs friction force data and load inertia data, such as friction force level 0.7 and load inertia level 0.8; the tensor generation unit further encapsulates the friction force data and load inertia data into a system capability tensor. In a simplified example, it can be represented as a binary tensor T=[0.7, 0.8]; in a multi-joint system, it can be expanded into a matrix form, such as generating [[0.7, 0.8], [0.3, 0.5]] in a two-joint scenario, where each row corresponds to a joint and each column corresponds to the friction and inertia components respectively; if the system also needs to reflect the torque margin, it can be further expanded in the column direction. The resulting tensor retains the capability differences between different joints and facilitates subsequent coupling operations with the manifold gradient matrix. It should be noted that if a certain type of data is lost within the synchronous sampling window, interpolation or the effective value from the previous cycle will be used to make up for it. If the data is lost continuously for more than a preset number of times, the tensor generation unit will output a degraded tensor, for example, by setting the missing components to a safe upper limit, so as to make subsequent control more conservative. If the position difference reaches a minimum value, causing the velocity to approach zero, the friction estimate is easily amplified. In this case, the system adds a stationary judgment flag and freezes the friction update when the joint is close to stationary, updating only the inertia trend. If the current saturation reaches the actuator limit, the system marks the current identification result as a limited result and does not directly use it as the basis for large manifold deformation, but only allows small step size correction. For example, before the inspection robot enters the pipe bend, the position of the propulsion motor is advanced from 120.0mm to 122.0mm, and the current increases from 1.8A to 2.4A; after entering the bend, the position is only advanced from 122.0mm to 122.8mm, while the current increases to 3.0A. After synchronous sampling, the differential calculation unit determines that the decrease in velocity difference exceeds the preset velocity difference threshold and the increase in current exceeds the preset current threshold. Therefore, the friction level is increased from 0.4 to 0.8 and the inertia level is increased from 0.5 to 0.7. The tensor generation unit outputs the system capability tensor of the current propulsion module for subsequent manifold curvature updates. The purpose of this step is to transform the raw sensor signals into structured capability information that can be used to control constraints, thereby enabling an online, stable, and computable representation of the robot's actual physical state. Furthermore, in order to maintain the same data axis definition with the subsequent manifold construction module and manifold optimization module, the tensor generation unit preferably adopts a fixed field order when outputting the system capability tensor. At least when the two basic components of friction and inertia are included, the tensor generation unit encapsulates the tensor in the order of friction first and inertia second. If subsequent expansions include torque margin and other components, they will continue to be added according to the preset field table order, without arbitrarily changing the column order between different paragraphs or joints. After this processing, subsequent modules can directly retrieve values ​​according to the predetermined field positions when reading the system capability tensor, avoiding ambiguity that the same symbol T represents different component arrangements in different positions.

[0020] In a preferred embodiment of the present invention, the manifold construction module includes: a parameter reading unit, used to read the reduction ratio parameter and rated torque parameter of the drive motor as physical parameters of the robot body; a matrix initialization unit, used to construct a three-dimensional manifold surface based on the reduction ratio parameter and the rated torque parameter, and convert the three-dimensional manifold surface into the knowledge manifold matrix; and a target definition unit, used to map the control state representing the deviation rate between the theoretical trajectory and the actual trajectory at the target position to zero as a minimum point in the knowledge manifold matrix.

[0021] This embodiment provides a manifold construction module refinement mechanism; specifically, in the aforementioned system, if the control parameters are only temporarily adjusted based on the online identification results without introducing the physical boundary of the robot body in advance, the system may still lead the particles to unsuitable parameter regions when the load changes abruptly. For example, the same scaling factor may be acceptable for a high reduction ratio, high torque motor, but may cause high-frequency jitter for a low torque, light reduction motor; therefore, this embodiment explicitly embeds the robot body structure constraints into the knowledge manifold through a parameter reading unit, a matrix initialization unit, and a target definition unit.

[0022] Specifically, the parameter reading unit reads the reduction ratio parameter and rated torque parameter from the device parameter table or the driver calibration register. For example, in a simplified example, suppose the reduction ratio of a certain joint is 100 and the rated torque is 12 N·m. Based on this, the system judges that the joint has a strong low-speed steady-state holding capability, but will be affected by both transmission smoothness and inertia amplification when switching quickly. The matrix initialization unit constructs a three-dimensional manifold surface based on these physical parameters; the three dimensions here can be understood as two control parameter dimensions plus a potential energy value dimension, or a position interval dimension, a capability level dimension and a potential energy dimension; for ease of explanation, it is assumed that 9 nodes are discretized on a two-dimensional plane of proportional coefficient Kp and feedforward coefficient Kf, and the system assigns an initial potential energy value to each node according to the reduction ratio and rated torque; For example, when the reduction ratio is high, the penalty value for excessive Kp is increased; when the rated torque is high, the reward value for moderate Kf is moderately increased; finally, a three-dimensional surface is obtained, which is then discretized and mapped into a knowledge manifold matrix. The target definition unit maps the overshoot-free control target to the minimum point in the knowledge manifold matrix. Taking the above 9 nodes as an example, if it is determined from experience and offline experiments that the combination of medium Kp and medium Kf is most likely to achieve the overshoot-free effect, then the node can be assigned a value of 1, while the surrounding nodes are assigned higher values ​​such as 2, 3, 5, etc., indicating deviation from the target state. Thus, the lowest point of the manifold is not an arbitrary mathematical optimum, but rather a high-precision, overshoot-free control preference already embedded; when subsequent particles move on the surface of this matrix, they will preferentially converge to this type of low-potential region. It should be noted that if the reduction ratio or rated torque is missing in the equipment parameter table, the system will enter a conservative initialization mode, setting the matrix as a whole to a higher damping level and expanding the minimum value area into a small range platform rather than a single point minimum value, in order to avoid overly aggressive parameter concentration. If the read rated torque is inconsistent with the limit reported by the real-time driver, the smaller value is used as the effective rated torque to prevent overestimation of system capability; if the parameters of different joints are significantly different, their own knowledge manifold matrices are constructed separately and read independently by joint in the subsequent decoding stage, without simple sharing. For example, at the maintenance station, the robotic arm needs to grasp the calibration fixture and load it onto the front end of the pipeline robot; the base joint reduction ratio is 120 and the rated torque is 18 N·m; the end joint reduction ratio is 50 and the rated torque is 6 N·m. The controller initializes the manifold matrix for both: the base joint matrix has a lower tolerance for high scaling factors to avoid heavy-load swaying; the end joint matrix allows for a more sensitive feedforward response so that the gripper target definition unit can quickly close the gripper and map the target without overshoot contact when the gripper approaches the tooling to the minimum point in the matrix, so that subsequent optimization always revolves around the safe gripping area. The purpose of this step is to bring the inherent physical boundaries of the robot body forward before parameter optimization, thereby realizing the physicalization and constraint of the control search space.

[0023] In a preferred embodiment of the present invention, the manifold construction module further includes a curvature update unit, which is configured to: determine a target location region according to the theoretical trajectory corresponding to the theoretical trajectory instruction when the load inertia data is higher than a preset inertia threshold, and increase the curvature value of the matrix element of the knowledge manifold matrix corresponding to the target location region according to a preset update step size; decrease the curvature value of the knowledge manifold matrix according to a preset update step size when the load inertia data is lower than the preset inertia threshold; and keep the current curvature value of the knowledge manifold matrix unchanged when the load inertia data is equal to the preset inertia threshold. The preset inertia threshold is determined by superimposing the calibrated inertia reference value of the robot in an unloaded state with the dynamic margin of the current working condition.

[0024] This embodiment provides a curvature update mechanism; specifically, after the aforementioned manifold initialization, if a fixed curvature is used throughout the entire control process, although the robot body boundary can be reflected, it cannot reflect changes in load during operation; for example, after the robotic arm grasps the front tooling, the end effector inertia will increase significantly. When the pipeline robot passes through the area of ​​accumulated attachments, the equivalent load will also increase. If the manifold surface does not change with the load state, the particles may still slide quickly towards the high-gain region along the original slope, thus causing the risk of overshoot when approaching the target position. Therefore, this embodiment introduces a curvature update unit to adaptively adjust the local curvature according to the load inertia data. Specifically, the curvature update unit reads the current load inertia data and compares it with a preset inertia threshold. The preset inertia threshold is determined based on the robot's calibration inertia reference value in an unloaded state, plus a dynamic margin calculated from the maximum allowable safe load of the current process segment. For ease of explanation, assume the inertia threshold is set to 0.6, and the initial elements of the knowledge manifold matrix corresponding to the target location region near the end of the theoretical trajectory are: [2, 1, 2]. When the online identification result shows that the current inertia is 0.8, which is higher than the threshold, the system determines the target location region based on the theoretical trajectory. For example, if the theoretical trajectory corresponds to the end moving from 50mm to 100mm, and the current position has reached 92mm, then 90mm to 100mm can be defined as the target position region; the curvature update unit increases the curvature value of the matrix elements in this region according to the preset update step size; wherein, the preset update step size is determined according to the limit value of the discrete parameter gradient change corresponding to the maximum allowable output torque of the drive motor in the current control cycle. If the step size is 1, then after the update, it can become: [3, 2, 3]; this means that a steeper potential barrier is formed in the region near the end point, and the particle enters the constraint boundary with a higher numerical gradient. After the particle enters this region, it no longer uses a high proportional gain coefficient, but tends to brake smoothly; When the load inertia is lower than a preset threshold, for example, an inertia value of 0.4, the system can reduce the curvature of the corresponding region; if the original matrix segment is [3, 2, 3] and the step size is still 1, it is reduced to [2, 1, 2]; the surface thus becomes smoother, and particles can more easily slide to the parameter region with faster response, thereby improving the positioning speed; when the inertia value is exactly equal to the threshold, the curvature remains unchanged, avoiding fluctuations around the threshold; to reduce frequent switching, the system can also introduce a small range of hysteresis region for the inertia data, for example, 0.58 to 0.62 are all considered as equal to the threshold processing range; If the previous scheme is adopted without setting a target location region, and only the curvature is uniformly increased across the entire matrix, a new defect will arise: the system will also be overly conservative when far from the target location, resulting in a slower initial acceleration phase; therefore, this embodiment specifically locates the target location region to be braked based on the theoretical trajectory, and only performs weighted deformation in the relevant local areas, thereby balancing rapid approach at the far end and smooth docking at the near end. If the theoretical trajectory itself has a very small span, such as only a minor calibration action, the target position area can be degenerated into the entire trajectory interval. In this case, the local update is equivalent to a slight global update. If the load inertia data continuously jitters and crosses the threshold, the data is first averaged and then the three-state judgment is performed to prevent the curvature from frequently fluctuating. If the current matrix element has reached the upper limit of the allowable range, the curvature is increased and clamped at the upper limit to avoid forming an insurmountable numerical cliff; if the current matrix element has dropped to the lower limit of the allowable range, the curvature is decreased and kept at the lower limit to avoid sacrificing stability for excessive speed. For example, at the maintenance station, after the robotic arm grips the tooling, it begins to move to the loading position of the pipeline robot; before gripping, the end-load inertia identification value is 0.45, and the curvature update unit reduces the curvature of the area near the loading position, so that the robotic arm can approach faster. After grabbing, the load inertia increases to 0.82. The system increases the curvature value within a 10mm range near the loading position, so that the end effector slows down significantly when approaching the positioning hole without overshooting the impact. The same control strategy is used in the process of the pipeline robot advancing towards the detection point in the pipe. When carrying the sampling module through a heavily fouled area, the inertia of the propulsion module increases and the curvature of the target position area is increased synchronously to suppress the oscillation when approaching the detection point. The purpose of this step is to enable the knowledge manifold to evolve synchronously with actual load changes, thereby achieving a dynamic balance between near-end overshoot prevention and far-end rapid approach.

[0025] In a preferred embodiment of the present invention, the manifold optimization module includes: a velocity update unit, used to initialize the current update velocity vector of the particle swarm, perform matrix multiplication of the current update velocity vector with the manifold gradient matrix, and perform element-wise weighted calculation of the product result with the system capability tensor to obtain the target update velocity; The parameter output unit is used to determine the spatial position of the particle swarm in the current control cycle based on the target update rate, and decode the spatial position into the target control parameters of the current control cycle, wherein the target control parameters include proportional coefficient, integral coefficient, derivative coefficient and feedforward coefficient.

[0026] This embodiment provides a manifold optimization module refinement mechanism; specifically, after the aforementioned manifold construction is completed, if the standard particle swarm optimization algorithm is still used to repeatedly calculate the individual optimum and the global optimum, it is often difficult to complete a sufficient number of iterations within a 1ms control cycle, resulting in insufficient real-time performance. Furthermore, the standard update formula is not sensitive to changes in physical capabilities and is prone to stagnation near local optima. Therefore, this embodiment introduces a velocity update unit and a parameter output unit to perform a single-step update within each control cycle, subject to both manifold and capability constraints. Specifically, the velocity update unit first initializes the current update velocity vector of the particle swarm. For ease of understanding, assume there are three particles, representing the positions of three sets of candidate control parameters P1, P2, and P3 respectively. Their current velocity vector can be simplified to: V=[1, -1, 0]; each value here represents the forward trend of the corresponding particle along the discrete coordinates of the manifold. After the system performs discrete differentiation on the target manifold matrix, it obtains the manifold gradient matrix. Assuming the gradient matrix is: G=[[2,0,0],[0,1,0],[0,0,-1]], then multiplying V and G yields the intermediate result VG=[2,-1,0]. This step reflects the modulation of particle velocity direction and amplitude by the manifold terrain. The system performs element-wise weighted calculations on VG and the system capability tensor; let the current capability tensor be T=[0.5, 1.0, 0.8], corresponding to the capability weights of the regions where the three particles are located, then the target update speed after shrinking can be expressed as: V'=[1.0, -1.0, 0]; if the capability of the region where the first particle is located is limited, its speed is compressed; if the second particle is in a region with sufficient capability, it can maintain a high mobility. Unlike traditional PSO, this system no longer emphasizes multi-round competition between groups. Instead, it directly embeds manifold gradients and physical capabilities into single-step updates, thus completing the search advancement for the current working condition with a single matrix operation. This system uses the current manifold gradient to directly replace the individual historical optimal and global historical optimal in traditional particle swarm optimization in terms of particle velocity, thereby adapting to the high real-time requirements within millisecond-level control cycles. Furthermore, to avoid misunderstanding of the meaning of the letters in the above simplified illustration, P1, P2, and P3 respectively represent the position identifiers of the candidate control parameters of three different particles; V represents the current updated velocity vector arranged by particle index; G represents the local gradient matrix corresponding to the current landing point of each particle; VG represents the intermediate result of the velocity vector after manifold gradient modulation; T represents the capability weight vector read by particle region; V' represents the target updated velocity vector after capability weight modulation, where the apostrophe is only used to indicate the updated or modulated result and does not indicate the meaning of differentiation; thus, the aforementioned letter symbols have a one-to-one correspondence in the objects they refer to in this set of examples; The parameter output unit determines the spatial position of the particle within the current control cycle based on the target update rate and decodes the spatial position into actual control parameters. For example, if a particle is currently located at coordinates 2,2,1,2, which correspond to the positions on the four discrete axes Kp, Ki, Kd, ​​and Kf respectively, and the target update rate is +1,0,-1,+1, then the position will be updated to 3,2,0,3 at the next moment. The system looks up the corresponding control parameters or interpolates them to obtain them, such as Kp=2.5, Ki=0.3, Kd=0.05, Kf=1.2. If multiple particles are updated at the same time, the one with the lowest potential energy, the one with the best residual prediction, or a weighted average can be used as the output parameter for this period. Furthermore, in order to avoid the aforementioned simplified example being misunderstood as a confusion between particle dimension and parameter dimension, there are two equivalent implementation methods in this embodiment. The first is a particle index implementation: each item in V corresponds to the overall update trend of a particle, G is used to characterize the local slope at the current landing point of each particle, and T is used to characterize the capability weight of the region where each particle is located. At this time, the result after shrinking is still a velocity value for each particle, and then the update is completed by the parameter coordinates bound to each particle itself. The second method is the parametric axis implementation: each particle is expanded separately, and its velocity vector is denoted as Vk, Vi, Vd, and Vf along the four axes Kp, Ki, Kd, ​​and Kf, respectively. The manifold gradient is also given according to the four parametric axes, and the capability tensor is also converted into capability weights on the four axes, so that the target update velocity of the particle on the four parametric axes can be obtained directly. The aforementioned examples of V=[1, -1, 0] and T=[0.5, 1.0, 0.8] are simplified illustrations of the first implementation, while the examples of coordinates 2, 2, 1, 2 and velocity +1, 0, -1, +1 are expanded descriptions of the internal parameter axes of a single particle under the second implementation. Both point to the same control concept and do not contradict each other. Furthermore, in the parametric axis implementation, the target update rate is preferably determined according to the axis-by-axis rule, that is, the target update amount on the d-th parametric axis is jointly determined by the current velocity of that axis, the local gradient of that axis, and the capability weight of that axis, and is subject to amplitude limiting after calculation; the following discrete update rule can be adopted:

[0027] Among them, the Each parameter axis preferably corresponds to one of the axes Kp, Ki, Kd, ​​and Kf; the current update rate on this axis is denoted as . The local gradient component of this axis at the current manifold node is denoted as The capability weight of this axis obtained from the system capability tensor mapping is denoted as... To ensure the consistency of dimensions at both ends of the update formula, the local gradient components... With the axis capability weight All of them have been pre-processed to be dimensionless and used as dimensionless pure scalar coefficients in the product operation; The target update amount after calculation and amplitude limiting is denoted as ,in This is a limit operation operator. Its operation logic is as follows: when the core operation value is less than the lower limit, it returns the lower limit value; when it is greater than the upper limit, it returns the upper limit value; when it is in between, it returns the core operation value itself. This represents scalar multiplication; the lower and upper limits of the allowed velocity range are denoted as . and ; According to this rule, if the gradient of a certain parameter axis is steep and the capability weight is low, the corresponding update amount will be significantly compressed; if the gradient direction is conducive to approximating the low potential energy region and the current physical capability allows it, the update amount of that axis can be kept at its original value or moderately increased. If only the manifold matrix from the previous layer is used without the three-level update of velocity vector, gradient matrix and capability tensor, there is still a defect: although the particles are constrained by the manifold, there is no fast suppression or amplification mechanism for velocity amplitude when capability changes suddenly, which means that although the update direction is correct, the step size may be inappropriate; this embodiment makes up for this step by calculating the capability element by element, so that the optimization step size is also shaped in real time by the current physical state. If the target update velocity calculation result of a certain particle is zero, it means that the current position is close to the stable point, and the system can keep the particle stationary for one cycle; if the velocity result is too large and exceeds the preset upper limit, it will be limited, for example, within the range of [-2, 2]; if the updated particle position falls below the minimum value or above the maximum value of the discrete axis, boundary clamping will be performed; if the potential energy values ​​of all particles are the same, the particle with the smaller trajectory residual history will be selected first; if one of the decoded parameters exceeds the write range of the driver, that parameter will be rolled back to the value of the previous cycle, and the other parameters will be updated as usual. For example, during the process of the robotic arm delivering the tooling into the front mounting position of the pipeline robot, when it approaches the mounting hole about 5mm in advance, the manifold gradient becomes steep in the high Kp region. At the same time, due to the large load inertia, the capability tensor exerts a compression effect on the velocity vector, causing the particle to switch from the original rapid approach mode to a steady-state convergence mode. After the assembly is completed, the control core switches to the pipeline robot propulsion module. The same single-step update mechanism continuously outputs different combinations of Kp, Ki, Kd, ​​and Kf in each stage of propulsion, turning, and stopping, without having to perform time-consuming multi-round offline searches. The purpose of this step is to transform particle swarm optimization from a high-iteration, weakly physically coupled form to a low-computation, strongly physically constrained single-step update form, thereby achieving implementable online optimization with millisecond-level control cycles. Furthermore, to ensure that the symbolic references in the entire text remain unique, when describing the schematic implementation by particle index, the current update velocity vector of the entire particle swarm is denoted as Vp, the corresponding local gradient matrix is ​​denoted as Gp, and the particle region capability weight is denoted as Tp; when describing a single particle by expanding it along four parameter axes Kp, Ki, Kd, ​​and Kf, the particle's parameter axis velocity vector is denoted as Va, the parameter axis gradient vector of the node where the particle is located is denoted as Ga, and the particle's parameter axis capability weight vector is denoted as Ta. The V, G, and T used above are abbreviations for ease of explanation. In formal implementation or subsequent explanation, they should fall into one of the two sets of notation mentioned above and should not be used interchangeably. In other words, Vp=[1, -1, 0] represents the overall update trend of three particles, while Va=+1, 0, -1, +1 represents the update trend of a single particle on the four parameter axes. Although both can be called velocity vectors, they correspond to the particle index space and parameter axis space respectively in data structure, thus avoiding the same symbol carrying two different physical meanings in the whole text.

[0028] In a preferred embodiment of the present invention, the drive closed-loop module includes: a waveform conversion unit, used to write the target control parameters into the controller register of the drive motor and convert them into a pulse width modulation waveform; a motor drive unit, used to drive the drive motor using the pulse width modulation waveform, and to enable the robot's joints to perform actions through the transmission components of the drive motor; and a residual extraction unit, used to collect the actual running trajectory corresponding to the joint's performance action, and to compare the actual running trajectory with the theoretical trajectory corresponding to the theoretical trajectory command to output the trajectory residual.

[0029] This embodiment provides a mechanism for refining the drive closed-loop module. Specifically, if the upstream has obtained the target control parameters, but these parameters only remain at the computation layer and do not establish a closed-loop connection with the underlying driver register and waveform output, the entire scheme still belongs to the computation layer optimization and cannot truly form controlled execution. To this end, this embodiment further refines the waveform conversion unit, motor drive unit and residual extraction unit, so that parameter updates can actually be applied to the drive waveform and then return to the trajectory error closed loop. Specifically, the waveform conversion unit writes the Kp, Ki, Kd, ​​and Kf output in the current control cycle into the controller register of the drive motor. This register can be an internal control word of the DSP or a mapped register of the FPGA. For example, if the decoded parameters for a given cycle are Kp=2.5, Ki=0.3, Kd=0.05, and Kf=1.2, the system converts them into a fixed-point format and writes them to the corresponding address driver to generate a pulse width modulation (PWM) waveform based on these parameters. For instance, in a simplified example, if the target torque corresponds to a duty cycle increase from 45% to 52%, the PWM generator outputs a new drive waveform, which is then sent to the power stage to drive the motor. After receiving the above waveform, the motor drive unit drives the corresponding joint to perform the action; for the robotic arm, the action can be joint rotation or gripper closure; for the pipeline robot, the action can be propulsion wheel rotation or steering actuator deflection; in this embodiment, the driving action directly corresponds to the parameters of the previous cycle, so the result of manifold optimization is no longer an abstract score, but is immediately transformed into the actual motion of the actuator. The residual extraction unit collects the actual running trajectory within the same closed loop and compares it with the theoretical trajectory, outputting the trajectory residual. For ease of explanation, it is assumed that the theoretical position sequence is [20.0mm, 22.0mm, 24.0mm] and the actual position sequence is [19.8mm, 21.6mm, 23.7mm], then the residuals at each time point are 0.2mm, 0.4mm, and 0.3mm, respectively. The system can extract a maximum residual of 0.4 mm, or a mean square residual of about 0.31 mm as the feedback quantity for the next cycle. If speed consistency is also considered, speed residuals can be extracted in parallel to more accurately determine whether there is insufficient early braking or excessive conservatism. If only the aforementioned optimized output parameters are used without constructing a residual extraction closed loop, new defects will arise: although the system can update the parameters according to the capability tensor, it cannot verify the actual effect of these parameters in the real drive chain. Especially under the influence of factors such as power device temperature rise, encoder quantization error, and mechanical backlash, the calculated optimal and the executed optimal are not always consistent. This embodiment uses residual extraction to pull the actual execution result back into the control chain, forming a correctable complete closed loop. If writing parameters to the register fails, the driver retains the parameters from the previous cycle and records a write error; if the PWM generator is in an abnormal state, the motor drive unit enters a safe stop or torque limiting mode; if the encoder loses lock during trajectory acquisition, the residual extraction unit switches to a low confidence mode, outputting only the trend error without directly triggering a large parameter correction. If the number of sampling points of the theoretical trajectory and the actual trajectory are inconsistent, the system will first align them according to the timestamp and then compare them; if the theoretical trajectory itself contains a jump command, a window comparison will be used near the jump point instead of a single point comparison to avoid misjudging normal command switching as residual abnormality. For example, during the loading stage at the workstation, the end effector of the robotic arm needs to insert the tooling into the front interface of the pipeline robot; the controller writes the currently optimized parameters into the joint drive register, driving the clamping joint and docking joint to move synchronously; after assembly, the control core continues to drive the pipeline robot to advance along the theoretical trajectory inside the pipe; when it advances to the detection point, the residual extraction unit finds that the theoretical position should reach 350mm, but it only reaches 347mm in reality, forming a 3mm residual; this residual is sent back to the subsequent damping fine-tuning link to suppress oscillations and accumulated errors in subsequent stages; The purpose of this step is to actually implement the parameter optimization results into the drive execution layer and feed the execution results back to the upstream control logic, thereby achieving closed-loop control from computation to physics and back from physics to computation.

[0030] In a preferred embodiment of the present invention, the drive closed-loop module further includes a damping fine-tuning unit, which is used to: increase the initial damping coefficient of the knowledge manifold matrix by a preset adjustment step size when the trajectory residual is higher than a preset residual threshold. If the trajectory residual is lower than a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​reduced by a preset adjustment step size; if the trajectory residual is equal to the preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​kept unchanged. This embodiment provides a damping fine-tuning mechanism. Specifically, in the aforementioned closed loop, the system can obtain the trajectory residual. However, if the residual is only used as an observation result and does not further act on the knowledge manifold matrix, the damping in the manifold is still mainly set by the initial experience, which cannot adapt to individual differences, device wear and temperature rise drift during long-term operation. Therefore, this embodiment introduces a damping fine-tuning unit so that the trajectory residual can be written back as the manifold damping adjustment amount. Specifically, the damping fine-tuning unit obtains the trajectory residual of the current cycle and compares it with a preset residual threshold; wherein, the preset residual threshold is a physical boundary value set based on the maximum positioning tolerance allowed by the execution end in the target process section, and the preset adjustment step size is proportional to the rate of change of the trajectory residual. For ease of explanation, we assume the preset residual threshold is 0.5mm, the initial damping coefficient of the current knowledge manifold matrix is ​​0.8, and the preset adjustment step size is 0.1. If the current residual is 0.9 mm, which is higher than the threshold, it indicates that the parameter combination of the previous cycle still has deviation or oscillation tendency in actual driving. The system will increase the damping coefficient from 0.8 to 0.9. After the damping coefficient is increased, the movement of particles on the manifold surface in the next cycle will be smoother and the parameter switching amplitude will be smaller, thereby suppressing overshoot and oscillation. If the current residual is below the threshold, for example, 0.2mm, it indicates that the current control is relatively stable and there may still be room for further improvement in response speed. At this point, the system lowers the damping coefficient from 0.8 to 0.7, giving the particle more flexible sliding ability in the next cycle and promoting the parameters to move closer to the faster response region; if the residual is exactly equal to the threshold, the damping coefficient is kept unchanged to avoid meaningless back-and-forth adjustments at the boundary point. If the previous closed-loop scheme is used without adding damping fine-tuning, a problem will be exposed: although the residual can be measured, the smoothness of the manifold will not be automatically corrected according to the actual execution effect. The system needs to rely on manual reset of the damping, which is difficult to adapt in the long term. This embodiment uses residual-damping mapping to directly convert execution deviation into manifold motion rule correction amount, so that the controller can gradually converge to a dynamic equilibrium point that is more suitable for the individual equipment state during long-term operation. If the residual exceeds the threshold for several consecutive periods and the damping coefficient has reached the upper limit, it indicates that the problem may no longer be insufficient manifold smoothness, but rather mechanical jamming or sensor failure. In this case, the system issues a maintenance alarm and limits further automatic adjustment. If the residual is too low for several consecutive cycles and the damping coefficient has dropped to the lower limit, the lower limit should be kept unchanged to prevent the parameter from being updated too quickly. If the residual fluctuates slightly around the threshold, a dead zone can be set. For example, 0.45mm to 0.55mm is considered equal to the threshold to reduce frequent changes in damping. If the sampling quality of the previous cycle is low, the damping fine-tuning in this cycle can be postponed and the original coefficient can be used. For example, after the pipeline robot enters the detection section, the theoretical endpoint of the propulsion trajectory is 500mm; the actual endpoints measured in three consecutive cycles are 498.8mm, 499.0mm and 498.7mm, respectively, with the corresponding residuals consistently higher than the threshold; the damping fine-tuning unit gradually increases the initial damping coefficient of the manifold from 0.8 to 1.0, making the particle update more stable, and the residuals of the propulsion module gradually converge in the subsequent cycles; After the robot passes through the high-friction area, the residual error drops below 0.2mm, and the system then slowly lowers the damping coefficient to restore a faster inspection and propulsion speed. The purpose of this step is to transform the execution error into an adaptive adjustment basis for the smoothness of the manifold motion, thereby achieving steady-state optimization under long-term operation and oscillation suppression under abnormal operating conditions.

[0031] In a preferred embodiment of the present invention, the system is applied to an industrial robotic arm control scenario. The physical parameters of the robot body include the arm span length parameter and joint mass parameter of the industrial robotic arm. The knowledge manifold matrix is ​​used to constrain the overshoot of the end effector of the industrial robotic arm during the grasping process. This embodiment provides an application mechanism in an industrial robotic arm scenario. Specifically, in the aforementioned refining and chemical plant maintenance process, before the pipeline robot enters the pipe, the workstation robotic arm usually needs to complete the tooling gripping, positioning loading, and docking calibration. If the manifold construction method that only considers the reduction ratio and rated torque is used, although it can cover the general dynamic constraints of the drive layer, the end effector gripping process may still experience overshoot when the arm span of the robotic arm changes significantly and the inertia amplification caused by different postures is large. Therefore, this embodiment introduces the arm span length parameter and joint mass parameter in the industrial robotic arm scenario, and uses the knowledge manifold matrix to constrain the end effector gripping overshoot. Specifically, the system reads the arm span and joint mass parameters of the robotic arm during initialization. For example, the equivalent arm span of the robotic arm in the retracted state is 0.4m, and the arm span increases to 1.1m when extended to the grasping position, with a significant increase in the equivalent inertia at the end effector. If the masses of the second and third joints are 12kg and 8kg respectively, the system can estimate the dynamic sensitive area under different postures. For ease of explanation, it is assumed that the arm spans of short, medium, and long, as well as the end-effector approach velocities of low, medium, and high, are discretized into a 3×3 matrix. For the region of long arm span and high approach velocity, the flow potential energy value is set to a higher value, for example, from 2 to 6, in order to suppress excessively large scaling factor and feedforward factor during optimization. For the region of short arm span and low approach velocity, a lower potential energy value can be set, for example, 1 or 2, to retain flexibility. During the grasping process, the knowledge manifold matrix plays a constraint role, especially on the trajectory of the end effector approaching the workpiece. If the gripper is only 5mm away from the fixture and the robotic arm is in a long-arm extension posture, the curvature update and particle optimization work together to make the control parameters converge towards the low overshoot region. In this way, the end effector can avoid hitting the workpiece or gripping overshoot without sacrificing too much cycle time. If the arm span and joint mass are not included in the industrial robotic arm scenario, and the model is still based on general drive parameters, a defect will be exposed: under the same motor output, the difference in end-effector behavior between long arm span posture and short arm span posture is underestimated, which makes it easier for end-effector swing and overshoot to occur in the extended grasping state; this embodiment fills this gap by injecting posture-related structural parameters into the manifold matrix. If the arm span parameter cannot be obtained accurately in real time under certain postures, the joint angle can be used to estimate the range value, and a conservative value can be assigned to the side with the larger arm span. If the joint mass parameter only has the design value and lacks the measured value after maintenance, the system will make a slight correction to the relevant weights in combination with the residual during the first run phase. If the mass of the object to be grasped is unknown, a manifold is constructed according to the no-load model before contact, and the capacity tensor and local curvature are quickly corrected based on the current increment after contact. If the mass of the object to be grasped is lower than the preset load lower limit, even if the arm span exceeds the preset extension threshold, the damping can be reduced through the capacity tensor to avoid system response lag. For example, at the maintenance station, an industrial robotic arm grabs a set of inspection components that need to be installed into the pipeline robot from the material rack; the robotic arm first approaches the material rack at high speed in a retracted posture, and then extends to a 1.1m arm span to grab it; The system recognized that the current working condition was a long reach and high approach speed, so it raised the potential barrier in the manifold matrix near the gripping point, so that the Kp and Kf of the optimization output would no longer increase. Finally, the gripper did not rebound significantly when it contacted the component, the end effector overshoot was constrained to an acceptable range, and the loading was successfully completed. The purpose of this step is to incorporate the unique posture and structural factors of industrial robotic arms into manifold constraints, thereby achieving end-effector overshoot suppression and stable contact control during the grasping process.

[0032] In a preferred embodiment of the present invention, the system is applied to a pipeline robot driving scenario, wherein the system capability tensor includes the real-time change in the friction coefficient of the pipeline inner wall, and the target manifold matrix is ​​used to adapt to sudden changes in pipeline load and suppress driving oscillations.

[0033] This embodiment provides an application mechanism for pipeline robots. Specifically, in the aforementioned maintenance process, after the robot completes the loading at the workstation, it enters the process pipeline to perform inspection, sampling, or light-load maintenance. Compared with industrial robotic arms, the prominent problem faced by pipeline robots is not the change in arm span, but the real-time change in the coefficient of friction caused by deposits on the inner wall of the pipeline, accumulated liquid, and transitions at bends, as well as sudden changes in load. If the fixed friction assumption under normal conditions is still used, the propulsion drive is prone to oscillation at the location where the friction coefficient increases or decreases abruptly; therefore, in this embodiment, the real-time friction coefficient change of the inner wall of the pipe is incorporated into the system capacity tensor, and the target manifold matrix is ​​used to adapt to sudden load changes. Specifically, the system continuously reads the driving current, propulsion displacement, and wheel speed differential data during the advancement process, and estimates the real-time change in the friction coefficient of the inner wall of the pipe from these data. For ease of explanation, it is assumed that the baseline value of the friction coefficient is 0.30 when the robot is running in the straight pipe cleaning section; after entering the scaling area, the estimated value rises to 0.55. When passing through the liquid accumulation section, the friction may drop to 0.20; the system defines the friction change as the offset relative to the reference value, such as +0.25 or -0.10, and uses it as a component of the system capacity tensor; if the current tensor is represented as [inertia 0.7, friction change +0.25, torque margin 0.4], it means that the robot is facing a heavy load and large friction, and the torque margin is limited; The target manifold matrix is ​​adjusted accordingly. Still using a discrete illustration, if the potential energy of a certain propulsion control parameter region was originally [2, 1, 2], when friction suddenly increases, the system can raise the potential energy in the low-to-medium feedforward region, while keeping the stable region suitable for high-friction propulsion at a lower value, for example, updating it to [4, 2, 3]. In this way, the particle is more likely to fall into the parameter region with stronger steady-state traction capability during single-step optimization, avoiding the output from swinging back and forth between high and low gain. If friction suddenly decreases, the system avoids continuing to maintain excessively high damping and excessively large traction output to prevent slippage and oscillation in the smooth section. If the pipeline robot is treated as a general mobile actuator without incorporating changes in pipe wall friction into the capability tensor, there will be obvious defects: the manifold's response to sudden load changes depends only on inertia and current changes, and it cannot distinguish between the two causes of actual load increase and deterioration of contact friction, which may result in inaccurate parameter correction direction; this embodiment introduces real-time friction coefficient changes to enable the controller to have higher recognition of changes in the pipeline environment. If the friction change fluctuates drastically in a short period of time, the system first performs time window smoothing and then decides whether to significantly deform the manifold; if the conversion deviation of wheel speed and displacement is found to exceed the preset tolerance range when estimating the friction coefficient, it indicates that there may be idling. At this time, the friction component is marked as high risk and the damping is forcibly increased. If the change in pipe diameter causes a change in the contact pattern, the reference friction value is reset when entering the new pipe diameter range to prevent the old range reference from misleading subsequent judgments; if the torque margin of the propulsion module is close to zero, even if the friction change is not significant, the feasible area of ​​the manifold is tightened first in accordance with the safe docking strategy. For example, after loading the detection components, the pipeline robot moves along the return pipeline of the refining unit; in the first 50m of clean straight pipe, the friction change is close to 0, the target manifold matrix remains relatively flat, and the robot moves forward at a relatively fast speed; when it reaches a section of curved pipe with severe deposition, the current increases significantly, the displacement slows down, the system estimates the friction change to rise to +0.22, and at the same time detects an increase in load inertia; The target manifold matrix then increases curvature and damping in the target area near the bend exit, making the particle optimization output more robust control parameters and significantly reducing propulsion oscillations; after passing through this section, the frictional change drops, the manifold gradually returns to smooth, and the robot re-enters the rapid inspection state; The purpose of this step is to directly incorporate changes in contact friction caused by the pipeline environment into the control constraints, thereby achieving rapid adaptation to sudden load changes and continuous suppression of driving oscillations.

[0034] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A robot drive control system based on particle swarm optimization algorithm, characterized in that, The system includes: The capability observation module is used to receive theoretical trajectory instructions from the host computer through the communication interface, collect the current data and position data of the robot's drive motor in real time through the configured current sensor and position encoder, and calculate the mechanical impedance characteristics based on the current data and position data to output a system capability tensor that includes physical dimensions such as friction force characteristics and load inertia characteristics. The manifold construction module is used to read the robot's physical parameters and initialize a knowledge manifold matrix based on the robot's physical parameters. This knowledge manifold matrix contains an initial damping coefficient pre-calibrated based on the ratio of the rated torque of the drive motor and the maximum allowable overload current. The module also uses the system capability tensor as an update weight to numerically update the local curvature of the knowledge manifold matrix to generate a target manifold matrix. The manifold optimization module is used to initialize the particle swarm on the surface of the target manifold matrix, extract the manifold gradient matrix by taking the derivative of the target manifold matrix, and perform single-step iterative optimization of the particle swarm within the current control cycle based on the manifold gradient matrix and the system capability tensor to output the target control parameters for the current control cycle. The drive closed-loop module is used to convert the target control parameters into drive waveforms to control the drive motor, collect the actual running trajectory of the drive motor, extract the trajectory residual between the actual running trajectory and the theoretical trajectory corresponding to the theoretical trajectory command, and numerically adjust the initial damping coefficient of the knowledge manifold matrix based on the trajectory residual.

2. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The capability observation module includes: A data synchronization unit is used to synchronously collect the current data and the position data at a preset frequency; The differential calculation unit is used to perform differential processing on the position data to obtain velocity differential data, and to perform online parameter identification based on the current data and the velocity differential data to calculate the friction force data and load inertia data. Tensor generation unit is used to encapsulate the friction force data and the load inertia data into the system capability tensor.

3. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The manifold construction module includes: The parameter reading unit is used to read the reduction ratio parameter and rated torque parameter of the drive motor as the physical parameters of the robot body; A matrix initialization unit is used to construct a three-dimensional manifold surface based on the reduction ratio parameter and the rated torque parameter, and to convert the three-dimensional manifold surface into the knowledge manifold matrix; The target definition unit is used to map the control state, which represents the deviation rate between the theoretical trajectory and the actual trajectory at the target position, to a minimum point in the knowledge manifold matrix.

4. The robot drive control system based on particle swarm optimization algorithm according to claim 2, characterized in that, The manifold construction module further includes a curvature update unit, which is used for: If the load inertia data is higher than the preset inertia threshold, the target location region is determined according to the theoretical trajectory corresponding to the theoretical trajectory instruction, and the curvature value of the matrix element corresponding to the knowledge manifold matrix in the target location region is increased according to the preset update step size. If the load inertia data is lower than a preset inertia threshold, the curvature value of the knowledge manifold matrix is ​​reduced according to a preset update step size; When the load inertia data is equal to a preset inertia threshold, the current curvature value of the knowledge manifold matrix remains unchanged; The preset inertia threshold is determined by superimposing the calibrated inertia reference value of the robot in an unloaded state with the dynamic margin of the current working condition.

5. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The manifold optimization module includes: The velocity update unit is used to initialize the current update velocity vector of the particle swarm, perform matrix multiplication of the current update velocity vector with the manifold gradient matrix, and perform element-wise weighted calculation of the product result with the system capability tensor to obtain the target update velocity. The parameter output unit is used to determine the spatial position of the particle swarm in the current control cycle based on the target update rate, and decode the spatial position into the target control parameters of the current control cycle, wherein the target control parameters include proportional coefficient, integral coefficient, derivative coefficient and feedforward coefficient.

6. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The drive closed-loop module includes: A waveform conversion unit is used to write the target control parameters into the controller register of the drive motor and convert them into pulse width modulation waveforms. A motor drive unit is used to drive the drive motor using the pulse width modulation waveform, and to make the robot's joints perform actions through the transmission components of the drive motor; The residual extraction unit is used to collect the actual running trajectory corresponding to the joint's execution action, and compare the actual running trajectory with the theoretical trajectory corresponding to the theoretical trajectory command to output the trajectory residual.

7. The robot drive control system based on particle swarm optimization algorithm according to claim 6, characterized in that, The drive closed-loop module further includes a damping fine-tuning unit, which is used for: If the trajectory residual is higher than a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​increased by a preset adjustment step size; If the trajectory residual is lower than a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix is ​​reduced by a preset adjustment step size; When the trajectory residual is equal to a preset residual threshold, the initial damping coefficient of the knowledge manifold matrix remains unchanged.

8. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The system is applied to industrial robotic arm control scenarios. The physical parameters of the robot body include the arm span length parameter and joint mass parameter of the industrial robotic arm. The knowledge manifold matrix is ​​used to constrain the overshoot of the end effector of the industrial robotic arm during the grasping process.

9. The robot drive control system based on particle swarm optimization algorithm according to claim 1, characterized in that, The system is applied to pipeline robot driving scenarios. The system capability tensor includes the real-time change in the friction coefficient of the pipeline inner wall. The target manifold matrix is ​​used to adapt to sudden changes in pipeline load and suppress driving oscillations.

Images