A method and device for precise assembly based on vision and force self-adaptive long axis centering

Abstract

The application discloses a kind of based on vision and force perception adaptive long axis precision centering assembly method and device, it is related to aero-engine assembly technical field, device includes linear guide, linear module, actuating mechanism, fixed seat, clamping mechanism, force sensor, binocular camera and coaxial installation first target and second target.Method includes driving shaft end part approaches hole end part, visual servo controller adjusts pose, generates helical search motion when force mutation, make shaft end slide into along chamfer conical surface, visual gain attenuation after lateral force disappears, carry out pose adjustment and provide basic damping, after visual gain attenuation to lower limit, force perception compliance controller provides dynamic damping and admittance stiffness and carries out pose closed-loop correction, axial force reaches target value and lateral force is zero when judging completion, loosen workpiece.The application solves the problem of pose deviation and damage in long axis centering process by the adaptive fusion of vision and force perception, and is suitable for high-precision assembly of long shaft parts of aero-engine.

Description

Technical Field

[0001] This invention relates to the field of aero-engine assembly technology, specifically an assembly method and apparatus based on vision and force-based adaptive long-axis precision alignment. Background Technology

[0002] With the continuous improvement of the performance requirements of aero engines, the assembly accuracy of its core components, such as long shaft parts, and the casing bores directly affects the engine's operational reliability and service life. Currently, the assembly of long shaft parts and casing bores mainly relies on the experience of technicians for alignment, or on using a binocular camera to measure the position before insertion, calculate the deviation, and then have a robot adjust its posture. After adjustment, the robot can perform pure position insertion without visual monitoring.

[0003] However, for long shafts with large length-to-diameter ratios and heavy weights, significant reference drift occurs during insertion, resulting in cumulative deviations between the actual and pre-adjusted poses during insertion, which can easily cause surface scratches. Since the deflection of the long shaft in the cantilever state follows the nonlinear distribution of the cantilever beam, the existing solution cannot adapt to the dynamic changes of the continuous shortening of the overhang length during insertion.

[0004] Furthermore, visual perception fails during deep hole insertion. After the long axis is inserted into the hole, the binocular vision is obstructed and the depth of field is insufficient, causing the target measurement signal-to-noise ratio to drop sharply. The existing solution is forced to switch to blind insertion mode, which cannot meet the precision matching requirements of aero-engines.

[0005] Therefore, in view of the technical problems existing in the prior art, there is an urgent need to develop an assembly method and device based on vision and force perception adaptive long axis precision alignment, which can realize closed-loop correction of dynamic posture and real-time compensation of cantilever deformation during the insertion of long axis, so as to meet the high-precision and zero-damage assembly requirements of long axis parts of aero-engines. Summary of the Invention

[0006] The purpose of this invention is to provide an assembly method and apparatus for long-axis precision alignment based on visual and force-based adaptive alignment, so as to solve the problems raised in the prior art.

[0007] To achieve the above objectives, the present invention provides an assembly device for long-axis precision alignment based on vision and force adaptive design, comprising a linear guide rail and a linear module. An actuator and a fixed base are slidably mounted on both ends of the linear guide rail via the linear module. A clamping mechanism and a force sensor are mounted on the actuator. A shaft-end part and a hole-end part are mounted on the clamping mechanism and the fixed base, respectively. A first target and a second target are coaxially mounted on the shaft-end part and the hole-end part, respectively. A binocular camera is provided laterally on the linear guide rail for positioning the first and second targets.

[0008] The linear guide rail has a structure in which the actuator and the fixed seat are slidably mounted on both ends of the linear module, which realizes the bidirectional active centering capability of the shaft end part and the hole end part. Compared with the single-sided motion scheme, it can shorten the motion stroke and improve the positioning accuracy. The first target and the second target are coaxially installed on the shaft end and the hole end respectively. The binocular camera is arranged laterally to simultaneously collect the pose of the two targets, realizing the direct measurement of the relative pose of the shaft and the hole, avoiding the cumulative error caused by separate measurement and coordinate transformation in the traditional scheme.

[0009] The first target is installed at the end of the shaft end part away from the end of the hole end part, and the second target is installed at the end of the hole end part away from the end of the shaft end part.

[0010] This layout allows the binocular camera to directly measure the six-dimensional deviation of the end of the long axis relative to the full-length outlet axis of the hole. It naturally includes the straightness and taper error of the hole along its entire length. Throughout the entire process of the long axis being inserted from the inlet to exiting or penetrating to the bottom of the blind hole, the side-mounted binocular camera can always simultaneously observe the moving first target and the stationary second target, eliminating the mutual occlusion problem found in the inlet layout.

[0011] The actuator is a six-degree-of-freedom parallel platform, and the force sensor is a six-dimensional force sensor.

[0012] Using a six-degree-of-freedom parallel platform as the actuator has advantages such as high rigidity, high precision, and fast dynamic response, making it suitable for high-precision alignment and assembly scenarios. The parallel platform and the linear module are connected in series to form a seven-degree-of-freedom redundant drive, realizing the decoupled control of attitude adjustment and axial feed. This avoids the coupling problem of attitude adjustment affecting insertion depth in traditional serial mechanisms. The use of a six-dimensional force sensor can simultaneously detect force and torque in three directions, providing complete contact force information for force-feed compliance control.

[0013] An assembly method for long-axis precision alignment based on vision and force perception includes the following steps: Step S1: Control the linear module to drive the shaft end part to approach the hole end part. The binocular camera collects the pose difference between the first target and the second target and calculates the pose deviation based on the elastic deformation. Adjust the pose by the pose deviation until the entry distance between the end face of the shaft end part and the end face of the hole end part reaches a preset threshold. The pose adjustment is controlled by the visual servo controller and the force compliance controller. Step S2: When the force sensor detects a sudden change in axial or lateral force, it generates a spiral search motion command and admittance compliance displacement according to the direction of the lateral force, drives the actuator to optimize in the horizontal plane, and makes the end face of the shaft end part slide into the chamfered conical surface of the hole end part. Step S3: After the lateral force continues to disappear for a preset time, the visual gain decreases with the change of insertion depth and lateral contact force. The visual servo controller adjusts the pose and provides basic damping based on the decreased visual gain. Step S4: As the insertion depth further increases, the visual gain decays to below the preset lower limit. The force compliance controller provides dynamic damping jointly tuned by the visual pose change rate and the force change rate, as well as admittance compliance stiffness modulated by the visual pose error, to perform closed-loop correction of the real-time pose. Step S5: When the force sensor detects that the axial force reaches the preset target value and the lateral force remains zero, the assembly is determined to be complete, and the clamping mechanism releases the workpiece.

[0014] This method divides the assembly process into five consecutive states, S1 to S5, with each state having a clear division of labor: S1 uses vision to achieve rapid, non-contact approach; S2 uses force sensing to guide safe passage through chamfering; S3 uses attenuated visual gain to achieve a smooth transition; S4 uses force sensing to dominate and complete precise alignment; and S5 uses force sensing to determine and complete assembly. This state machine design balances efficiency and safety, ensuring both the speed of free space and the safety of the contact process, and solving the problem that traditional single control modes cannot meet the needs of multiple stages.

[0015] In step S1, the theoretical end deflection and rotation angle are calculated based on the cantilever beam bending theory, and the elastic deformation component is subtracted from the pose difference detected by vision to obtain the pose deviation, and then the output control command is obtained. The calculation formula is as follows: ; in, To output control commands; For visual servo gain matrix; This is a pose deviation; The basic damping gain matrix; This represents the rate of change of visual pose error.

[0016] A cantilever beam deformation compensation algorithm is introduced in the free approach state to subtract the elastic deformation component from the total pose deviation detected by vision, thus obtaining the pure rigid pose deviation. This solution addresses the problem that the measured pose of the long axis due to its own weight is not equal to the actual rigid pose. Traditional solutions directly use the deformed pose measured by vision as the control input, which leads to incorrect correction direction and further deviation with each correction. This solution calculates and deducts the deflection and rotation angle theoretically, so that the control command is only for rigid deviation. This fundamentally solves the deformation interference problem in the cantilever state of the long axis and improves the initial alignment accuracy.

[0017] In step S2, when the force sensor detects that the rate of change of axial force or lateral force exceeds a preset abrupt change threshold, the visual servo controller switches to a silent state, recording only pose data. The force compliance controller generates control commands based on the lateral and axial forces fed back by the force sensor. The calculation formula is as follows: ; in, To output control commands to the pose adjustment mechanism; This is the Jacobian transpose matrix; This is the lateral force compliance coefficient; , The horizontal and vertical lateral forces are measured by the force sensor. This is the torque compliance coefficient; The torque about the X-axis; The virtual stiffness matrix is ​​modulated by visual error; For admittance control, the compliant displacement output; The control command The linear drive module and actuator make a small spiral motion in the YZ plane and slide in along the chamfered conical surface. At the same time, they rotate around the X-axis to find the optimal position. The pitch and yaw degrees of freedom are passively guided by the geometry of the orifice chamfer.

[0018] In the chamfered contact state, the vision servo switches to a silent state, only recording data. It is completely dominated by the force-feedback compliant controller, which uses a spiral search strategy to make a small spiral motion in the YZ plane, while rotating around the X-axis to optimize, so that the axis end naturally slides into the chamfered cone surface. The pitch and yaw degrees of freedom are passively guided by the chamfer geometry. This design avoids unreliable feedback from vision due to occlusion or reflection during the contact phase. At the same time, it uses the chamfer's own guiding characteristics to achieve passive centering, reducing the complexity of active control and the risk of oscillation.

[0019] In steps S3 and S4, visual servoing and force compliance are nonlinearly fused through state-time-based dominance and bidirectional parameter coupling to generate control commands. The calculation formula is as follows: ; in, For outputting control commands to the actuator; The visual servo gain matrix varies with contact force. Dual attenuation with insertion depth d; This is a pose deviation; The dynamic damping matrix is ​​composed of the superposition of the basic damping term, the visual motion velocity modulation term, and the contact force change rate modulation term. The current speed of the end effector; The virtual stiffness matrix increases quadratically with visual pose error. The compliant displacement output is for admittance control.

[0020] By employing a nonlinear fusion control of vision and force, the visual gain decreases exponentially with insertion depth and lateral force, the dynamic damping is jointly tuned by the visual velocity and the rate of change of force, and the virtual stiffness increases quadratically with visual error. This design achieves seamless switching between visual servoing and force compliance, the dynamic damping suppresses mechanical oscillations, and the secondary stiffness modulation prevents visual errors from introducing excessive correction forces. This fusion strategy solves the problem of precise alignment after vision fails in the hole.

[0021] In step S5, when the force sensor detects that the axial force reaches the preset target bearing force and the lateral force remains zero, a control command is generated. The calculation formula is as follows: ; in, This is the axial control command output to the seventh-axis linear module; Force-controlled proportional gain; The target axial bearing force is preset; The axial force measured by the force sensor; The direction vector indicates that the control command only applies to the insertion direction X-axis; When axial force Achieving target sitting force When the lateral force remains zero, it is determined that the end of the long shaft has entered the bottom of the hole or exited the end of the hole, and the clamping mechanism releases the workpiece.

[0022] By setting a target sitting force When the measured axial force reaches the target value and the lateral force remains zero, it is determined that the shaft end has reached the bottom or passed through smoothly. The condition that the lateral force remains zero ensures that the assembly posture is correct and avoids misjudgment caused by jamming or tilting. The judgment logic is clear and robust, and can reliably distinguish between normal contact and abnormal jamming, providing sufficient safety basis for the release of the clamping mechanism.

[0023] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention uses a binocular camera to collect the six-dimensional pose of the target at the shaft end and the hole end in real time. Combined with cantilever beam deformation compensation to deduct the elastic deformation component, it achieves high-precision visual servo approximation in the free approach state. In the cylindrical mating state, it maintains precise alignment even if the visual part fails. Compared with the existing intermittent visual solution, it avoids the problem of pre-alignment but failure to insert caused by accumulated deviation. 2. This invention adopts a visual and force-sensory adaptive fusion strategy: in free space, it relies on visual rapid approach, and in the contact critical zone, it relies on force-sensory compliant guidance. Through spiral search motion and admittance compliance control, the shaft end slides naturally into the chamfered cone surface. At the same time, it achieves predictive defense before contact through dynamic force threshold monitoring. Compared with the blind search of pure force control and post-collision detection, it takes into account both efficiency and safety, and avoids microscopic damage to the core components of the aero-engine. 3. In the cylindrical mating state, the visual servo gain of this invention decreases exponentially with the insertion depth and lateral contact force. At the same time, the dynamic damping is jointly tuned by the visual pose change rate and the force change rate. The virtual stiffness increases twice with the visual pose error. When the target is blocked by the hole wall, causing the measurement signal-to-noise ratio to drop sharply, the force compliance controller smoothly takes over the control and continues to provide high-precision pose correction. Compared with the existing solution, which completely switches to the blind insertion mode with deteriorated accuracy after visual failure, this invention achieves a seamless transition from vision to force, always maintaining the H7 / g6 level precision mating requirements. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the assembly device for long-axis precision alignment based on vision and force perception according to the present invention. Figure 2 This is a flowchart of an assembly method for long-axis precision alignment based on vision and force perception, according to the present invention.

[0025] The following are the labels in the diagram: 1. Linear guide rail; 2. Linear module; 3. Actuator; 4. Fixing base; 5. Shaft end part; 6. Hole end part; 7. First target; 8. Second target; 9. Force sensor; 10. Clamping mechanism; 11. Binocular camera. Detailed Implementation

[0026] 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.

[0027] Example 1: This example provides an assembly device for long shaft precision alignment based on vision and force, which can realize closed-loop correction of dynamic posture and real-time compensation of cantilever deformation during the insertion of long shaft, and is suitable for high-precision assembly of long shaft parts of aero-engines.

[0028] like Figure 1 As shown, the device includes a linear guide rail 1, a linear module 2, an actuator 3, a fixed base 4, a clamping mechanism 10, a force sensor 9, and a binocular camera 11.

[0029] Specifically, the linear guide rail 1 is fixedly installed on the workbench or foundation in the horizontal direction, and the two ends of the linear guide rail 1 are respectively slidably installed with the actuator 3 and the fixed seat 4 through the linear module 2.

[0030] The linear module 2 drives the actuator 3 and the fixed seat 4 to slide along the linear guide rail 1. The fixed seat 4 is used to fix the hole end part 6, and the actuator 3 is used to drive the shaft end part 5 to achieve six-degree-of-freedom pose adjustment.

[0031] The actuator 3 is a six-degree-of-freedom parallel platform, and its moving platform is connected in series with the linear module 2 to form a seven-degree-of-freedom redundant drive structure.

[0032] The six-degree-of-freedom parallel platform includes an upper platform, a lower platform, and six electrically operated telescopic chains connecting the upper and lower platforms. Each electrically operated telescopic chain has a built-in position encoder and drive motor.

[0033] The six-degree-of-freedom parallel platform can achieve precise motion in six degrees of freedom, including translation along the X, Y, and Z axes and rotation around the X, Y, and Z axes. Its positioning accuracy is better than ±0.01mm and its angular accuracy is better than ±0.01°. The linear module 2 is installed on the moving platform of the six-degree-of-freedom parallel platform and is used to drive the clamping mechanism 10 to move linearly along the X-axis.

[0034] The clamping mechanism 10 is a self-centering gripper, installed at the output end of the linear module 2, and is used to clamp long shaft parts.

[0035] The self-centering gripper is pneumatically or electrically driven, with adjustable clamping force. It can automatically center shaft end parts of different diameters, ensuring that the axis of the shaft end part 5 coincides with the center line of the clamping mechanism 10.

[0036] The force sensor 9 is a six-dimensional force sensor, installed between the six-degree-of-freedom parallel platform and the clamping mechanism 10, specifically, between the output end of the linear module 2 and the self-centering gripper.

[0037] The six-dimensional force sensor can detect force components Fx, Fy, and Fz in three orthogonal directions and torque components Mx, My, and Mz in three orthogonal directions in real time, with a sampling frequency of no less than 1000Hz and a measurement accuracy better than ±0.1N.

[0038] The shaft end part and the hole end part 6 are respectively provided with a first target 7 and a second target 8. The first target 7 is fixed to the end face of the shaft end part away from the hole end part 6, that is, the end face of the shaft end part 5, and is fixed to the end face of the hole end part 6 away from the shaft end part 5, that is, the outlet face of the hole end part 6.

[0039] The geometric centers of the first target 7 and the second target 8 are both located on the axis of the insertion direction. The two targets are coaxially installed and their normals are parallel to the insertion direction, which facilitates high-precision pose calculation by the binocular camera 11.

[0040] The binocular camera 11 is mounted on the side of the linear guide rail 1, and its optical axis is at a certain angle to the insertion direction. This allows the binocular camera 11 to simultaneously observe the moving first target 7 and the stationary second target 8 throughout the entire process of the shaft end part being inserted from the inlet to exiting or penetrating to the bottom of the blind hole, thus avoiding the problem of the targets blocking each other.

[0041] The binocular camera 11 has a frame rate of no less than 60fps, a resolution of no less than 5 megapixels, and a pose measurement accuracy better than ±0.02mm.

[0042] Example 2: This example describes the overall process of precisely aligning and assembling the shaft end part 5 and the hole end part 6 using the above-mentioned assembly device.

[0043] like Figure 2 As shown, the assembly method in this embodiment is divided into five consecutive stages: workpiece loading and calibration, free approach state, chamfer contact state, cylindrical mating state, bottom surface contact state, and completion determination. The following is a detailed description using the assembly of a long shaft of a certain type of aero-engine as an example.

[0044] Workpiece loading and calibration: The operator or automatic loading robot places the shaft end part 5 on the clamping mechanism 10. The clamping mechanism 10 clamps itself with 0.4MPa air pressure to fix the hole end part 6 on the fixed base 4. The first target 7 and the second target 8 are installed in the predetermined positions to ensure that the two targets are coaxial and their normal is parallel to the X-axis.

[0045] The linear module 2 drives the actuator 3 to move slowly along the positive X direction, so that the end face of the shaft end part 5 is about 50mm away from the inlet of the hole end part 6. The binocular camera 11 begins to acquire images of the first target 7 and the second target 8 in real time. After image processing, the six-dimensional pose difference between the two targets is calculated. ; At this time, since the shaft end part 5 is in a cantilever state, its own weight will cause elastic sag, and the sag deformation compensation algorithm is executed.

[0046] Specifically, the overhang length from the clamping point to the end face of the first target 7 is Loverhang = 800 mm, the self-weight of the shaft end part is W = 25 × 9.8 = 245 N, the material elastic modulus is E = 200 GPa, and the moment of inertia of the section is I = π × (40 4 ) / 64 = 2.01 × 10 5 mm 4 Calculate the end deflection δmodel and rotation angle θdeflection according to the cantilever beam bending theory: δmodel=(W×Loverhang 3 ) / (3E×I)=(245×800 3 ) / (3×200×10 3 ×2.01×10 5 )≈0.104mm; θdeflection=(W×Loverhang 2 ) / (2E×I)=(245×800 2 ) / (2×200×10 3 ×2.01×10 5 )≈0.000195rad≈0.011°; The controller subtracts the elastic deformation component from the total pose difference ΔPmeasured measured visually, obtaining the pure rigid pose deviation ΔPrigid = ΔPmeasured - [δmodel, θdeflection] T .

[0047] For example, if visual measurement shows: ΔPmeasured=[0.5mm,0.3mm,0.1mm,0.02°,0.03°,0.01°] T ; After deducting deformation: ΔPrigid=[0.396mm,0.3mm,0.1mm,0.009°,0.019°,0.01°] T .

[0048] Free approach state: The linear control module 2 drives the shaft end part 5 to rapidly approach the hole end part 6 at a speed of v0=15mm / s, while the binocular camera 11 continuously acquires the target pose at a frame rate of 60fps and calculates ΔPrigid in real time.

[0049] At this stage, the shaft end part 5 and the hole end part 6 are not in contact, and the contact force Fext measured by the six-dimensional force sensor is approximately 0. Therefore, the vision servo controller is completely dominant, and the control law degenerates into position proportional and velocity damping PD control, outputting control commands. : ; Where Kp(Fext) is the visual servoing gain matrix; Since Fext≈0, Kp=diag(120,120,80,25,25,15), the unit is N / μm, which is pre-tuned based on the parallel platform servo bandwidth of 20Hz and load inertia of 25kg; ΔPrigid is a purely rigid pose deviation vector; Bd is the basic damping gain matrix, Bd=diag(0.8,0.8,0.6,0.3,0.3,0.2), used to suppress mechanical overshoot oscillations caused by the proportional term; The visual pose error change rate is obtained by dividing the pose difference between two consecutive frames by the sampling time of 16.67 ms.

[0050] Suppose at a certain moment; ΔPrigid=[0.4mm,0.25mm,0.08mm,0.01°,0.015°,0.008°] T , , Therefore, τS1≈[48,30,6.4,2.5,3.75,1.2] is calculated. T According to the instruction, the actuator 3 adjusts its posture synchronously in six degrees of freedom, so that the axis of the shaft end part 5 gradually approaches the axis of the hole end part 6.

[0051] During this process, the force control loop is disconnected, and the admittance controller becomes rigid to ensure rapid approach. When the distance between the end face of the shaft end part 5 and the entrance of the hole end part 6 reaches the preset threshold d0=2mm, the vision servo controller detects a displacement Δz=48mm, which takes about 3.2 seconds. The S1 stage ends, and the state machine is ready to switch to S2.

[0052] Chamfer contact state: When the end face of the shaft end part 5 enters within 2mm of the hole opening, the shaft end may touch the chamfer of the hole opening. The chamfer size is C1.5 and the angle is 45°. The six-dimensional force sensor begins to detect the sudden change of axial force Fx or lateral force Fy, Fz.

[0053] In this embodiment, when the rate of change of axial force Fx exceeds a preset threshold of 0.5 N / ms or the lateral force modulus value ‖Fyz‖ exceeds 1 N, the state switches to the chamfered contact state.

[0054] For example, the sensor's measured data are: horizontal to the right Fy=2.3N, vertical downward Fz=-1.8N, and the torque around the axis Mx=0.15N·m.

[0055] Upon entering S2, the visual servo controller switches to a silent state, recording pose data at a low power consumption with a frequency of 10Hz, and no longer actively participates in control. The force compliance controller takes over the control and generates spiral search motion commands based on force feedback.

[0056] The control command τS2 generated by the force-feed compliance controller is: Among them: J T The Jacobian transpose matrix of the parallel platform maps forces / torques to joint driving forces; kf is the lateral force compliance coefficient, taken as 0.15 mm / N; kγ is the torque compliance coefficient, which is taken as 0.8 rad / (N·m); Fy=2.3N, Fz=-1.8N, Mx=0.15N·m; Kd(ΔP) is the virtual stiffness matrix of the visual error modulation; The initial value is set to diag(0,0.5,0.5,0,0.3,0.3), in N / mm.

[0057] Xf is the compliance displacement output of admittance control, derived from the admittance control equation: Real-time solution; Where M = diag(0,10,10,0,5,5) kg or kg·m², D = diag(0,50,50,0,20,20) N·s / m, and K = diag(0,2,2,0,1,1) N / μm.

[0058] Substituting into the calculation, we get: kf×Fy=0.15×2.3≈0.345mm; kf×Fz=0.15×(-1.8)=-0.27mm; kγ×Mx=0.8×0.15=0.12rad≈6.9°.

[0059] τS2 drives actuator 3 to perform a small-amplitude spiral motion in the YZ plane. The spiral motion parameters are: amplitude 0.5mm, frequency 0.5Hz, pitch 0.2mm / turn. At the same time, it rotates 6.9° around the X-axis in the positive direction of Mx, so that the shaft end face slides into the chamfered cone surface.

[0060] During the helical search, the pitch and yaw degrees of freedom are passively guided by the orifice chamfer geometry, and the controller does not actively intervene.

[0061] For example, when the end face of the shaft contacts the lower edge of the chamfer, the chamfered cone surface generates an upward component force, guiding the end face of the shaft to the center. This process takes about 0.8 to 1.5 seconds until the lateral force gradually disappears.

[0062] When the six-dimensional force sensor detects a lateral force modulus value ‖Fyz‖ < 0.2N for a duration of 200ms, it indicates that the shaft end has successfully passed through the chamfer and entered the straight hole inlet, and the state machine switches from S2 to S3.

[0063] Cylindrical fit state: After entering the cylindrical fit state, the shaft end part 5 has entered the straight hole section, but it is not yet fully aligned.

[0064] At this time, as the shaft end extends into the hole, the relative angle between the target and the camera deteriorates, and the reflection interference from the hole wall increases, causing the measurement signal-to-noise ratio of the binocular camera 11 to begin to decrease. Therefore, the controller adopts an exponential decay strategy to dynamically adjust the visual servo gain.

[0065] The formula for calculating the visual servoing gain matrix Kp(Fext,d) is as follows: Where: Kp0 is the reference gain matrix, the same as in stage S1: diag(120,120,80,25,25,15); γ is the force attenuation coefficient, taken as 0.3 N⁻¹; ‖Fyz‖ is the lateral force modulus, currently measured to be 0.15N; μ is the depth attenuation coefficient, taken as 0.005mm⁻¹; d is the current insertion depth, which is calculated from the X-direction displacement of the first target 7 from S2 to the current position. Assume that the current d = 15 mm. diag(0,1,1,0,1,1) is the degree-of-freedom selection matrix, removing the X translation and Rx rotation degrees of freedom.

[0066] Substituting the values ​​into the equation: exp(-γ×‖Fyz‖)=exp(-0.3×0.15)≈0.956; exp(-μ×d)=exp(-0.005×15)≈0.928.

[0067] Therefore, Kp≈Kp0×0.956×0.928≈Kp0×0.887, which means that the visual gain has decreased by about 11.3% when inserted at 15mm.

[0068] At this point, visual and force sensing are combined for control, and the control command τS3 is: ; The parameters are as follows: ΔPrigid: Current pure rigid pose deviation. Assume ΔPrigid = [0mm, 0.08mm, -0.05mm, 0°, 0.02°, 0.01°] T .

[0069] : Dynamic damping matrix, composed of basic damping terms: Bd0 = diag(0.8, 0.8, 0.6, 0.3, 0.3, 0.2) is formed by adding a visual motion speed modulation term with a coefficient α = 0.05 and a force change rate modulation term with a coefficient β = 0.02.

[0070] Suppose the current =[0mm / s,0.2mm / s,-0.1mm / s,0° / s,0.05° / s,0.03° / s] T ; =[0N / s,0.1N / s,-0.05N / s,0.02N×m / s,0,0] T ; but .

[0071] The current speed of the end effector is obtained by encoder differential. =[5mm / s,0.1mm / s,-0.05mm / s,0.01° / s,0.02° / s,0.01° / s] T .

[0072] Kd(ΔP): Virtual stiffness matrix, which increases quadratically with visual pose error. Let the reference stiffness be Kd0=diag(0,1,1,0,0.5,0.5), and the translational error modulus of ΔPrigid be 0.094mm and the rotational error modulus be 0.022°. Then Kd=Kd0×(1+10×‖ΔP‖²)≈Kd0×1.09.

[0073] Xf: Admittance compliance displacement output, calculated in real time by the admittance controller; Currently, Xf ≈ [0mm, 0.02mm, -0.01mm, 0°, 0.005°, 0.003°] T .

[0074] Substituting the above data into the τS3 formula, the control command is calculated. The actuator 3 adjusts its position and posture in real time according to the command, while the linear module 2 advances slowly at a speed of v=3mm / s.

[0075] During this stage, the visual gain continues to decrease exponentially with increasing insertion depth: when d=30mm, exp(-0.005×30)=0.861, and it decreases to 82.3% of Kp0; when d=50mm, it decreases to 76.5% of Kp0.

[0076] Cylindrical mating state: When the insertion depth d is further increased to 80mm, the visual gain has decayed to below the preset lower limit, i.e. Kp<0.2·Kp0. At this time, due to the target being blocked by the hole wall and insufficient depth of field, the measurement signal-to-noise ratio of the binocular camera 11 drops sharply, and the pose measurement error increases to more than ±0.1mm. The controller automatically switches to the force-dominated mode.

[0077] In this mode, the visual servo gain matrix Kp(Fext,d) is clamped to the lower limit Kpmin=0.15×Kp0, and vision only provides low-frequency, low-gain pose reference. The force compliance controller takes over and provides high dynamic damping and admittance compliance stiffness.

[0078] The control commands still follow the general formula τS3, but the parameter weights have changed significantly: The coefficient β of the force feedback rate modulation term in the dynamic damping matrix Bd is increased from 0.02 to 0.15, which significantly increases the contribution of force feedback to damping.

[0079] The quadratic growth coefficient in the virtual stiffness matrix Kd(ΔP) is increased from 10 to 30, making visual errors more sensitive to stiffness modulation and avoiding excessive correction forces introduced by incorrect poses.

[0080] Assuming the insertion depth is d = 85 mm, the lateral force Fyz = [0.05 N, -0.03 N]. T The axial force Fx = 8N is caused by slight interference from the hole wall. Visual inspection: ΔPrigid=[0mm,0.12mm,-0.08mm,0°,0.03°,0.02°] T .

[0081] The force-feedback compliant controller generates an Xf displacement output based on force feedback, approximately [0mm, 0.05mm, -0.02mm, 0°, 0.01°, 0.005°]. T Prioritize the guidance of compliance.

[0082] Meanwhile, the speed of linear module 2 is reduced to v=1mm / s to ensure smooth advancement within narrow gaps.

[0083] During this stage, actuator 3 continuously provides pose closed-loop correction. When the axis of the shaft end coincides with the axis of the hole, the lateral force Fyz gradually approaches zero, and the axial force Fx stabilizes at around 5N. When the insertion depth d reaches 100mm, the state machine prepares to enter S5.

[0084] Bottom surface contact state and completion determination: When the shaft end part 5 approaches the bottom of the hole, the force sensor 9 detects that the axial force Fx rises rapidly.

[0085] In this embodiment, the bottom of the hole is a planar structure. When the shaft end contacts the bottom surface, the axial force will change abruptly. The controller monitors the axial force Fz and the lateral force Fyz in real time. When the axial force Fz reaches the preset target sitting force Ftarget=50N, and the lateral force Fyz remains zero and exceeds the stable time threshold of 200ms, the assembly is determined to be complete.

[0086] At this point, the controller outputs the final control command τS4 to linear module 2: ; Where Kf is the force control proportional gain, taken as 0.05mm / N; Ftarget = 50N; Fz is the measured axial force, assumed to be 48.5N; =[0,0,1,0,0,0] T , direction vector, indicating that it only acts on the X-axis.

[0087] The calculated value of τS4 is 0.05×(50-48.5)=0.075mm. That is, the linear module 2 continues to feed micro-feed by 0.075mm. When Fz reaches 50.2N and stabilizes for 50ms, it is determined that the shaft end has been firmly seated at the bottom of the hole.

[0088] Subsequently, the controller issues a command: the linear module 2 stops feeding and maintains its current position; the clamping mechanism 10 releases air, the grippers open, releasing the shaft end part 5; the binocular camera 11 stops acquiring data; and the force sensor 9 is reset to standby mode.

[0089] Special scene handling: Dynamic compensation for cantilever deformation: During the workpiece loading and calibration stage, the overhang length changes with the clamping point. If a shaft of different length is replaced, the controller automatically reads the parameters such as the overhang and moment of inertia I of the new shaft, and recalculates the theoretical deflection δmodel and rotation angle θdeflection to achieve one-click compensation.

[0090] Abnormal contact retraction: In the chamfered contact state or cylindrical mating state, if the lateral force Fyz suddenly increases to more than 9N, for example, if it encounters a foreign object in the hole, switch back to the free approach state and execute the original path retraction command. The linear module 2 retracts 50mm in the reverse direction and re-performs the spiral search optimization to avoid damaging the workpiece.

[0091] Visual failure emergency handling: If the binocular camera 11 loses the target for more than 500ms due to reflection, obstruction or other reasons, the controller will automatically set the visual gain Kp to zero and switch to force compliance mode. It will then rely on admittance control and dynamic damping to complete the remaining insertion process, ensuring that the assembly is not interrupted.

[0092] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. An assembly method for adaptive long-axis precision alignment based on vision and force perception, characterized in that, An assembly device based on vision and force adaptive long axis precision alignment is applied. The assembly device includes a linear guide rail (1) and a linear module (2). At both ends of the linear guide rail (1), an actuator (3) and a fixed seat (4) are slidably mounted through the linear module (2). A clamping mechanism (10) and a force sensor (9) are mounted on the actuator (3). A shaft end part (5) and a hole end part (6) are mounted on the clamping mechanism (10) and the fixed seat (4), respectively. A first target (7) and a second target (8) are coaxially mounted on the shaft end part (5) and the hole end part (6), respectively. A binocular camera (11) is provided on the side of the linear guide rail (1) for positioning the first target (7) and the second target (8). The assembly method includes the following steps: Step S1: Control the linear module (2) to drive the shaft end part (5) to approach the hole end part (6). The binocular camera (11) collects the pose difference between the first target (7) and the second target (8) and calculates the pose deviation based on the elastic deformation. The pose is adjusted by the pose deviation until the entry distance between the end face of the shaft end part (5) and the end face of the hole end part (6) reaches the preset threshold. The pose adjustment is controlled by the visual servo controller and the force compliance controller. Step S2: When the force sensor (9) detects a sudden change in axial or lateral force, it generates a spiral search motion command and admittance compliance displacement according to the direction of the lateral force, driving the actuator (3) to optimize in the horizontal plane, so that the end face of the shaft end part (5) slides into the chamfered conical surface of the hole end part (6); wherein, when the force sensor (9) detects that the rate of change of axial or lateral force exceeds the preset sudden change threshold, the visual servo controller switches to a silent state and only records the pose data, and the force compliance controller generates control commands according to the lateral and axial forces fed back by the force sensor (9). The calculation formula is as follows: ； in, To output control commands to the pose adjustment mechanism; This is the Jacobian transpose matrix; This is the lateral force compliance coefficient; , The horizontal and vertical lateral forces measured by the force sensor (9); This is the torque compliance coefficient; The torque about the X-axis; The virtual stiffness matrix is ​​modulated by visual error; For admittance control, the compliant displacement output; The control command The drive linear module (2) and the actuator (3) make a small spiral motion in the YZ plane and slide in along the chamfered cone surface. At the same time, they rotate around the X axis to find the optimal position. The pitch and yaw degrees of freedom are passively guided by the chamfer geometry of the orifice. Step S3: After the lateral force continues to disappear for a preset time, the visual gain decreases with the change of insertion depth and lateral contact force. The visual servo controller adjusts the pose and provides basic damping based on the decreased visual gain. Step S4: As the insertion depth further increases, the visual gain decays to below the preset lower limit. The force compliance controller provides dynamic damping jointly tuned by the visual pose change rate and the force change rate, as well as admittance compliance stiffness modulated by the visual pose error, to perform closed-loop correction of the real-time pose. Step S5: When the force sensor (9) detects that the axial force reaches the preset target value and the lateral force remains zero, the assembly is determined to be complete, and the clamping mechanism (10) releases the workpiece.

2. The assembly method for adaptive long-axis precision alignment based on vision and force perception according to claim 1, characterized in that, The first target (7) is installed at the end of the shaft end part (5) away from the hole end part (6), and the second target (8) is installed at the end of the hole end part (6) away from the shaft end part (5).

3. The assembly method for adaptive long-axis precision alignment based on vision and force perception according to claim 1, characterized in that, The actuator (3) is a six-degree-of-freedom parallel platform, and the force sensor (9) is a six-dimensional force sensor.

4. In the assembly method for long-axis precision alignment based on vision and force perception according to claim 1, in step S1, the theoretical end deflection and rotation angle are calculated according to the cantilever beam bending theory to obtain the elastic deformation component, and the elastic deformation component is subtracted from the pose difference detected by vision to obtain the pose deviation, and thus the output control command is obtained. The calculation formula is as follows: ； in, To output control commands; For visual servo gain matrix; This is a pose deviation; The basic damping gain matrix; This represents the rate of change of visual pose error.

5. The assembly method for adaptive long-axis precision alignment based on vision and force perception according to claim 1, characterized in that: In steps S3 and S4, visual servoing and force compliance are nonlinearly fused through state-time-based dominance and bidirectional parameter coupling to generate control commands. The calculation formula is as follows: ； in, For outputting control commands to the actuator (3); The visual servo gain matrix varies with contact force. Dual attenuation with insertion depth d; This is a pose deviation; The dynamic damping matrix is ​​composed of the superposition of the basic damping term, the visual motion velocity modulation term, and the contact force change rate modulation term. The current speed of the end effector; The virtual stiffness matrix increases quadratically with visual pose error. The compliant displacement output is for admittance control.

6. The assembly method for adaptive long-axis precision alignment based on vision and force perception according to claim 1, characterized in that: In step S5, when the force sensor (9) detects that the axial force reaches the preset target sitting force and the lateral force remains zero, a control command is generated. The calculation formula is as follows: ； in, For outputting axial control commands to the seventh-axis linear module (2); Force-controlled proportional gain; The target axial bearing force is preset; The axial force measured by the force sensor (9); The direction vector indicates that the control command only acts on the insertion direction X-axis; when the axial force... Achieving target sitting force When the lateral force remains zero, it is determined that the end of the long shaft enters the bottom of the hole or exits the end of the hole, and the clamping mechanism (10) releases the workpiece.

Images