Force-vision based intelligent anti-interference control method for robot assembly of automobile terminal

Abstract

The application relates to the technical field of intelligent control, and discloses an intelligent anti-interference control method for robot assembly of a vehicle terminal based on force-vision. The method comprises the following steps: dividing an assembly contact phase and generating a continuous phase indication value according to a normal contact force, a normal contact force change rate and a residual distance of a visual target; generating a phase perception compensation gain through mapping of the continuous phase indication value; outputting an assembly disturbance estimation quantity through a neural network disturbance observer from joint state quantity and the normal contact force; and superimposing to generate an assembly anti-interference control torque through first layer compensation of the assembly disturbance estimation quantity and second layer compensation of the phase perception compensation gain adjustment approximation residual error. The application solves the problem that neural network approximation residual compensation gain cannot be adaptively adjusted according to the assembly contact phase, resulting in that fixed gain cannot be well adjusted at different assembly stages.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and in particular to an intelligent anti-interference control method for assembling automotive terminal robots based on force-vision. Background Technology

[0002] In automotive final assembly and sub-assembly stations, industrial robots performing precision assembly tasks need to simultaneously handle multi-source coupled disturbances such as workpiece dimensional tolerances, conveyor line vibrations, sudden changes in surface friction of parts, and the flexibility of the robot's own joints. Existing technologies, force-vision fusion-based robot assembly control methods typically employ impedance control or hybrid force / position control strategies, combining a vision system for coarse workpiece positioning, force sensors for contact force feedback, and introducing a disturbance observer to estimate and compensate for unknown disturbances. Some solutions further employ neural networks to replace traditional linear disturbance observers to approximate nonlinear disturbances.

[0003] Traditional linear disturbance observers, based on the assumption of a single linear model, suffer from severe model mismatch when dealing with disturbance signals exhibiting strong nonlinear characteristics due to multi-source coupling in automotive final assembly, resulting in insufficient disturbance estimation accuracy. Although introducing neural network disturbance observers can improve the approximation capability of nonlinear disturbances, the limited number of nodes and the lag in online weight updates of the neural network itself inevitably result in an approximation residual between its output and the actual disturbance. Existing technologies uniformly use fixed gain to compensate for this approximation residual, without distinguishing the differences in residuals at different physical stages of the assembly process.

[0004] However, in the actual process of automotive final assembly, the amplitude characteristics of the residual approximation by the neural network exhibit significant stage differences in different physical stages: from the free approach of the workpiece to the moment of impact with the target surface, and then to steady-state insertion. The residual is small and stable in the free approach stage. At the moment of initial contact, due to the sudden change in friction force and the unmodeled dynamics triggered by the impact, the residual exhibits a transient peak several times the steady-state value. In the steady-state insertion stage, the residual falls back to a moderate level. Fixed gain compensation cannot simultaneously meet the different needs of these three stages: if the gain is set according to the residual peak at the moment of initial contact, overcompensation will occur in the free approach and steady-state insertion stages, leading to control signal jitter and high-frequency oscillation of the end effector. If the gain is set according to the residual mean value in the steady-state stage, insufficient compensation will occur at the moment of initial contact, causing the transient impact force to exceed the safety threshold of the part and resulting in surface damage. Therefore, the existing technology lacks a mechanism that can automatically adjust the residual compensation intensity according to the physical stage of assembly contact, and also lacks a method to deeply integrate force and visual signals at the control gain level to identify the assembly contact phase online. Summary of the Invention

[0005] This application provides an intelligent anti-interference control method for automotive terminal robot assembly based on force-vision. It solves the problem that the residual compensation gain of neural network approximation cannot be adaptively adjusted with the assembly contact phase, resulting in the fixed gain being inconsistent in different assembly stages. Through the contact phase driven adaptive residual compensation mechanism of force-vision fusion perception, the peak value of transient collision force at the moment of initial contact during assembly is controlled within the safety threshold of the part. At the same time, it eliminates control chattering caused by overcompensation during free approach and steady-state insertion stages.

[0006] This application provides a force-vision-based intelligent anti-interference control method for automotive terminal robot assembly, the force-vision-based intelligent anti-interference control method for automotive terminal robot assembly includes:

[0007] Step S1: Based on the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target, the assembly process is divided online into the free approach phase, the initial contact phase, and the steady-state insertion phase, and continuous phase indication values ​​are generated.

[0008] Step S2: Based on the continuous phase indication value, generate phase-sensing compensation gain according to the mapping relationship between the maximum value corresponding to the initial contact phase, the minimum value corresponding to the free approach phase, and the intermediate value corresponding to the steady-state insertion phase.

[0009] Step S3: Input the joint state quantities of the assembled workpiece and the normal contact force into the neural network disturbance observer, and output the assembly disturbance estimate. The deviation between the assembly disturbance estimate and the actual disturbance is the approximation residual.

[0010] Step S4: Using the assembly disturbance estimate as the first layer of feedforward compensation and the phase-sensing compensation gain to adjust the compensation strength of the approximation residual as the second layer of compensation, the two layers of compensation are superimposed to generate the assembly anti-interference control torque.

[0011] The technical solution provided in this application fuses three heterogeneous sensing signals—the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target—to divide the assembly process online into a free approach phase, an initial contact phase, and a steady-state insertion phase, generating a continuous phase indication value. The beneficial effect of this technical feature is that it overcomes the deficiency in existing technologies where relying solely on a single sensing mode cannot accurately determine the physical stage of the assembly. The normal contact force reflects whether physical contact has been established between the workpiece and the target and the magnitude of the contact force; the rate of change of the normal contact force reflects whether the contact process is in a state of violent collision or impact or has stabilized; and the remaining distance of the visual target reflects the spatial distance between the workpiece and the target assembly surface. All three are indispensable to uniquely determine the physical phase of the assembly. The continuous phase indication value uses a smooth, continuous transition method instead of the traditional discrete hard-switching judgment, avoiding the disturbance to subsequent control links caused by repeated jumps in the phase judgment result near the threshold. This force-vision fusion phase perception mechanism achieves deep fusion of force and visual information at the control gain level, rather than just at the pose estimation level, providing a driving signal with a clear physical causal relationship for subsequent adaptive compensation gain adjustment. Phase-sensing compensation gain is generated based on the mapping relationship between the continuous phase indication value and the maximum value corresponding to the initial contact phase, the minimum value corresponding to the free approach phase, and the intermediate value corresponding to the steady-state insertion phase. The beneficial effect of this technology is that it quantitatively matches the compensation strength with the actual statistical characteristics of the approximation residual at each assembly stage. The residual peak is the largest at the initial contact phase, so the compensation gain is set to the maximum value to quickly suppress collision impact. The residual is the smallest at the free approach phase, so the compensation gain is set to the minimum value to avoid over-responding to sensor noise. The residual is in the middle at the steady-state insertion phase, so the compensation gain is set to the intermediate value to balance the compensation effect and control stability. This phase-sensing gain adjustment mechanism is different from the fixed gain or purely mathematically driven adaptive gain schemes in the prior art. Its gain change is driven by the stage characteristics of the physical process of assembly contact rather than by abstract mathematical variables, and has clear engineering interpretability.

[0012] The joint state variables and normal contact force of the assembled workpiece are input into a neural network disturbance observer to output an assembly disturbance estimate. The beneficial effect of this technology is that it uses the nonlinear approximation capability of the neural network to replace the linear assumption of the traditional linear disturbance observer, and performs online approximation of nonlinear disturbances coupled from multiple sources such as workpiece tolerance, conveyor line vibration, friction force abrupt change and joint flexibility in automotive final assembly. By using the normal contact force as one of the input components of the neural network, the network can sense changes in the assembly contact state and thus make a more accurate approximation response to the disturbance caused by the contact. Using the assembly disturbance estimate as the first-layer feedforward compensation and the phase-sensing compensation gain adjusting the approximation residual compensation strength as the second-layer compensation, the assembly anti-interference control torque is generated by superimposing the two layers of compensation. The beneficial effect of this technology is that it forms a well-defined two-layer anti-interference architecture. The first-layer feedforward compensation undertakes the coarse compensation task for the main part of the lumped disturbance, while the second-layer compensation is specifically for fine compensation of the approximation residual that the first layer cannot eliminate, according to the phase-matching strength. The superposition of the two layers ensures that the total control torque will not lose its ability to suppress collision impacts when the residual is at a peak due to only neural network compensation, nor will it generate unnecessary control chatter when the residual is stable due to only fixed gain switching compensation. In the application scenario of automotive terminal assembly, which has strict requirements for cycle time and part protection, this two-layer architecture enables the assembly anti-interference control torque to automatically adapt to the disturbance characteristics of different contact stages throughout the entire assembly cycle, without requiring engineers to manually adjust the compensation parameters for each model or workpiece, and has the flexible applicability to multi-variety mixed-line production. Attached Figure Description

[0013] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of one embodiment of the intelligent anti-interference control method for assembling automotive terminal robots based on force-vision in this application. Detailed Implementation

[0015] This application provides a force-vision based intelligent anti-interference control method for automotive terminal robot assembly. The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in a sequence other than that illustrated or described herein. Furthermore, the terms "comprising" or "having" and any variations thereof are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or device that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0016] For ease of understanding, the specific process of the embodiments of this application is described below. Please refer to [link / reference]. Figure 1 One embodiment of the force-vision-based intelligent anti-interference control method for automotive terminal robot assembly in this application includes:

[0017] Step S1: Based on the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target, the assembly process is divided online into the free approach phase, the initial contact phase, and the steady-state insertion phase, and continuous phase indication values ​​are generated.

[0018] Specifically, the normal contact force refers to the pure contact force value obtained after the Z-axis component of the six-dimensional force / torque sensor is compensated for by the gravity of the workpiece and the gripper. This sensor is installed between the robot's end flange and the workpiece gripper, with a sampling frequency of 1000Hz, a range of ±200N, and a resolution of 0.05N. The normal contact force change rate refers to the first-order difference result obtained by subtracting the normal contact force of two adjacent control cycles and then dividing by the control cycle of 0.001s, with the unit being N / s. The remaining distance of the visual target refers to the Euclidean distance between the current pose of the workpiece and the pre-taught target assembly pose, calculated by the industrial camera installed at the end of the robot arm acquiring grayscale images at a frequency of 30fps, extracting the coordinates of the workpiece feature points through Gaussian filtering, Canny edge detection, and template matching algorithms, and then converting them to the robot's base coordinate system through the camera intrinsic parameter matrix and the hand-eye transformation matrix, with the unit being mm. This distance value is interpolated from 30Hz to 1000Hz through a zero-order hold to match the force control loop frequency. The three phases are divided as follows: the free approach phase corresponds to the stage before the workpiece has contacted the target surface, at which time the normal contact force is lower than the force contact threshold of 3.0N and the remaining visual target distance is greater than the distance contact threshold of 6.0mm; the initial contact phase corresponds to the instant when the workpiece just contacts the target surface, at which time the remaining visual target distance is less than 6.0mm and the absolute value of the normal contact force change rate exceeds the force change rate impact threshold of 500N / s, which is determined based on the typical collision mechanics characteristics of metal-metal contact in automotive final assembly; the steady-state insertion phase corresponds to the stage where the workpiece has established stable contact and continues to be inserted, at which time the normal contact force has stabilized above 3.0N but the normal contact force change rate has fallen back to below 500N / s. The force contact threshold of 3.0N is set based on 60 times the noise level of the six-dimensional force sensor (0.05N) to ensure that sensor noise will not misjudge contact. The distance contact threshold of 6.0mm is set based on the distance of the coarse positioning trajectory termination point from the target surface to 5mm plus the maximum visual positioning error of 1mm. The force change rate impact threshold of 500N / s is set based on the average change rate when a desired contact force of 30N is established within a typical collision time window of 0.06s. The continuous phase indication value is a closed interval from 0.1 to 1.0, where 0.1 corresponds to the free approach phase, 1.0 corresponds to the initial contact phase, and 0.4 corresponds to the steady-state insertion phase. The three phases are smoothly transitioned through a hyperbolic tangent mapping rather than hard switching, avoiding sudden torque changes in the control signal due to discrete jumps in phase judgment.

[0019] Step S2: Based on the continuous phase indication value, generate phase sensing compensation gain according to the mapping relationship between the maximum value corresponding to the initial contact phase, the minimum value corresponding to the free approach phase, and the intermediate value corresponding to the steady-state insertion phase.

[0020] Specifically, the phase-sensing compensation gain refers to the real-time gain value obtained by performing a power-law nonlinear mapping based on the continuous phase indication value, followed by a first-order low-pass filter, with units of N·m. The specific calculation process for the power-law nonlinear mapping is as follows: Using the continuous phase indication value as the base and a shape parameter of 1.5 as the exponent, perform a power operation. Multiply the result by the difference between the upper gain limit of 10.0 N·m and the lower gain limit of 1.0 N·m (9.0 N·m), and add the lower gain limit of 1.0 N·m to obtain the original compensation gain. The lower gain limit of 1.0 N·m is set based on a safety margin of 2 times the root mean square value of the approximation residual of 0.5 N·m under the free approach phase; the upper gain limit of 10.0 N·m is set based on a safety margin of 1.25 times the peak value of the approximation residual of 8 N·m under the initial contact phase; the shape parameter of 1.5 makes the mapping curve convex upwards. When the continuous phase indication value is 0.4, i.e., the steady-state insertion phase, the calculated original compensation gain is... This value is 1.64 times the mean residual value of 2 N·m in the steady-state stage, covering residual fluctuations without introducing excessive chattering. The original compensation gain is output as a phase-sensing compensation gain after being filtered by a first-order low-pass filter. The filter coefficient is set to 0.05, corresponding to a cutoff frequency of approximately 8 Hz. This frequency is higher than the characteristic frequency of phase switching during assembly (below 2 Hz) to preserve the dynamic trend of gain change with phase, while being lower than the main frequency components of sensor noise (above 50 Hz) to filter out high-frequency gain spikes caused by noise.

[0021] Step S3: Input the joint state quantities and normal contact forces of the assembled workpiece into the neural network disturbance observer, and output the assembly disturbance estimate. The deviation between the assembly disturbance estimate and the actual disturbance is the approximation residual.

[0022] Specifically, the neural network perturbation observer adopts a radial basis function network structure. Its function is to approximate online multi-source coupled perturbations caused by workpiece dimensional tolerances during assembly, base perturbations caused by conveyor line vibration, frictional abrupt changes caused by differences in part surface roughness, and unmodeled dynamics introduced by robot joint flexibility and transmission clearance. The joint state variables received by the input layer include 12 components: 6 joint angles and 6 joint angular velocities collected by the encoders of each robot joint. These are concatenated with the 6-dimensional projection component of the normal contact force along the assembly direction to form an 18-dimensional input state vector. The hidden layer contains 50 Gaussian radial basis function nodes, each with an 18-dimensional center vector. Initial values ​​are obtained by performing K-means clustering on 500 sets of offline assembly process data, with the width parameter uniformly set to 2.0. The output layer is a linear layer with an output dimension of 6, corresponding one-to-one with the robot's 6 joint degrees of freedom. The output result is the assembly perturbation estimate. The weight matrix is ​​50 rows and 6 columns. Based on the sliding surface variables, it is updated online every 1ms control cycle according to an adaptive law. The update direction is the negative direction of the cross product of the hidden layer output vector and the sliding surface variables, multiplied by the learning rate of 50. A weight decay term is also added, with a decay coefficient of 0.01 multiplied by the current weight, to prevent the weights from growing indefinitely over long periods. The deviation between the assembly disturbance estimate and the actual disturbance is the approximation residual. This residual has a root mean square value of 0.5 N·m in the free approach phase, exhibits a transient spike with an amplitude of 3 to 8 N·m in the initial contact phase, and falls back to 1 to 2 N·m in the steady-state insertion phase.

[0023] Step S4: Using the assembly disturbance estimate as the first layer of feedforward compensation and the phase-sensing compensation gain adjustment to approximate the residual compensation force as the second layer of compensation, the two layers of compensation are superimposed to generate the assembly anti-interference control torque.

[0024] Specifically, the assembly anti-interference control torque consists of four summed terms. The first term is the nominal dynamic compensation amount, which calculates the inertia matrix, Coriolis force matrix, and gravitational torque vector based on the robot's joint angles and angular velocities. This is then combined with the product of the Jacobian matrix transpose and the normal contact force to compensate for the known inertial force, Coriolis force, gravity, and contact force. The second term, the first-layer feedforward compensation amount, directly takes the assembly disturbance estimate output from step S3 and performs feedforward compensation on the nonlinear portion of the lumped disturbance that the neural network can fit. The third term, the second-layer compensation, is composed of the phase-sensing compensation gain output in step S2 multiplied by a saturation switching term based on the integral sliding surface. The integral sliding surface is obtained by adding the joint tracking error multiplied by a proportional coefficient of 20, the first-order difference of the joint tracking error, and the time accumulation of the joint tracking error multiplied by an integral coefficient of 5. The saturation switching term takes a linear value when the absolute value of each component of the integral sliding surface divided by the boundary layer thickness of 0.5 does not exceed 1, and takes a positive or negative value when it exceeds 1. This design transforms discontinuous switching into continuous linear output within the boundary layer to reduce control chattering. The fourth term is the approach law control quantity, obtained by multiplying the approach law gain of 2.0 N·m by the saturation switching term, driving the system state to approach the sliding surface. The assembly anti-interference control torque obtained after summing the four terms is sent to each joint servo driver in a 1ms period. According to Lyapunov stability analysis, under the condition that the phase sensing compensation gain is always not less than the actual amplitude of the approximation residual, the joint tracking error converges to the neighborhood centered on the origin with a radius determined by the boundary layer thickness and weight attenuation coefficient. During the assembly process, the force tracking error is controlled within ±3N, the peak value of the transient collision force at the initial contact instant is controlled within 45N and below the part safety threshold of 50N, and the position tracking error is controlled within ±0.15mm.

[0025] In one specific embodiment, step S1 divides the assembly process online into a free approach phase, an initial contact phase, and a steady-state insertion phase based on the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target, generating continuous phase indication values, including:

[0026] The difference between the normal contact force and the preset force contact threshold is processed by hyperbolic tangent smoothing to obtain a force contact index with a value range of zero to one.

[0027] The difference between the remaining distance to the visual target and the preset distance contact threshold is processed by hyperbolic tangent smoothing to obtain a distance contact index with a value range of zero to one.

[0028] The difference between the absolute value of the normal contact force change rate and the preset force change rate impact threshold is processed by hyperbolic tangent smoothing to obtain a force change rate impact index with a value range of zero to one.

[0029] Specifically, the force difference is obtained by subtracting the force contact threshold of 3.0N from the normal contact force obtained in the Z-axis direction of the 6-dimensional force / torque sensor after gravity compensation. This force difference is then multiplied by the force contact sensitivity coefficient of 2.0. The dimensionless scaling factor is obtained. Taking the hyperbolic tangent of this scaling factor yields an intermediate value ranging from -1 to 1. Adding 1 to this intermediate value and dividing by 2 gives the force contact index, ranging from 0 to 1. The force contact threshold of 3.0N is set based on 60 times the sensor noise level of 0.05N to ensure that the background noise in the non-contact state does not cause the force contact index to deviate from 0. The force contact sensitivity coefficient is 2.0. The setting is based on the following: when the normal contact force exceeds the threshold of 2N, reaching 5N, the force contact index output value reaches 0.98, which is very close to 1. When the normal contact force is 0, the force difference is -3.0N, the scaling is -6.0, and the hyperbolic tangent value is close to -1. After adding 1 and dividing by 2, the force contact index approaches 0. When the normal contact force reaches the desired contact force of 30N, the force difference is 27N, the scaling is 54, and the hyperbolic tangent value is exactly equal to 1. After adding 1 and dividing by 2, the force contact index is exactly equal to 1.

[0030] The distance difference is obtained by subtracting the remaining visual target distance from the contact threshold of 6.0 mm. Here, the difference is the threshold minus the actual distance, so that the smaller the distance, the larger the index. This distance difference is then multiplied by the contact sensitivity coefficient of 0.5. The dimensionless scaling factor is obtained. After taking the hyperbolic tangent of this scaling factor, 1 is added, and the result is divided by 2 to obtain a distance contact index ranging from 0 to 1. The distance contact threshold of 6.0 mm is set based on the sum of the distance from the end point of the visually guided coarse positioning trajectory to the target assembly surface being 5 mm, plus the maximum visual positioning error of 1 mm. The distance contact sensitivity coefficient is 0.5. The setting is based on the fact that when the remaining distance to the visual target decreases from 6mm to 2mm, the distance difference is 4mm, the zoom is 2.0, and the distance contact index is 0.98, which is close enough to 1. The specific calculation of the force change rate impact index is to subtract the force change rate impact threshold of 500N / s from the absolute value of the normal contact force change rate to obtain the impact difference, and then multiply this impact difference by the force change rate impact sensitivity coefficient of 0.02. The dimensionless scaling factor is obtained. After taking the hyperbolic tangent of this scaling factor, adding 1 and dividing by 2 yields the force change rate impact index, ranging from 0 to 1. The force change rate impact threshold of 500 N / s is set based on the average rate of change of the desired contact force of 30 N within a typical impact setup time of 0.06 s. The force change rate impact sensitivity coefficient is 0.02. The settings are based on the fact that when the absolute value of the force change rate reaches 600 N / s, the impact difference is 100 N / s, the scaling factor is 2.0, and the force change rate impact index is 0.98, which is sufficient to characterize the occurrence of collision impact. All three indices use hyperbolic tangent smooth mapping instead of step judgment. The advantage of this is that the hyperbolic tangent function is continuously differentiable over the entire domain, allowing the three indices to transition smoothly near the threshold instead of producing jumps from 0 to 1. This avoids repeated jumps in index values ​​caused by small jitters in the sensor signal near the threshold, which could lead to high-frequency oscillations in the subsequent phase sensing compensation gain.

[0031] In one specific embodiment, step S1 divides the assembly process online into a free approach phase, an initial contact phase, and a steady-state insertion phase, including:

[0032] When the distance to the contact index approaches 0, the current assembly process is divided into a free approach phase;

[0033] When the distance to contact index approaches 1 and the force change rate impact index approaches 1, the current assembly process is divided into the initial contact phase.

[0034] When the distance contact index approaches 1, the force contact index approaches 1, and the force change rate impact index approaches 0, the current assembly process is divided into a steady-state insertion phase.

[0035] Specifically, the free approach phase corresponds to the motion stage where the assembled workpiece has not yet made physical contact with the target assembly surface. At this time, the remaining distance of the visual target is still greater than the distance contact threshold of 6.0 mm. After hyperbolic tangent smoothing mapping, the output value of the distance contact index approaches 0. At the same time, since no contact has occurred, the normal contact force is lower than the sensor's background noise level, and the force contact index also approaches 0. The normal contact force change rate fluctuates within the sensor noise range, and the force change rate impact index approaches 0. Therefore, the condition that the distance contact index approaches 0 can uniquely determine the free approach phase without the need for other indices. The initial contact phase corresponds to the transient stage when the assembled workpiece just collides and contacts the target assembly surface. At this time, the remaining distance to the visual target is less than the contact threshold of 6.0 mm, and the contact index approaches 1, indicating that the workpiece has entered the close range of the target assembly surface. At the same time, due to the impact of the collision, the normal contact force rises sharply in a very short time. The absolute value of the rate of change of the normal contact force far exceeds the impact threshold of 500 N / s, and the impact index of the rate of change of force approaches 1. Therefore, the initial contact phase is determined by the two conditions that the contact index approaches 1 and the impact index of the rate of change of force approaches 1. The duration of this phase depends on the decay rate of the collision impact. In the metal-metal contact condition of automotive end assembly, the typical duration is 0.05 s to 0.15 s.

[0036] The steady-state insertion phase corresponds to the stage where the assembled workpiece has established stable contact with the target assembly surface and continues to be inserted along the assembly direction. At this time, the remaining distance of the visual target is continuously less than the distance contact threshold of 6.0 mm, the distance contact index remains close to 1, the normal contact force has been stably established and exceeds the force contact threshold of 3.0 N, the force contact index is close to 1, and the collision impact has subsided. The normal contact force change rate has fallen back to below the force change rate impact threshold of 500 N / s, and the force change rate impact index is close to 0. Therefore, the steady-state insertion phase is determined by the combination of the three conditions: the distance contact index is close to 1, the force contact index is close to 1, and the force change rate impact index is close to 0. The three phase determination conditions are mutually exclusive and complete: when the distance to contact index approaches 0, it must be a free approaching phase; when the distance to contact index approaches 1, the force change rate impact index distinguishes between the initial contact phase and the steady-state insertion phase; when the force change rate impact index approaches 1, it is the initial contact phase; when the force change rate impact index approaches 0 and the force contact index approaches 1, it is the steady-state insertion phase. There is no overlap or omission in the determination. The beneficial effect of this combined determination rule is that relying solely on the normal contact force cannot distinguish between the moment when the workpiece has not yet made contact and the moment when the force is temporarily 0 after contact; relying solely on the remaining distance of the visual target cannot determine whether physical contact has occurred and the intensity of the contact. All three indicators are indispensable to uniquely determine the physical phase of the assembly.

[0037] In one specific embodiment, generating continuous phase indication values ​​in step S1 includes:

[0038] The free approach phase, initial contact phase, and steady-state insertion phase are respectively mapped to preset free approach indicator constants, initial contact indicator constants, and steady-state insertion indicator constants;

[0039] The three indicator constants are weighted and interpolated based on the distance contact index, force contact index, and force change rate impact index to obtain continuous phase indicator values. The numerical range of the continuous phase indicator values ​​is limited to the closed interval formed by the free approach indicator constant and the initial contact indicator constant, and there is a continuous transition between the free approach phase, the initial contact phase, and the steady-state insertion phase.

[0040] Specifically, the free approach indicator constant is set to 0.1. This value corresponds to the state where the root mean square value of the approximation residual at the free approach phase is 0.5 N·m, which is at its lowest level. Choosing 0.1 instead of 0 ensures that the output value of subsequent gain mapping does not degenerate into the lower bound of the gain itself, preserving a minimum residual compensation capability. The initial contact indicator constant is set to 1.0. This value corresponds to the state where the approximation residual at the initial contact phase exhibits a transient peak of 3 to 8 N·m. Taking the upper bound of the closed interval, 1.0, ensures that subsequent gain mapping outputs the maximum compensation gain at this phase. The steady-state insertion indicator constant is set to 0.4. This value corresponds to the state where the approximation residual at the steady-state insertion phase falls back to the level of 1 to 2 N·m. 0.4 is located slightly below 0.1 and 1.0, consistent with the physical characteristic that the residual amplitude in the steady-state stage is lower than that in the initial contact stage but higher than that in the free approach stage. The specific calculation for weighted interpolation is as follows:

[0041]

[0042] in For continuous phase indication values, For distance contact indicators, For force contact index, The force change rate impact index, The initial contact indication constant is set to 1.0. The steady-state insertion indicator constant is set to 0.4. The free approach indicator constant is set to 0.1.

[0043] Under the free approach phase Approaching 0, Approaching 1:

[0044]

[0045] Approaching the free approach indicator constant Under the initial contact phase Approaching 1 and Approaching 1:

[0046]

[0047] Approaching the initial contact indication constant Under steady-state insertion phase Approaching 1 Approaching 1 and Approaching 0:

[0048]

[0049] Approaching steady-state insertion indication constant .because , and The values ​​of all three indicators are in the range of 0 to 1, and the three indicator constants satisfy the following conditions: That is, given the size relationship of 0.1 < 0.4 < 1.0, the weighted interpolation result will necessarily fall within the range of 0.1 < 0.4 < 1.0. and Within the closed interval [0.1, 1.0], no out-of-bounds errors will occur. All three indices are continuous outputs of a hyperbolic tangent smoothing map; their weighted combination operation is a composite of multiplication and addition, preserving continuous differentiability. Therefore, the continuous phase indicator values... The transition between the free approach phase, the initial contact phase, and the steady-state insertion phase is smooth and continuous, without any abrupt breaks. This beneficial effect avoids the sudden torque changes in the subsequent phase sensing compensation gain caused by the step changes in the continuous phase indication values ​​at the moment of phase switching.

[0050] In one specific embodiment, step S2 generates a phase-sensing compensation gain based on the continuous phase indication value, according to the mapping relationship between the initial contact phase and the maximum value, the free approach phase and the minimum value, and the steady-state insertion phase and the intermediate value, including:

[0051] The continuous phase indicator values ​​are subjected to power-law nonlinear mapping to obtain the original compensation gain. The lower bound of the power-law nonlinear mapping is the preset minimum gain value, the upper bound is the preset maximum gain value, and the power exponent is the preset shape parameter.

[0052] The original compensation gain is subjected to a first-order low-pass filter to obtain the phase-aware compensation gain. The filter coefficient of the first-order low-pass filter is determined by the preset cutoff frequency and control period.

[0053] Specifically, using the continuous phase indication value as the base and the shape parameter 1.5 as the exponent, a power operation is performed to obtain the power result. This power result is then multiplied by the difference between the maximum gain of 10.0 N·m and the minimum gain of 1.0 N·m (9.0 N·m), and the minimum gain of 1.0 N·m is added to obtain the original compensation gain. The formula for calculating the power-law nonlinear mapping is:

[0054]

[0055] in For the original compensation gain, This is the minimum gain.

[0056] This represents the maximum gain. The continuous phase indication value is p=1.5, which is the shape parameter. The minimum gain of 1.0 N·m is set based on a safety margin of 2 times the root mean square value of the approximate residual of 0.5 N·m under the free approach phase, ensuring that basic compensation capability is retained even in the stage of minimum residual. The maximum gain of 10.0 N·m is set based on a safety margin of 1.25 times the peak value of the approximate residual of 8 N·m under the initial contact phase, ensuring that the compensation force can fully cover the residual amplitude in the stage of maximum residual. The shape parameter of 1.5 is set to make the mapping curve convex upwards. When the continuous phase indication value is 0.1, i.e., the free approach phase:

[0057]

[0058] When the continuous phase indication value is 0.4, i.e., the steady-state insertion phase:

[0059]

[0060] This value is 1.64 times the mean residual value of 2 N·m in the steady-state stage. When the continuous phase indication value is 1.0, i.e., the initial contact phase:

[0061]

[0062] The formula for calculating a first-order low-pass filter is:

[0063]

[0064] in The phase-sensing compensation gain for the current control cycle. This is the phase-sensing compensation gain for the previous control cycle.

[0065] These are the filter coefficients.

[0066] To control the cycle, This is the original compensation gain for the current control cycle. The corresponding cutoff frequency is:

[0067]

[0068] in The cutoff frequency is set at 8Hz. The rationale for this setting is as follows: the characteristic frequency of phase switching during assembly is below 2Hz, and 8Hz is significantly higher than this characteristic frequency, ensuring that the phase sensing compensation gain can follow the dynamic trend of phase changes without causing response lag; simultaneously, the main frequency components of the 6-dimensional force / torque sensor noise are above 50Hz, and 8Hz is significantly lower than this noise frequency, filtering out high-frequency spikes in the original compensation gain caused by sensor noise. The beneficial effect of this first-order low-pass filter is that, during the transition from the initial contact phase to the steady-state insertion phase, as the original compensation gain drops from 10.0 N·m to 3.27 N·m, the filter causes the gain to decrease smoothly in an exponential decay manner rather than a step-like drop, avoiding transient force tracking errors during assembly caused by a sudden decrease in compensation force.

[0069] In one specific embodiment, step S3 inputs the joint state quantities and normal contact forces of the assembled workpiece into a neural network perturbation observer, and outputs an assembly perturbation estimate, including:

[0070] The joint angles, joint angular velocities, and the projected components of the normal contact force along the assembly direction are concatenated into an input state vector, which is then fed into the input layer of the radial basis function network.

[0071] Based on the Euclidean distance between the input state vector and the center vector of each Gaussian radial basis function node in the hidden layer, the output vector of the hidden layer is obtained through Gaussian activation operation. The center vector of each Gaussian radial basis function node is determined by offline clustering preset.

[0072] The hidden layer output vector is linearly weighted by the weight matrix of the output layer to obtain the assembly disturbance estimate. The weight matrix is ​​updated online according to the sliding surface variables using an adaptive law, with an additional weight decay term added during the online update.

[0073] Specifically, the input state vector is composed of three parts. The first part consists of six components of joint angles acquired by the encoders of the robot's six joints. The second part consists of six components of joint angular velocities obtained by first-order difference of the joint angles over a control cycle of 0.001s. The third part consists of six components of the projection of the normal contact force along the assembly direction, with only the sixth component taking the value of the normal contact force and the other five components taking 0. After concatenation, the total dimension of the input state vector is 18. The number of nodes in the input layer of the radial basis function network is 18, consistent with the dimension of the input state vector. The hidden layer contains 50 Gaussian radial basis function nodes, and the number of nodes in the output layer is 6, consistent with the number of degrees of freedom of the robot joints. The specific process of Gaussian activation is as follows: Calculate the Euclidean distance between the 18-dimensional input state vector and the 18-dimensional center vector of the j-th hidden layer node. Divide the square of this Euclidean distance by 2 (in brackets) and multiply by the square of the node width parameter (in brackets), then take the negative value. Perform natural exponential operation on this negative value to obtain the scalar activation value of the j-th node. The activation values ​​of the 50 nodes are arranged vertically to form the 50-dimensional hidden layer output vector. The initial values ​​of the 50 center vectors are obtained by performing K-means clustering on 500 sets of offline automotive terminal assembly process data. The number of clusters is set to 50, consistent with the number of hidden layer nodes. Each cluster center is the 18-dimensional center vector of the corresponding node. The width parameter of the 50 nodes is uniformly set to 2.0. This value is based on 0.5 times the average Euclidean distance between the cluster centers, ensuring that the Gaussian activation response regions of adjacent nodes have a moderate overlap to guarantee the continuity of the network's coverage of the input space.

[0074] The weight matrix of the output layer has a dimension of 50 rows and 6 columns. Multiplying the 50-dimensional hidden layer output vector by the transpose of the weight matrix yields a 6-dimensional assembly disturbance estimate. This operation involves linearly weighting and summing the 50 hidden layer activation values ​​according to their weight coefficients over the six joint components. The weight matrix is ​​updated online according to an adaptive law every 0.001s control cycle. The update direction is the negative of the cross product of the hidden layer output vector and the sliding surface variable, i.e., the current weight matrix plus a negative learning rate multiplied by the hidden layer output vector multiplied by the transpose of the sliding surface variable. The learning rate is set to 50, a value chosen to ensure that the weights can generate an effective correction within 10 to 20 control cycles when the sliding surface variable deviates from the zero plane due to the approximation of the residuals. The weight decay term is a negative decay coefficient multiplied by the learning rate and the current weight matrix. The decay coefficient is set to 0.01 and added to the adaptive law along with the update direction term. This term applies a small contraction to the weight matrix towards zero in each control cycle, preventing the weights from growing infinitely over long periods and causing the network output to diverge. The beneficial effect of the weight decay term is that when assembly disturbances disappear or decrease, the weight matrix automatically decays back to its previous high value instead of remaining at a historical high, avoiding overcompensation by the neural network for disappeared disturbances.

[0075] In one specific embodiment, step S4 uses the assembly disturbance estimate as the first-layer feedforward compensation amount and the phase-sensing compensation gain adjustment to approximate the residual compensation force as the second-layer compensation amount. The two layers of compensation amounts are superimposed to generate the assembly anti-interference control torque, including:

[0076] An integral sliding surface is constructed based on a linear combination of three terms: joint tracking error, the derivative of joint tracking error, and the time integral of joint tracking error.

[0077] The phase-sensing compensation gain is multiplied by the saturation switching term based on the integral sliding surface to obtain the second-layer compensation amount. The saturation switching term is a linear output when the integral sliding surface is within the preset boundary layer and a symbolic output when the integral sliding surface exceeds the preset boundary layer.

[0078] The assembly anti-interference control torque is obtained by summing the four terms: the nominal dynamic compensation, the first-level feedforward compensation, the second-level compensation, and the approach law control.

[0079] Specifically, the joint tracking error is the 6-dimensional joint space expected angle vector obtained by inverse kinematics solution of the Cartesian space expected trajectory output by the impedance controller in step S2, minus the 6-dimensional joint space actual angle vector collected by the robot's 6 joint encoders. The specific calculation of the integral sliding surface is as follows: the joint tracking error is obtained by first-order difference over a 0.001s control cycle to obtain the derivative of the joint tracking error; the joint tracking error is obtained by cycle-by-cycle accumulation over a 0.001s control cycle to obtain the time integral of the joint tracking error; the derivative of the joint tracking error is multiplied by the identity matrix, the joint tracking error is multiplied by the positive definite diagonal scaling factor matrix, and the time integral of the joint tracking error is multiplied by the positive definite diagonal integral factor matrix; the three terms are added together to obtain the 6-dimensional integral sliding surface.

[0080] All six diagonal elements of the scaling factor matrix are set to 20. This value is set to ensure that the sliding surface contributes a proportional contribution of about 0.2 rad / s when the joint tracking error is 0.01 rad, matching the magnitude of the joint angular velocity. All six diagonal elements of the integral coefficient matrix are set to 5. This value is set based on the principle that, during the steady-state phase, the integral term eliminates the small steady-state tracking deviation caused by constant disturbances such as friction. This allows the integral term to accumulate to a correction amount of the same order of magnitude as the proportional term within 0.2 seconds.

[0081] The specific calculation of the saturation switching term is as follows: each component of the integral sliding surface is divided by the boundary layer thickness of 0.5 to obtain the normalized sliding mode quantity. When the absolute value of the normalized sliding mode quantity does not exceed 1, the saturation switching term takes the normalized sliding mode quantity itself; when the absolute value of the normalized sliding mode quantity exceeds 1, the saturation switching term takes a positive 1 or a negative 1. The boundary layer thickness of 0.5 is set to cover the range of small jitters in the sliding surface caused by sensor noise and weight update lag within the boundary layer. Within the boundary layer, discontinuous sign switching is converted into continuous linear output to eliminate high-frequency jitter of the control torque.

[0082] The specific calculation of the second-layer compensation quantity involves multiplying the phase-sensing compensation gain by the six components of the saturation switching term component by component to obtain a 6-dimensional second-layer compensation quantity vector. The nominal dynamic compensation quantity is calculated based on the 6th-order joint spatial inertia matrix determined by the robot's joint angles, the 6th-order Coriolis force matrix determined by the joint angles and angular velocities, the 6-dimensional gravitational torque vector determined by the joint angles, and the product of the transpose of the 6th-order Jacobian matrix determined by the joint angles and the normal contact force, compensating for the known inertial forces, Coriolis forces, gravity, and contact forces during the robot's motion. The first-layer feedforward compensation quantity is directly taken from the 6-dimensional assembly disturbance estimate output in step S3.

[0083] The approach law control quantity is a 6-dimensional vector obtained by multiplying the approach law gain of 2.0 N·m by the six components of the saturation switching term component by component. The approach law gain of 2.0 N·m is set based on providing a constant driving torque to make the system state approach the sliding surface when the integral sliding surface deviates from the zero surface. 2.0 N·m is within the range of 1% to 3% of the rated torque of each joint of the robot, and will not cause overload to the joint servo driver. The assembly anti-interference control torque is a 6-dimensional vector obtained by adding the nominal dynamic compensation quantity, the first layer feedforward compensation quantity, the second layer compensation quantity, and the approach law control quantity component by component, and is sent to each joint servo driver with a period of 0.001s. The beneficial effect of this double-layer superposition structure is that the first layer feedforward compensation quantity compensates for the main nonlinear part that the neural network can fit in the lumped disturbance, and the second layer compensation quantity is specifically for the approximation residual that the neural network cannot fit, and performs precise compensation with phase matching force. The two layers have clear division of labor and do not interfere with each other.

[0084] In one specific embodiment, an integral sliding surface is constructed based on a linear combination of three terms: joint tracking error, the derivative of joint tracking error, and the time integral of joint tracking error, including:

[0085] The joint tracking error is obtained by subtracting the expected joint space angle from the impedance controller output and solved by inverse kinematics from the actual joint space angle acquired by the encoder. The joint tracking error is multiplied by a preset positive definite diagonal proportional coefficient matrix, the derivative of the joint tracking error obtained by the first difference of the joint tracking error over the control cycle is multiplied by the identity matrix, and the time integral of the joint tracking error obtained by the accumulation of the joint tracking error over the control cycle is multiplied by a preset positive definite diagonal integral coefficient matrix. The three terms are added together to obtain the integral sliding surface.

[0086] In this process, the desired joint space angle is calculated by the impedance controller in step S2 based on the deviation between the desired and actual contact forces. This result is superimposed onto the target assembly pose to obtain the desired Cartesian space trajectory. The Cartesian space pose is then converted into a 6-dimensional joint space angle vector through inverse kinematics of the 6-DOF robot. The actual joint space angle is the 6-dimensional angle vector collected by the robot's six joint encoders within each 0.001s control cycle. Subtracting the actual joint space angle from the desired joint space angle yields the 6-dimensional joint tracking error vector. The joint tracking error is then multiplied by a positive definite diagonal scaling factor matrix, which is a 6th-order diagonal matrix with all six diagonal elements set to 20. This value ensures that the sliding surface contributes a proportional contribution of approximately 0.2 rad / s when the joint tracking error is 0.01 rad, matching the order of magnitude of the joint angular velocity. The derivative of the joint tracking error is obtained by subtracting the joint tracking error of the previous control cycle from the current control cycle's joint tracking error and then dividing by the control cycle of 0.001 s. Multiplying this derivative by a 6th-order identity matrix preserves the original value. The time integral of the joint tracking error is obtained by accumulating the joint tracking error cycle by cycle at 0.001 s control cycles from the assembly start time. This time integral is then multiplied by a positive definite diagonal integral coefficient matrix, which is a 6th-order diagonal matrix with all six diagonal elements set to 5. This value allows the integral term to accumulate to a correction amount of the same order of magnitude as the proportional term within 0.2s, thus eliminating the small steady-state tracking deviation caused by constant disturbances such as friction. The three 6-dimensional vectors—the proportional term, the derivative term, and the integral term—are added component-by-component to obtain the 6-dimensional integral sliding surface vector.

[0087] In one specific embodiment, the phase-sensing compensation gain is multiplied by a saturation switching term based on the integral sliding surface to obtain the second-layer compensation amount, including:

[0088] The normalized sliding mode quantity is obtained by dividing each component of the integral sliding mode surface by the preset boundary layer thickness. When the absolute value of the normalized sliding mode quantity does not exceed 1, the saturation switching term takes the normalized sliding mode quantity itself. When the absolute value of the normalized sliding mode quantity exceeds 1, the saturation switching term takes the same positive or negative sign value as the normalized sliding mode quantity. The phase sensing compensation gain is multiplied by the saturation switching term component by component to obtain the second layer compensation quantity.

[0089] The six components of the integral sliding surface are divided by the boundary layer thickness of 0.5 to obtain six normalized sliding surface quantities. The boundary layer thickness of 0.5 is set to cover the range of small jitter of the integral sliding surface caused by sensor noise and weight update lag. This range is estimated based on the amplitude of sliding surface fluctuation caused by the 0.05N resolution of the 6-dimensional force / torque sensor after dynamic transmission in the joint space.

[0090] For each normalized sliding mode quantity, a saturation judgment is performed: when the absolute value of the normalized sliding mode quantity does not exceed 1, it indicates that the component of the integral sliding surface is inside the boundary layer, and the saturation switching term takes the normalized sliding mode quantity itself, that is, the output is the linear value of the component of the integral sliding surface divided by the boundary layer thickness of 0.5. At this time, the saturation switching term is proportional to the integral sliding surface, and the control action is continuous and chatter-free; when the absolute value of the normalized sliding mode quantity exceeds 1, it indicates that the component of the integral sliding surface has exceeded the boundary layer range, and the saturation switching term takes the same positive or negative sign value as the normalized sliding mode quantity, that is, positive 1 or negative 1. At this time, the saturation switching term is a constant sign output, maintaining the strong robustness of sliding mode control to disturbances. After the above saturation judgment, the six components form a 6-dimensional saturation switching term vector. The phase-sensing compensation gain is multiplied by each component of the 6-dimensional saturation switching term vector to obtain a 6-dimensional second-layer compensation vector. Each component of this vector is equal to the phase-sensing compensation gain multiplied by the saturation switching term value of the corresponding joint. Under the free approach phase, the phase-sensing compensation gain is 1.28 N·m, which makes the amplitude of the second-layer compensation smaller. Under the initial contact phase, the phase-sensing compensation gain is 10.0 N·m, which increases the amplitude of the second-layer compensation to quickly suppress the peak of the approaching residual. Under the steady-state insertion phase, the phase-sensing compensation gain is 3.27 N·m, which keeps the amplitude of the second-layer compensation at a medium level to smoothly compensate for the residual.

[0091] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A force-vision based intelligent anti-interference control method for automotive terminal robot assembly, characterized in that, The method includes: Step S1: Based on the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target, the assembly process is divided online into the free approach phase, the initial contact phase, and the steady-state insertion phase, and continuous phase indication values ​​are generated. Step S2: Based on the continuous phase indication value, and according to the mapping relationship between the maximum value corresponding to the initial contact phase, the minimum value corresponding to the free approach phase, and the intermediate value corresponding to the steady-state insertion phase, generate the phase-aware compensation gain. Step S3: Input the joint state quantities of the assembled workpiece and the normal contact force into the neural network disturbance observer, and output the assembly disturbance estimate. The deviation between the assembly disturbance estimate and the actual disturbance is the approximation residual. Step S4: Using the assembly disturbance estimate as the first layer of feedforward compensation and the phase-sensing compensation gain to adjust the compensation strength of the approximation residual as the second layer of compensation, the two layers of compensation are superimposed to generate the assembly anti-interference control torque.

2. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 1, characterized in that, In step S1, based on the normal contact force of the assembled workpiece, the rate of change of the normal contact force, and the remaining distance of the visual target, the assembly process is divided online into a free approach phase, an initial contact phase, and a steady-state insertion phase, generating continuous phase indication values, including: The difference between the normal contact force and the preset force contact threshold is subjected to hyperbolic tangent smoothing mapping to obtain a force contact index with a value range of zero to one. The difference between the remaining distance of the visual target and the preset distance contact threshold is subjected to hyperbolic tangent smoothing mapping to obtain a distance contact index with a value range of zero to one. The difference between the absolute value of the normal contact force change rate and the preset force change rate impact threshold is processed by hyperbolic tangent smoothing to obtain a force change rate impact index with a value range of zero to one.

3. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 2, characterized in that, In step S1, the assembly process is divided online into a free approach phase, an initial contact phase, and a steady-state insertion phase, including: When the distance contact index approaches 0, the current assembly process is divided into the free approach phase; When the distance contact index approaches 1 and the force change rate impact index approaches 1, the current assembly process is divided into the initial contact phase. When the distance contact index approaches 1, the force contact index approaches 1, and the force change rate impact index approaches 0, the current assembly process is divided into the steady-state insertion phase.

4. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 3, characterized in that, The step S1 of generating continuous phase indication values ​​includes: The free approach phase, the initial contact phase, and the steady-state insertion phase are respectively mapped to preset free approach indicator constants, initial contact indicator constants, and steady-state insertion indicator constants; The three indicator constants are weighted and interpolated based on the distance contact index, the force contact index, and the force change rate impact index to obtain the continuous phase indicator value. The numerical range of the continuous phase indicator value is limited to the closed interval formed by the free approach indicator constant and the initial contact indicator constant, and it transitions continuously between the free approach phase, the initial contact phase, and the steady-state insertion phase.

5. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 1, characterized in that, In step S2, based on the continuous phase indication value, and according to the mapping relationship between the maximum value corresponding to the initial contact phase, the minimum value corresponding to the free approach phase, and the intermediate value corresponding to the steady-state insertion phase, a phase-sensing compensation gain is generated, including: The continuous phase indication value is subjected to power-law nonlinear mapping to obtain the original compensation gain. The lower bound of the power-law nonlinear mapping is a preset minimum gain value, the upper bound is a preset maximum gain value, and the power exponent is a preset shape parameter. The original compensation gain is subjected to a first-order low-pass filter to obtain the phase-sensing compensation gain. The filter coefficient of the first-order low-pass filter is determined by a preset cutoff frequency and control period.

6. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 1, characterized in that, In step S3, the joint state quantities of the assembled workpiece and the normal contact force are input into the neural network disturbance observer, and the assembly disturbance estimate is output, including: The angles of each joint, the angular velocities of each joint, and the projection components of the normal contact force along the assembly direction are concatenated into an input state vector, which is then fed into the input layer of the radial basis function network. Based on the Euclidean distance between the input state vector and the center vector of each Gaussian radial basis function node in the hidden layer, the output vector of the hidden layer is obtained through Gaussian activation operation. The center vector of each Gaussian radial basis function node is determined by offline clustering preset. The hidden layer output vector is linearly weighted by the weight matrix of the output layer to obtain the assembly disturbance estimate. The weight matrix is ​​updated online according to the sliding surface variables using an adaptive law, and a weight decay term is added in the online update.

7. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 1, characterized in that, In step S4, the assembly disturbance estimate is used as the first-layer feedforward compensation amount, and the phase-sensing compensation gain is used to adjust the compensation strength of the approximation residual as the second-layer compensation amount. The two layers of compensation amounts are superimposed to generate the assembly anti-interference control torque, including: An integral sliding surface is constructed based on a linear combination of three terms: joint tracking error, the derivative of joint tracking error, and the time integral of joint tracking error. The phase-sensing compensation gain is multiplied by the saturation switching term based on the integral sliding surface to obtain the second layer compensation amount. The saturation switching term is a linear output when the integral sliding surface is within the preset boundary layer and a sign output when the integral sliding surface exceeds the preset boundary layer. The assembly anti-interference control torque is obtained by summing the nominal dynamic compensation amount, the first layer feedforward compensation amount, the second layer compensation amount, and the approach law control amount.

8. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 7, characterized in that, The integral sliding surface is constructed based on a linear combination of three terms: joint tracking error, the derivative of joint tracking error, and the time integral of joint tracking error. This includes: The joint tracking error is obtained by subtracting the expected joint space angle obtained from the impedance controller output and solved by inverse kinematics from the actual joint space angle obtained by the encoder. The joint tracking error is multiplied by a preset positive definite diagonal proportional coefficient matrix, the derivative of the joint tracking error obtained by the first difference of the joint tracking error according to the control cycle is multiplied by the identity matrix, and the time integral of the joint tracking error obtained by the accumulation of the joint tracking error according to the control cycle is multiplied by a preset positive definite diagonal integral coefficient matrix. The three terms are added together to obtain the integral sliding surface.

9. The intelligent anti-interference control method for automotive terminal robot assembly based on force-vision as described in claim 8, characterized in that, The step of multiplying the phase-sensing compensation gain by a saturation switching term based on the integral sliding surface to obtain the second-layer compensation amount includes: The normalized sliding mode quantity is obtained by dividing each component of the integral sliding mode surface by a preset boundary layer thickness. When the absolute value of the normalized sliding mode quantity does not exceed 1, the saturation switching term is taken as the normalized sliding mode quantity itself. When the absolute value of the normalized sliding mode quantity exceeds 1, the saturation switching term is taken as the same positive or negative sign value as the normalized sliding mode quantity. The phase sensing compensation gain is multiplied by the saturation switching term component by component to obtain the second layer compensation quantity.

Images