A method and system for controlling the timing of a screw clamp of a welding robot manipulator

Abstract

The application relates to the technical field of welding clamping, and discloses a welding robot manipulator screw clamping timing control method and system. The method comprises the following steps: driving a bidirectional screw to rotate forward to release a first clamping piece from a second clamping piece pre-contacting a clamped body; injecting a sinusoidal disturbance into the screw to collect a motor q-axis current response; performing inversion network embedding through a beam bending priori to obtain a load distribution ratio by regression with beam deflection and micro-amount difference constraints; and triggering release according to the load distribution ratio and clamping by a self-locking lead angle and a wedge-shaped conical surface. The application improves the state observability of the manipulator switching process and the intrinsic safety of the failure mode.

Description

Technical Field

[0001] This application relates to the field of welding clamping technology, and in particular to a welding robot manipulator screw clamping timing control method and system. Background Technology

[0002] In continuous welding operations involving multiple stations and multiple workpieces, welding robots need to frequently switch their end effector between a fixed parking mechanism and the robot wrist to adapt to different weld seam shapes and workpiece sizes. For example, invention patent CN109849042B discloses a quick-change device for a robot end effector. This device includes a frame, a tool connector fixed to the frame, a robot end effector pneumatically connected to the tool connector, and a pneumatic device located on the frame. The pneumatic device fixes the tool connector to the frame via a cylinder rod passing through a through-hole in a limiting plate. A photoelectric sensor or inductive sensor on the positioning plate senses the position of the long pin, and the control box sends an extension command to the cylinder and a locking command to the pneumatic connector, thereby enabling the robot's execution end to switch between peripheral tools such as welding torches, grinding tools, and laser cleaners. For example, the utility model patent with publication number CN212886349U discloses an automatic clamping structure for parts. The clamping structure is driven by a motor to rotate a bidirectional transmission screw in both directions. The outer surface of the bidirectional transmission screw is threaded with two sliding blocks. The same rotation angle input of a single motor is synchronously converted into a mirror motion of the two sliding blocks moving closer or further apart through the left and right rotation sections, thereby driving the clamping columns at the top of the two sliding blocks to clamp or release the workpiece. As can be seen from the above solutions, existing welding robot manipulator switching solutions generally adopt a transmission structure that uses a bidirectional lead screw to drive two sets of opposing sliders. A single servo motor drives the two sets of clamping plates to achieve mirror opening and closing by rotating the threaded sections at both ends in opposite directions. This is combined with pre-programmed sequential timing logic or pneumatic locking elements to complete the transfer of the manipulator between the fixed parking mechanism and the robot wrist. The clamping status at both ends of the clamped body is usually fed back to the control cabinet by a single-point arrival signal given by the limit switch, proximity switch, or photoelectric and inductive sensors at the end of the slider stroke. The control cabinet uses the arrival of the preset delay or limit signal as the trigger condition for the next action. The entire switching process is executed sequentially according to a fixed time sequence of release, transfer, and re-clamping.

[0003] The above solutions can meet the switching needs of production lines with stable cycle times and uniform operating conditions. However, in actual welding workshops, disturbances such as decreased lubrication of the lead screw, uneven wear of the clamping plates, atypical distribution of the robot's posture on different workpieces, and cumulative increase in motor housing temperature due to continuous operation coexist for a long time. The preset delay open-loop timing triggers the next action before the previous clamping mechanism has completely disengaged from the force end or before the motor has reached the predetermined clamping force, causing the clamped body to fall while both clamping mechanisms are in a semi-clamped state. The single-point limit signal can only reflect the slider position and cannot reflect the actual load distribution between the clamping plates and the clamped body. The lead screw transmission clearance, changes in slider guide friction, and torque deviation of the robot's own weight under different postures all cause deviations between the limit position and the clamping position. While adding force sensors to the clamping plates to obtain the actual clamping force is a feasible approach, the clamping plates are affected by welding spatter, arc radiation, and high temperatures. The cable routing and protective structure of the external sensor cause geometric interference with the original clamping structure, making long-term reliability difficult to guarantee. Solutions using electromagnets or pneumatic components as locking actuators immediately lose clamping force in the event of power or gas outages. Even if a purely data-driven regression model is introduced to infer the load state from the electrical signals of the driving components without adding external sensors, the lack of verifiable physical consistency anchors corresponding to the actual deformation of the clamped object makes it difficult to quantify and constrain the output value drift in conditions not covered by the training set.

[0004] To address the issue that preset delays and single-point positioning signals cannot characterize the load distribution of the clamped body, the timing trigger condition can be changed from delay to load state determination. However, if the acquisition of the load state relies on an external clamping force sensor, it faces failure and geometric interference problems in welding environments. This leads to a new problem of reverse-engineering the load distribution from the original drive components themselves without adding any external sensors. Furthermore, even if the load distribution can be reverse-engineered from the motor's own electrical signals, this reverse-engineering process still faces new problems of extrapolation failure under conditions such as screw lubrication attenuation, increased clamp wear, and atypical robot posture. If it relies solely on a pure data-driven model or uses only general partial differential equations as soft constraints superimposed on the end of the loss function, it will still fail under conditions not covered by the training set. Moreover, even if the load distribution reverse-engineering can provide reasonable results within the training set coverage, if the reverse-engineering process fails under extreme conditions, the robot will still face the risk of falling. This leads to the final problem of how the switching system should passively complete the bottom clamping at the physical level when the reverse-engineering result is wrong, avoiding reliance on software-layer algorithms for compensation. Summary of the Invention

[0005] This application provides a welding robot manipulator screw clamping timing control method and system, which solves the problems that the load distribution of the clamping body during the switching process between the fixed parking mechanism and the robot wrist of the welding robot manipulator cannot be accurately obtained without adding external sensors, and that the risk of the manipulator falling off cannot be passively eliminated at the physical level when the inversion result is wrong under extreme working conditions. It improves the state observability of the manipulator switching process and the inherent safety of failure modes.

[0006] In a first aspect, this application provides a timing control method for the lead screw clamping of a welding robot manipulator, the timing control method for the lead screw clamping of a welding robot manipulator includes: Step S1: Drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the pre-contacting body with the second clamping piece, and obtains the pre-contact state; Step S2: Inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, collect the q-axis current response of the drive motor, and obtain the current response sequence; Step S3: By using an inversion network with embedded beam bending priors, the fundamental frequency and second harmonic component of the current response sequence, the screw angle and the motor housing temperature are spliced ​​together as joint features. The difference between the beam deflection analysis value and the measured micromomentum of the clamped body under loads at both ends is used as a physical residual term to constrain the regression output of the joint features, thereby obtaining the load distribution ratio of the clamped body between the first clamp and the second clamp. Step S4: Trigger the release of the first clamping plate according to the load distribution ratio, and use the self-locking lead angle of the bidirectional screw and the wedge-shaped cone of the second clamping plate to clamp the object.

[0007] Secondly, this application provides a timing control system for the lead screw clamping of a welding robot manipulator, the timing control system for the lead screw clamping of the welding robot manipulator includes: The drive module is used to drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the clamped body that has been in pre-contact with the second clamping piece, and obtains a pre-contact state. The acquisition module is used to inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, acquire the q-axis current response of the drive motor, and obtain the current response sequence. The analysis module is used to combine the fundamental frequency and second harmonic component of the current response sequence, the lead screw angle and the motor housing temperature into a joint feature by an inversion network with embedded beam bending priors. The difference between the analytical value of the beam deflection of the clamped body under loads at both ends and the measured micromomentum is used as a physical residual term to constrain the regression output of the joint feature, so as to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp. The clamping module is used to trigger the release of the first clamping piece according to the load distribution ratio, and the self-locking lead angle of the bidirectional screw and the wedge-shaped cone surface of the second clamping piece perform bottom clamping on the clamped body.

[0008] Thirdly, a welding robot manipulator screw clamping timing control device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor invokes the instructions in the memory to cause the welding robot manipulator screw clamping timing control device to execute the above-described welding robot manipulator screw clamping timing control method.

[0009] Fourthly, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, cause the computer to execute the above-described welding robot manipulator screw clamping timing control method.

[0010] In the technical solution provided in this application, the synchronous mirror action of the first clamping plate releasing and the second clamping plate pre-contacting the clamped body by the forward rotation of the bidirectional lead screw is used as the starting point of the switching process. The same rotation angle input of a single drive motor is directly converted into the symmetrical displacement of the two sliders at the hardware level through the left and right rotation segments, eliminating the position synchronization control link required in the dual-motor scheme. The transmission chain is compact and the motion consistency is guaranteed by geometric relationship rather than control algorithm. The establishment of the pre-contact state incorporates the first clamping mechanism and the second clamping mechanism into the contact relationship with the clamped body at the same time, so that the key critical state of the switching process becomes a neutral state that can be observed later rather than a transient state that is difficult to distinguish. In this critical state, a low-amplitude sinusoidal disturbance is injected into the bidirectional lead screw and the q-axis current response of the drive motor is collected. This allows the drive element itself to serve as both an excitation source and an observation source. There is no need to install a force sensor at the clamp or to lay a tactile array on the clamped body to obtain the electrical characteristics associated with load distribution. This avoids the geometric interference and lifespan risks caused by the high temperature and arc light environment of the welding workshop to the external sensor cables and protective structures. The sensing link of the method is compressed to the q-axis current sampling circuit that already exists inside the servo driver at the physical level. The observability of the state of the switching process is changed from relying on external measuring elements to relying on the electrical signals of the original drive element itself.

[0011] By embedding a beam bending prior inversion network, the joint features of the fundamental frequency and second harmonic components of the current response sequence, the lead screw angle, and the motor housing temperature are regressed. The difference between the midpoint deflection analytical value derived from the Euler-Bernoulli beam equation and the measured micromomentum synchronously detected by the inertial measurement unit is used as the regression output of the physical residual term to constrain the joint features. This ensures that the inversion network is still physically constrained by the beam bending analytical relationship even under three conditions not covered by the training set: lead screw lubrication decay, increased clamp wear, and atypical robot posture. The regression value drift of the pure data-driven model under extrapolation conditions is converged to a range consistent with the actual deformation of the clamped body. This demonstrates the substantial contribution of the algorithm features to the solution when introducing the algorithm in the specific application field of welding robot switching. The beam bending prior is not superimposed on the end of the loss function as a general regularization term, but rather forms a verifiable physical consistency closed loop based on the mechanical modeling of the clamped body as a simply supported beam and the measured micromomentum of the robot wrist. The final load distribution ratio triggers the release of the first clamping plate. Combined with the self-locking inequality that the lead angle of the bidirectional lead screw is strictly less than the secondary friction angle, and the geometric constraint of the wedge-shaped conical surface of the second clamping plate, this ensures that even if the inversion result shows a deviation, the clamped body can still automatically convert the displacement trend into an increase in radial clamping force through wedge engagement. This transforms the potential for immediate loss of lockout in traditional electromagnetic or pneumatic locking schemes under power or air shortage conditions into a passive, bottom-line clamping mechanism based on geometric relationships. The root mean square error of the load distribution ratio inversion is under conditions not covered by the three training sets. The range of 0.128 to 0.165 in the pure data-driven model decreased to the range of 0.046 to 0.058. Under the action of its own weight, the additional radial clamping force obtained by the wedge-shaped conical surface transformation of the clamped body reached 51.2N. After being superimposed on the clamping force established before release, the total radial clamping force exceeded the critical level of 75N to prevent falling. The inversion robustness and failure fallback margin of the welding robot manipulator screw clamping timing control method and system under the conditions of screw lubrication, clamp wear and deviation of manipulator posture are quantitatively guaranteed. Attached Figure Description

[0012] 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.

[0013] Figure 1 This is a schematic diagram of an embodiment of the welding robot manipulator screw clamping timing control method in this application. Figure 2 This is a schematic diagram comparing the amplitude of the spectral components of the q-axis current response sequence under different load distribution ratios in the embodiments of this application; Figure 3This is a schematic diagram of the training loss curve of the inversion network with embedded beam bending prior in an embodiment of this application; Figure 4 This is a schematic diagram comparing the root mean square error of load distribution ratio inversion for three inversion methods under four typical working conditions in the embodiments of this application. Detailed Implementation

[0014] This application provides a method and system for timing control of the lead screw clamping mechanism of a welding robot. 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 used 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 apparatus that comprises 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 apparatus.

[0015] 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 welding robot manipulator screw clamping timing control method in this application includes: Step S1: Drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the pre-contacting body with the second clamping piece, and obtains the pre-contact state; Specifically, the pre-contact state refers to the critical state where geometric contact occurs between the V-shaped conical surface of the second clamping piece and the outer conical surface of the clamped object, but a complete clamping force has not yet been established. The boundary for this determination is when the axial preload applied by the second clamping piece to the clamped object falls within a narrow range of 5N to 10N. The lower limit of this range is 5N to ensure that the contact is a valid physical contact rather than a false trigger caused by encoder drift, and the upper limit is 10N to ensure that the main load-bearing component of the clamped object is still the first clamping piece, and the load shared by the second clamping piece does not exceed 12% of the total weight. In this sense, the pre-contact state becomes a neutral critical state where the load distribution ratio can be inverted by subsequent steps. The alignment error threshold is 0.05mm in the axial direction and 0.1mm in the radial direction. This threshold is derived from the 12-degree half-cone angle of the V-shaped conical surface. Exceeding this threshold will cause unilateral contact between the outer conical surface of the clamped object and the V-shaped conical surface, and the current response signal will be contaminated by off-center load noise.

[0016] Step S2: Inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, collect the q-axis current response of the drive motor, and obtain the current response sequence; Specifically, the amplitude of the low-amplitude sinusoidal disturbance is determined within the electrical angle range of 0.3 to 0.5 degrees, and the frequency is determined within the process bandwidth of 5 Hz to 20 Hz. The lower limit of the amplitude is limited by the encoder resolution and the noise floor of the current loop, while the upper limit of the amplitude is limited by the self-locking condition of the bidirectional lead screw—the axial micro-motion of the slider corresponding to the disturbance must not exceed the static friction boundary between the slider nut and the lead screw pair, to avoid the pre-contact state being destroyed by the disturbance itself. The frequency selection avoids the 50 Hz power frequency of the machine tool foundation and its harmonics, and avoids the high-frequency region outside the servo position loop bandwidth, so that the energy of the q-axis current response sequence is concentrated on the fundamental frequency and its second harmonic. The power frequency notch parameter refers to the coefficients of a second-order infinite impulse response digital filter with a stopband center frequency of 50 Hz and a stopband width of 4 Hz, which is implemented by the servo driver firmware in the sampling circuit. The power frequency coupling component has been removed from the current response sequence before it is sent to subsequent spectrum processing.

[0017] Step S3: By using the inversion network with embedded beam bending prior, the fundamental frequency and second harmonic component of the current response sequence, the screw angle and the motor housing temperature are spliced ​​together as joint features. The difference between the beam deflection analysis value and the measured micromomentum of the clamped body under load at both ends is used as the regression output of the physical residual term constrained joint features to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp. Specifically, the beam bending prior in the inversion network is derived from the analytical expression of midpoint deflection derived from the Euler-Bernoulli beam equation after the clamped body is modeled as a simply supported beam. This analytical expression takes the weight distribution ratio of the clamped body at both ends as input and the midpoint deflection as output. The prior embedding path is not used as a network structure layer, but as a physical residual term in the loss function, obtained by subtracting the analytical deflection from the measured micromotor synchronously detected by the inertial measurement unit of the welding robot's wrist. This embedding method ensures that the network output is still constrained by the analytical relationship of beam bending even under conditions not covered by the training set, avoiding pure data-driven regression value drift under three extrapolation conditions: deterioration of screw lubrication, increased clamp wear, and atypical robot posture. This shifts the risk of falling during switching caused by the separation of state feedback and timing control from the signal layer to the physical consistency layer. The extraction frequencies of the fundamental frequency and the second harmonic component correspond to the injection disturbance frequency of 10Hz and its second harmonic frequency of 20Hz, respectively. The second harmonic component carries the nonlinear offset information of the algebraic sum of the counter-torques of the left and right rotating segments of the bidirectional screw, and is a key observation for load distribution ratio inversion.

[0018] Step S4: The first clamp is released according to the load distribution ratio, and the self-locking lead angle of the bidirectional screw and the wedge-shaped cone of the second clamp are used to clamp the receiving body.

[0019] Specifically, the inequality between the self-locking lead angle and the friction angle is determined by the lead angle of the bidirectional screw (4mm, outer diameter 20mm, and friction coefficient 0.15). The lead angle of 3.64 degrees is strictly less than the friction angle of 8.53 degrees, ensuring that the screw pair remains locked under any external force along the axial direction. This self-locking characteristic is the prerequisite for the wedge-shaped conical surface bottom clamping—the second clamping plate is backed by the second slider, the second slider is geometrically locked by the self-locking screw, and the displacement trend of the clamped body relative to the second clamping plate along the axial direction under its own weight is converted into an increase in radial clamping force through the 12-degree half-cone angle of the V-shaped conical surface. The value of the half-cone angle of 12 degrees comes from the trade-off between the complementary angle of the friction angle and the wedge amplification ratio. If the angle is too small, the increase in radial clamping force will be insufficient to support the weight of the clamped body; if the angle is too large, the axial reaction force generated by the wedge will break through the self-locking boundary of the screw. The risk of immediate loss of lock when using electromagnets or pneumatic components as locking actuators under power or air shortage conditions is eliminated passively by this geometric backstop mechanism.

[0020] In one specific embodiment, step S1 includes: Based on the predetermined pick-up and place posture of the welding robot wrist and the centering error threshold of the fixed parking mechanism, the welding robot wrist is centered to obtain the wrist posture after centering. Based on the orientation of the wrist position, a forward rotation command is applied to the drive motor to obtain the synchronous mirror displacement of the bidirectional lead screw. Based on the synchronous mirror displacement, the wedging state between the first clamping piece and the outer conical surface of the clamped body is released, and the second clamping piece is brought into contact with the outer conical surface of the clamped body to obtain a light contact state of the second clamping piece. Based on the step signal of the q-axis current of the drive motor, the light contact state of the second clamping piece is processed to obtain the pre-contact state.

[0021] Specifically, the centering error thresholds are set at 0.05 mm in the axial direction and 0.1 mm in the radial direction. The axial threshold is determined by converting the 4 mm lead per revolution of the bidirectional lead screw to the controllable displacement accuracy under the 23-bit resolution of the servo encoder. The radial threshold is determined by converting the 12-degree half-cone angle of the V-shaped conical surface of the second clamping piece to the eccentricity tolerance after converting to the contact plane of the outer conical surface of the clamped body. Exceeding these thresholds will cause one-sided contact between the outer conical surface of the clamped body and the V-shaped conical surface, resulting in an off-center load component in the q-axis current response of the motor during the disturbance injection stage, which contaminates the input features of the subsequent inversion network. The wrist position after centering is calculated by the welding robot control cabinet in conjunction with the SLAM navigation map and the teaching coordinates, ensuring that the central axis of the second clamping piece and the axis of the clamped body fall within the above-mentioned threshold window. Synchronous mirror displacement refers to the amount of motion in which the distance the first slider moves backward along the screw axis is exactly equal to the distance the second slider moves forward along the screw axis when a single drive motor is rotating in the forward direction. Its physical basis is that the first threaded segment and the second threaded segment are divided by the midpoint of the screw axis, have opposite directions of rotation, and have the same lead. The same rotation angle input of a single motor is synchronously converted into the symmetrical motion of the two sliders on the axis through the left and right rotation segments. The mirror motion of the two clamping plates is guaranteed at the hardware level without any synchronization control algorithm.

[0022] The forward rotation command of the drive motor causes the first slider to retract axially along the lead screw from the end of its stroke in the initial clamping state. The initial radial clamping force borne by the first clamping plate then decreases linearly with the amount of retraction. When the amount of retraction reaches the critical displacement that completely opens the radial gap between the V-shaped cone surface and the outer cone surface of the clamped object, the wedging state is released. At the same motor rotation angle, the second slider moves forward in the opposite direction with the same displacement. The inner distance between the two second clamping plates gradually contracts from the initial opening value of 90mm. The outer edge of the V-shaped cone surface of the second clamping plate approaches the outer cone surface of the clamped object in a parallel and asymptotic manner, resulting in a light contact state of the second clamping plates. The endpoint criterion for this contact approach process is derived from the step signal of the q-axis current of the drive motor. During the idle stroke phase, the motor only overcomes the frictional torque of the lead screw pair, and the q-axis current remains at a low plateau between 0.2A and 0.4A. When the V-shaped conical surface of the second clamping plate makes physical contact with the outer conical surface of the clamped object and begins to establish preload, the contact reaction torque from the second clamping plate is superimposed on the motor load torque. The q-axis current exhibits a step of more than 0.15A within two main control cycles. The controller uses this step as the determination flag for the occurrence of contact and latches the current lead screw main rotation angle as the zero-point reference for the pre-contact state. The axial preload corresponding to the pre-contact state falls within a narrow range of 5N to 10N. The lower limit ensures that the contact is an effective physical contact rather than a false trigger caused by current noise. The upper limit ensures that the main load-bearing body of the clamped object is still the first clamping plate, and the load sharing ratio of the second clamping plate does not exceed 12% of the self-weight of the clamped object, making the pre-contact state a neutral critical state in which the load distribution ratio can be inverted.

[0023] In one specific embodiment, step S2 includes: Based on the reference value of the main rotation angle of the lead screw in the pre-contact state, the position ring reference angle of the drive motor is sinusoidally superimposed to obtain a low-amplitude sinusoidal disturbance signal. Based on the low-amplitude sinusoidal disturbance signal, the bidirectional lead screw is subjected to disturbance injection processing to obtain the torsional response state of the lead screw pair under disturbance excitation. Based on the torsional response state of the lead screw pair, the q-axis current loop of the drive motor is subjected to high-frequency synchronous sampling processing to obtain the original q-axis current sequence; Based on the power frequency notch parameters, the original q-axis current sequence is subjected to band-stop filtering to obtain the current response sequence.

[0024] Specifically, the reference value of the lead screw main rotation angle is taken from the latched value of the lead screw main rotation angle read by the absolute encoder in the pre-contact state. The encoder resolution is 23 bits, corresponding to 8,388,608 counting scales per revolution. This reference value is kept frozen throughout the entire disturbance injection window and serves as the zero-phase reference point for sinusoidal superposition processing. The sinusoidal superposition processing of the reference angle of the drive motor position loop is performed at the input end of the servo drive position loop, that is, a sinusoidal time function with an amplitude A of 0.4 electrical degrees and a frequency f of 10Hz is superimposed on the original position command to obtain a low-amplitude sinusoidal disturbance signal. The disturbance amplitude is determined under triple boundary conditions. The lower limit is constrained by the effective angle change threshold of the servo encoder at 23-bit resolution being no less than 0.05 degrees. The upper limit is constrained by the self-locking condition of the bidirectional lead screw—the disturbance amplitude, after being converted into the slider axial micro-motion by the lead angle of 3.64 degrees, must be lower than the critical displacement corresponding to the static friction torque between the lead screw nut and the lead screw. The value of 0.4 degrees ensures that the slider axial micro-motion falls within 0.004 mm, far below the elastic deformation of the initial contact indentation between the V-shaped conical surface and the outer conical surface of the clamped body, ensuring that the pre-contact state is not destroyed by the disturbance itself. The disturbance frequency of 10 Hz avoids the high-frequency region outside the 100 Hz bandwidth boundary of the servo position loop, allowing the q-axis current response to be fully presented under the closed-loop tracking of the position loop. At the same time, it avoids the 50 Hz power frequency interference band of the machine tool foundation and its lower sideband, ensuring that the subsequent spectrum processing has clear energy peaks at the fundamental frequency and the second harmonic 20 Hz frequency point.

[0025] The disturbance injection process is executed by the servo driver with a position loop main cycle of 0.5ms. Each main cycle sends a sinusoidal superposition command to the current loop reference. The drive motor responds to the disturbance with a small torsional amplitude corresponding to 0.4 degrees within the self-locking range of the bidirectional lead screw, obtaining the torsional response state of the lead screw pair. This response state represents the cumulative elastic torsional deformation of the bidirectional lead screw within the self-locking range. In this response state, the first and second sliders do not experience observable axial displacement; only alternating load modulation occurs along the lead screw axially through the left and right helical segments on the reaction torque transmission path of the clamped body. During the duration of the disturbance, the q-axis current loop performs high-frequency synchronous sampling of the q-axis current at a frequency of 2kHz. The sampling rate of 2kHz is an integer multiple of the servo position loop main cycle frequency, ensuring that 200 equally spaced current samples are collected within 100ms of each disturbance cycle. The disturbance duration of 0.5s covers 5 complete disturbance cycles, and a total of 1000 current samples are collected and written to the circular buffer as the original q-axis current sequence. The power frequency notch parameters are set as follows: notch center frequency 50Hz, stopband half-width 2Hz, and stopband attenuation not less than 40dB. The notch filter is a second-order infinite impulse response digital filter. Its transfer function zeros and poles are mapped from the analog notch filter prototype to the 2kHz sampling domain through bilinear transformation. The filter coefficients are fixed in the servo driver firmware and executed in the sampling loop. After band-stop filtering of the original q-axis current sequence, the current response sequence is obtained. The energy of the current response sequence in the 50Hz and lower sideband range of 48Hz to 52Hz is reduced by at least two orders of magnitude compared with the original sequence. The energy of the two frequency points corresponding to the perturbation injection, namely the 10Hz fundamental frequency and the 20Hz second harmonic, is not affected by the filtering. The complete load distribution information is retained for feature extraction of the subsequent inversion network.

[0026] In one specific embodiment, step S3 involves using an inversion network with embedded beam bending priors to concatenate the fundamental frequency and second harmonic component of the current response sequence, the lead screw angle, and the motor housing temperature into a joint feature, including: The current response sequence is processed by Fast Fourier Transform to obtain the fundamental frequency amplitude, fundamental frequency phase, second harmonic amplitude, and second harmonic phase; Based on the temperature coefficient of copper resistance and the temperature of the motor housing, the root mean square value of the current response sequence is subjected to copper resistance temperature compensation processing to obtain the temperature compensation current value. The fundamental frequency amplitude, fundamental frequency phase, second harmonic amplitude, second harmonic phase, temperature compensation current value, lead screw angle, and motor housing temperature are spliced ​​together in a predetermined order to obtain a joint feature vector. Based on the pre-stored training set statistics, the joint feature vector is standardized with zero mean and unit variance to obtain the joint features.

[0027] Specifically, the current response sequence has a length of 1000 sample points, corresponding to the cumulative number of samples at a sampling rate of 2000 Hz under a 0.5-second perturbation duration. During frequency domain processing, zeros are padded to the end of the sequence to 1024 points to match the requirement of the Fast Fourier Transform (FFT) radix-2 algorithm for data length to be an integer power of 2. A Hanning window is used to suppress spectral leakage. The frequency resolution is obtained by dividing the sampling rate by the number of transform points, i.e., 1.95 Hz per spectral line. The fundamental frequency amplitude is taken as the peak amplitude of the spectral line corresponding to the perturbation frequency of 10 Hz and its two adjacent spectral lines. The fundamental frequency phase is obtained by the arctangent operation of the complex value of the spectral line and reduced to the range of 0 to 2 times pi. The second harmonic amplitude and phase are extracted in the same way at the 20 Hz spectral line position. The copper resistance temperature coefficient is set at 0.00393 Kelvin, which is the nominal temperature coefficient of electrical pure copper at a reference temperature of 20 degrees Celsius. The reference temperature is set at 25 degrees Celsius, corresponding to the calibrated temperature of the winding resistance in the motor's factory condition. The reference winding resistance is set at 2.4 ohms, corresponding to the DC resistance of the selected servo motor's U-phase winding at the reference temperature. The copper resistance temperature compensation process is based on the relationship that the current winding resistance is equal to the reference winding resistance multiplied by 1 plus the copper resistance temperature coefficient multiplied by the difference between the motor housing temperature and the reference temperature. The temperature compensation current value is obtained by multiplying the root mean square value of the current response sequence by the ratio of the reference winding resistance to the current winding resistance. This process normalizes the current amplitude deviation caused by the copper resistance drift under different heating conditions to the 25-degree Celsius reference condition, ensuring that the combined characteristics maintain temperature invariance during long-term continuous switching operations.

[0028] The joint feature vector is concatenated into a 7-dimensional column vector in the following order: fundamental frequency amplitude, fundamental frequency phase, second harmonic amplitude, second harmonic phase, temperature compensation current value, lead screw angle, and motor housing temperature. The predetermined order follows the rule of prioritizing frequency domain observations over time domain observations, electrical quantities over mechanical quantities, and dynamic quantities over slow variables. This order is strictly consistent with the alignment order of the label data during the training phase to avoid distortion of the inversion output caused by dimensional misalignment during controller inference. The training set statistics include the mean vector and standard deviation vector of each dimension of the training samples. These statistics are obtained by statistically analyzing 72,000 samples collected under 144 working conditions, including 4 robot arm postures, 4 clamp wear degrees, 3 lead screw lubrication states, and 3 ambient temperatures, dimension by dimension. After the statistics are completed, they are fixed in the controller's non-volatile memory area in the form of floating-point constants. After the controller is powered on, they are loaded into a read-only register and will not be updated throughout the entire product lifecycle. Zero-mean unit variance standardization involves subtracting the mean vector from the current joint feature vector dimension by dimension and then dividing by the standard deviation vector to obtain the joint features. The standardized joint features have comparable numerical scales across all dimensions, preventing the regression head gradient from concentrating on a few dimensions with large magnitudes due to excessively large differences in input amplitudes during the training and inference phases of the inversion network. The controller re-executes the complete chain of Fourier transform, copper resistance temperature compensation, vector concatenation, and standardization in each inference cycle, with the entire process completed within a 0.5 millisecond master control cycle.

[0029] Figure 2 This is a schematic diagram comparing the amplitude of the spectral components of the q-axis current response sequence under different load distribution ratios in the embodiments of this application. Figure 2 As shown, the horizontal axis represents the frequency points corresponding to the current response sequence after fast Fourier transform processing, and the vertical axis represents the amplitude values ​​at each frequency point. In the figure, the bars filled with diagonal lines represent the amplitude under the load distribution ratio of 0.30, the bars filled with grids represent the amplitude under the load distribution ratio of 0.55, and the bars filled with dots represent the amplitude under the load distribution ratio of 0.85. The amplitude of the 10 Hz fundamental frequency corresponding to the disturbance injection is 0.085 A, 0.098 A, and 0.110 A under the three load distribution ratios, respectively. The amplitude increases approximately linearly with the load distribution ratio from 0.30 to 0.85. The amplitude of the 20 Hz second harmonic corresponding to the disturbance injection is 0.012 A, 0.032 A, and 0.058 A under the three load distribution ratios, respectively. The amplitude increases non-linearly with the load distribution ratio from 0.30 to 0.85, and the amplitude increment is greater than that corresponding to the 10 Hz fundamental frequency. The amplitudes at the three non-disturbance frequency points of 5 Hz, 15 Hz, and 25 Hz do not exceed 0.011 A under the three load distribution ratios. Figure 2It can be seen that the offset law of the 20 Hz second harmonic amplitude under different load distribution ratios is consistent with the energy coupling at this frequency point of the nonlinear change of the algebraic sum of the counter-torques of the left and right rotating segments of the bidirectional screw. After simultaneously incorporating the 10 Hz fundamental frequency amplitude and the 20 Hz second harmonic amplitude into the predetermined order of the joint feature vector, the input dimension of the inversion network simultaneously carries the linear offset information and nonlinear offset information of the load distribution.

[0030] In one specific embodiment, step S3 involves regressing the joint features using an inversion network that embeds beam bending priors, including: The current response sequence is downsampled at equal intervals to obtain the current timing input. The current timing input is fed into a two-layer one-dimensional convolutional neural network for spectral morphology feature extraction to obtain a local spectral feature map. The local spectral feature map is input into the gated cyclic unit for phase evolution modeling during the perturbation cycle to obtain the temporal hidden state vector; After concatenating the temporal hidden state vector and joint features through channels, the input is given to a two-layer fully connected regression head for nonlinear mapping to obtain the candidate regression output.

[0031] Specifically, the current timing input is obtained by downsampling a current response sequence with a length of 1000 sample points at equal intervals. The downsampling factor is 2, that is, one sample is retained every other sample, resulting in a current timing input with 500 sample points. Before downsampling, a finite impulse response low-pass filter with a cutoff frequency of 0.4 times the Nyquist frequency after sampling is applied to the current response sequence to suppress aliasing. The 500 sample points correspond to uniformly distributed observation samples within a 0.5-second perturbation window, covering 5 complete perturbation cycles, and retaining 100 samples in a single perturbation cycle to ensure the discriminability of waveform details within the cycle in subsequent convolutional layers. The first layer of a two-layer one-dimensional convolutional neural network has a kernel length of 7, 16 output channels, ReLU activation function, and a stride of 1. No pooling is applied after convolution to preserve the temporal length. The second layer has a kernel length of 5, 32 output channels, ReLU activation function, and a stride of 2. After the second convolution, the temporal length is reduced to 250, and the number of channels is 32, resulting in a local spectral feature map. The kernel lengths of 7 and 5 are determined according to the principle of covering the fundamental frequency half-cycle and the second harmonic half-cycle of the disturbance. The first layer covers the rising or falling waveform within the fundamental frequency half-cycle, and the second layer covers the waveform details within the second harmonic half-cycle. The cascading of the two layers allows the local spectral feature map to simultaneously carry the fundamental frequency envelope shape and the second harmonic phase shift information.

[0032] The gated recurrent unit (ROU) employs a two-layer stacked structure, with each layer having a hidden dimension of 32. The input consists of 250 time steps of the local spectral feature map expanded along the temporal dimension, with each time step having an input feature dimension of 32. The ROU outputs only the hidden state of the second layer at the final time step as the temporal hidden state vector, with a vector dimension of 32. During the perturbation cycle, phase evolution modeling is performed by passing the accumulated phase offset between adjacent time steps through the update and reset gates of the ROU, allowing the temporal hidden state vector to accumulate phase evolution information over five perturbation cycles at the final moment. Channel splicing is performed along the feature dimension, concatenating the 32-dimensional temporal hidden state vector with the 7-dimensional joint feature obtained from weight 4 to obtain a 39-dimensional fused feature vector. The two fully connected regression heads have two layers: the first layer has an input dimension of 39, an output dimension of 32, and uses ReLU activation; the second layer has an input dimension of 32, an output dimension of 1, and uses Sigmoid activation. Sigmoid activation compresses the output to an open / closed interval of 0 to 1, yielding candidate regression outputs. These candidate outputs physically correspond to a preliminary estimate of the weight distribution of the clamped body at the second clamping point. The entire regression chain takes 0.6 milliseconds for one inference iteration on the controller. All weights are stored in the controller's non-volatile memory as 32-bit floating-point constants after convergence via backpropagation during the training phase and are not updated during the inference phase.

[0033] Figure 3 This is a schematic diagram of the training loss curve of the inversion network with embedded beam bending priors in an embodiment of this application. Figure 3 As shown, the horizontal axis represents the number of training epochs of the inversion network under the backpropagation algorithm. A single epoch refers to the iteration cycle corresponding to the completion of one forward inference and one backward parameter update for all samples in the training set. The vertical axis represents the value of the loss function of the inversion network in the current training epoch. The loss function value is obtained by weighting and summing the data-driven loss term and the physical residual loss term according to a predetermined weight coefficient. In the figure, the solid line represents the total loss value of the inversion network in the current training epoch, the dashed line represents the individual value of the data-driven loss term of the inversion network in the current training epoch, and the dotted line represents the individual value of the physical residual loss term of the inversion network in the current training epoch after multiplying by the weight coefficient 0.2.

[0034] In the initial training epochs, the total loss starts at 0.205 and decreases rapidly with each epoch. The initial value of the data-driven loss term is close to the initial value of the total loss. Their decay slopes are essentially the same from epoch 1 to epoch 30, jointly dominating the decrease in the total loss. This stage reflects the initial fitting of the mapping relationship between the joint features and the candidate regression output by the weights of the convolutional layers, gated recurrent units, and fully connected regression heads of the inversion network under backpropagation. The value of the physical residual loss term in epoch 1... The value has fallen to a lower range below 0.018, maintaining a steady and slow decline throughout the training process without exhibiting a rapid decline phase of the same scale as the data-driven loss term. This pattern reflects that the physical residual term, consisting of the difference between the analytical value of the beam deflection under the load at both ends and the measured micromomentum, has imposed an effective physical consistency constraint on the candidate regression output in the initial training stage. This constraint guides the candidate regression output to a value range consistent with the analytical relationship of the Euler-Bernoulli beam equation in the early stage before the inversion network parameters have completed the data-driven fitting. Between rounds 30 and 150, the total loss value gradually converged from 0.05 to around 0.02. During this period, the data-driven loss term and the physical residual loss term decreased synchronously, and their numerical differences gradually narrowed. After round 150, the total loss value, the data-driven loss term value, and the physical residual loss term value all entered a stable convergence range and stabilized around 0.015. There was no phenomenon where one loss term value increased inversely due to a decrease in another loss term value. Figure 3 It can be seen that the inversion network can effectively fit the nonlinear mapping relationship between the joint feature vector and the candidate regression output under the selected sample size and training hyperparameters. The data-driven loss term and the physical residual loss term converge collaboratively without mutual inhibition throughout the training process. This verifies that the method of embedding the difference between the beam deflection analysis value and the measured micromomentum of the clamped body under the load at both ends as the physical residual term into the loss function can both preserve the inversion network's ability to extract the data-driven pattern from the joint features and maintain the physical consistency between the candidate regression output and the actual deformation of the clamped body during the parameter optimization process of the inversion network.

[0035] In one specific embodiment, step S3 uses the difference between the analytical value of the beam deflection of the clamped body under loads at both ends and the measured micromomentum as the regression output of the joint feature of the physical residual constraint, to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp, including: Based on the candidate regression output and the self-weight of the clamped body, the load components at the first clamp and the load components at the second clamp are allocated and calculated to obtain the first end load component and the second end load component. Based on the Euler-Bernoulli beam equation and the moment of inertia, elastic modulus and effective length of the clamped body, the midpoint deflection of the first end load component and the second end load component is calculated to obtain the analytical value of beam deflection. The displacement of the midpoint of the clamped body is synchronously detected and processed by the inertial measurement unit on the wrist of the welding robot to obtain the measured micro-motion. Based on the physical residual term formed by the difference between the analytical value of beam deflection and the measured micromomentum, the candidate regression output is subjected to physical consistency constraint processing to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp.

[0036] Specifically, the weight of the clamped body is taken as 83.4 N, corresponding to the total load of the spring clamp and welding gun assembly carried by the selected robot arm under the gravitational field. This value is weighed at the factory and written into the controller configuration area before switching operations. The allocation calculation is performed according to the algebraic relationship that the second end load component is equal to the candidate regression output multiplied by the weight of the clamped body, and the first end load component is equal to 1 minus the candidate regression output multiplied by the weight of the clamped body. When the candidate regression output value is 0.78, the second end load component is 65.1 N and the first end load component is 18.3 N. The mechanical modeling of the clamped body under loads at both ends adopts a simply supported beam model. The first end load component acts on the end of the clamped body held by the first clamping plate, and the second end load component acts on the other end of the clamped body held by the second clamping plate. The moment of inertia of the section is taken as 2.1 x 10^-8 cubic meters to the power of 4, corresponding to the geometric parameters of the circular section of the clamped body. The elastic modulus is taken as 200 gigapascals, corresponding to the nominal elastic modulus of the alloy steel material selected for the clamped body at room temperature. The effective length is taken as 0.32 meters, corresponding to the axial distance between the center lines of the clamping ends of the clamped body. The analytical expression of the midpoint deflection of the Euler-Bernoulli beam equation under concentrated loads at both ends is derived from beam theory. The midpoint deflection is equal to the product of the first end load component and the second end load component, multiplied by the cube of the effective length, divided by 48 times the product of the elastic modulus and the moment of inertia of the section, and then divided by the self-weight of the clamped body. Substituting the above values ​​into the analytical expression, the analytical value of the beam deflection is obtained.

[0037] The inertial measurement unit (IMU) is installed on the wrist of the welding robot near the center of the second clamping plate. It includes a 3-axis accelerometer and a 3-axis gyroscope. The accelerometer range is set to twice the gravitational acceleration, the zero-bias stability is set to 10 microgravity accelerations per hour, and the sampling rate is set to 1000 Hz. The wrist-mounted coordinate system is factory-calibrated, and the alignment error with the axis of the clamped body is within 0.1 degrees. The displacement offset is calculated by performing two time integrations on the acceleration time series data within the disturbance injection window along the direction perpendicular to the axis of the clamped body. Before integration, a high-pass filter with a cutoff frequency of 1 Hz is applied to the acceleration time series data to remove the gravity component and low-frequency drift. After two integrations, the displacement time series data at the midpoint of the clamped body along the vertical direction is obtained. The peak-to-peak value of this displacement time series data within the disturbance window is taken as the measured micromotor. The physical residual term is obtained by subtracting the measured micromomentum from the analytical value of the beam deflection. During the inference phase, the controller feeds back the physical residual term as a correction signal to the candidate regression output. The correction is calculated by adding the physical residual term to the candidate regression output and the conversion factor of the weight of the clamped body (calculated as 3 times the effective length cubed divided by 48 times the product of the elastic modulus and the moment of inertia of the section) to obtain the corrected regression value. This corrected regression value is then hard-limited in the 0-1 range to obtain the load distribution ratio of the clamped body between the first and second clamps. The physical consistency constraint processing ensures that the purely data-driven candidate regression output is still constrained by the analytical relationship of the Euler-Bernoulli beam equation under three types of working conditions not covered by the training set: lubrication attenuation of the lead screw, increased clamp wear, and atypical robot posture. The output range of the load distribution ratio is consistent with the physical reachable boundary.

[0038] Figure 4 This diagram illustrates the comparison of the root mean square error of load distribution ratio inversion using three inversion methods under four typical operating conditions in this application embodiment. Figure 4As shown, the horizontal axis represents the inversion working condition category, from left to right: the baseline working condition, the lead screw lubrication attenuation working condition, the clamp wear increase working condition, and the non-typical working condition of the robot's posture. The vertical axis represents the root mean square error of the load distribution ratio inversion result relative to the true value under each working condition. In the figure, the diagonally filled cylinders represent the root mean square error of the pure data-driven model, the grid-filled cylinders represent the root mean square error of the general partial differential equation soft constraint method, and the dot-filled cylinders represent the root mean square error of the scheme of this application. Under the baseline operating condition, the root mean square errors (RMS) of the three inversion methods are 0.045, 0.041, and 0.038, respectively, and the values ​​are close. Under the condition of decreased lubrication of the lead screw, the RMS of the three inversion methods are 0.128, 0.092, and 0.046, respectively. The RMS of the pure data-driven model increases by approximately 1.84 times compared to the baseline operating condition, while the RMS of the proposed solution increases by approximately 0.21 times compared to the baseline operating condition. Under the condition of increased wear of the clamping plate, the RMS of the three inversion methods are 0.142, 0.103, and 0.052, respectively. Under the condition of atypical robot posture, the RMS of the three inversion methods are 0.165, 0.118, and 0.058, respectively. Figure 4 It is known that the pure data-driven model exhibits a significant increase in the root mean square error of extrapolation drift when the training set is not covered. The general partial differential equation soft constraint method shows some convergence in extrapolation drift under the same conditions, but the convergence range is limited. In this application, after using the difference between the analytical value of the beam deflection under the load at both ends and the measured micromomentum as the physical residual term to constrain the regression output of the joint feature, the root mean square error of the load distribution ratio inversion falls within the range of 0.06 under the three types of training set uncovered conditions, and the regression consistency of the inversion network under extrapolation conditions is maintained.

[0039] In one specific embodiment, step S4 includes: Based on the preset load allocation threshold and the number of consecutive trigger cycles, the load allocation ratio is subjected to timing stability determination processing to obtain the first clip release trigger signal; Based on the release trigger signal of the first clamping piece, a reverse command is applied to the drive motor to obtain the disengagement state between the first clamping piece and the outer conical surface of the clamped body; Based on the self-locking inequality relationship between the lead angle and the secondary friction angle of the bidirectional lead screw, the support position of the second clamping piece against the bidirectional lead screw in the disengaged state is subjected to self-locking treatment to obtain the axial locking state of the second clamping piece. Based on the displacement trend of the clamped body relative to the second clamping piece along the axial direction under its own weight, the wedge-shaped conical surface of the second clamping piece and the outer conical surface of the clamped body are wedge-fitted and pressed together to obtain the clamped body in a bottom-clamping state.

[0040] Specifically, the preset load distribution threshold is set to 0.85, and the number of consecutive trigger cycles is set to 20 main control cycles totaling 10 milliseconds. The timing stability judgment process outputs the first clamp release trigger signal when the load distribution ratio is not lower than 0.85 within 20 consecutive main control cycles. The threshold of 0.85 ensures that not less than 85% of the weight of the clamped body is borne by the second clamp, leaving a safety margin of no more than 15% to cope with instantaneous fluctuations in the inversion network. The number of consecutive trigger cycles of 20 ensures that a single threshold crossing caused by instantaneous noise will not immediately trigger release. When the load distribution ratio falls between 0.60 and 0.85, the lead screw keeps the current rotation angle unchanged and re-enters the disturbance injection retest. When the load distribution ratio is lower than 0.60 and shows a downward trend within 10 consecutive main control cycles totaling 5 milliseconds, the backtracking process is triggered, causing the drive motor to rotate 18 degrees forward, corresponding to the second clamp advancing 0.2 mm axially to strengthen the clamping. The drive motor reverse command has a rotation speed of 60 revolutions per minute and a rotation angle of 2 revolutions, or 720 degrees. This corresponds to the first slider retracting 8 millimeters along the screw axis. The radial gap between the V-shaped conical surface of the first clamping piece and the outer conical surface of the clamped piece opens from 0 in the wedge-engaged state to a level greater than 0.1 millimeters, which is greater than the elastic recovery amount of the initial contact indentation between the outer conical surface of the clamped piece and the conical surface of the first clamping piece. This results in the disengagement state between the first clamping piece and the outer conical surface of the clamped piece. The controller uses the main rotation angle of the screw reaching the disengagement target angle as the determination node for the disengagement state.

[0041] The lead screw has a lead of 4 mm, an outer diameter of 20 mm, and a friction coefficient of 0.15. The lead angle is obtained by dividing the lead by pi and multiplying by the arctangent of the outer diameter, resulting in 3.64 degrees. The friction angle is obtained by calculating the arctangent of the friction coefficient, resulting in 8.53 degrees. The self-locking inequality relationship states that the lead angle of 3.64 degrees is strictly less than the friction angle of 8.53 degrees. The lead screw pair remains locked and does not rotate under any external force applied along the axial direction to the first or second slider. The self-locking locking process relies on this geometric relationship to lock the axial position of the second slider by the lead screw pair itself in the disengaged state. The second clamp does not move axially backward against the support position of the second slider, thus achieving the axial locking state of the second clamp. The wedge-shaped conical surface of the second clamping piece has a half-cone angle of 12 degrees. The outer conical surface of the clamping body and the wedge-shaped conical surface are fitted with the same half-cone angle of 12 degrees. The half-cone angle is determined by the trade-off between the complementary angle of the secondary friction angle and the amplification ratio of the wedge engagement. When the angle is less than 10 degrees, the radial clamping force increment is insufficient to support the 83.4 Newtons of self-weight of the clamping body. When the angle is greater than 15 degrees, the axial reaction force generated by the wedge engagement breaks through the self-locking boundary of the screw, causing the second slider to undergo a reverse displacement. Under its own weight, the clamped body tends to displace outwards from the large-diameter end of the second clamping piece along its own axis. This displacement tendency is measured to be 0.3 mm. The radial clearance change is obtained by multiplying the displacement tendency by the tangent of the half-cone angle of 12 degrees. The contact stiffness between the outer conical surface of the clamped body and the wedge-shaped conical surface is taken as 800 N per millimeter. The additional radial clamping force is obtained by multiplying the contact stiffness by the radial clearance change, resulting in 51.2 N. This additional radial clamping force is superimposed on the insufficient clamping force established by the second clamping piece before disengagement, so that the total radial clamping force on the clamped body is restored to a level exceeding the critical clamping force of 75 N required for the clamped body not to fall off, thus obtaining the bottom clamping state of the clamped body. The entire wedge clamping process is completed passively at the physical level by relying on the bidirectional screw self-locking geometry and V-shaped wedge geometry.

[0042] The above describes the timing control method for the lead screw clamping of the welding robot manipulator in the embodiments of this application. The following describes the timing control system for the lead screw clamping of the welding robot manipulator in the embodiments of this application. One embodiment of the timing control system for the lead screw clamping of the welding robot manipulator in the embodiments of this application includes: The drive module is used to drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the clamped body that has been in pre-contact with the second clamping piece, and obtains a pre-contact state. The acquisition module is used to inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, acquire the q-axis current response of the drive motor, and obtain the current response sequence. The analysis module is used to combine the fundamental frequency and second harmonic component of the current response sequence, the lead screw angle and the motor housing temperature into a joint feature by an inversion network with embedded beam bending priors. The difference between the analytical value of the beam deflection of the clamped body under loads at both ends and the measured micromomentum is used as a physical residual term to constrain the regression output of the joint feature, so as to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp. The clamping module is used to trigger the release of the first clamping piece according to the load distribution ratio, and the self-locking lead angle of the bidirectional screw and the wedge-shaped cone surface of the second clamping piece perform bottom clamping on the clamped body.

[0043] This invention also provides a timing control device for the lead screw clamping of a welding robot, which can be a server. The device includes a processor, memory, display screen, input device, network interface, and database connected via a system bus. The processor, designed as a computer, provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores the data corresponding to this embodiment. The network interface communicates with external terminals via a network connection. The computer program, when executed by the processor, implements the above-described method.

[0044] The present invention also provides a computer-readable storage medium, which can be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium, wherein the computer-readable storage medium stores instructions that, when the instructions are executed on a computer, cause the computer to perform the steps of the welding robot manipulator screw clamping timing control method.

[0045] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0046] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a welding robot's manipulator screw clamping timing control device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0047] 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 method for timing control of lead screw clamping in a welding robot manipulator, characterized in that, The method includes: Step S1: Drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the pre-contacting body with the second clamping piece, and obtains the pre-contact state; Step S2: Inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, collect the q-axis current response of the drive motor, and obtain the current response sequence; Step S3: By using an inversion network with embedded beam bending priors, the fundamental frequency and second harmonic component of the current response sequence, the screw angle and the motor housing temperature are spliced ​​together as joint features. The difference between the beam deflection analysis value and the measured micromomentum of the clamped body under loads at both ends is used as a physical residual term to constrain the regression output of the joint features, thereby obtaining the load distribution ratio of the clamped body between the first clamp and the second clamp. Step S4: Trigger the release of the first clamping plate according to the load distribution ratio, and use the self-locking lead angle of the bidirectional screw and the wedge-shaped cone of the second clamping plate to clamp the object.

2. The welding robot manipulator screw clamping timing control method according to claim 1, characterized in that, Step S1 includes: Based on the predetermined pick-up and place posture of the welding robot wrist and the centering error threshold of the fixed parking mechanism, the welding robot wrist is centered to obtain the centered wrist posture. Based on the orientation of the wrist position after centering, a forward rotation command is applied to the drive motor to obtain the synchronous mirror displacement of the bidirectional lead screw; Based on the synchronous mirror displacement, the wedging state between the first clamping piece and the outer conical surface of the clamped body is released, and the second clamping piece is brought into contact with the outer conical surface of the clamped body to obtain a light contact state of the second clamping piece. Based on the step signal of the q-axis current of the drive motor, the second clamp light contact state is processed to obtain the pre-contact state.

3. The welding robot manipulator screw clamping timing control method according to claim 1, characterized in that, Step S2 includes: Based on the reference value of the main rotation angle of the lead screw in the pre-contact state, the position ring reference angle of the drive motor is sinusoidally superimposed to obtain a low-amplitude sinusoidal disturbance signal. Based on the low-amplitude sinusoidal disturbance signal, the bidirectional lead screw is subjected to disturbance injection processing to obtain the torsional response state of the lead screw pair under disturbance excitation. Based on the torsional response state of the lead screw pair, the q-axis current loop of the drive motor is subjected to high-frequency synchronous sampling processing to obtain the original q-axis current sequence. Based on the power frequency notch parameters, the original q-axis current sequence is subjected to band-stop filtering to obtain the current response sequence.

4. The welding robot manipulator screw clamping timing control method according to claim 1, characterized in that, In step S3, an inversion network embedding the beam bending prior is used to concatenate the fundamental frequency and second harmonic component of the current response sequence, the lead screw angle, and the motor housing temperature into a joint feature, including: The current response sequence is processed by Fast Fourier Transform to obtain the fundamental frequency amplitude, fundamental frequency phase, second harmonic amplitude, and second harmonic phase; Based on the temperature coefficient of copper resistance and the temperature of the motor housing, the root mean square value of the current response sequence is subjected to copper resistance temperature compensation processing to obtain the temperature compensation current value. The fundamental frequency amplitude, the fundamental frequency phase, the second harmonic amplitude, the second harmonic phase, the temperature compensation current value, the lead screw angle, and the motor housing temperature are spliced ​​together in a predetermined order to obtain a joint feature vector; Based on the pre-stored training set statistics, the joint feature vector is standardized with zero mean and unit variance to obtain the joint features.

5. The welding robot manipulator screw clamping timing control method according to claim 4, characterized in that, In step S3, the joint features are regressed using an inversion network that embeds beam bending priors, including: The current response sequence is downsampled at equal intervals to obtain the current timing input; The current timing input is fed into a two-layer one-dimensional convolutional neural network for spectral morphology feature extraction to obtain a local spectral feature map. The local spectral feature map is input into a gated cyclic unit for perturbation-period phase evolution modeling to obtain a temporal hidden state vector; After concatenating the temporal hidden state vector and the joint feature through channels, the result is input into a two-layer fully connected regression head for nonlinear mapping to obtain the candidate regression output.

6. The welding robot manipulator screw clamping timing control method according to claim 5, characterized in that, In step S3, the difference between the analytical value of the beam deflection of the clamped body under loads at both ends and the measured micromomentum is used as the physical residual term to constrain the regression output of the joint feature, thereby obtaining the load distribution ratio of the clamped body between the first clamp and the second clamp, including: Based on the candidate regression output and the weight of the clamped body, the load components at the first clamping plate and the load components at the second clamping plate are allocated and calculated to obtain the first end load component and the second end load component. Based on the Euler-Bernoulli beam equation and the moment of inertia, elastic modulus, and effective length of the clamped body, the midpoint deflection of the first end load component and the second end load component is calculated to obtain the analytical value of the beam deflection. The displacement offset of the midpoint of the clamped body is synchronously detected and processed by the inertial measurement unit of the welding robot's wrist to obtain the measured micro-motion. Based on the physical residual term formed by the difference between the beam deflection analytical value and the measured micromomentum, the candidate regression output is subjected to physical consistency constraint processing to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp.

7. The welding robot manipulator screw clamping timing control method according to claim 1, characterized in that, Step S4 includes: Based on a preset load allocation threshold and the number of consecutive trigger cycles, the load allocation ratio is subjected to timing stability determination processing to obtain the first clip release trigger signal; Based on the release trigger signal of the first clamping piece, a reverse command is applied to the drive motor to obtain the disengagement state between the first clamping piece and the outer conical surface of the clamped body; Based on the self-locking inequality relationship between the lead angle and the secondary friction angle of the bidirectional screw, the support position of the second clamping piece against the bidirectional screw in the disengaged state is subjected to self-locking locking treatment to obtain the axial locking state of the second clamping piece. Based on the displacement trend of the clamped body relative to the second clamping piece along the axial direction under its own weight, the wedge-shaped conical surface of the second clamping piece and the outer conical surface of the clamped body are wedge-fitted and pressed together to obtain the bottom clamping state of the clamped body.

8. A timing control system for the lead screw clamping of a welding robot manipulator, characterized in that, For implementing the welding robot manipulator screw clamping timing control method as described in any one of claims 1-7, the welding robot manipulator screw clamping timing control system comprises: The drive module is used to drive the bidirectional lead screw to rotate forward, so that the first clamping piece releases from the clamped body that has been in pre-contact with the second clamping piece, and obtains a pre-contact state. The acquisition module is used to inject a low-amplitude sinusoidal disturbance into the bidirectional lead screw, acquire the q-axis current response of the drive motor, and obtain the current response sequence. The analysis module is used to combine the fundamental frequency and second harmonic component of the current response sequence, the lead screw angle and the motor housing temperature into a joint feature by an inversion network with embedded beam bending priors. The difference between the analytical value of the beam deflection of the clamped body under loads at both ends and the measured micromomentum is used as a physical residual term to constrain the regression output of the joint feature, so as to obtain the load distribution ratio of the clamped body between the first clamp and the second clamp. The clamping module is used to trigger the release of the first clamping piece according to the load distribution ratio, and the self-locking lead angle of the bidirectional screw and the wedge-shaped cone surface of the second clamping piece perform bottom clamping on the clamped body.

9. A timing control device for the lead screw clamping of a welding robot manipulator, characterized in that, The method includes a memory and a processor, wherein the memory stores a computer program that can run on the processor, and the processor executes the computer program to implement the welding robot manipulator screw clamping timing control method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is run by the processor, it causes the processor to execute the welding robot manipulator screw clamping timing control method as described in any one of claims 1 to 7.

Images