An arc pulsation suppression servo control method for a welding process

Abstract

The application provides a welding process-oriented arc pulsation suppression servo control method. The method collects a welding current waveform signal in real time and extracts a pulsation feature, calculates a pulsation disturbance torque estimation value from an alternating current increment signal of the current based on a pre-calibrated transfer function model of the welding current to the joint disturbance torque, and generates a same-frequency opposite-phase feedforward compensation torque instruction; in a short-circuit transition stage, the method switches to an impact torque template compensation and dynamically improves servo stiffness parameters; and the feedforward compensation torque, the feedback torque and the gravity compensation torque are superimposed and sent to the current loop for execution. The method changes the servo system from a passive feedback correction mode to an active feedforward compensation mode based on the welding condition, and realizes the advance suppression of the arc pulsation disturbance without changing the closed-loop stability.

Description

Technical Field

[0001] This application relates to the field of welding automation control technology, and in particular to a servo control method for suppressing arc pulsation in welding processes. Background Technology

[0002] In pulsed MIG welding, the periodic high and low pulse currents output by the welding power source, along with the arc reaction force generated by the short-circuit transition of the molten droplets, are transmitted to the joints of the welding robot through the welding torch and tool flange, forming periodic disturbance torques. Existing welding robot servo systems employ a three-loop cascaded PID control architecture (position-speed-current), using a position encoder to detect joint angular displacement deviations and perform feedback correction. However, the inherent delay in the detection-computation-execution link causes a phase lag between the correction torque and the disturbance torque, which cannot completely cancel out the periodic disturbances under high-frequency pulse welding conditions, resulting in steady-state residual vibration. This residual vibration causes minute trajectory jitter at the tip of the welding torch, leading to uneven weld surface texture and excessive weld reinforcement fluctuations. Summary of the Invention

[0003] In a first aspect, this application provides a servo control method for arc pulsation suppression in welding processes, applied to a welding robot equipped with a servo drive system, comprising the following steps: real-time acquisition of the current waveform signal output by the welding power source, followed by analog-to-digital conversion and digital pre-filtering to obtain a welding current numerical sequence; performing arc pulsation feature extraction and condition identification on the welding current numerical sequence to obtain the instantaneous frequency and locked phase of the welding pulse, and detecting the start and end states of the droplet short-circuit transition event; removing the DC bias component from the welding current numerical sequence, extracting the AC incremental signal, and applying it to the joint disturbance torque based on a pre-calibrated welding current. The transfer function model is used to generate a feedforward compensation torque command that is in the same frequency but opposite in phase to the arc pulsation disturbance torque in real time. Specifically, during the normal pulse phase, the estimated disturbance torque values ​​of each joint are obtained from the AC incremental signal through the discretized difference equation of the transfer function model. When a short-circuit transition event is detected, an impact compensation torque command is generated according to a pre-calibrated impact torque pulse template. The feedforward compensation torque command is then superimposed in real time with the feedback torque command output by the servo controller and the gravity compensation torque command to generate a composite torque command, which is then sent to the current loop of the servo driver for execution. This method, by superimposing a feedforward compensation channel based on the welding current waveform before the current loop of the servo driver, utilizes the mechanical transmission delay between the change in welding current and the transmission of arc force to the joint. A reverse compensation torque is pre-injected before the disturbance torque actually acts on the joint, transforming the servo system from a passive feedback correction mode to an active feedforward compensation mode. This achieves proactive suppression of the joint disturbance torque caused by arc pulsation without changing the original closed-loop control characteristics.

[0004] Optionally, the transfer function model is obtained through offline calibration. This offline calibration includes independently performing frequency domain scanning and system identification at multiple current amplitude operating points to obtain the transfer function parameter set corresponding to each operating point. During runtime, linear interpolation is performed between the transfer function parameters of adjacent operating points based on the current welding current amplitude. This amplitude-based calibration and interpolation strategy ensures that the modeling error of the transfer function model across the entire operating current range is below a preset threshold.

[0005] Optionally, the transfer function model adopts a second-order form and is discretized into difference equations using the Tustin bilinear transform. The coefficients of the difference equations are pre-calculated and stored during the offline calibration stage. This discretization method enables the transfer function model to complete the real-time calculation of the disturbance torque through a small number of multiplication and addition operations in each control cycle.

[0006] Optionally, the acquisition of the locked phase of the welding pulse is achieved using a digital phase-locked loop (PLL) algorithm, which includes four stages: normalization preprocessing, phase detection, loop filtering, and numerically controlled oscillation. This PLL can extract accurate phase information of the pulse in real time from noisy welding current waveforms.

[0007] Optionally, the detection of the droplet short-circuit transition event is achieved through a threshold determination of the current change rate, including short-circuit initiation determination, short-circuit end determination, and duration validity verification. This detection logic can identify the event at the first moment of the short-circuit transition and filter out false events caused by noise through an anti-false triggering mechanism.

[0008] Optionally, the generation of the impact compensation torque command includes scaling the pre-calibrated template according to the current drop amplitude and replaying it according to the control cycle. This mechanism enables the feedforward channel to automatically adjust the impact compensation force for short-circuit transition events of different intensities.

[0009] Optionally, an adaptive stiffness adjustment step is also included, which increases the position loop gain during the short-circuit transition and recovers it through exponential decay after the short circuit ends. This mechanism enables the servo system to exhibit higher equivalent mechanical stiffness during the short-circuit transition while avoiding secondary oscillations caused by step parameter switching.

[0010] Optionally, after the impact compensation torque playback is completed, a cosine-gradient weighting function is used to smoothly switch to the normal feedforward mode. This gradual transition mechanism ensures that the feedforward torque command remains continuous during mode switching, avoiding the introduction of new disturbances by a step transition.

[0011] Optionally, the system also includes an online monitoring step for the feedforward effect, which adaptively adjusts the feedforward gain coefficient by tracking the root mean square value of the error and sets a lock-in observation period to prevent parameter oscillation. This adaptive mechanism allows the feedforward gain coefficient to automatically converge to an equilibrium value that matches the current operating conditions.

[0012] Optionally, the integrator of the digital phase-locked loop is saturated and limited, and a reset operation is performed when the lock is continuously lost. This protection mechanism prevents frequency divergence caused by integrator overflow and can automatically restore the locked state under extreme conditions such as process switching.

[0013] Optionally, the DC bias component and the half-amplitude of the current pulse are determined at the end of each complete pulse cycle by acquiring the peak and valley values ​​of the welding current within that cycle. This cycle-by-cycle update strategy enables the DC bias estimation to track the gradual drift of the welding conditions.

[0014] The arc pulsation suppression servo control method for welding processes provided in this application constructs a feedforward compensation channel for the disturbance torque from the welding current to the joint. Utilizing the mechanical delay time difference between the welding current waveform and the transmission of arc force to the joint, a reverse compensation torque is injected before the disturbance reaches the joint. Compared to traditional methods that increase PID gain or add notch filters, this feedforward channel does not alter the transfer function and stability margin of the original closed-loop control system. It can simultaneously handle periodic disturbances in the normal pulse phase and nonlinear impact disturbances in the short-circuit transition phase. Furthermore, it adapts to different welding current amplitude conditions through amplitude calibration and online interpolation, and adaptively adjusts the feedforward gain through an online monitoring mechanism, thereby reducing weld trajectory jitter amplitude and weld reinforcement fluctuation. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the overall process of the arc pulsation suppression servo control method for welding processes provided in the embodiments of this application.

[0016] Figure 2 This is a schematic diagram of the module architecture and data flow of the arc pulsation suppression servo control system provided in an embodiment of this application.

[0017] Figure 3 This is a flowchart illustrating the system initialization and offline calibration stages provided in an embodiment of this application.

[0018] Figure 4 This is a flowchart illustrating the arc pulsation feature extraction and operating condition identification stage provided in an embodiment of this application.

[0019] Figure 5 This is a flowchart illustrating the feedforward compensation torque generation stage provided in an embodiment of this application. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with the accompanying drawings and specific embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0021] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0022] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in detail below with reference to the accompanying drawings.

[0023] In one embodiment, this application provides a servo control method for suppressing arc pulsation in welding processes, applied to an automated welding station consisting of a six-axis welding robot and a digital pulse MIG welding power source. The welding robot is equipped with an AC servo drive system, which employs a DSP-based current-speed-position three-loop cascade control architecture. This method superimposes a dedicated feedforward compensation channel before the current loop input of the servo drive. Using the real-time current waveform signal of the welding power source as input, and after calculation using a pre-calibrated transfer function model of the welding current to the joint disturbance torque, a compensation torque command with the same frequency and opposite phase as the arc pulsation torque is generated in real time. This command is then superimposed on the traditional PID feedback torque and gravity compensation torque at the torque command addition node and sent to the current loop for execution. This method solves the technical problem that the traditional three-loop cascade PID control architecture cannot adequately suppress the periodic joint disturbance torque caused by welding arc pulsation due to the inherent delay in the detection-computation-execution link. It achieves the beneficial effect of advancing the suppression timing of the disturbance torque from after the deviation occurs to before the disturbance reaches the joint without changing the original closed-loop control system's transfer function and stability margin.

[0024] like Figure 1 As shown, this method includes the following steps:

[0025] The system involved in the method consists of five core functional modules, such as... Figure 2As shown. The five functional modules include a real-time welding current acquisition module, an arc pulsation feature extraction and operating condition identification module, a current-pulsation torque mapping model and feedforward torque generation module, an adaptive stiffness adjustment module, and a torque command synthesis and output module. The real-time welding current acquisition module acquires the current waveform signal in real time from the current feedback output port of the welding power source through a high-speed analog interface and outputs a digitized welding current numerical sequence. The arc pulsation feature extraction and operating condition identification module performs real-time signal processing on the digitized current sequence and outputs the pulse instantaneous frequency, lock-in phase, short-circuit transition flag, and current amplitude estimate. The current-pulsation torque mapping model and feedforward torque generation module takes the incremental AC welding current signal as input, calculates the disturbance torque estimate of each joint through the difference equation of the transfer function model, and generates an inverse feedforward compensation torque command. The adaptive stiffness adjustment module dynamically adjusts the position loop gain parameter of the servo controller during short-circuit transition events. The torque command synthesis and output module superimposes the feedforward compensation torque, PID feedback torque, and gravity compensation torque in real time and sends them into the current loop.

[0026] S000, System Initialization and Offline Calibration.

[0027] like Figure 3 As shown, this stage completes the calibration of the transfer function, the calibration of the impact torque template, and the initialization of the control parameters.

[0028] S010, Offline transfer function calibration: Perform frequency domain scanning and system identification independently at multiple current amplitude operating points to obtain the transfer function parameter set corresponding to each operating point.

[0029] The welding robot is stationary with all joints locked in a typical welding posture. The welding power supply is set to calibration mode, outputting a frequency-adjustable sinusoidal current excitation signal. The frequency scanning range covers 10Hz to 1kHz, with 5Hz steps. The amplitude of the current excitation signal covers multiple calibration levels; for example, the number of calibration amplitude operating points is set. The four corresponding calibration current amplitude ranges are as follows: A, A, A and A.

[0030] The calibration employs an independent identification strategy for the operating point of each amplitude value. Arc reaction force. With welding current The relationship between them is nonlinear, with the electromagnetic force component approximately proportional to the square of the current, and the plasma jet force component approximately proportional to the current. At different amplitude levels, the equivalent linear gain from welding current to joint disturbance torque can differ by several times, meaning a fixed set of linear transfer function parameters cannot simultaneously and accurately fit the frequency response across the entire amplitude range. Therefore, at each amplitude level… ( It can independently perform a complete frequency domain scan and system identification.

[0031] At each amplitude level Next, apply the following to each frequency point one by one. The current excitation signal is a sinusoidal wave with a center amplitude. The real-time welding current acquisition module acquires the welding current waveform at a sampling rate of not less than 100kHz. Each joint encoder synchronously records the joint angular displacement deviation. ,in Number the joints. and Fast Fourier Transform (FFT) is performed to extract the amplitude ratio and phase difference at the excitation frequency. Then, the angular displacement frequency response is converted into the equivalent disturbance torque frequency response. In the frequency domain, the joint... The equation of motion is ,in For joints The equivalent moment of inertia under the calibrated attitude, in units of... The value is determined by the robot's dynamic parameters. Therefore, the frequency response amplitude of the welding current to the equivalent disturbance torque of the joint is:

[0032] G_j

[0033] Phase is G_j ,because Introduced in the frequency domain Phase shift. After traversing all frequency points under this amplitude range, the equivalent disturbance torque frequency response dataset of each joint at this operating point is obtained.

[0034] For each amplitude level The frequency response data at the operating point were systematically identified using the least squares method, and the transfer function parameter set of each joint at the operating point was obtained by independent fitting. K_j The transfer function model adopts a second-order element form:

[0035]

[0036] in For joints In the The static gain at each amplitude operating point, in units of ; The natural frequency, in units of ; The damping ratio is dimensionless. The calibration results are stored as a parameter table in the servo driver's non-volatile memory, totaling [number missing]. Each parameter value.

[0037] The choice of the second-order transfer function model form is based on the following physical analysis. The physical source of the arc reaction force consists of two parts: electromagnetic force and plasma jet force. The electromagnetic force is determined by the Lorentz force generated by the welding current flowing in the arc column, and its amplitude is approximately proportional to the square of the current, expressed as: ,in This is a coefficient related to the geometry and permeability of the arc cylinder, with a typical value of approximately N / A². The plasma jet force is determined by the recoil force generated by the directional injection of high-temperature gas from the electric arc, and its amplitude is approximately proportional to the current, expressed as: Typical value N / A. The above two coefficients are only used for physical background explanation and are not involved in online calculation. The arc reaction force is transmitted to each joint through the elastic transmission path of the welding torch rigid body and the robot linkage. The dynamic characteristics of this transmission path can be approximated by a second-order linear system, corresponding to the mass-spring-damping model. The transfer function model of this scheme is not obtained by analytically modeling the arc force and then deriving it step by step, but is directly obtained through end-to-end frequency domain identification in the S010 offline calibration stage. The nonlinear characteristics of the arc force and the dynamic characteristics of the mechanical transmission path are implicitly absorbed into the transfer function parameters. Therefore, the direct metric for model accuracy is the end-to-end Bode plot fitting accuracy in S010, rather than the linearization accuracy of the arc force itself.

[0038] For example, with joint 1 in Taking the calibration results of working point A as an example: N·m / A, rad / s, The static gain The range of values ​​is N·m / A, the closer the joint is to the end, the greater its The larger the value, the longer the lever arm. (This refers to the natural frequency.) The range of values ​​is rad / s, depends on the link stiffness and the end load mass, higher This indicates high stiffness of the transmission path. Damping ratio. The typical range of values ​​is The damping ratio of most robot joints is in The range indicates that vibration decay is relatively slow. Typical parameters for joint 6 are... N·m / A, rad / s, During calibration, the excitation duration at each frequency point must be no less than 20 times the frequency period to ensure a steady-state response. A full-frequency scan of each amplitude level takes approximately 30 minutes. The total time for each gear is approximately 2 hours. The calibration accuracy check standard is: the relative amplitude error between the fitted transfer function model Bode plot and the measured frequency response data should be less than 5%, and the phase error should be less than 10°.

[0039] An alternative implementation of the transfer function calibration is an online adaptive identification model based on recursive least squares. This scheme does not pre-calibrate the transfer function; instead, it uses the welding current and the angular displacement fed back by the encoder as input-output pairs in real time during welding, and identifies the transfer function coefficients online using a recursive least squares algorithm. Its advantages include automatic adaptation to attitude changes and operating condition drift, eliminating the need for offline calibration; its disadvantages include a convergence requirement of approximately 100 to 200 pulse cycles, during which compensation accuracy is poor, and the computational load of matrix operations is approximately 50 times that of the lookup table method. Another alternative implementation is a frequency domain lookup table and sine synthesis method. In the S030 stage, a lookup table for the amplitude and phase of the transfer function is pre-calculated in 1Hz steps within the 20 to 600Hz range. During runtime, the lookup table is performed using the instantaneous frequency as an index, and the feedforward torque is directly synthesized. This method has extremely low computational load, but it assumes the welding current is a pure sine waveform, and accuracy decreases when the current waveform is distorted before and after the short-circuit transition. This embodiment selects the time-domain difference equation path as the primary option. Its core advantage lies in its ability to accurately track any waveform without relying on the assumption that the input is a pure sine signal.

[0040] S020, Short-circuit transition impact torque pulse template calibration: Obtain a standardized impact torque pulse waveform template through statistical analysis.

[0041] The welding power source is set to normal pulsed MIG welding mode. For example, the pulse frequency is 200Hz, the peak current is 350A, and the base current is 80A. Actual welding is performed using a standard test plate, with ER70S-6 welding wire (1.2mm diameter) and a shielding gas mixture of 80% Ar and 20% CO2.

[0042] The real-time welding current acquisition module continuously acquires the welding current waveform at a sampling rate of 100kHz, while high-precision encoders at each joint synchronously record angular displacement data. The encoder resolution is no less than 20 bits. The welding current change rate is then used to... Threshold determination ( di / dt The start and end times of all short-circuit transition events are automatically marked (A / ms). For each short-circuit transition event, the joint angular displacement pulse response waveform is extracted within a time window from 1ms before the short circuit occurs to 2ms after the short circuit ends. This is then analyzed using the equivalent moment of inertia. The angular displacement response is converted into an equivalent disturbance torque impulse response using second-order numerical differentiation, and the conversion method is the same as in S010. Simultaneously, the current drop amplitude for each short-circuit event is recorded. ,in This is the sequence number of the short circuit event.

[0043] The equivalent torque impulse response of no less than 100 short-circuit transition events was statistically averaged to obtain the standard impact torque impulse template for each joint. The unit is N·m, corresponding to the reference drop range. The absolute torque value at the reference drop amplitude. The arithmetic mean of the current drop amplitudes from 100 short-circuit events is taken. The pulse template... The duration is approximately 3ms, corresponding to about 300 sampling points at a 100kHz sampling rate, stored as a discrete array in the non-volatile memory of the servo driver. Simultaneously, statistical characteristic parameters of the short-circuit transition event are recorded, including the average short-circuit duration, peak short-circuit current, and peak current change rate, for reference during runtime short-circuit detection threshold calibration.

[0044] The physical mechanism upon which the impact torque pulse template calibration is based is as follows: During the droplet short-circuit transition phase of pulsed MIG welding, the welding current drops from the peak current to the short-circuit current in a very short time, the arc voltage drops instantaneously to near zero, and then the current rises again to re-ignite the arc. The arc reaction force generated in this process exhibits high amplitude and short-pulse impact characteristics, with a typical torque mutation amplitude reaching 3 to 5 times that of normal welding pulsating torque, and a half-sine pulse width of approximately 1 to 2 ms. The ability of traditional PID controllers to suppress such pulsed disturbances is limited by gain-stability constraints, requiring extremely high gain, but high gain can cause system oscillations. The impact torque pulse template, through offline statistical calibration, solidifies the time-domain characteristics of this nonlinear mutation event's torque into replayable template data, and adapts to short-circuit events of different intensities during operation by amplitude scaling.

[0045] One online learning and expansion mechanism for the impact torque template is as follows: After each short-circuit transition event, the pulse template is updated using an exponentially weighted moving average based on the deviation between the actual angular displacement pulse response fed back by the encoder and the template prediction value. The update formula is as follows: ,in This is the forgetting factor. After approximately 20 short-circuit events, the template automatically adapts to the statistical characteristics of the current welding condition, making it suitable for scenarios where offline calibration cannot be immediately re-executed after changes in wire diameter or shielding gas ratio.

[0046] S030, Feedforward Gain and Stiffness Parameter Initialization: Set initial values ​​for control parameters based on the calibration results of S010 and S020.

[0047] The static gain of the transfer function obtained from S010 Set the initial value of the feedforward gain for each joint. The range of values ​​is Slightly less than 1.0 to retain margin and prevent overcompensation from causing reverse oscillation. Based on the peak amplitude of the impact torque template obtained from S020, set the gain boost factor for adaptive stiffness adjustment. The range of values ​​is Set the stiffness recovery time constant. ms, the range of values ​​is ms.

[0048] For the output of S010 The transfer function parameters are grouped, and the coefficient set of the difference equations for each operating point is pre-calculated using the Tustin bilinear transform. The Tustin bilinear transform converts continuous domain variables... Replace with ,in μs is the control period, corresponding to a 16kHz current loop frequency. Let the auxiliary variable... Substitute it into After expanding the denominator polynomial, let the normalized denominator constant be... This yields explicit expressions for the coefficients of the five difference equations:

[0049]

[0050]

[0051] For example, for joint 1 in Work location A N·m / A, rad / s, , μs. The calculation process is as follows: s -1 , , , , Therefore, we can conclude that: (dimensionless) (dimensionless) N·m / A, N·m / A, Total at all work sites Each float32 coefficient value is stored in the DSP's on-chip RAM.

[0052] Set the current pulse half-amplitude estimate. The initial value is A corresponds to the intermediate operating point of the calibration gear. Before the end of the first complete pulse cycle after system startup, the feedforward torque generation module uses the pre-stored difference equation coefficients corresponding to this initial value; the first online update is valid. Then, it automatically switches to online interpolation mode. The DC bias component is then set. The initial value is A. After system startup, at the end of the first complete pulse cycle, the feature extraction module automatically updates the value to the actual estimate. The historical state variables of the difference equation are initialized to zero: for each joint... N·m, A.

[0053] The initialization parameters for the digital phase-locked loop are set as follows: center frequency. The initial value is set to a typical pulse frequency of 200Hz, and this value is automatically updated by the zero-crossing detection result of S210 after system startup. (The local oscillation phase of the phase detector is also mentioned.) Initialized to zero. Loop filter integrator / accumulator. Initialize to zero. The initial frequency of the numerically controlled oscillator is equal to... The normalized parameter and During the coarse locking stage, the maximum and minimum values ​​of the sliding window are estimated online. They are initially set to 350A and 80A as default values, respectively, and are automatically overwritten by the actual sampled data after the welding power supply starts outputting.

[0054] The cold start transition behavior of the system from initialization to steady-state operation is as follows: Within approximately 5ms of the first pulse cycle when the welding power source begins outputting pulse current, the feedforward torque generation module performs calculations using the initial difference equation coefficients and zero-valued historical states. At this time, the difference equation output is still in the startup transient phase, and the accuracy of the feedforward compensation is limited but will not produce harmful reverse disturbances because the initial state variable being zero corresponds to the direction where the feedforward torque output amplitude is relatively small. Approximately 5ms after the end of the first complete pulse cycle, the S200 output becomes effective for the first time. and The estimated values ​​and coefficients of the difference equation automatically switch to the pre-stored coefficients or interpolation coefficients corresponding to the operating point. Historical state variables naturally converge to the correct values ​​during the recursive process of the difference equation. The digital phase-locked loop completes locking within approximately 18ms, or 4 pulse cycles. Therefore, the cold start transition time of the entire system is approximately 20ms, during which the effective suppression rate of the feedforward compensation gradually increases from approximately 50% to a steady-state value of 85%. This cold start time is much shorter than the stable arc establishment time of approximately 200ms, and does not affect the arc initiation quality during the welding process.

[0055] After initialization, the system enters online operation mode.

[0056] The S100 acquires the current waveform signal output by the welding power source in real time, and obtains the welding current numerical sequence through analog-to-digital conversion and digital pre-filtering.

[0057] S110, Analog Signal Conditioning and Analog-to-Digital Conversion: Converts the analog voltage signal output from the welding power supply current feedback port into a digital quantity.

[0058] The current feedback output port of the welding power supply provides an analog voltage signal from 0 to 10V, linearly corresponding to a welding current from 0 to 500A. This analog voltage signal undergoes impedance matching and common-mode noise suppression via a differential operational amplifier, which has a common-mode rejection ratio (CMRR) of at least 80dB and a bandwidth of at least 500kHz. The signal-conditioned voltage signal is then fed into a 12-bit successive approximation analog-to-digital converter (ADC) with a sampling rate of 100kHz (i.e., a sampling interval of 10μs). The ADC has a reference voltage of 10V and a quantization resolution of [missing information]. mV / LSB, corresponding to a current resolution of A / LSB.

[0059] The selection of the 100kHz sampling rate is based on the following: the typical range of welding pulse frequency is 30 to 500Hz, and the digital phase-locked loop and short-circuit detection algorithm in S200 require a sufficient oversampling rate to ensure phase accuracy and rate of change calculation accuracy. According to the Nyquist theorem, for the 10th harmonic of a 500Hz pulse frequency at 5kHz, a sampling rate of at least 10kHz is required. Considering... The calculations utilize a 4-point spacing to reduce noise. A 100kHz sampling rate provides 200 samples per pulse cycle at 500Hz and 500 samples per pulse cycle at 200Hz, offering sufficient time resolution for phase accuracy. Simultaneously, the 100kHz sampling rate enables the 2kHz low-pass filter to achieve a 50x oversampling ratio, providing a flat amplitude-frequency response within the passband. The group delay introduced by the filter is approximately 0.15ms, significantly smaller than the pulse period.

[0060] The signal conditioning circuit employs a single-point grounding design, merging the ground wires of the analog signal conditioning circuit and the DSP digital circuit at the power input to prevent digital switching noise from coupling to the analog signal channel through the ground loop. The current feedback signal from the welding power supply is transmitted to the servo driver via a shielded twisted-pair cable, with the cable shield grounded at one end on the servo driver side. Electrical isolation between the signal conditioning circuit board and the DSP motherboard is achieved through an optocoupler with a withstand voltage of at least 1500V to prevent high-current interference from the welding circuit from being conducted to the control circuit through the signal cable.

[0061] The analog-to-digital converter outputs a 12-bit unsigned integer. The data is stored in a ping-pong buffer within the DSP. This ping-pong buffer has a double-buffered structure, with each buffer storing 64 sampling points corresponding to a 640μs time window. The DMA controller automatically fills the currently active buffer with the output data from the analog-to-digital converter, switching the active flag every 640μs. Subsequent signal processing modules read data from the inactive buffer, avoiding data contention. The DMA transfer employs a priority arbitration mechanism, with the DMA channel of the analog-to-digital converter having higher priority than other peripheral channels, ensuring that sampled data is not lost.

[0062] S120, Digital Pre-Filtering: Performs calibration transformation and low-pass filtering on the original sampled data.

[0063] First, a linear calibration conversion is performed to convert the raw digital value output by the analog-to-digital converter into a physical current value: The unit is A. Then... A second-order IIR low-pass filter is performed. This second-order IIR low-pass filter is a Butterworth type with a cutoff frequency of 2kHz and a sampling rate of 100kHz, used to filter out quantization noise from the analog-to-digital converter and high-frequency electromagnetic interference. The difference equation of the filter is:

[0064]

[0065] The filter coefficients are derived from the second-order Butterworth low-pass design formula based on the cutoff frequency. rad / s and sampling rate Calculated at kHz. For example, the value of the coefficient is: , , , , The above coefficients are pre-calculated in the S030 stage and fixed in the DSP's on-chip ROM.

[0066] The output welding current numerical sequence is obtained after calibration conversion and low-pass filtering. The unit is ampere-ampere (A), and the update rate is 100 kHz. The welding current numerical sequence is accessible to subsequent modules through a shared RAM area on the DSP chip, employing a single producer-single consumer model without the need for bus arbitration.

[0067] The selection of the 2kHz cutoff frequency is based on the following: the typical upper limit of welding pulse frequency is 500Hz, and its fundamental frequency component and first three harmonic components are all below 2kHz. Therefore, setting the cutoff frequency to 2kHz can completely preserve the time-domain waveform characteristics of the welding pulse. Meanwhile, the high-frequency components above 2kHz are mainly quantization noise from the analog-to-digital converter and high-frequency electromagnetic interference generated by the welding power supply IGBT switching, and do not contain useful pulsation information. If the cutoff frequency is set too low, such as 500Hz, it will truncate the sharpness of the rising and falling edges of the high-frequency pulse, affecting short-circuit detection. Calculation accuracy; if set too high, such as 10kHz, it cannot adequately suppress switching noise, leading to an increased false trigger rate in subsequent feature extraction. The historical state variables of the filter. , , , The filter is initialized to zero when the system starts up. The output of the first two sampling points has a brief startup transient. After about 20μs, the filter enters a steady-state working state.

[0068] For example, when the peak welding current is 350A, the analog-to-digital converter reading is LSB. Obtained through calibration conversion. A. After being low-pass filtered at 2kHz, the signal retains the time-domain waveform characteristics of the welding pulse (the pulse frequency typically ranges from 30 to 500Hz, far below the 2kHz cutoff frequency), while suppressing analog-to-digital converter quantization noise and high-frequency electromagnetic interference components.

[0069] S200, perform arc pulsation feature extraction and working condition identification on the welding current numerical sequence, obtain the instantaneous frequency and locked phase of the welding pulse, and detect the start and end status of the droplet short-circuit transition event.

[0070] like Figure 4 As shown, this stage consists of three parallel sub-steps: real-time pulse frequency tracking, pulse phase locking, and short-circuit transition event detection. The feature extraction operation is deployed on the CPU2 core of the DSP, running independently at a sampling rate of 100kHz, and updating its internal state every 10μs. At the beginning of each 16kHz control cycle, CPU2 synchronously transmits the latest feature vector to CPU1 via IPC message RAM for use by the feedforward torque generation module.

[0071] S210, Real-time Pulse Frequency Tracking: Extracts the instantaneous frequency of the welding pulse through zero-crossing detection.

[0072] Calculate welding current Relative to base current deviation ,in Estimated using a sliding window minimum filter, the window length is 5 pulse periods. Detection. The positive zero-crossing point, i.e. and At that moment, the zero-crossing time is accurately calculated using linear interpolation. Spacing between adjacent zero-crossing points Instantaneous frequency The unit is Hz.

[0073] When the instantaneous frequency Beyond the valid range At Hz, an abnormal flag is set. The lower limit of the effective frequency range, 20Hz, corresponds to a pulse period of up to 50ms. Frequency below this usually indicates the welding power supply is in DC output mode rather than pulse mode, in which case feedforward compensation is not activated. The upper limit of 600Hz covers the highest pulse frequency specification of current industrial pulsed MIG welding power supplies. Frequency signals exceeding 600Hz are usually electromagnetic interference or harmonic false detections. The calculation method for the zero-crossing time of the linear interpolation is as follows: T_s This interpolation improves the time resolution of the zero-crossing moments from a sampling interval of 10 μs to approximately 0.1 μs, resulting in an instantaneous frequency measurement accuracy better than 0.01% at a pulse frequency of 200 Hz. For example, in a welding operation with a pulse frequency of 200 Hz, the distance between adjacent zero-crossing points is... ms, Hz.

[0074] S220, Pulse Phase Locking: Employs a digital phase-locked loop algorithm to lock the precise phase of the welding pulse.

[0075] First, the welding current signal is preprocessed by normalization: ,in This represents the DC bias component of the current. This represents the half-amplitude of the current pulse. The normalization extracts the AC component of the welding current for phase detection.

[0076] The normalization parameter is determined using a two-stage startup sequence to break the circular dependency between frequency information and the normalization parameter. During the coarse-locking phase, i.e., the first 20ms after system startup, and The maximum and minimum values ​​of the welding current sequence are estimated online using a sliding window. The sliding window length is fixed at 500 sampling points (5ms), covering at least one typical pulse cycle. During the fine-locking stage, after the digital phase-locked loop (PLL) meets the locking condition, the sliding window length is set to the precise locking period. Within each complete pulse cycle, the maximum and minimum values ​​of the welding current sequence are updated accordingly. and .

[0077] The digital phase-locked loop consists of three subunits: a phase detector, a loop filter, and a digitally controlled oscillator. The phase detector outputs the phase error. The output of the multiplication phase detector contains a value proportional to... The DC component and the second harmonic component, among which This represents the phase difference between the fundamental frequency component of the input signal and the local oscillator. The second harmonic component is implicitly suppressed by a subsequent loop filter, whose equivalent bandwidth is... rad / s, for a typical 200Hz pulse frequency doubled to 400Hz, the attenuation is approximately dB can effectively suppress it.

[0078] The loop filter is a second-order PI filter, outputting a frequency correction value. T_s ,in This is an accumulator with saturation limiting. The proportional gain... Integral gain For example, let the damping ratio be... natural frequency rad / s, then s⁻¹, s⁻². Sampling period μs, consistent with the welding current sampling rate.

[0079] The integral accumulator is set with a saturation limit value. ,in Hz represents the maximum permissible frequency deviation. This saturation limiting prevents integrator overflow: when the center frequency deviates significantly from the actual pulse frequency, the integral term is limited to [a certain value]. Within the specified range, ensure that the frequency offset of the CNC oscillator does not exceed [the specified range]. To avoid locking at harmonic frequencies or divergence, when the root mean square value of the phase error of the digital phase-locked loop exceeds the unlock threshold of 0.3 rad for 50 ms consecutively, it is determined to be in an unlocked state, and an unlock reset operation is performed: the integrator accumulator is cleared, and the latest output of S210 is used. Reinitialize the center frequency and revert the normalization parameters to the coarse-locking stage.

[0080] The numerically controlled oscillator updates the local frequency: Cumulative phase: T_s The lockout time is approximately ms, meaning locking is completed within approximately 4 pulse cycles. The locking criterion is the root mean square value of the phase error. rad. After locking This corresponds to the zero-phase point of the fundamental frequency component of the input current waveform. For a pulsed square wave current with a duty cycle of approximately 30%, there is a fixed phase offset between the zero-phase point of the fundamental frequency component and the rising edge of the pulse. This phase offset does not affect the calculation results in the S310 time-domain difference equation path because this path is independent of... .

[0081] For actual pulsed MIG welding current waveforms, the normalized signal contains a fundamental frequency component and higher harmonic components. The multiplicative phase detector extracts the phase information of the fundamental frequency component by multiplying it with a local cosine signal. The locking accuracy depends on the signal-to-noise ratio of the fundamental frequency component rather than the waveform symmetry. Under the condition that the typical signal-to-noise ratio of the pulsed welding current is greater than 20dB, the digital phase-locked loop (PLL) can reliably lock. The cycle-by-cycle update strategy of the normalization parameters directly extracts the peak and valley values ​​in the time domain, without relying on the correspondence between the peak time and the fundamental sinusoidal phase. Therefore, it is completely robust to pulsed square wave waveforms with asymmetrical duty cycles. For example, under an asymmetrical pulse waveform with a pulse frequency of 200Hz, a peak duration of 1.5ms, and a base duration of 3.5ms, the PLL completes locking within approximately 18ms, or 4 pulse cycles, and the root mean square value of the locked phase error stabilizes within 0.03rad.

[0082] S230, Short-circuit transfer event detection: Identify droplet short-circuit transfer events by determining the threshold of the rate of change of welding current.

[0083] Real-time calculation of the first derivative of welding current Use a 4-point spacing to reduce the impact of noise. μs.

[0084] The short circuit initiation determination condition is: and ,in A / ms, with a value range of A / ms. The short-circuit transition flag is set when both of the above conditions are met simultaneously. Record the start time of the short circuit. .

[0085] The short circuit termination condition is: when At that time, if and ,in A / ms, with a value range of A / ms, then set Record the time when the short circuit ends. .

[0086] The threshold and The recommended values ​​are based on statistical analysis of the current change rate over at least 100 short-circuit events under ER70S-6 welding wire, 80% Ar and 20% CO2 shielding gas conditions, using the 95th percentile of the current change rate distribution. For different welding wire diameters and shielding gas ratios, it is recommended to adjust the above thresholds based on actual statistical results during the SO20 calibration phase.

[0087] Anti-false triggering logic: Short circuit duration Must meet Otherwise, it is judged as a noise false trigger and the event is discarded. The lower limit of the duration verification window of 0.5ms excludes instantaneous false triggers caused by analog-to-digital converter quantization noise, and the upper limit of 5ms excludes long-term short circuit events caused by abnormalities such as welding wire adhesion during the welding process (such events need to be handled by the protection logic of the welding power supply itself and are not normal droplet transition behavior).

[0088] The internal state of the short-circuit transition detection logic can be described by a finite state machine containing two states: a normal state and a short-circuit detection state. In the normal state, the system continuously monitors the rate of change of current. When both the rate of change is below a negative threshold and the current value is below a current threshold, the state transitions to the short-circuit detection state and is set. Under short-circuit detection conditions, when both the rate of change is higher than the positive threshold and the current value is higher than the recovery threshold, the state returns to normal and is reset. The recovery transition after the short circuit ends is managed by the cosine gradient window of S320, and the short circuit detection state machine is only responsible for state detection.

[0089] The threshold and The selection of a detector needs to balance detection sensitivity and false trigger rate. Too small a value can lead to false triggering of the normal pulse falling edge, while too large a value can cause low-amplitude short-circuit events to be missed. For example, under ER70S-6 welding wire diameter of 1.2mm and shielding gas conditions of 80%Ar and 20%CO2, the peak current change rate of the normal pulse falling edge is approximately 200 to 350 A / ms, while the peak current change rate of the short-circuit transition is approximately 500 to 1200 A / ms, with a clear dividing range between the two. Setting the threshold to 500 A / ms ensures a distribution spacing greater than 150 A / ms, corresponding to a detection accuracy exceeding 99%. For scenarios with different welding wire diameters, such as 0.8mm diameter wires which experience more frequent short-circuit transitions and a higher rate of change, the threshold can be appropriately increased to 600 to 800 A / ms; while 1.6mm diameter wires which experience fewer short-circuit transitions and a lower rate of change, the threshold can be appropriately decreased to 300 to 500 A / ms.

[0090] For example, in the 17th cycle of the welding pulse, exist ms time reached A / ms, less than The threshold for A / ms, and at the same time A, less than The current threshold of A is used to determine... Record the lowest current value. A. After the short circuit lasted for approximately 1.8ms, Rise to A / ms, greater than the positive threshold of 300 A / ms, and A, greater than A's recovery threshold Restored to 0. ms in This event is valid within the millisecond range.

[0091] The feature vector structure output by S200 includes the following fields: instantaneous pulse frequency. (float32, unit Hz), phase locked (float32, unit: rad), short-circuit transition flag (uint8), minimum current during short circuit (float32, unit A), peak rate of change of current (float32, unit A / ms), Phase-Locked Loop Locking Status Flag (uint8), estimated half-amplitude value of current pulse (float32, in A), and DC bias component. (float32, unit A). The aforementioned At the end of each complete pulse cycle, CPU2 calculates the value within that cycle. Half the difference between the maximum and minimum values ​​of the sequence serves two purposes: the interpolation index of the operating point of the difference equation coefficients in the feedforward torque generation module and the online update of the normalized parameters of the digital phase-locked loop. Synchronized updates within this period and The mean value is used to calculate the AC incremental input of the feedforward torque generation module and to perform normalization preprocessing for the digital phase-locked loop. and The initial values ​​are set to 150A and 100A respectively in stage S030, and are automatically updated to the actual estimated values ​​at the end of the first complete pulse cycle after system startup. The above feature vectors are synchronously transmitted to CPU1 at the beginning of each control cycle via IPC message RAM, and data consistency is ensured by hardware semaphores.

[0092] S300: Remove the DC bias component from the welding current numerical sequence, extract the AC incremental signal, and generate a feedforward compensation torque command that is in the same frequency and opposite phase to the arc pulsation disturbance torque in real time based on the pre-calibrated transfer function model of welding current to joint disturbance torque.

[0093] like Figure 5 As shown, this stage is based on the short-circuit transition flag. The state switches between the feedforward path of the difference equation in the normal pulse phase and the impact template compensation path in the short-circuit transition phase. The feedforward torque generation operation is deployed in the high-priority interrupt service routine of the DSP CPU1 core, triggered by a 16kHz current loop timer interrupt, occupying approximately 15% of the CPU1's computing resources.

[0094] S310, Calculation of feedforward torque during normal pulse phase: The estimated value of the disturbance torque of each joint is obtained from the AC incremental signal of welding current through the discretized difference equation of the transfer function model.

[0095] exist This sub-step is executed during the non-short-circuit transition phase. The real-time welding current acquisition module outputs at a sampling rate of 100kHz. The sequence, and this difference equation is executed with a control period of 16 kHz, the control period μs. At the start of each control cycle. Read the latest data from the shared register. The value, used as the input for the current control cycle, is equivalent to zero-order hold downsampling. The subscripts of the difference equations in this section... To control the cycle Step size, , , These all refer to the sampled values ​​latched from the shared register at the start of the continuous control cycle. This downsampling does not introduce aliasing because the second-order IIR low-pass filter of the S120 has limited the signal bandwidth to within 2kHz, far below the Nyquist frequency of 8kHz for a 16kHz sequence.

[0096] The transfer function This describes the mapping relationship between current pulsation components and joint disturbance torque. The constant arc force generated by the steady-state DC current has been absorbed by the PID feedback loop and gravity compensation, and does not constitute a pulsation disturbance. Therefore, the input to the difference equation must be an incremental AC current signal. ,in The DC bias component of the current is calculated by S200 at the end of each complete pulse cycle and synchronously transmitted to CPU1 via IPC message RAM.

[0097] The operating point selection of the difference equation coefficients during runtime is based on the current welding current amplitude. .according to Determine the adjacent calibration working point interval where it is located. For S030 pre-stored Linear interpolation is performed on the coefficients of the set of difference equations. Let the interpolation weights be... Then the real-time coefficient , , , , Similarly. When Below or higher When this is the case, the pre-stored coefficient of the nearest gear is used. The update rate is equal to the pulse frequency, and the coefficients of the difference equation remain unchanged between two updates.

[0098] For each joint ( ), through the pass function The Tustin discretized difference equation is used to calculate the estimated value of the arc pulsation torque:

[0099]

[0100] The computational complexity of this difference equation involves 5 multiplications and 4 additions, totaling approximately 54 CPU cycles for each of the 6 joints in each control cycle. Dimensional verification of the difference equation is as follows: The dimensions are , The dimensions are All dimensions are consistent.

[0101] Generate inverse feedforward torque command: ,in This is the feedforward gain coefficient, with an initial value of 0.85 and a range of values ​​of [value missing]. It can be adjusted online adaptively via S520.

[0102] For example, a complete numerical calculation is performed using joint 1 at a pulse frequency of 200Hz and an operating point of 150A. The transfer function is... Frequency response at Hz: rad / s, The molecule is . In the denominator: , The denominator is . Current pulse amplitude A. Estimation of peak disturbance moment N·m. Feedforward compensation torque N·m. Residual disturbance torque N·m, which is only 15% of the original perturbation fundamental frequency component, and the fundamental frequency component suppression rate is about 85%.

[0103] Further verification of the transient tracking capability of the difference equation for the leading edge of a real pulse square wave. Let... This corresponds to the moment when the welding current jumps from a base value of 80A to a peak value of 350A. A. Initial state N·m, A. Substitute the coefficients into the difference equation , , , , :

[0104] : A. N·m.

[0105] : A. N·m.

[0106] : A. N·m.

[0107] The above stepwise calculations show that the torque estimate oscillates and rises in the first few control cycles after a step input, reflecting the underdamped transient response characteristics of a second-order system to a step input. If the input remains in a step state for a sufficiently long time... It will asymptotically converge to the steady-state gain value. N·m, the convergence time constant is approximately In actual pulse welding, within the 1.5ms peak duration (approximately 24 control cycles), the difference equation is still in the transient response stage, and the output has not yet fully reached its steady-state value. This is precisely the accuracy advantage of the time-domain difference equation over the pure frequency-domain method when dealing with pulses of finite duration—the difference equation faithfully tracks the transient transition process, rather than assuming the input is an infinitely continuous sinusoidal signal.

[0108] Further verification of the suppression rate under 350Hz high-frequency conditions. The transfer function in... Frequency response at Hz: rad / s, The molecule is . in the denominator , , The denominator is . G_1 N·m / A. Peak value of disturbance torque estimation. N·m. Feedforward compensation torque N·m. Residual disturbance torque The amplitude of the transfer function at 350 Hz is only 15% of the original perturbation, and the suppression rate is also approximately 85%. This result indicates that although the amplitude of the transfer function at 350 Hz is attenuated compared to 200 Hz, the feedforward channel simultaneously estimates the attenuated perturbation torque by the same proportion. The fixed feedforward gain has a theoretical suppression rate of 85% for all frequency components, regardless of frequency.

[0109] The stability of the direct coefficient interpolation strategy is guaranteed as follows: Since the coefficients of the difference equations at each calibration operating point correspond to stable discrete poles located within the unit circle in the z-plane, and the pole positions between adjacent operating points are small, the damping ratio change is less than 0.1, the natural frequency change is less than 20%, and the poles corresponding to the coefficients after convex combination are still located within the unit circle, the system remains stable. When the welding posture changes significantly, i.e., the joint angle change exceeds 30°, the transfer function parameters may shift. It is recommended to re-execute the S010 calibration or enable the attitude-dependent parameter interpolation table. After calibration at 3 to 5 typical attitude points during the S010 calibration phase, linear interpolation can be performed between multiple sets of parameters based on the current joint angle during runtime to achieve attitude adaptation. When the joint angle is between two calibration attitude points, the interpolation error is usually less than 10%, still ensuring a disturbance suppression rate of over 70%.

[0110] S320, Impact compensation torque generation during short-circuit transition: Generates impact compensation torque command based on pre-calibrated impact torque pulse template.

[0111] when The sub-step is triggered when the current changes from 0 to 1. This is based on the current drop magnitude of the current short-circuit event. Reference drop amplitude compared to S020 standard Perform scaling: .by At the starting moment, the pre-stored pulse template is read in steps according to the control cycle. ( , (Point corresponding to 3ms), generate impact compensation torque:

[0112]

[0113] in This represents the control cycle number corresponding to the short-circuit initiation time. During the short-circuit transition, impact compensation replaces normal feedforward.

[0114] After the template playback is complete, in the transition window Within milliseconds, a cosine-gradient weighting function is used to smoothly switch back to normal feedforward mode. Let the template playback end time be... The transition weights are:

[0115] ,

[0116] Within the transition window, the feedforward torque is ,in The normal feedforward torque is calculated synchronously for the S310 difference equation. This is the last output value of the template. The torque command decreases continuously and without abrupt changes from 1 cosine to 0. After the transition is complete, switch to normal feedforward mode completely.

[0117] For example, in a short-circuit transition event, A, A, during calibration A. A, Set the peak value of the calibrated joint 1 pulse template. N·m, the peak value of the impact compensation torque is N·m. During the short-circuit transition, the peak angular displacement of joint 1 decreased from about 0.005° without compensation to about 0.0007°, and the corresponding end displacement decreased from about 0.35 mm to about 0.05 mm, with a suppression rate of about 86%.

[0118] S400, in response to the change of the short-circuit transition flag, dynamically adjusts the stiffness parameters of the servo controller.

[0119] S410, Stiffness Enhancement During Short-Circuit Transition: Instantly boosts the position loop gain to a preset multiple of its normal value.

[0120] When detected When the value changes from 0 to 1, a parameter switch is performed immediately. The position loop proportional gain is increased to... K_{p,0} The differential gain is increased to K_{d,0} ,in , and These represent the proportional gain and derivative gain of the position loop under normal operating conditions. Simultaneously, the integral limit value of the velocity loop is... Increase to normal value To prevent integral saturation, the parameter switching is completed within one control cycle, i.e., 62.5μs.

[0121] For example, suppose under normal operating conditions s⁻¹, s, Improvement during short-circuit transition s⁻¹, The increased gain enables the joint to exhibit higher equivalent mechanical stiffness during short-circuit transition, resulting in a stronger restoring force against angular displacement deviations caused by impact torque.

[0122] The stiffness enhancement factor The selection criteria are as follows: If the value is too small, the effect of stiffness improvement will be not obvious, and the gain on impact suppression will be limited; If the bandwidth is too large, the equivalent bandwidth of the position loop will approach the bandwidth of the speed loop, thus disrupting the bandwidth separation condition of cascade control and causing coupled oscillations between loops. Typical position loop bandwidths are approximately 50 to 100 Hz, and speed loop bandwidths are approximately 300 to 500 Hz. Increasing the equivalent bandwidth of the position loop from approximately 75Hz to approximately 187Hz is still less than half the bandwidth of the velocity loop, thus satisfying the bandwidth separation condition. If Above 4.0, the equivalent bandwidth of the position loop will approach 300Hz, at which point coupling with the velocity loop begins. Therefore... The safe range is The specified value of 2.5 is a compromise between stiffness gain and stability margin.

[0123] The speed loop integral limit value Synchronized upgrade to normal value The reason for the multiple is that the position loop gain increases during the short-circuit transition, which leads to an increase in the speed command amplitude. If the speed loop integral limit is not increased accordingly, the integrator will reach saturation during the short-circuit impact, causing the integrator to need extra time to exit saturation after the impact ends, resulting in an overshoot phenomenon similar to integral saturation in an air conditioning system.

[0124] S420, Stiffness Smooth Recovery: After the short-circuit transition, an exponential decay function is used to restore the gain to its normal value.

[0125] when When restoring from 1 to 0, starting from the beginning of the restoration... Initially, an exponential decay function is used to smoothly recover the PID parameters:

[0126]

[0127] when and When the relative difference is less than 1%, approximately after ms, lock . The same recovery logic is executed. The discrete implementation of the exponential recovery function is as follows: ,in T_c These are pre-calculated constants. This smooth recovery mechanism avoids transient oscillations caused by step parameter changes. The physical significance of this recovery process lies in simulating the natural damping characteristics of a mechanical system: after a short-circuit impact, the servo system gradually recovers from a high-stiffness state to a normal-stiffness state, with the recovery rate changing from a time constant. control. The range of values The selection of ms is based on the following criteria: the recovery time constant should be greater than the typical duration of the short-circuit transient event by 1 to 2 ms to avoid the stiffness from decreasing before the short circuit ends, and should be less than half of the pulse period to ensure that recovery is completed before the next short-circuit event may occur. The 5-fold recovery time corresponding to ms is 15ms, which is 3 pulse cycles at a pulse frequency of 200Hz, providing a sufficient stiffness recovery window between consecutive short-circuit events.

[0128] The entire system operates under a unified control clock, with a control cycle of μs corresponds to a 16kHz current loop frequency. The arc pulsation feature extraction module is deployed on the CPU2 core of the DSP, independently running the digital phase-locked loop and zero-crossing detection algorithm at a 100kHz sampling rate, updating its internal state every 10μs. At the beginning of each 16kHz control cycle, CPU2 synchronously transmits the latest feature vector to CPU1 via IPC message RAM. The execution order of each module on the CPU1 side within each control cycle is as follows: At any given moment, the welding current real-time acquisition module completes one analog-to-digital converter sampling and acquires... Meanwhile, CPU1 reads the latest feature vector of CPU2, which is now ready, through hardware semaphores; μs, the feedforward torque generation module completes the difference equation calculation and outputs... ; μs, the adaptive stiffness adjustment module completes the stiffness parameter adjustment; μs, the torque command synthesis module completes torque synthesis. Feed into the current loop DAC; μs, the next control cycle begins. The total delay from welding current sampling to compensation torque output is 20μs, excluding the asynchronous processing time of the feature extraction module on CPU2, as the feature vector is pre-ready at the start of the control cycle. This 20μs delay only applies to the S310 difference equation path, which directly uses... The input is independent of the frequency and phase information output by the feature extraction module. The maximum delay of the feature extraction module output does not exceed one control cycle, i.e., 62.5μs, and does not affect the time-domain accuracy of the feedforward torque.

[0129] The on-chip RAM data storage layout of the feedforward torque generation module comprises two parts. The difference equation coefficients and state storage occupy approximately 576 bytes, of which... The impact template data storage occupies approximately 1.2KB, consisting of 120 float32 values ​​(approximately 480 bytes) multiplied by 6 joints and 5 coefficients, and approximately 24 float32 values ​​(approximately 96 bytes) multiplied by 6 joints, 2 historical states, and 2 historical inputs. The data includes 288 float32 values ​​(6 joints multiplied by 48 sampling points). The feedforward torque generation module outputs a feedforward command vector in each control cycle. It can be directly written into the register of the torque synthesis module with zero delay.

[0130] S500, the feedforward compensation torque command is superimposed in real time with the feedback torque command and gravity compensation torque command output by the servo controller to generate a composite torque command, and the composite torque command is sent to the current loop of the servo driver for execution.

[0131] S510, three-way torque superposition and amplitude limiting protection.

[0132] Traditional PID controller output feedback torque The reference position is determined by the position loop and velocity loop based on the trajectory planning. With the encoder measured position The deviation is calculated. The S300 output feedforward compensation torque. Gravity compensation module output Perform joint-by-joint vector addition:

[0133]

[0134] Then torque limiting protection is applied: ,in For joints The rated peak torque is determined by the motor parameters. The limited composite torque command is converted into an analog current command by a DAC and sent to the current loop, or directly sent as a digital quantity to the digital current loop controller to drive the IGBT inverter bridge to output the corresponding motor phase current.

[0135] For example, under welding conditions with a pulse frequency of 200Hz and a peak current of 350A, the torque components of joint 1 are as follows: N·m (mainly used for gravity compensation and trajectory tracking). N·m (feedforward compensation torque). N·m. Resultant torque The torque is N·m, which is much smaller than the rated peak torque of 100 N·m for joint 1, and is within the safe range. The feedforward torque accounts for only about 1% of the resultant torque, and has almost no impact on the power consumption of the system.

[0136] S520 features online monitoring of feedforward effects and adaptive adjustment of feedforward gain.

[0137] Real-time calculation of tracking error for each joint .right In one pulse cycle Calculate the root mean square value within the sliding window e_j .Will e_j Compared with the preset performance threshold For comparison, this threshold corresponds to a displacement of approximately 0.02 mm at the end.

[0138] when e_j When the duration exceeds 10 pulse cycles, the feedforward gain increases by 5%. The maximum value is 1.0. When e_j When the duration exceeds 20 pulse cycles, the feedforward gain coefficient decreases by 2%. The lower limit is 0.5. After each adjustment, a lockout observation period of 10 pulse cycles is entered, during which no further actions are taken. Adjustment.

[0139] The adaptive mechanism possesses inherent negative feedback stability. When If slight overcompensation occurs when the value approaches 1.0, e_j It will increase in the opposite direction and exceed This triggers subsequent downward adjustments, making It automatically converges to the equilibrium point corresponding to the disturbance moment estimation error. Because the increment step size of 5% is greater than the decrement step size of 2%, and the down-adjustment trigger condition is more stringent (requiring the root mean square value to be below 50% of the threshold for 20 consecutive cycles). Will tend to make e_j Stable at Within the dead zone. This dead zone ensures that the parameters will not continuously oscillate during fine-tuning. Under the constraint of locked observation period, The maximum single oscillation amplitude is 5%, and it is subject to hard limiting. Within the specified range, it will not cause system instability.

[0140] The quantitative convergence rate is estimated as follows. Each adjustment requires two stages: a decision waiting period and a lock-in observation period. The decision waiting period for the increase direction is 10 pulse cycles, and for the decrease direction it is 20 pulse cycles. The lock-in observation period is 10 pulse cycles for both. Therefore, the shortest interval for each adjustment is 20 to 30 pulse cycles. The maximum number of adjustments from the initial value of 0.85 to the lower limit of 0.5 is approximately 18, as each adjustment is a 2% decrease with multiplicative decay. The longest convergence time is approximately s. The maximum number of adjustments required to raise the value from the initial 0.85 to the maximum of 1.0 is approximately 4. Upon reaching the 1.0 hard limit, the maximum is approximately s. Therefore The adaptive convergence time does not exceed 3 seconds in any direction.

[0141] when When set to 1.0 (full compensation), theoretically 100% cancellation of disturbance torque can be achieved if the transfer function calibration accuracy is perfect. However, in actual operating conditions, the transfer function calibration has residual deviations of less than 5% amplitude error and less than 10° phase error. Full compensation will introduce approximately 5% to 10% overcompensation components into the reverse disturbance. The upper limit is set to 1.0 with an initial value of 0.85, and after adaptive convergence... It typically stabilizes within the range of 0.80 to 0.95, corresponding to an actual inhibition rate of 80% to 95%, thus avoiding the risk of overcompensation while maintaining sufficient inhibitory effect. If the value is below 0.5, meaning the compensation effect is insufficient to reduce the residual vibration below the threshold, check whether the transfer function calibration accuracy has degraded or whether the welding posture has changed significantly, and consider re-performing the S010 calibration.

[0142] Under non-welding conditions, when the welding power supply is in standby mode and the output current is zero, even with power frequency interference noise resulting in a 50Hz sine wave signal of approximately 2.5A, the estimated amplitude of the feedforward torque is only about 0.002 N·m, far less than the typical value of 0.01 N·m for motor cogging torque, and will not have an observable impact on the system. The digital phase-locked loop cannot meet the locking condition due to its low signal-to-noise ratio, and the system automatically enters an unlocked state. This characteristic prevents the system from generating false triggers or abnormal torque output under non-welding conditions.

[0143] In this embodiment, the total delay from welding current sampling to compensation torque output in the entire signal processing chain is 20 μs, including the analog-to-digital converter sampling time and the S310 difference equation calculation time. This delay is much smaller than the mechanical delay of 500 μs to 2 ms for the arc pulsating torque to be transmitted from the arc to the joint, ensuring that the feedforward compensation torque is injected before the disturbance torque reaches the joint, thus achieving the timing conditions for advance compensation.

[0144] The effectiveness of the method was verified in an automated welding station equipped with a six-axis industrial welding robot and a digital pulse MIG welding power source. The test plate was Q235 carbon steel plate, 6mm thick, with a V-groove and a weld length of 300mm. The test conditions covered three pulse frequencies: 100Hz, 200Hz, and 350Hz, with a welding speed of 0.6m / min, and welding postures including flat and horizontal welding. The control group had feedforward compensation disabled, while the experimental group had it enabled. Three welds were welded under each condition, and the average values ​​were compared.

[0145] Evaluation indicators include: weld trajectory jitter amplitude, measured by a laser displacement sensor, is the peak-to-peak value of the welding torch displacement fluctuation in the weld normal direction, with a criterion of not exceeding 0.05 mm; weld reinforcement uniformity is measured by a 3D profile scanner, which measures the standard deviation of the longitudinal section reinforcement. The judgment standard is not more than 0.1mm; the jitter suppression rate is the difference between the jitter amplitude when uncompensated and when compensated divided by the jitter amplitude when uncompensated, and the judgment standard is not less than 80%; the feedforward response delay is the time difference from the rising edge of the welding current pulse to the feedforward torque output measured synchronously by an oscilloscope, and the judgment standard is not more than 50μs; the short-circuit transition impulse suppression rate judgment standard is not less than 75%.

[0146] Under flat welding conditions with a 200Hz pulse frequency, after enabling feedforward compensation, the angular displacement jitter amplitude of joint 1 decreased from approximately 0.0015° without compensation to approximately 0.0002°, corresponding to a reduction in end displacement from approximately 0.1mm to approximately 0.015mm, with a suppression rate of approximately 85%, meeting the 80% criterion. The suppression rates for joints 2 to 6 were approximately 82%, 87%, 84%, 83%, and 80%, respectively, all meeting the criterion. The measured feedforward response delay was 20μs, meeting the 50μs criterion. The standard deviation of weld reinforcement decreased from 0.18mm without compensation to 0.06mm, meeting the 0.1mm criterion. During the short-circuit transition phase, the impact suppression rate was approximately 86%, meeting the 75% criterion.

[0147] Under the flat welding condition with a pulse frequency of 100Hz, due to the lower pulse frequency and the peak duration of each pulse being approximately 3ms, the difference equation has more sufficient time to track the transient characteristics of the input waveform, resulting in higher accuracy in perturbation torque estimation. The measured jitter suppression rate of joint 1 is approximately 88%, slightly higher than under the 200Hz condition. Under the high-frequency condition of 350Hz, due to the feedforward gain... As the linear gain coefficient at the output of the difference equation, the proportionally scaled transfer function estimates the perturbation torque at each frequency, and the theoretical suppression rate of each frequency component is... The actual overall suppression rate was measured to be approximately 81%, slightly lower than that at 200Hz, but still meeting the 80% criterion. The slightly lower suppression rate at 350Hz is due to the slightly larger calibration error of the transfer function model at high frequencies compared to low frequencies.

[0148] Under horizontal welding conditions, the equivalent rotational inertia of each joint changes because the angle between the direction of gravity and the weld normal differs from that under flat welding conditions. The joint angle difference between the typical welding posture and the horizontal welding posture selected during S010 calibration is approximately 15 to 25°, with the transfer function parameter offset being less than 10%. In actual measurements at 200Hz, the suppression rate of joint 1 under horizontal welding conditions is approximately 80%, slightly lower than the 85% under flat welding conditions, but still meeting the judgment criteria. If the difference between the horizontal welding posture and the calibration posture exceeds 30°, it is recommended to add horizontal welding posture calibration points in the S010 stage to improve accuracy.

[0149] The impact of the feedforward compensation channel on the stability of the original closed-loop control system is analyzed as follows. Feedforward torque At the addition node and feedback torque After being superimposed, it is fed into the current loop, but It depends only on the welding current input and not on the position output feedback from the encoder. Therefore, the feedforward channel does not participate in the closed-loop circuit of the position-velocity-current three-loop system, and does not change the characteristic equation and pole positions of the original closed-loop system. The closed-loop transfer function, gain margin, and phase margin remain unchanged. This characteristic is the core advantage of the feedforward compensation scheme compared to increasing the PID gain or adding a notch filter—the latter two require modification of the closed-loop transfer function, which may reduce the stability margin of the system or introduce new resonance peaks while improving the suppression capability at specific frequencies.

[0150] The engineering applicability of the method is as follows. In terms of hardware resources, the method only requires adding an analog acquisition channel to the servo driver to receive the current feedback signal from the welding power source, without requiring any modifications to the mechanical structure or sensor configuration of the robot body. The computational workload of the difference equation is 5 multiplications and 4 additions per joint, totaling approximately 54 CPU cycles for 6 joints, occupying about 15% of the DSP computing resources, without affecting the real-time performance of the original three-loop control algorithm. In terms of storage resources, the difference equation coefficients and the impact template require approximately 1.8KB of RAM space, which is less than 1% of the typical 64KB to 256KB on-chip RAM of contemporary DSP chips. The method is suitable for scenarios employing periodic arc welding processes such as pulsed MIG, pulsed TIG, and CMT, where the welding power source provides a current feedback interface and the pulse frequency is in the range of 20 to 600Hz. For non-arc welding processes such as laser welding or friction stir welding, since there is no arc pulsation disturbance, the method is not applicable.

[0151] The above description is merely one embodiment of this application and does not constitute a limitation on the scope of protection of this application. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this application should be included within the scope of protection of this application.

[0152] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit described above can be implemented in hardware.

[0153] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A servo control method for suppressing arc pulsation in welding processes, applied to a welding robot equipped with a servo drive system, characterized in that, Includes the following steps: The current waveform signal output by the welding power source is acquired in real time, and after analog-to-digital conversion and digital pre-filtering, the welding current value sequence is obtained. Arc pulsation feature extraction and working condition identification are performed on the welding current numerical sequence to obtain the instantaneous frequency and locked phase of the welding pulse, and the start and end status of the droplet short-circuit transition event is detected. The DC bias component is removed from the welding current numerical sequence, the AC incremental signal is extracted, and a feedforward compensation torque command with the same frequency and opposite phase as the arc pulsation disturbance torque is generated in real time based on a pre-calibrated transfer function model of welding current to joint disturbance torque. Specifically, when in the normal pulse phase, the disturbance torque estimate of each joint is obtained from the AC incremental signal through the discretized difference equation of the transfer function model. When the short-circuit transition event is detected, an impact compensation torque command is generated according to the pre-calibrated impact torque pulse template. The feedforward compensation torque command is superimposed in real time with the feedback torque command output by the servo controller and the gravity compensation torque command to generate a composite torque command, which is then sent to the current loop of the servo driver for execution.

2. The method according to claim 1, characterized in that, The transfer function model is obtained through offline calibration, which includes: At multiple current amplitude operating points, frequency domain scanning and system identification are performed independently to obtain the transfer function parameter set corresponding to each operating point; During operation, linear interpolation is performed between the transfer function parameters of adjacent working points based on the current welding current amplitude to obtain the real-time transfer function parameters.

3. The method according to claim 2, characterized in that, The transfer function model adopts a second-order element form and is discretized into difference equations through the Tustin bilinear transform. The coefficients of the difference equations are pre-calculated and stored during the offline calibration stage by the transfer function parameters and control period of each operating point, and can be directly called or obtained by interpolation based on the interpolated operating point parameters during runtime.

4. The method according to claim 1, characterized in that, The acquisition of the locked phase of the welding pulse is implemented using a digital phase-locked loop algorithm, including: The welding current numerical sequence is preprocessed by normalization to remove the DC bias component and normalize it according to the half amplitude of the current pulse. The phase error between the normalized signal and the local oscillation signal is obtained by a phase detector; The phase error is smoothed by a loop filter, and the frequency correction is output. The local frequency and accumulated phase are updated by a numerically controlled oscillator according to the frequency correction amount.

5. The method according to claim 1, characterized in that, The start and end states of the detected droplet short-circuit transition event include: The rate of change of the welding current numerical sequence is acquired in real time; When the rate of change is lower than the negative threshold and the current value is lower than the preset current threshold, a short-circuit transition event is determined to have started. When the rate of change is higher than the positive threshold and the current value is higher than the preset recovery threshold, the short-circuit transition event is determined to have ended. The validity of the duration of the short-circuit transition event is verified, and falsely triggered events that do not meet the preset duration range are discarded.

6. The method according to claim 1, characterized in that, The step of generating the impact compensation torque command based on the pre-calibrated impact torque pulse template includes: The impact torque pulse template is scaled based on the ratio of the current drop amplitude of the current short-circuit event to the reference drop amplitude during calibration. Based on the short circuit initiation time, the scaled template is played back step by step according to the control cycle to generate the impact compensation torque command.

7. The method according to claim 1, characterized in that, It also includes an adaptive stiffness adjustment step: In response to the start of the short-circuit transition event, the position loop proportional gain and derivative gain of the servo controller are increased to a preset multiple of their normal values; In response to the end of the short-circuit transition event, the proportional gain and differential gain are smoothly restored to their normal values ​​using an exponential decay function.

8. The method according to claim 6, characterized in that, After the impact compensation torque command is played back, within a preset transition time window, a cosine gradient weighting function is used to smoothly switch the feedforward torque from the impact compensation mode to the normal feedforward mode.

9. The method according to claim 1, characterized in that, It also includes an online monitoring step for the feedforward effect: The tracking error of each joint is acquired in real time, and the root mean square value of the tracking error is obtained within a sliding window; When the root mean square value continuously exceeds the performance threshold for a first preset number of cycles, the feedforward gain coefficient is increased. When the root mean square value is continuously lower than the performance threshold by a preset percentage for a second preset number of cycles, the feedforward gain coefficient is reduced. After each adjustment, a lockout observation period is entered, during which no adjustment of the feedforward gain coefficient is performed.

10. The method according to claim 1, characterized in that, The integrator in the loop filter is set with a saturation limit value to constrain the frequency offset of the numerically controlled oscillator from exceeding the maximum allowable frequency deviation. When the root mean square value of the phase error continues to exceed the unlock threshold for a preset duration, an unlock reset operation is performed to clear the integrator and reinitialize the center frequency with the current zero-crossing detection frequency.

Images