Double-frequency composite pulse reinforced swinging electric arc additive repairing device and repairing method

Abstract

The invention discloses a double-frequency composite pulse reinforced swinging electric arc additive repairing device and method, and belongs to the technical field of fuse wire additive manufacturing. The device comprises a GTA welding system, a pendulum oscillator, an ultrahigh-frequency excitation power supply and a multi-thread synchronous digital regulation and control system, and regulation and control of the heat input gradient and the molten drop impact force are achieved through low-frequency and ultrahigh-frequency pulse current composition and dynamic matching with a swinging arc path. The method comprises the steps of surface treatment, pulse compounding, welding gun positioning and repairing operation, heat distribution of center heat flow focusing and edge heat blocking is formed through different effects of double-frequency pulses in the swing stage and the stay stage, and the flowing and solidification characteristics of a molten pool are optimized. The problems of coarse grains, poor forming precision and the like in a traditional process are solved, the uniformity, interface bonding strength and repairing efficiency of a deposition layer are improved, and the method is suitable for in-service repairing of complex components such as nuclear power equipment.

Description

Technical Field

[0001] The present invention belongs to the technical field of fuse additive manufacturing, and in particular relates to a dual-frequency composite pulse enhanced swing arc additive repair device and a repair method. Background Art

[0002] At present, the additive repair technologies for large and complex structural parts mainly include laser additive repair, electron beam cladding, and arc fuse additive repair. Compared with laser additive repair and electron beam cladding, among the arc fuse additive repair technologies, the gas tungsten arc (GTA) fuse additive repair technology has shown some unique advantages, such as stable arc shape, low equipment cost, and easy remote control. It is widely applicable to a variety of nuclear power materials such as stainless steel and nickel-based alloys, and has broad prospects in the field of in-service repair applications of nuclear power equipment. Although a series of progress has been made in the research on arc additive repair of nuclear power equipment, there are still some technical bottlenecks that need to be solved urgently. For example, the traditional GTA arc has insufficient dynamic control of the heat input gradient, and the additive repair process often adopts a constant heat input mode, which can easily lead to serious local heat accumulation in the molten pool, exacerbating welding deformation and crack propagation. Furthermore, in existing GTA additive repair processes, the transfer behavior of the filler wire droplet is difficult to precisely control, and there is a lack of a dynamic control mechanism for the droplet volume and frequency. The morphology of the central and edge molten pools differs significantly, which can easily lead to defects such as lack of fusion and undercutting, directly affecting the uniformity of the deposited layer and the interfacial bonding strength. To address these issues, Chinese Publication No. CN115106613A discloses an arc-fuse additive manufacturing system and method. This system improves wire feeding stability through the integrated design of the welding torch and wire feed mechanism. However, it is still limited to a single heat input control mode and fails to effectively address issues such as heat accumulation in the molten pool, droplet transfer, and the coordinated optimization of interlayer thermal cycles. Furthermore, Chinese Publication No. CN115415640A discloses a soft device for arc additive repair under spatial constraints. This device utilizes a soft robotic arm and a rope-driven structure to enhance the flexibility and multi-degree-of-freedom adjustment capabilities of the welding torch in confined spaces. However, its technical focus is on spatial adaptability and does not address the core contradiction between dynamic heat input control and optimized droplet behavior. China Publication No. CN117102629A discloses a method that enhances deposit stability through the use of a swinging contact rod and a threaded adjustment mechanism. However, this method still lacks the technical difficulties of controlling droplet transfer and precisely regulating thermal gradients. Consequently, conventional swinging arc additive repair technology is prone to causing unfused deposit edges or undercut defects during additive repair of large components due to excessive swing amplitude. Summary of the Invention

[0003] Based on the above shortcomings, the present invention proposes a dual-frequency composite pulse enhanced swing arc additive repair device and repair method. The present invention combines low-frequency pulse current with ultra-high frequency excitation pulse current to achieve dynamic matching control with the motion path of the swing arc, significantly optimizing the molten pool flow behavior and solidification characteristics, so as to solve the problems of coarse grains of the deposited layer, poor forming accuracy, frequent element segregation defects, and low repair efficiency in the traditional arc additive repair process.

[0004] The technical solution adopted by the present invention is as follows: a dual-frequency composite pulse enhanced swing arc additive repair device, including a GTA welding system, the GTA welding system includes a GTA welding gun, a wire feeding device, a welding walking platform, a shielding gas system and a welding power supply, and also includes a pendulum oscillator, an ultra-high frequency excitation power supply and a multi-threaded synchronous digital control system; the oscillator drives the welding gun to swing periodically laterally through a flexible electrical signal; the ultra-high frequency pulse current of the ultra-high frequency excitation power supply and the low-frequency pulse current of the welding power supply are combined through an isolation coupling device and then introduced into the welding circuit; the multi-threaded synchronous digital control system is used to realize the coordinated control of the composite pulse current waveform and the swinging arc path, thereby realizing on-demand control of the heat input gradient and the molten droplet impact force.

[0005] Furthermore, the welding power supply is connected to the shielding gas through a pipeline, and the positive pole of the power supply is connected to the welding walking platform by a DC positive connection method; the pendulum oscillator is installed above the platform through a bracket, the pendulum oscillator is driven by a stepper motor, and the GTA welding gun and wire feeding device are fixed to the end of the pendulum oscillator by a clamp; the ultra-high frequency excitation power supply is connected in parallel with the welding circuit through an isolation coupling device to introduce high-frequency pulse current into the arc; the multi-threaded synchronous digital control system electrical signal is connected to the drive motor of the pendulum oscillator, and integrates a Hall current sensor, a Hall voltage sensor and an electromagnetic relay to realize full electrical signal closed-loop control, and through dynamic analysis of the correlation between the electrical signal characteristics of the swinging arc and the swing trajectory, the dual-frequency pulse parameters are adjusted in real time to achieve microsecond-level synchronous matching of the welding current waveform and the arc motion trajectory.

[0006] Furthermore, the multi-threaded synchronous digital control system includes a Hall current / voltage sensor, an analog-to-digital converter, and a PLC controller containing an FPGA module. The PLC controller obtains the parameters of the swinging arc instantaneous voltage signal and pulse waveform characteristics in real time through the Hall current / voltage sensor. The PLC controller analyzes the dynamic correlation relationship by constructing a multi-layer perception neural network model, inputs the processed timing signal into the dual-output neural network, generates time compensation and energy adjustment instructions, drives the dual-frequency pulse current waveform parameters of the ultra-high frequency pulse and the welding low-frequency pulse to be corrected, and at the same time, uses synchronous error detection to feedback the control deviation in real time, triggering the control logic based on the composite pulse current waveform to form a closed loop, which can adapt to the nonlinear changes of the arc trajectory and dynamically update the network weights through the online learning mechanism during continuous operation, wherein the current change rate is not less than 50A / μs.

[0007] Furthermore, the neural network model is a three-layer hybrid network including an input layer, a hidden layer and a dual output layer. The input layer includes 128 neurons, receives 100ms time window time series data, and includes four-channel signals: instantaneous value of arc voltage, instantaneous value of composite current, ultra-high frequency pulse phase mark, and swing position encoder signal; the feature extraction layer of the hidden layer is 64 convolution neurons, using SeLU activation function, and the convolution kernel time width is 15 sampling points, which is responsible for extracting pulse edge features and swing period features; the timing processing layer of the hidden layer is 48 bidirectional LSTM neurons, using tanh activation function to process long-term dependencies of time series; the time compensation branch of the output layer includes 3 linear activation neurons, namely rising edge compensation amount, falling edge compensation amount and pulse width compensation amount, with a compensation range of ±50μs, and the energy adjustment branch is 2 sigmoid activation neurons, including ultra-high frequency amplitude coefficient = 0.8-1.2 and low frequency peak coefficient = 0.9-1.1.

[0008] Furthermore, the input processing flow of the dual-output neural network includes: after the arc voltage signal is subjected to wavelet noise reduction, the swing cycle characteristics of the pendulum oscillator are extracted through peak detection, namely the left / right side dwell point identification and swing midpoint positioning, by detecting the current signal, analyzing the pulse characteristics, including the low-frequency pulse duty cycle and ultra-high frequency pulse count, fusing the swing position encoder signal with the voltage signal, calculating the arc length change rate, and outputting the instruction generation logic including the time compensation instruction and the energy adjustment instruction.

[0009] In the time compensation instruction,

[0010] Rising edge compensation = K1 × (actual left endpoint arrival time - target time) + K2 × arc stability index during the left dwell phase; falling edge compensation = K3 × (actual right endpoint voltage fluctuation integral) + K4 × swing acceleration; pulse width compensation = K5 × (maximum arc length change during the dwell phase) × current duty cycle; where K1-K5 are dynamic coefficients obtained through network learning.

[0011] In the energy adjustment instruction,

[0012] Ultra-high frequency amplitude coefficient = 0.8 + 0.4 × sigmoid (residence time × material evaporation prediction + arc length change rate); low frequency peak coefficient = 0.9 + 0.2 × tanh (swing speed × melt pool oscillation amplitude + heat accumulation correction term).

[0013] Furthermore, the synchronization error detection includes a synchronization error detection method and an energy error detection method. The synchronization error detection method is to use a timing unit of an FPGA module with a resolution of 10ns to continuously measure the actual time difference between the rising edge / falling edge of the low-frequency and ultra-high-frequency pulses. The energy error detection method is to calculate the integral power deviation in each half swing cycle. In the judgment criteria, the synchronization error thresholds are: warning ±5μs, adjustment ±10μs, emergency ±20μs; the energy error thresholds are: warning ±3%, adjustment ±5%, emergency ±8%; weighted comprehensive error is used for deviation fusion. When the weighted comprehensive error is greater than 0 .7 or when a single error reaches the adjustment threshold, the neural network is activated to generate new parameters. When a single error reaches the emergency threshold, the preset safety strategy is started, including cutting off high-frequency pulses and performing time compensation by directly modifying the phase register of the PWM controller. The new rising edge time is based on the original set value, and the basic compensation amount output by the neural network is superimposed: the incremental adjustment value after the corrected Δ rising edge and the current synchronization error are dynamically weighted, and the falling edge synchronization feedforward compensation is triggered by the advance compensation Δ falling edge time. Energy adjustment includes ultra-high frequency voltage and low-frequency peak current adjustment. When any parameter exceeds the safety boundary, it is forcibly locked at the boundary value and an alarm is triggered.

[0014] Furthermore, the online learning mechanism adopts a dual-modal weight update strategy, which is triggered every 50 swing cycles in fine-tuning mode, and uses recursive least squares to dynamically adjust the output layer weights, completing real-time compensation within 15ms; deep update is initiated when material changes or continuous performance degradation are detected, and the full network weights are updated through the Adam optimizer, combined with 1,000 groups of samples from the abnormal database for 2-second training; a total of five protection mechanisms are implemented through pre-update weight snapshots, virtual environment verification, post-update online testing, dynamic attenuation of learning rate, and gradient clipping to ensure that the learning process is stable and controllable; 100 groups of real-time working condition data are continuously rolled and stored, and the learning process is automatically triggered in the event of an abnormality.

[0015] Furthermore, the welding power supply provides direct current or pulsed direct current with a current range of 0-500A, a low-frequency pulse current frequency of 0.5-15Hz, and a duty cycle of 10%-90% continuously adjustable; the ultra-high frequency excitation power supply outputs a square wave pulse current with a frequency of 20-100kHz, an output voltage of 0-75V, a pulse width of 0-50%, and a maximum output power of 3000W; the oscillator is connected to the GTA welding gun and controls its periodic swing, with the maximum swing amplitude on one side being 100mm, the swing frequency being 0-25Hz, and the maximum dwell time of swinging to the left and right sides being 2.5s.

[0016] The present invention also provides a dual-frequency composite pulse enhanced swing arc additive repair method based on the above device, comprising the following steps:

[0017] S1: Grind, remove rust and clean the surface of the repaired part;

[0018] S2: Connect the ultra-high frequency excitation power supply in parallel with the arc circuit through an isolation coupling device to achieve the combination of ultra-high frequency pulse current and low frequency pulse current;

[0019] S3: Make the welding gun perpendicular to the surface of the component to be repaired, fix the welding wire clamp and move it synchronously with the welding gun;

[0020] S4: Introduce welding shielding gas, start the welding machine to strike the arc, turn on the ultra-high frequency excitation power supply and wire feeding mechanism to perform additive repair; when the repair is completed, first turn off the ultra-high frequency excitation power supply and wire feeding mechanism, and then turn off the welding machine to extinguish the arc.

[0021] Furthermore, the above method further includes:

[0022] (a) An ultrahigh-frequency and low-frequency composite pulse current is applied to the swinging arc. The ultrahigh-frequency pulse compresses the arc through the electromagnetic confinement effect and generates periodic thermal shocks, which enhances the kinetic energy of the droplet transfer and forms micro-turbulent perturbations in the molten pool. The peak segment of the composite pulse is matched to the arc swing phase to maintain the stability of the molten pool flow field, and the base segment is matched to the dwell phase on the left and right sides of the arc to reduce heat accumulation.

[0023] (b) When the arc swings left and right, large-volume, high-frequency droplet transfer is used to form an axis-focused molten pool with large melt depth in the middle and small melt depth on both sides, which is used to enhance the metallurgical reaction; when the arc swings left and right, small-volume, low-frequency droplet transfer is used to form a shallow and wide edge molten pool, which is used to suppress lateral heat conduction;

[0024] (c) Through periodic dynamic heat distribution of central heat flux focusing and edge heat isolation, the interlayer temperature gradient is reduced, and directional control of solidification structure and high-precision forming of the deposited layer are achieved.

[0025] The principle of the present invention: During the swing arc additive repair process, after introducing the ultra-high frequency pulse current, the arc is compressed under the influence of the electromagnetic confinement effect, and the arc energy density and stability are significantly improved. The periodic thermal effect stimulated by its ultrasonic frequency characteristics enhances the kinetic energy of the droplet transition, continuously impacts the molten pool, and drives the formation of micro-turbulence disturbances inside the molten pool, effectively breaking the solute diffusion boundary layer at the front of the solidification interface, promoting equiaxed grain growth, and helping to eliminate element segregation defects. At the same time, the base value segment of the low-frequency pulse can suppress the heat accumulation rate during the dwell period, avoiding deformation and cracking caused by overheating. This composite technology not only solves the inherent defects of the swing arc additive repair process, but also significantly improves the contour accuracy, dimensional consistency and interlayer thermal cycle controllability of the repair layer through the coordinated regulation of dual-frequency pulses, providing reliable protection for the safe operation of nuclear power equipment under extreme working conditions such as high temperature and high pressure.

[0026] The advantages and beneficial effects of the present invention are as follows: First, the introduction of an ultrahigh-frequency pulse current compresses the arc due to the electromagnetic confinement effect, significantly improving its energy density and stability. The periodic thermodynamic effect stimulated by its ultrasonic characteristics enhances the kinetic energy of the droplet transfer, continuously impacts the molten pool, and drives the formation of microturbulent disturbances within the molten pool. This effectively breaks the solute diffusion boundary layer at the solidification interface front, promotes equiaxed grain growth, and helps eliminate pores and element segregation defects. Second, by precisely matching the peak / base segment timing of the composite pulse with the arc oscillation motion / dwelling, on-demand regulation of the heat input gradient and droplet impact force is achieved. The pulse peak segment maintains the stability of the molten pool flow field during the oscillating arc additive repair process, while the base segment reduces the heat accumulation rate during the non-oscillation dwell process. The continuous impact of the ultrahigh-frequency pulse optimizes the solidification nucleation conditions of the molten pool, facilitating high-precision shaping of the deposited layer profile, dimensional consistency of complex geometric features, and refined control of interlayer thermal cycles, thereby improving forming accuracy and repair process efficiency. Unlike the traditional GTA additive repair process, which involves continuous superposition of heat and force, this process decouples heat input from arc force into a dynamically alternating "heat accumulation-heat relaxation" cycle through spatiotemporal control of the composite pulse current and the swinging arc position interval. Furthermore, through process parameter matching, during the GTA arc swing phase, the droplets are forced to transition in a large-volume, high-frequency mode, forming a central molten pool with a relatively large depth and width. This strengthens the metallurgical reaction in the central molten pool of the deposited layer, constrains heat input to the axial region of the deposited layer, and achieves directional control of the solidification structure. When the arc swings to the two-sided dwelling stage, the droplets transition in a small-volume, low-frequency mode, forming a gradient-transition, shallow, and wide edge molten pool. This suppresses lateral heat conduction by reducing the molten pool volume and extending the heat diffusion path. This periodic dynamic heat flow distribution strategy of central heat flow focusing and edge heat blocking significantly reduces the interlayer temperature gradient. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1This is the overall equipment layout diagram of the dual-frequency composite pulse enhanced swing arc additive repair device of the present invention;

[0028] Figure 2 Schematic diagram of the swing structure of the dual-frequency composite pulse enhanced swing arc additive repair method of the present invention;

[0029] Figure 3 Automatic control circuit diagram for multi-threaded synchronous digital control system;

[0030] Figure 4 Schematic diagram of the waveform control principle of the dual-frequency composite pulse enhanced swing GTA arc additive repair method;

[0031] Figure 5 The actual electrical signals during the additive repair welding process, including (a) Example 2, (b) Example 3, (c) Example 4, (d) Example 5, and (e) Example 6;

[0032] Figure 6 1 is the morphology of the additive repair deposition layer in the examples of the present invention, wherein (a) Example 2, (b) Example 3, (c) Example 4, (d) Example 5, and (e) Example 6;

[0033] Figure numerals: 1. gas cylinder, 2. welding power supply, 3. Hall voltage sensor, 4. wire feeder, 5. pendulum oscillator, 6. GTA welding gun, 7. swing control box, 8. AC contactor, 9. isolation coupling device, 10. ultra-high frequency excitation power supply, 11. workpiece to be repaired, 12. Hall current sensor, 13. PLC controller, 14. electromagnetic relay, 15. analog-to-digital converter.

[0034] Specific implementation method In order to more clearly illustrate the purpose, technical solutions and advantages of the present invention, the specific implementation methods of the present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0035] Example 1

[0036] like Figure 1-3As shown, a dual-frequency composite pulse enhanced swing arc additive repair device includes a GTA welding system, which includes a GTA welding gun, a wire feeding device, a welding walking platform, a shielding gas system and a welding power supply, and also includes a pendulum oscillator, an ultra-high frequency excitation power supply and a multi-threaded synchronous digital control system; the oscillator drives the welding gun to swing periodically laterally through a flexible electrical signal; the ultra-high frequency pulse current of the ultra-high frequency excitation power supply and the low-frequency pulse current of the welding power supply are combined through an isolation coupling device and then introduced into the welding circuit; the multi-threaded synchronous digital control system is used to achieve coordinated control of the composite pulse current waveform and the swinging arc path. The welding power supply is connected to the shielding gas through a pipeline, and the positive pole of the power supply is connected to the welding walking platform by a DC positive connection method; the pendulum oscillator is installed above the platform through a bracket, the pendulum oscillator is driven by a stepping motor, and the GTA welding gun and wire feeding device are fixed to the end of the pendulum oscillator by a clamp; the ultra-high frequency excitation power supply is connected in parallel with the welding circuit through an isolation coupling device to introduce high-frequency pulse current into the arc; the multi-threaded synchronous digital control system electrical signal is connected to the drive motor of the pendulum oscillator, and integrates a Hall current sensor, a Hall voltage sensor and an electromagnetic relay to realize full electrical signal closed-loop control, and through dynamic analysis of the correlation between the electrical signal characteristics of the swinging arc and the swing trajectory, the dual-frequency pulse parameters are adjusted in real time to achieve microsecond-level synchronous matching of the welding current waveform and the arc motion trajectory.

[0037] The multi-threaded synchronous digital control system includes a Hall current / voltage sensor, an analog-to-digital converter, and a PLC controller containing an FPGA module. The PLC controller obtains the parameters of the swing arc instantaneous voltage signal and pulse waveform characteristics in real time through the Hall current / voltage sensor. The PLC controller analyzes the dynamic mutual correlation by constructing a multi-layer perception neural network model, inputs the processed timing signal into a dual-output neural network, generates time compensation and energy adjustment instructions, drives the dual-frequency pulse current waveform parameters of the ultra-high frequency pulse and the welding low-frequency pulse to be corrected, and at the same time, uses synchronous error detection to feedback control deviations in real time, triggering control logic based on the composite pulse current waveform to form a closed loop. The system can adapt to the nonlinear changes of the arc trajectory and dynamically update the network weights through an online learning mechanism during continuous operation, wherein the current change rate is not less than 50A / μs.

[0038] The welding power supply provides direct current or pulsed direct current with a current range of 0-500A, a low-frequency pulse current frequency of 0.5-15Hz, and a duty cycle continuously adjustable from 10% to 90%; the ultra-high frequency excitation power supply outputs a square wave pulse current with a frequency of 20-100kHz, an output voltage of 0-75V, a pulse width of 0-50%, and a maximum output power of 3000W; the oscillator is connected to the GTA welding gun and controls its periodic oscillation, with a maximum unilateral oscillation amplitude of 100mm, an oscillation frequency of 0-25Hz, and a maximum dwell time of 2.5s when oscillating to the left or right sides.

[0039] The neural network model is a three-layer hybrid network including an input layer, a hidden layer and a dual output layer. The input layer includes 128 neurons and receives 100ms time window time series data, including four-channel signals: instantaneous value of arc voltage, instantaneous value of composite current, ultra-high frequency pulse phase mark, and swing position encoder signal; the feature extraction layer of the hidden layer is 64 convolution neurons, using SeLU activation function, and the convolution kernel time width is 15 sampling points, which is responsible for extracting pulse edge features and swing period features; the time series processing layer of the hidden layer is 48 bidirectional LSTM neurons, using tanh activation function to process long-term dependencies of time series; the time compensation branch of the output layer includes 3 linear activation neurons, namely rising edge compensation amount, falling edge compensation amount and pulse width compensation amount, with a compensation range of ±50μs, and the energy adjustment branch is 2 sigmoid activation neurons, including ultra-high frequency amplitude coefficient = 0.8-1.2 and low frequency peak coefficient = 0.9-1.1.

[0040] The input processing flow of the dual-output neural network includes: after the arc voltage signal is subjected to wavelet noise reduction, the swing cycle characteristics of the pendulum oscillator are extracted through peak detection, namely the left / right side dwell point identification and swing midpoint positioning, by detecting the current signal, analyzing the pulse characteristics, including the low-frequency pulse duty cycle and ultra-high frequency pulse count, fusing the swing position encoder signal with the voltage signal, calculating the arc length change rate, and outputting the instruction generation logic including the time compensation instruction and the energy adjustment instruction.

[0041] Among them, in the time compensation instruction,

[0042] Rising edge compensation = K1 × (actual left endpoint arrival time - target time) + K2 × arc stability index in the left dwell phase; falling edge compensation = K3 × (actual right endpoint voltage fluctuation integral) + K4 × swing acceleration; pulse width compensation = K5 × (maximum arc length change in the dwell phase) × current duty cycle, where K1-K5 are dynamic coefficients obtained through network learning.

[0043] Among them, in the energy adjustment instruction,

[0044] Ultra-high frequency amplitude coefficient = 0.8 + 0.4 × sigmoid (residence time × material evaporation prediction + arc length change rate); low frequency peak coefficient = 0.9 + 0.2 × tanh (swing speed × melt pool oscillation amplitude + heat accumulation correction term).

[0045] The synchronization error detection method includes a synchronization error detection method and an energy error detection method. The synchronization error detection method uses a timing unit of an FPGA module with a resolution of 10ns to continuously measure the actual time difference between the rising edge / falling edge of a low-frequency and ultra-high-frequency pulse. The energy error detection method is to calculate the integral power deviation in each half swing cycle. In the judgment standard, the synchronization error thresholds are: warning ±5μs, adjustment ±10μs, emergency ±20μs; the energy error thresholds are: warning ±3%, adjustment ±5%, emergency ±8%; weighted comprehensive error is used for deviation fusion. When the weighted comprehensive error is greater than 0.7 Or when a single error reaches the adjustment threshold, the neural network is activated to generate new parameters. When a single error reaches the emergency threshold, the preset safety strategy is started, including cutting off high-frequency pulses and performing time compensation by directly modifying the phase register of the PWM controller. The new rising edge time is based on the original set value, and the basic compensation amount output by the neural network is superimposed: the incremental adjustment value after the corrected Δ rising edge and the current synchronization error are dynamically weighted, and the falling edge synchronization feedforward compensation is triggered by the advance compensation Δ falling edge time. Energy adjustment includes ultra-high frequency voltage and low-frequency peak current adjustment. When any parameter exceeds the safety boundary, it is forcibly locked at the boundary value and an alarm is triggered.

[0046] The online learning mechanism adopts a dual-modal weight update strategy, which is triggered every 50 swing cycles in fine-tuning mode, and uses recursive least squares to dynamically adjust the output layer weights, completing real-time compensation within 15ms; deep update is initiated when material changes or continuous performance degradation are detected, and the Adam optimizer is used to update the weights of the entire network, and 1,000 groups of samples from the abnormal database are combined for 2-second training; a total of five protection mechanisms are implemented, including pre-update weight snapshots, virtual environment verification, post-update online testing, dynamic attenuation of the learning rate, and gradient clipping, to ensure that the learning process is stable and controllable; 100 groups of real-time working condition data are continuously rolled and stored, and the learning process is automatically triggered in the event of an abnormality.

[0047] Among them, the training method of the neural network model in this embodiment adopts a three-stage hybrid training strategy: ① Offline pre-training: Based on historical welding data of different materials, swing frequencies and energy combinations, deep reinforcement learning is used to optimize the initial weights, the reward function integrates synchronization error and energy deviation, and 10 million sets of virtual samples are generated through the simulation environment to enhance generalization; ② Online fine-tuning: After deployment, recursive least squares fast iteration is performed through real-time sensor data (10kHz sampling), the output layer weights are updated every 50 swing cycles, and the memory of key working conditions is strengthened by combining priority experience replay; ③ Adversarial training: Generate adversarial networks to simulate arc disturbances (such as sudden changes in airflow) to improve the model's anti-interference ability, and dynamically adjust the Gaussian noise exploration coefficient (σ=0.01~0.1) during online learning.

[0048] Among them, the dataset construction method of this embodiment includes data acquisition, feature engineering and data enhancement. ① Data acquisition: synchronously record arc voltage, composite current, encoder position and pulse marker signal synchronized with FPGA nanosecond level, cooperate with infrared thermal imager to calibrate the molten pool state, covering 500 sets of typical working conditions of materials such as steel, nickel-based alloy, titanium-based alloy at different swing frequencies; ② Feature engineering: use wavelet transform to extract pulse edge features, fuse multi-sensor data through Kalman filtering, and construct a 256-dimensional input vector containing time alignment error, energy integration deviation and trajectory acceleration; ③ Data enhancement: use timing distortion (±5% stretching) and bandwidth-limited noise (10-100kHz) to simulate on-site interference to ensure the robustness of the model under ±0.1mm trajectory jitter and ±5% power fluctuation.

[0049] This embodiment achieves collaboration through a layered real-time control architecture. In terms of hardware, an FPGA processes nanosecond pulse synchronization, a PLC runs a neural network, and a stepper driver interprets swing acceleration commands. Regarding data flow, Hall effect sensor signals are filtered by an isolation conditioning module and converted to digital quantities by an analog-to-digital converter. The FPGA extracts timing characteristics and transmits them to the PLC via EtherCAT (100μs cycle). The time compensation and energy coefficients output by the neural network are directly written to the PWM registers of the welding / high-frequency power supply. Regarding fault tolerance, when the network output exceeds a limit (e.g., the high-frequency voltage exceeds a certain limit), a switch is made to a preset PID controller, triggering an online learning process. The RTC (DS3231) timestamp is used to accurately trace back fault data.

[0050] The physical meaning and determination method of the coefficients in the time compensation instruction in this embodiment are as follows:

[0051] ①Δ rising edge (μs): Corrects the time deviation between the rising edge of the low-frequency pulse and the ultra-high-frequency pulse. The actual edge time difference measured in real time by the FPGA is generated through neural network mapping. The weight coefficient K1 is determined by historical data regression to ensure linear adjustment within the range of ±50μs.

[0052] ②Δ falling edge (μs): compensates for the phase lag of the low-frequency falling edge. Its value is proportional to the integral of the voltage fluctuation at the right dwell point. The coefficient K2 is calibrated through arc stability experiments (usually 0.15-0.3).

[0053] ③ Pulse width compensation (μs): Dynamically adjust the low-frequency pulse width to match the dwell time. The calculation formula is K3×(maximum arc length change rate)×current duty cycle. K3 is optimized through the molten pool solidification experiment (usually 0.05-0.1).

[0054] The physical meaning and determination method of the coefficients in the energy adjustment instruction in this embodiment are as follows:

[0055] ①UHF amplitude coefficient (0.8-1.2): controls the normalized ratio of the UHF pulse voltage (0-75V), which is output by the sigmoid function based on the predicted value of the evaporation rate in the dwell phase and the arc length change rate. The slope parameter is determined by calculating the material vaporization heat.

[0056] ② Low-frequency peak coefficient (0.9-1.1): Adjust the peak ratio of the low-frequency pulse current (0-500A), which has a tanh function relationship with the swing acceleration and the oscillation amplitude of the molten pool. The gain coefficient is obtained by inverse deduction of the heat input-penetration model.

[0057] The specific calculation method for the synchronization error and energy error in this embodiment is as follows:

[0058] ① Synchronization error: The FPGA hardware timing unit precisely measures the edge timing difference between low-frequency and ultra-high-frequency pulses. This time difference (μs) is captured by a delay chain (resolution 78ps). Weighting factors (0.6 / 0.4) prioritize rising-edge alignment. The calculated result is updated every 100μs, triggering emergency compensation if the error exceeds 10μs.

[0059] ② Energy error is evaluated based on the percentage of power integral deviation, which is calculated every 5ms. If it is >5% for three consecutive times, energy adjustment is initiated.

[0060] The specific content and triggering conditions of the preset security policy in this embodiment are as follows:

[0061] The pre-set safety strategy details: When the system detects an emergency-level error (such as a synchronization error exceeding ±20μs or an energy limit exceeding ±8%), it immediately activates three levels of protection. The first level of protection involves forcibly reducing the UHF voltage to 50V (66% of the original 75V) and limiting the low-frequency current to 300A (60% of the original 500A) to mitigate the risk of energy runaway. The second level of protection switches to a backup PID controller, bypassing the neural network and directly outputting fixed parameters (such as a 10Hz low-frequency pulse and a 20kHz UHF reference). If the first two levels of protection fail, the third level of protection disconnects the welding / high-frequency power supply main circuit (via a solid-state relay) within 2ms and triggers a mechanical emergency stop (stepper motor brake). All actions are executed directly via FPGA hardware logic (independent of the PLC cycle), ensuring a response latency of less than 100μs.

[0062] The safety policy trigger conditions are as follows: ① Synchronization error trigger: The FPGA detects an edge deviation greater than ±20μs within three consecutive pulse cycles (possibly due to electromagnetic interference or clock desynchronization); ② Energy error trigger: The power integral value exceeds ±8% of the set value for 10ms (indicating weld pool overheating or arc failure); ③ Hardware fault trigger: This includes analog-to-digital converter signal loss (no data for 1ms), network communication interruption (EtherCAT slave timeout of 500μs), and heat dissipation abnormality (IGBT temperature greater than 85°C). Once triggered, the system enters a safe state and locks down, requiring a manual reset and complete fault diagnosis before restarting.

[0063] The specific implementation process of the online learning mechanism in this embodiment is as follows:

[0064] ① Data preprocessing: Real-time sensor data is normalized and subjected to wavelet denoising through a sliding window (length 100ms, step size 10ms). Time / frequency domain features are extracted and stored in a circular buffer, with priority given to retaining key data for the 10 cycles before and after the error trigger. ② Incremental learning acceleration: FPGA hardware is used to accelerate the matrix inversion operation in the RLS algorithm, compressing the fine-tuning delay from 15ms to 8ms. ③ Adversarial sample generation: During the deep update phase, extreme working condition data (such as a sudden arc offset of ±5mm) is synthesized through a conditional generative adversarial network to enhance model robustness. ④ Weight update arbitration: When the performance difference between the new and old weights is less than 2%, a momentum smoothing transition strategy is adopted to avoid control parameter jumps. ⑤ Resource management: The second CPU core of the PLC is dynamically allocated exclusively for training tasks, and memory mapping is used to ensure that the real-time control thread always has the highest priority.

[0065] like Figure 4As shown, the welding power supply provides low-frequency pulse current, while an external high-frequency excitation power supply provides ultra-high-frequency pulse current. The swinging arc swings perpendicular to the direction of the additive repair motion and can be divided into two parts: the middle swing phase and the left and right stop phases. During the dual-frequency composite pulse-enhanced swing GTA arc additive repair process, the ultra-high-frequency pulse current acts on the entire arc swing process, while the base current segment of the low-frequency pulse current corresponds to the left and right stop phases of the swinging arc, and the peak current segment corresponds to the middle swing phase of the swinging arc.

[0066] Example 2

[0067] A Q235 steel plate with a thickness of 5mm (within 0.2mm tolerance) was used as a simulated workpiece for repair. Before repair, the workpiece surface was polished with sandpaper and the oxide film removed with a wire brush until the base metal surface was smooth and free of marks. The polished surface and the surrounding 20mm area were then cleaned with a 75% alcohol solution to remove any oil or contaminants. The repair phase began after the surface was completely dry. ER120S-G steel welding wire with a diameter of 1.2mm was used as the filler material. GTA welding was performed using a DC positive connection with a constant current of 235A. 99.99% pure argon was used as the shielding gas at a flow rate of 15L / min, a synchronized wire feed speed of 30mm / s, and a tungsten electrode diameter of 3.2mm. During the simulated repair, the welding torch axis was maintained perpendicular to the workpiece surface. Before beginning the repair, ensure a good working environment with adequate ventilation to facilitate cooling of the deposited layer. After the simulated repair, the repaired part was allowed to cool naturally to room temperature.

[0068] Example 3

[0069] A Q235 steel plate with a thickness of 5mm (within 0.2mm tolerance) was used as a simulated workpiece for repair. Before repair, the workpiece surface was polished with sandpaper, and the oxide film was removed with a wire brush. The base metal surface was then polished to a smooth, flawless finish. The polished surface and the surrounding 20mm area were then cleaned with a 75% alcohol solution to remove any oil contamination. The repair process began after the surface was completely dry. ER120S-G steel welding wire with a diameter of 1.2mm was used as the filler material. GTA welding was performed using a DC positive connection with a constant current of 235A. 99.99% pure argon was used as the shielding gas at a flow rate of 15L / min, a synchronized wire feed speed of 30mm / s, and a tungsten electrode diameter of 3.2mm. During the simulated repair process, the welding torch was driven by an oscillator with a 25mm amplitude and a 9Hz frequency, with a 0.15s dwell time on each side. The torch axis was always maintained perpendicular to the workpiece's top surface. Before the repair begins, ensure the working environment is good and maintain good ventilation to help the deposit layer cool down. After the simulated repair is completed, the repaired part should be allowed to cool naturally to room temperature.

[0070] Example 4

[0071] A Q235 steel plate with a thickness of 5mm (within 0.2mm) was used as a simulated workpiece for repair. Before repair, the workpiece surface was polished with sandpaper, and the oxide film was removed with a wire brush. The base material surface was then polished to a smooth, mark-free finish. The polished surface and the surrounding 20mm area were then cleaned with a 75% alcohol solution to remove oil stains. The repair phase began after the surface was completely dry. A 1.2mm diameter ER120S-G steel welding wire was used as the filler material. GTA welding was performed using a DC positive connection with a constant current of 235A. 99.99% pure argon was used as the shielding gas at a flow rate of 15L / min. The wire feed speed was synchronized to 30mm / s, and the tungsten electrode diameter was 3.2mm. Immediately after the repair arc was struck, the ultrahigh frequency excitation power supply was activated, with a pulse frequency of 30kHz and a pulse width duty cycle of 50%. During the repair process, the welding torch is driven by an oscillator to oscillate horizontally with an amplitude of 25 mm and a frequency of 9 Hz, pausing for 0.15 seconds on each side. The torch axis remains perpendicular to the workpiece surface. Before the simulated repair begins, ensure a good working environment with adequate ventilation to aid in cooling the weld. After the repair is complete, the UHF excitation power supply is turned off, followed by the welding machine power supply, and the repaired part is allowed to cool naturally to room temperature.

[0072] Example 5

[0073] A Q235 steel plate with a thickness of 5mm (with an error of less than 0.2mm) was used as a simulated workpiece to be repaired. Before repair, the workpiece surface was polished with sandpaper, and the oxide film was removed with a wire brush. The surface of the parent material was polished until it was smooth and free of marks. Then, a 75% alcohol solution was used to clean the polished surface and the surrounding 20mm range to remove oil stains. After the surface was completely dry, the repair phase began. An ER120S-G steel welding wire with a diameter of 1.2mm was used as the filling material. The GTA welding method used was direct current positive connection, pulsed direct current, a base current of 210A, a peak current of 260A, a pulse width duty cycle of 50%, and 99.99% purity argon as the shielding gas. The flow rate was 15L / min, the synchronous matching wire feed speed was 30mm / s, and the tungsten electrode diameter was 3.2mm. During the simulated repair process, the welding torch was driven by an oscillator to oscillate horizontally with an amplitude of 25 mm and a frequency of 9 Hz. The left and right dwell times were 0.15 seconds, and the welding torch axis remained perpendicular to the workpiece surface. Before the simulated repair began, a good working environment with adequate ventilation was required to facilitate cooling of the weld. After the repair was completed, the repaired part was allowed to cool naturally to room temperature.

[0074] Example 6

[0075] A Q235 steel plate with a thickness of 5mm (within 0.2mm tolerance) was used as a simulated workpiece for repair. Before repair, the workpiece surface was polished with sandpaper and wire brushed to remove oxide film. The base metal surface was then polished to a smooth, mark-free finish. The polished surface and the surrounding 20mm area were then cleaned with a 75% alcohol solution to remove oil and contaminants. The repair process began after the surface was completely dry. ER120S-G steel welding wire with a diameter of 1.2mm was used as the filler material. GTA welding was performed using a direct current (DC) positive connection and pulsed DC with a base current of 210A and a peak current of 260A. 99.99% pure argon was used as the shielding gas at a flow rate of 15L / min. The wire feed speed was synchronized to 30mm / s, and the tungsten electrode diameter was 3.2mm. Immediately after the repair arc was initiated, the ultrahigh frequency excitation power supply was activated, with a pulse frequency of 30kHz and a pulse width duty cycle of 50%. During the simulated repair process, the welding torch was driven by an oscillator to oscillate horizontally with an amplitude of 25 mm and a frequency of 9 Hz. The dwell time on each side was 0.15 seconds, and the welding torch axis remained perpendicular to the workpiece surface. Before the simulated repair began, a good working environment with adequate ventilation was required to facilitate cooling of the weld. After the repair was completed, the UHF excitation power supply was turned off, followed by the welding machine power supply, and the repaired part was allowed to cool naturally to room temperature.

[0076] The experimental results are as follows Figure 5 As shown, the average current is about 235A, and the heat input of each process is the same. The deposition current / voltage characteristics of each process are as follows: Figure 5 In (a), the conventional DC mode deposition current / voltage has no significant fluctuation; Figure 5 In (b), the swinging DC deposition current changes steadily, while the deposition voltage changes periodically. Its voltage value varies in the range of about 13.4V to 14.6V. Its fluctuation is related to the welding torch swing trajectory and the change of tungsten electrode height. When swinging to the edge, the arc is long and the deposition voltage reaches the peak value. When swinging to the middle, the arc is short and the deposition voltage reaches the valley value, forming a trend of first decreasing and then increasing. Figure 5 In (c), after superimposing the UHF excitation current, the deposition current waveform becomes wider, generating a disturbance of about ±12A. This current is the peak current of the UHF excitation. The deposition voltage also fluctuates by about ±5V, which is approximately the peak voltage of the UHF excitation. Further analysis Figure 5 (d) and (e) show that after the swing pulse current and its superposition of ultra-high frequency excitation, the deposition current / voltage disturbance amplitude remains at about ±12A and ±5V, and both are superimposed with the same ultra-high frequency excitation.

[0077] Comparison of repair morphologies using different techniques Figure 6 As shown: the maximum melting depth of traditional DC deposition is about 7mm, and the narrowest melting width is about 10mm. Figure 6(a), this is mainly because the arc energy is concentrated and it is easy to produce unfused defects; the swing DC disperses the arc energy by lateral swing, reducing the melting depth to about 3.5mm and increasing the melting width to about 25mm. The single-pass repair area is increased, which can reduce the stacking process. Figure 6 (b) After superimposing ultra-high frequency excitation, the melting depth is stable at about 3.5 mm, the melting width is maintained at about 25 mm without undercut defects, and the solidification morphology is more uniform. Figure 6 (c) Comparison Figure 5 (b) and Figure 5 (c) Constant DC and Figure 6 (d) and Figure 6 (e) The pulsed DC additive repair process, through alternating base / peak value regulation, can reduce the average heat input, reduce the width of the heat-affected zone, improve the bonding density of the repair layer, further refine the grain size of the composite process deposited layer, and enhance the equiaxed degree of the grains in the heat-affected zone. The above results, to a certain extent, can illustrate that the swinging arc additive repair process under the combined regulation of low-frequency pulses and ultra-high-frequency pulses can achieve dynamic decoupling of heat input and arc force, that is, the peak segment of the low-frequency pulse matches the welding gun swing to maintain the stability of the molten pool, while the base segment reduces the accumulation of dwelling heat. Experimental results show that the dual-frequency composite pulse enhanced swinging arc additive repair technology of the present invention, through the micro-turbulence disturbance excited by the ultra-high-frequency pulse current, effectively breaks the solute diffusion boundary layer at the front of the solidification interface, significantly promotes the equiaxed growth of the metal grains in the molten pool, and simultaneously suppresses the formation of coarse columnar crystals and dendrites. This composite process significantly optimizes the solidification nucleation conditions of the molten pool, significantly improves the contour accuracy of the deposited layer, reduces the interlayer thermal cycle control error, improves the forming efficiency, and significantly outperforms the traditional GTA process in comprehensive performance.

[0078] In summary, the present invention achieves the coordinated regulation of ultra-high frequency excitation pulses and oscillating arcs in time and space. Under the dynamic decoupling of thermal-mechanical coupling, on the one hand, it promotes the formation of a uniform and fine equiaxed crystal structure in the molten pool; on the other hand, through the precise matching of the pulse peak / base current and the oscillation trajectory (movement / dwelling), the flow behavior of the molten pool is optimized, gas escape and impurity floating are accelerated, and the density of the deposited layer is greatly improved.

[0079] Compared to existing arc additive repair technologies, the present invention solves the problems mentioned in the background art, such as loss of control over heat input gradients, insufficient droplet transition stability, and difficulty in regulating interlayer thermal cycles. Its advantages lie in not only significantly improving the forming accuracy, metallurgical quality (grain refinement, segregation suppression), and mechanical properties (density, fatigue resistance) of the repair layer, but also featuring low equipment modification costs, convenient operation, and strong process adaptability. This enables efficient and highly reliable in-service repair of geometric defects in complex components such as nuclear power equipment, providing strong technical support for ensuring the safe operation of key equipment and extending its service life.

Claims

1. A dual-frequency composite pulse enhanced swing arc additive repair device, comprising a GTA welding system, the GTA welding system comprising a GTA welding gun, a wire feeder, a welding travel platform, a shielding gas system, and a welding power supply, characterized by: It also includes a pendulum oscillator, an ultra-high frequency excitation power supply and a multi-threaded synchronous digital control system; the oscillator drives the welding gun to swing periodically laterally through a flexible electrical signal; the ultra-high frequency pulse current of the ultra-high frequency excitation power supply and the low-frequency pulse current of the welding power supply are combined through an isolation coupling device and then introduced into the welding circuit; the multi-threaded synchronous digital control system is used to achieve coordinated control of the composite pulse current waveform and the swinging arc path.

2. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 1, characterized in that: The welding power supply is connected to the shielding gas through a pipeline, and the positive pole of the power supply is connected to the welding walking platform by a DC positive connection method; the pendulum oscillator is installed above the platform through a bracket, the pendulum oscillator is driven by a stepping motor, and the GTA welding gun and wire feeding device are fixed to the end of the pendulum oscillator by a clamp; the ultra-high frequency excitation power supply is connected in parallel with the welding circuit through an isolation coupling device to introduce high-frequency pulse current into the arc; the multi-threaded synchronous digital control system electrical signal is connected to the drive motor of the pendulum oscillator, and integrates a Hall current sensor, a Hall voltage sensor and an electromagnetic relay to realize full electrical signal closed-loop control, and through dynamic analysis of the correlation between the electrical signal characteristics of the swinging arc and the swing trajectory, the dual-frequency pulse parameters are adjusted in real time to achieve microsecond-level synchronous matching of the welding current waveform and the arc motion trajectory.

3. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 2, characterized in that: The multi-threaded synchronous digital control system includes a Hall current / voltage sensor, an analog-to-digital converter, and a PLC controller containing an FPGA module. The PLC controller obtains the parameters of the swing arc instantaneous voltage signal and pulse waveform characteristics in real time through the Hall current / voltage sensor. The PLC controller analyzes the dynamic mutual correlation by constructing a multi-layer perception neural network model, inputs the processed timing signal into a dual-output neural network, generates time compensation and energy adjustment instructions, drives the dual-frequency pulse current waveform parameters of the ultra-high frequency pulse and the welding low-frequency pulse to be corrected, and at the same time, uses synchronous error detection to feedback control deviations in real time, triggering control logic based on the composite pulse current waveform to form a closed loop. The system can adapt to the nonlinear changes of the arc trajectory and dynamically update the network weights through an online learning mechanism during continuous operation, wherein the current change rate is not less than 50A / μs.

4. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 3, characterized in that: The neural network model is a three-layer hybrid network including an input layer, a hidden layer and a dual output layer. The input layer includes 128 neurons and receives 100ms time window time series data, including four-channel signals: instantaneous value of arc voltage, instantaneous value of composite current, ultra-high frequency pulse phase mark, and swing position encoder signal; the feature extraction layer of the hidden layer is 64 convolution neurons, using SeLU activation function, and the convolution kernel time width is 15 sampling points, which is responsible for extracting pulse edge features and swing period features; the time series processing layer of the hidden layer is 48 bidirectional LSTM neurons, using tanh activation function to process long-term dependencies of time series; the time compensation branch of the output layer includes 3 linear activation neurons, namely rising edge compensation amount, falling edge compensation amount and pulse width compensation amount, with a compensation range of ±50μs, and the energy adjustment branch is 2 sigmoid activation neurons, including ultra-high frequency amplitude coefficient = 0.8-1.2 and low frequency peak coefficient = 0.9-1.

1.

5. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 4, characterized in that: The input processing flow of the dual-output neural network includes: after the arc voltage signal is subjected to wavelet noise reduction, the swing cycle characteristics of the pendulum oscillator are extracted through peak detection, namely the left / right side dwell point identification and swing midpoint positioning, by detecting the current signal, analyzing the pulse characteristics, including the low-frequency pulse duty cycle and ultra-high frequency pulse count, fusing the swing position encoder signal with the voltage signal, calculating the arc length change rate, and outputting the instruction generation logic including the time compensation instruction and the energy adjustment instruction. Among them, in the time compensation instruction, Rising edge compensation amount = K1 × (actual left endpoint arrival time - target time) + K2 × arc stability index during the left dwell phase; Falling edge compensation amount = K3 × (actual right endpoint voltage fluctuation integral) + K4 × swing acceleration; Pulse width compensation amount = K5 × (maximum arc length change in the dwell phase) × current duty cycle, where K1-K5 are the dynamic coefficients learned by the network. Among them, in the energy adjustment instruction, UHF amplitude coefficient = 0.8 + 0.4 × sigmoid (residence time × material evaporation prediction + arc length change rate); Low-frequency peak coefficient = 0.9 + 0.2 × tanh (swing speed × molten pool oscillation amplitude + heat accumulation correction term).

6. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 5, characterized in that: The synchronization error detection method includes a synchronization error detection method and an energy error detection method. The synchronization error detection method uses a timing unit of an FPGA module with a resolution of 10ns to continuously measure the actual time difference between the rising edge / falling edge of a low-frequency and ultra-high-frequency pulse. The energy error detection method is to calculate the integral power deviation in each half swing cycle. In the judgment standard, the synchronization error thresholds are: warning ±5μs, adjustment ±10μs, emergency ±20μs; the energy error thresholds are: warning ±3%, adjustment ±5%, emergency ±8%; weighted comprehensive error is used for deviation fusion. When the weighted comprehensive error is greater than 0.7 Or when a single error reaches the adjustment threshold, the neural network is activated to generate new parameters. When a single error reaches the emergency threshold, the preset safety strategy is started, including cutting off high-frequency pulses and performing time compensation by directly modifying the phase register of the PWM controller. The new rising edge time is based on the original set value, and the basic compensation amount output by the neural network is superimposed: the incremental adjustment value after the corrected Δ rising edge and the current synchronization error are dynamically weighted. The falling edge synchronization feedforward compensation is triggered by the advance compensation falling edge time. The energy adjustment includes ultra-high frequency voltage and low-frequency peak current adjustment. When any parameter exceeds the safety boundary, it is forcibly locked at the boundary value and an alarm is triggered.

7. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 6, characterized in that: The online learning mechanism adopts a dual-modal weight update strategy, which is triggered every 50 swing cycles in fine-tuning mode, and uses recursive least squares to dynamically adjust the output layer weights, completing real-time compensation within 15ms; deep update is initiated when material changes or continuous performance degradation are detected, and the Adam optimizer is used to update the weights of the entire network, and 1,000 groups of samples from the abnormal database are combined for 2-second training; a total of five protection mechanisms are implemented, including pre-update weight snapshots, virtual environment verification, post-update online testing, dynamic attenuation of the learning rate, and gradient clipping, to ensure that the learning process is stable and controllable; 100 groups of real-time working condition data are continuously rolled and stored, and the learning process is automatically triggered in the event of an abnormality.

8. The dual-frequency composite pulse enhanced swing arc additive repair device according to claim 1, characterized in that: The welding power supply provides direct current or pulsed direct current with a current range of 0-500A, a low-frequency pulse current frequency of 0.5-15Hz, and a duty cycle continuously adjustable from 10% to 90%; the ultra-high frequency excitation power supply outputs a square wave pulse current with a frequency of 20-100kHz, an output voltage of 0-75V, a pulse width of 0-50%, and a maximum output power of 3000W; the oscillator is connected to the GTA welding gun and controls its periodic oscillation, with a maximum unilateral oscillation amplitude of 100mm, an oscillation frequency of 0-25Hz, and a maximum dwell time of 2.5s when oscillating to the left or right sides.

9. A dual-frequency composite pulse enhanced swing arc additive repair method based on the device according to any one of claims 1 to 8, characterized in that: The following steps are involved: S1: Grind, remove rust and clean the surface of the repaired part; S2: Connect the ultra-high frequency excitation power supply in parallel with the arc circuit through an isolation coupling device to achieve the combination of ultra-high frequency pulse current and low frequency pulse current; S3: Make the welding gun perpendicular to the surface of the component to be repaired, fix the welding wire clamp and move it synchronously with the welding gun; S4: Introduce welding shielding gas, start the welding machine to strike the arc, turn on the ultra-high frequency excitation power supply and wire feeding mechanism to perform additive repair; when the repair is completed, first turn off the ultra-high frequency excitation power supply and wire feeding mechanism, and then turn off the welding machine to extinguish the arc.

10. The dual-frequency composite pulse enhanced swing arc additive repair method according to claim 9, characterized in that: Also includes: (a) An ultrahigh-frequency and low-frequency composite pulse current is applied to the swinging arc. The ultrahigh-frequency pulse compresses the arc through the electromagnetic confinement effect and generates periodic thermal shocks, which enhances the kinetic energy of the droplet transfer and forms micro-turbulent perturbations in the molten pool. The peak segment of the composite pulse is matched to the arc swing phase to maintain the stability of the molten pool flow field, and the base segment is matched to the dwell phase on the left and right sides of the arc to reduce heat accumulation. (b) When the arc swings left and right, large-volume, high-frequency droplet transfer is used to form an axis-focused molten pool with large melt depth in the middle and small melt depth on both sides; when the arc swings left and right, small-volume, low-frequency droplet transfer is used to form a shallow and wide edge molten pool; (c) Through periodic dynamic heat distribution of central heat flux focusing and edge heat isolation, the interlayer temperature gradient is reduced, and directional control of solidification structure and high-precision forming of the deposited layer are achieved.

Images