A method and system for intelligent power control and management of handheld welding equipment

Abstract

This invention belongs to the field of intelligent control and big data processing technology, specifically relating to a power intelligent control and management method and system based on handheld welding equipment. The method includes: real-time acquisition of arc physical state signals, welding torch operation posture signals, and material interface feature signals of the workpiece during the welding process using multimodal sensors; millisecond-level time-series modeling of the arc physical state signals to extract dynamic indicators of arc stability; spatial motion trajectory analysis of the welding torch operation posture signals to generate dynamic features of the welding torch posture; and joint frequency and spatial domain analysis of the material interface feature signals to identify the type of dissimilar metal interface region where the current welding point is located. This system, by integrating a multimodal millisecond-level sensing system of arc spectrum, welding torch posture, and material interface, achieves real-time, accurate, and comprehensive characterization of the welding conditions.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control and big data processing technology, specifically relating to a power intelligent control and management method and system based on handheld welding equipment. Background Technology

[0002] With the development of intelligent manufacturing and flexible manufacturing systems, handheld welding equipment is increasingly widely used in automobile manufacturing, aerospace, and precision machinery. The welding process places extremely high demands on the dynamic response capability of the power supply output, especially when handling the joining of dissimilar metals (such as aluminum-steel and copper-stainless steel). Significant differences in material thermal conductivity, melting point, and arc stability necessitate millisecond-level adaptive adjustment of power supply parameters based on real-time operating conditions. Traditional handheld welding equipment generally employs a control strategy that presets fixed current, voltage, or waveform parameters, essentially an open-loop control mode based on a static process database. This mode cannot detect sudden changes in the arc state caused by material interface variations, contact resistance fluctuations, or environmental disturbances during the welding process, easily leading to frequent arc extinction, excessive metal spatter, and poor weld formation, severely affecting joint strength and appearance quality.

[0003] Big data-driven intelligent welding control methods have attracted attention in recent years. These methods typically collect vast amounts of historical welding data, train complex models in the cloud to predict optimal power parameters, and then send the optimization results to the device for execution. However, existing solutions have fundamental limitations: firstly, model inference relies on high-performance cloud platforms, making deployment on resource-constrained handheld devices impossible; secondly, the latency in the data upload, cloud processing, and command feedback chain is typically hundreds of milliseconds or more, far from meeting the demands of the rapid evolution of the arc state on a millisecond-scale timescale during welding. Therefore, even with high model accuracy, their offline or near-real-time adjustment mechanisms are still insufficient to cope with sudden disturbances in actual welding conditions.

[0004] The lack of lightweight, low-latency real-time decision-making capabilities at the edge, coupled with the limitations of cloud-based big data models due to communication latency and device computing power, leads to a break in the "perception-decision-execution" closed loop. This contradiction is particularly pronounced in highly dynamic scenarios such as dissimilar metal welding: it requires both the high-dimensional feature understanding capabilities provided by big data models and the generation and issuance of control commands within an extremely short time. How to achieve millisecond-level dynamic optimization of power control strategies at the device level while maintaining the intelligence level of the model has become a pressing technical challenge in the field of intelligent control of handheld welding power supplies. Summary of the Invention

[0005] This invention provides a power intelligent control and management method and system for handheld welding equipment, which aims to solve the technical problem that the electric arc frequently extinguishes or the spatter rate exceeds the standard when the power parameters of the handheld welding equipment are fixed during the welding of dissimilar metals, thus causing welding defects. It also overcomes the limitation of existing historical offline adjustment schemes based on big data that cannot respond to millisecond-level dynamic changes in welding conditions.

[0006] This invention provides a power intelligent control and management method for handheld welding equipment, comprising:

[0007] The arc physical state signal, welding torch operation posture signal, and material interface characteristic signal of the workpiece being welded are collected in real time by a multimodal sensor.

[0008] The arc physical state signal is modeled in millisecond time series to extract dynamic indicators of arc stability; the welding torch operation posture signal is analyzed in spatial motion trajectory to generate dynamic features of welding torch posture.

[0009] The frequency and spatial domain joint analysis of the material interface feature signals is performed to identify the type of dissimilar metal interface region where the current welding point is located;

[0010] The arc stability dynamic index, the welding torch posture dynamic characteristics, and the dissimilar metal interface region type are fused using multi-source features to construct a real-time characterization vector for welding conditions.

[0011] Based on the real-time characterization vector of the welding condition, a power output command matching the current millisecond-level welding condition is generated through a pre-trained power parameter dynamic control neural network model. The power output command includes the output current amplitude, voltage slope, pulse frequency, and base value time ratio. The power output command is then sent to the welding power execution unit to achieve closed-loop dynamic control of the welding power output parameters.

[0012] In one embodiment of the present invention, the multimodal sensor includes an arc spectral sensor disposed at the front end of the welding torch, a six-axis inertial measurement unit inside the welding torch handle, and a high-frequency eddy current probe array integrated around the welding torch nozzle. The arc spectral sensor is used to capture the intensity distribution of the arc radiation spectrum in the 300 nm to 1100 nm band, with a sampling frequency of not less than 20 kHz. The six-axis inertial measurement unit is used to synchronously acquire the linear acceleration and angular velocity data of the welding torch in three-dimensional space, with a sampling frequency of not less than 1000 Hz. The high-frequency eddy current probe array consists of eight differential eddy current coils evenly distributed in a circle, with an operating frequency of 1.5 MHz, used to detect the spatial gradient changes of the electrical conductivity and magnetic permeability of the metal material within a 10 mm range in front of the weld.

[0013] As one embodiment of the present invention, the step of performing millisecond-level time-series modeling of the electric arc physical state signal to extract dynamic indicators of electric arc stability specifically includes: performing sliding slices on the electric arc spectral signal according to time windows, with each time window having a length of 5 milliseconds and a step size of 1 millisecond; performing wavelet packet decomposition on the spectral intensity sequence within each time window to extract the energy entropy of the third-level detail coefficients and the root mean square value of the fourth-level approximation coefficients; combining the energy entropy and the root mean square value to form a two-dimensional electric arc stability feature vector; when the energy entropy exceeds a preset threshold of 0.85 and the root mean square value is lower than a preset threshold of 0.3, it is determined that the current electric arc is in an unstable critical state.

[0014] As one embodiment of the present invention, the step of analyzing the spatial motion trajectory of the welding torch operation posture signal to generate welding torch posture dynamic features specifically includes: performing zero-bias correction and temperature drift compensation on the raw acceleration and angular velocity data output by the six-axis inertial measurement unit; converting the corrected angular velocity data into a sequence of attitude angles of the welding torch relative to the initial coordinate system using a quaternion integration algorithm; calculating the first derivative of the attitude angle sequence to obtain the rate of change of angular velocity, and the second derivative to obtain the angular acceleration; and statistically aggregating the attitude angles, the rate of change of angular velocity, and the angular acceleration within a time window to generate a nine-dimensional posture dynamic feature vector containing the mean, variance, and peak factor.

[0015] As one embodiment of the present invention, the step of performing joint frequency and spatial domain analysis on the material interface feature signals to identify the type of dissimilar metal interface region where the current welding point is located specifically includes: performing Fast Fourier Transform on the eight channel signals output by the high-frequency eddy current probe array to obtain the amplitude-frequency response curve of each channel in the frequency band from 0.5 MHz to 2.5 MHz; calculating the cross-correlation coefficient between the amplitude-frequency response curves of adjacent channels to form a seven-dimensional spatial correlation vector; performing principal component analysis on the spatial correlation vector and retaining the top three principal components with a cumulative contribution rate of not less than 95%; inputting the three principal components into a pre-trained support vector machine classifier to output the type of dissimilar metal interface region to which the current welding point belongs, wherein the type includes one of aluminum-steel transition zone, copper-stainless steel fusion line, and titanium-nickel diffusion zone.

[0016] As one embodiment of the present invention, the multi-source feature fusion to construct a real-time representation vector of welding conditions specifically includes: concatenating a two-dimensional arc stability feature vector, a nine-dimensional pose dynamic feature vector, and a material interface feature vector composed of three principal components to form a 14-dimensional original feature vector; normalizing the 14-dimensional original feature vector so that the numerical range of each dimension is mapped to the interval between 0 and 1; and inputting the normalized vector into a gating attention mechanism module, which dynamically allocates the importance of each feature dimension through learnable weights and outputs a weighted 14-dimensional real-time representation vector of welding conditions.

[0017] In one embodiment of the present invention, the pre-trained power parameter dynamic control neural network model is a three-layer fully connected neural network with 14 input layer nodes, 64 hidden layer nodes, and 4 output layer nodes. The activation function of the neural network is a modified linear unit. The loss function of the neural network is defined as a weighted multi-objective function, wherein the arc stability index error weight is 0.5, the spatter rate prediction error weight is 0.3, and the power consumption deviation weight is 0.2. During the training phase, the neural network is trained end-to-end using 100,000 sets of labeled data collected from real welding experiments. Each set of data contains 14-dimensional input features and corresponding four-dimensional power parameter labels.

[0018] As one embodiment of the present invention, after the power output command is sent to the welding power execution unit, a feedback verification step is also included: the actual output current and voltage waveforms are monitored in real time by a current Hall sensor and a voltage divider circuit; the actual waveforms are compared with the command waveforms point by point to calculate the mean square error; when the mean square error exceeds the preset tolerance value of 0.05, the command retransmission mechanism is triggered to regenerate and send the corrected power output command.

[0019] This invention provides a power intelligent control and management system based on a handheld welding device, comprising:

[0020] The multimodal sensing data acquisition unit is used to acquire in real time the physical state signals of the arc, the operating posture signals of the welding torch, and the material interface characteristic signals of the workpiece being welded during the welding process through an arc spectrum sensor, a six-axis inertial measurement unit, and a high-frequency eddy current probe array.

[0021] The arc stability dynamic index extraction unit is used to perform millisecond-level time-series modeling on the arc physical state signal to extract arc stability dynamic indices.

[0022] The welding torch posture dynamic feature generation unit is used to perform spatial motion trajectory analysis on the welding torch operation posture signal to generate welding torch posture dynamic features.

[0023] A dissimilar metal interface region type identification unit is used to perform frequency domain and spatial domain joint analysis on the material interface feature signals to identify the dissimilar metal interface region type where the current welding point is located.

[0024] The welding condition real-time characterization vector construction unit is used to fuse the arc stability dynamic index, the welding torch posture dynamic features and the dissimilar metal interface region type through multi-source feature fusion to construct the welding condition real-time characterization vector.

[0025] The power parameter dynamic control command generation unit is used to generate a power output command that matches the current millisecond-level welding condition based on the real-time characterization vector of the welding condition and through a pre-trained power parameter dynamic control neural network model.

[0026] The power supply execution and feedback verification unit is used to send the power output command to the welding power supply execution unit and perform feedback verification of the actual output waveform through a current Hall sensor and a voltage divider circuit.

[0027] In one embodiment of the present invention, the multimodal sensing data acquisition unit includes a back-illuminated complementary metal-oxide-semiconductor image sensor with a pixel size of 3.5 micrometers and a frame rate of 20,000 frames per second as the photosensitive element of the arc spectral sensor; a six-axis inertial measurement unit comprising a three-axis microelectromechanical system accelerometer and a three-axis microelectromechanical system gyroscope, with an accelerometer range of ±16 times the gravitational acceleration and a gyroscope zero-bias instability of less than 0.5 degrees per hour; and a high-frequency eddy current probe array where each differential eddy current coil is wound with 120 turns of enameled copper wire with a diameter of 0.2 mm, has an outer diameter of 5 mm, and an effective excitation current of 50 mA.

[0028] As one embodiment of the present invention, the wavelet packet decomposition performed by the arc stability dynamic index extraction unit adopts a biorthogonal wavelet basis function, and the decomposition level is four layers; the energy entropy calculation formula is -1 multiplied by the sum of the products of the energy proportion of each sub-band and its logarithm; the root mean square value calculation formula is the square root of the sum of the squares of the amplitudes of each sampling point divided by the number of sampling points.

[0029] As one embodiment of the present invention, the quaternion integration algorithm executed by the welding torch posture dynamic feature generation unit adopts the Runge-Kutta fourth-order numerical integration method with an integration step size of 1 millisecond; the attitude angles include roll angle, pitch angle and yaw angle; the peak factor is defined as the ratio of the signal peak value to the root mean square value.

[0030] In one embodiment of the present invention, the fast Fourier transform performed by the dissimilar metal interface region type identification unit adopts the radix-2 algorithm with 1024 transformation points; the support vector machine classifier adopts a radial basis function kernel with a penalty parameter C value of 10 and a kernel function width parameter gamma value of 0.01.

[0031] As one embodiment of the present invention, the gated attention mechanism module in the welding condition real-time representation vector construction unit includes two fully connected layers. The first layer has an output dimension of 14, the second layer has an output dimension of 14, the activation function is a hyperbolic tangent function, and finally the attention weights of each dimension are generated by an sigmoid function.

[0032] As one embodiment of the present invention, the three-layer fully connected neural network in the power parameter dynamic control instruction generation unit is embedded in the non-volatile memory of the main control chip of the welding equipment during deployment. The main control chip is a 32-bit reduced instruction set architecture processor with a main frequency of 400 MHz and a built-in floating-point arithmetic unit.

[0033] In one embodiment of the present invention, the current Hall sensor in the power supply execution and feedback verification unit has a range of ±500 amperes and a linearity error of less than 5%; the voltage divider circuit uses a high-precision metal film resistor to form a 10:1 attenuation network with a temperature drift coefficient of less than 25 ppm per degree Celsius; and the time window length for calculating the mean square error is 10 millimeters.

[0034] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0035] This invention achieves real-time, accurate, and comprehensive characterization of welding conditions by integrating a multimodal millisecond-level sensing system that integrates arc spectrum, welding torch posture, and material interface.

[0036] It abandons the traditional fixed parameter or offline historical data driven control mode and constructs a closed-loop millisecond-level dynamic control link from perception to decision-making to execution; the multi-source feature fusion and gating attention mechanism adopted can adaptively focus on the most critical working condition features at present, which significantly improves the recognition accuracy of dissimilar metal interface areas.

[0037] The pre-trained neural network model directly outputs four-dimensional power parameter commands, ensuring a high degree of matching between the control actions and transient operating conditions; the feedback verification mechanism further guarantees the accuracy of command execution.

[0038] This invention effectively suppresses arc extinction and excessive spatter, reducing the welding defect rate to below 0.8%, while controlling power consumption fluctuations to within 3%, achieving high reliability, high consistency and high energy efficiency operation of handheld welding equipment in complex dissimilar metal welding scenarios. Attached Figure Description

[0039] Figure 1 This is a schematic diagram of the overall technical solution architecture of the intelligent power control and management method and system based on handheld welding equipment proposed in this invention;

[0040] Figure 2 This is a schematic diagram of the core principle framework of the power parameter dynamic control neural network driven by multi-source feature fusion and gating attention mechanism in this invention;

[0041] Figure 3 This is a flowchart illustrating the logic of dual-channel feature extraction for millisecond-level timing modeling of electric arc physical state signals and spatial trajectory analysis of welding torch operation posture in this invention.

[0042] Figure 4 This is a flowchart illustrating the logical flow of the present invention, which uses joint frequency and spatial domain analysis of material interface characteristic signals to identify the type of dissimilar metal interface regions.

[0043] Figure 5 This is a flowchart illustrating the closed-loop control logic framework for constructing real-time welding condition representation vectors and generating power output commands in this invention.

[0044] Figure 6 This is a schematic diagram of the multi-level interaction relationship and data flow between the handheld welding device, the multimodal sensing unit, and the welding power supply execution unit in this invention. Detailed Implementation

[0045] Please refer to Figures 1 to 6 This invention provides a power intelligent control and management method and system for handheld welding equipment. It aims to solve the technical problem of frequent arc extinction or excessive spatter rate caused by fixed power parameters during the welding of dissimilar metals, leading to welding defects. It also overcomes the limitations of existing offline adjustment schemes based on big data, which cannot respond to millisecond-level dynamic changes in welding conditions. The specific embodiments of this invention will be described in detail below with reference to the accompanying drawings.

[0046] The method includes: acquiring arc physical state signals, welding torch operation posture signals, and material interface feature signals of the workpiece in real time during the welding process using multimodal sensors; performing millisecond-level time-series modeling on the arc physical state signals to extract arc stability dynamic indicators; performing spatial motion trajectory analysis on the welding torch operation posture signals to generate welding torch posture dynamic features; performing frequency and spatial domain joint analysis on the material interface feature signals to identify the type of dissimilar metal interface region where the current welding point is located; fusing the arc stability dynamic indicators, the welding torch posture dynamic features, and the dissimilar metal interface region type using multi-source features to construct a real-time welding condition representation vector; based on the real-time welding condition representation vector, generating a power output command matching the current millisecond-level welding condition through a pre-trained power parameter dynamic control neural network model, wherein the power output command includes output current amplitude, voltage slope, pulse frequency, and base value time ratio; and sending the power output command to the welding power execution unit to achieve closed-loop dynamic control of the welding power output parameters.

[0047] In the method described, step S1 involves real-time acquisition of arc physical state signals, welding torch operation posture signals, and material interface characteristic signals of the workpiece during the welding process using multimodal sensors. The multimodal sensors include an arc spectral sensor located at the front end of the welding torch, a six-axis inertial measurement unit inside the welding torch handle, and a high-frequency eddy current probe array integrated around the welding torch nozzle. The arc spectral sensor captures the intensity distribution of the arc radiation spectrum in the 300 nm to 1100 nm band, with a sampling frequency of at least 20 kHz. The six-axis inertial measurement unit synchronously acquires the linear acceleration and angular velocity data of the welding torch in three-dimensional space, with a sampling frequency of at least 1000 Hz. The high-frequency eddy current probe array consists of eight circumferentially distributed differential eddy current coils, operating at a frequency of 1.5 MHz, used to detect the spatial gradient changes in conductivity and permeability of the metal material within a 10 mm range in front of the weld. Data acquisition from all sensors is strictly synchronized, and a hardware triggering mechanism ensures that the timestamp alignment error of each channel does not exceed 50 microseconds. After analog-to-digital conversion, the raw data is stored in the local cache as a 16-bit signed integer, awaiting call from subsequent processing modules.

[0048] Step S2 involves performing millisecond-level time-series modeling on the arc physical state signal to extract dynamic indices of arc stability. This step specifically includes sliding slices of the arc spectral signal according to time windows, with each time window being 5 milliseconds long and a step size of 1 millisecond. Since the sampling frequency is 20 kHz, each time window contains 100 sampling points. Wavelet packet decomposition is performed on the spectral intensity sequence within each time window, with four decomposition levels using bioorthogonal wavelet basis functions. The third-level detail coefficients reflect the high-frequency perturbation characteristics of the arc, and the fourth-level approximation coefficients characterize the low-frequency trend. The energy entropy E of the third-level detail coefficients and the root mean square value of the fourth-level approximation coefficients are extracted. Energy entropy The calculation formula is:

[0049]

[0050] in For the first The proportion of energy carried by each cell to the total energy. This represents the total number of sub-bands. The root mean square value. The calculation formula is:

[0051]

[0052] in The fourth level of the approximation coefficient sequence One sampling point, Where is the sequence length. and The combination forms a two-dimensional arc stability feature vector. When Exceeding the preset threshold of 0.85 and When the value falls below a preset threshold of 0.3, the current electric arc is determined to be in an unstable critical state. This determination serves as one of the important bases for subsequent control.

[0053] Step S3 involves analyzing the spatial motion trajectory of the welding torch's operating attitude signal to generate dynamic characteristics of the welding torch's posture. This step first performs zero-bias correction and temperature drift compensation on the raw acceleration and angular velocity data output from the six-axis inertial measurement unit. Zero-bias correction uses static calibration data, while temperature drift compensation uses an internal temperature sensor to read the ambient temperature and correct it using a lookup table. The corrected angular velocity data is then converted into a sequence of attitude angles of the welding torch relative to the initial coordinate system using a quaternion integration algorithm.

[0054] Quaternion integration employs the Runge-Kutta fourth-order numerical integration method with an integration step size of 1 millisecond. Attitude angles include roll, pitch, and yaw. The first derivative of the attitude angle sequence is then calculated to obtain the rate of change of angular velocity, and the second derivative to obtain angular acceleration. Within a continuous 10-millimeter time window, the mean, variance, and peak factor are calculated for the attitude angles, rate of change of angular velocity, and angular acceleration, respectively, where the peak factor is defined as the ratio of the signal peak value to the root mean square value. This generates a nine-dimensional pose dynamic feature vector containing nine statistical measures.

[0055] Step S4 involves performing joint frequency and spatial domain analysis on the material interface feature signals to identify the type of dissimilar metal interface region where the current welding point is located. This step first performs Fast Fourier Transform (FFT) on the eight channels of the high-frequency eddy current probe array output, with 1024 transform points, using the radix-2 algorithm. The amplitude-frequency response curves of each channel in the 0.5 MHz to 2.5 MHz frequency band are obtained. The cross-correlation coefficients between the amplitude-frequency response curves of adjacent channels are calculated, forming a seven-dimensional spatial correlation vector. Principal component analysis is performed on the spatial correlation vector, retaining the top three principal components with a cumulative contribution rate of not less than 95%. These three principal components are input into a pre-trained support vector machine (SVM) classifier, which uses a radial basis function kernel with a penalty parameter C of 10 and a kernel width gamma parameter of 0.01. The classifier outputs the type of dissimilar metal interface region to which the current welding point belongs, including one of the following: aluminum-steel transition zone, copper-stainless steel fusion line, or titanium-nickel diffusion zone. The classification confidence score is also output simultaneously for subsequent feature fusion weight adjustment.

[0056] Step S5 involves fusing the arc stability dynamic index, the welding torch pose dynamic features, and the dissimilar metal interface type using multi-source features to construct a real-time welding condition representation vector. This step first concatenates the two-dimensional arc stability feature vector, the nine-dimensional pose dynamic feature vector, and the material interface feature vector composed of three principal components to form a 14-dimensional original feature vector. The 14-dimensional original feature vector is then normalized using a minimum-maximum scaling method, mapping the numerical range of each dimension to the 0-1 interval. The normalized vector is then input into a gated attention mechanism module. This module contains two fully connected layers: the first layer has an output dimension of 14 and uses a hyperbolic tangent function as its activation function; the second layer also has an output dimension of 14 and generates attention weights for each dimension using a sigmoid function. The final output is a weighted 14-dimensional real-time welding condition representation vector. The learnable parameters of the gated attention mechanism are jointly optimized with the main neural network during model training to ensure dynamic focusing on key features.

[0057] Step S6 involves generating power output commands that match the current millisecond-level welding condition based on the real-time representation vector of the welding condition and a pre-trained power parameter dynamic control neural network model. The neural network model is a three-layer fully connected neural network with 14 nodes in the input layer, 64 nodes in the hidden layer, and 4 nodes in the output layer. The activation function is a modified linear unit. The loss function is defined as a weighted multi-objective function L:

[0058]

[0059] in This is due to the error in the arc stability index. For the splash rate prediction error, To account for power consumption deviation, the neural network was trained end-to-end using 100,000 sets of labeled data collected from real welding experiments during the training phase. Each set of data contained 14-dimensional input features and corresponding four-dimensional power parameter labels. After training, the model was stored in the non-volatile memory of the welding equipment's main control chip. The main control chip is a 32-bit RISC processor with a clock frequency of 400 MHz, a built-in floating-point unit, and a single inference time of no more than 200 microseconds.

[0060] Step S7 involves sending the power output command to the welding power supply execution unit to achieve closed-loop dynamic control of the welding power supply output parameters. The power output command includes the output current amplitude, voltage slope, pulse frequency, and base time ratio. The command is sent to the welding power supply execution unit via the controller area network bus at a communication baud rate of 500 kilobits per second. The execution unit adjusts the drive signals of the power semiconductor devices in real time according to the command, changing the output waveform. Immediately after the command is sent, a feedback verification step is initiated: the actual output current and voltage waveforms are monitored in real time using a current Hall sensor and a voltage divider circuit.

[0061] The current Hall sensor has a range of ±500 amps and a linearity error of less than 5%. The voltage divider circuit uses a high-precision metal film resistor to form a 10:1 attenuation network, with a temperature drift coefficient of less than 25 ppm per degree Celsius. The actual waveform and the command waveform are compared point-by-point within a 10 mm time window to calculate the mean square error. When the mean square error exceeds the preset tolerance value of 0.05, the command retransmission mechanism is triggered, and a corrected power output command is regenerated and sent. The retransmission mechanism allows a maximum of three attempts; if the error still exceeds the tolerance, a safe power reduction mode is entered.

[0062] The above steps constitute a complete millisecond-level closed-loop control chain. The total latency from data acquisition to command execution is controlled within 5 milliseconds, meeting the response requirements for dynamic changes in welding conditions. The entire system continuously monitors the status of each stage during operation to ensure the accuracy and safety of control.

[0063] At the system level, this invention provides a power intelligent control and management system based on handheld welding equipment. The system includes a multimodal sensor data acquisition unit, an arc stability dynamic index extraction unit, a welding torch pose dynamic feature generation unit, a dissimilar metal interface region type identification unit, a welding condition real-time characterization vector construction unit, a power parameter dynamic adjustment command generation unit, and a power execution and feedback verification unit.

[0064] In the multimodal sensing data acquisition unit, the photosensitive element of the arc spectral sensor is a back-illuminated complementary metal-oxide-semiconductor (CMOS) image sensor with a pixel size of 3.5 micrometers and a frame rate of 20,000 frames per second. The six-axis inertial measurement unit includes a three-axis microelectromechanical system (MEMS) accelerometer and a three-axis MEMS gyroscope. The accelerometer has a range of ±16 times the acceleration due to gravity, and the gyroscope's zero-bias instability is less than 0.5 degrees per hour. Each differential eddy current coil in the high-frequency eddy current probe array is wound with 120 turns of 0.2 mm diameter enameled copper wire, has an outer diameter of 5 mm, and an effective excitation current of 50 mA.

[0065] The wavelet packet decomposition performed by the arc stability dynamic index extraction unit employs bioorthogonal wavelet basis functions, with a decomposition level of four. The calculation of energy entropy and root mean square value strictly follows the aforementioned formulas, and the calculation module is deployed on a dedicated digital signal processor to ensure real-time performance.

[0066] The quaternion integration algorithm executed by the welding torch pose dynamic feature generation unit adopts the Runge-Kutta fourth-order numerical integration method with an integration step size of 1 millisecond. The attitude angle calculation accuracy is better than 0.1 degrees, and the pose feature update frequency is 1000 Hz.

[0067] The Fast Fourier Transform (FFT) performed by the dissimilar metal interface region type identification unit uses the radix-2 algorithm with 1024 transformation points. The Support Vector Machine (SVM) classifier is deployed on an embedded coprocessor, with a classification latency of no more than 300 microseconds.

[0068] The gated attention mechanism module in the real-time representation vector construction unit for welding conditions comprises two fully connected layers. The first and second layers have an output dimension of 14, and the activation function is the hyperbolic tangent function. Finally, the attention weights for each dimension are generated through a sigmoid function. This module shares some low-level feature extraction parameters with the main neural network, reducing redundant computation.

[0069] The three-layer fully connected neural network in the power parameter dynamic control instruction generation unit is embedded in the non-volatile memory of the welding equipment's main control chip during deployment. The main control chip is a 32-bit RISC processor with a clock speed of 400 MHz and a built-in floating-point unit. Fixed-point quantization technology is used for model inference to improve computational efficiency while ensuring accuracy.

[0070] The current Hall sensor in the power supply execution and feedback verification unit has a range of ±500 amps and a linearity error of less than 5%. The voltage divider circuit uses a high-precision metal film resistor to form a 10:1 attenuation network with a temperature drift coefficient of less than 25 ppm per degree Celsius. The mean square error calculation time window is 10 mm long, and the verification period is 1 millisecond.

[0071] Through the collaborative work of the aforementioned units, this system achieves a complete closed loop from multimodal perception, feature extraction, working condition identification, parameter decision-making to execution feedback. In actual welding tests, for typical dissimilar metal combinations such as aluminum-steel, copper-stainless steel, and titanium-nickel, the system reduces the number of arc extinctions to below 0.2 times per minute, controls the spatter rate to below 1.5%, reduces the welding defect rate to below 0.8%, and controls power consumption fluctuations to below 3%. The system possesses high reliability, high consistency, and high energy efficiency, making it suitable for complex industrial environments.

Claims

1. A power intelligent control and management method for handheld welding equipment, characterized in that, include: The arc physical state signal, welding torch operation posture signal, and material interface characteristic signal of the workpiece being welded are collected in real time by multimodal sensors. The arc physical state signal is subjected to time-series modeling with a time resolution of milliseconds to extract dynamic indicators of arc stability. The spatial motion trajectory of the welding torch operation posture signal is analyzed to generate dynamic features of the welding torch posture. The frequency and spatial domain joint analysis of the material interface feature signals is performed to identify the type of dissimilar metal interface region where the current welding point is located; The arc stability dynamic index, the welding torch posture dynamic characteristics, and the dissimilar metal interface region type are fused using multi-source features to construct a real-time characterization vector for welding conditions. Based on the real-time characterization vector of the welding condition, a power output command matching the current millisecond-level welding condition is generated by a pre-trained power parameter dynamic control neural network model. The power output command includes the output current amplitude, voltage slope, pulse frequency, and base value time ratio. The power output command is sent to the welding power supply execution unit to achieve closed-loop dynamic control of the welding power supply output parameters.

2. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The multimodal sensor includes an arc spectral sensor located at the front end of the welding torch, a six-axis inertial measurement unit inside the welding torch handle, and a high-frequency eddy current probe array integrated around the welding torch nozzle. The arc spectral sensor is used to capture the intensity distribution of the arc radiation spectrum in the 300 nm to 1100 nm band, with a sampling frequency of not less than 20 kHz. The six-axis inertial measurement unit is used to simultaneously acquire the linear acceleration and angular velocity data of the welding torch in three-dimensional space, with a sampling frequency of not less than 1000 Hz. The high-frequency eddy current probe array consists of eight differential eddy current coils evenly distributed in a circle, with an operating frequency of 1.5 MHz, used to detect the spatial gradient changes of electrical conductivity and magnetic permeability of the metal material within a 10 mm range in front of the weld.

3. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The arc physical state signal is subjected to millisecond-level time-series modeling to extract dynamic indicators of arc stability, including: The arc spectrum signal is sliced ​​by time window, with each time window being 5 milliseconds long and a step size of 1 millisecond; Wavelet packet decomposition is performed on the spectral intensity sequence within each time window to extract the energy entropy of the third-level detail coefficients and the root mean square value of the fourth-level approximation coefficients; The energy entropy and the root mean square value are combined to form a two-dimensional arc stability feature vector; When the energy entropy exceeds a preset threshold of 0.85 and the root mean square value is lower than a preset threshold of 0.3, the current electric arc is determined to be in an unstable critical state.

4. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The spatial motion trajectory of the welding torch operation posture signal is analyzed to generate dynamic features of the welding torch posture, including: Zero bias correction and temperature drift compensation are performed on the raw acceleration and angular velocity data output by the six-axis inertial measurement unit; The corrected angular velocity data is converted into a sequence of attitude angles of the welding torch relative to the initial coordinate system using a quaternion integration algorithm. Calculate the first derivative of the attitude angle sequence to obtain the rate of change of angular velocity, and the second derivative to obtain the angular acceleration; The attitude angle, rate of change of angular velocity, and angular acceleration are statistically aggregated within a time window to generate a nine-dimensional pose dynamic feature vector containing mean, variance, and peak factor.

5. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The frequency and spatial domain joint analysis of the material interface feature signals is performed to identify the type of dissimilar metal interface region where the current welding point is located, including: Fast Fourier transform was performed on the eight channel signals output by the high-frequency eddy current probe array to obtain the amplitude-frequency response curves of each channel in the frequency band from 0.5 MHz to 2.5 MHz. Calculate the cross-correlation coefficients of the amplitude-frequency response curves between adjacent channels to form a seven-dimensional spatial correlation vector; Principal component analysis was performed on the spatial correlation vectors, and the top three principal components with a cumulative contribution rate of not less than 95% were retained. The three principal components are input into a pre-trained support vector machine classifier, which outputs the type of the dissimilar metal interface region to which the current welding point belongs. The type includes one of aluminum-steel transition zone, copper-stainless steel fusion line, and titanium-nickel diffusion zone.

6. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The arc stability dynamic index, the welding torch pose dynamic characteristics, and the dissimilar metal interface region type are fused using multi-source features to construct a real-time welding condition representation vector, including: The two-dimensional arc stability feature vector, the nine-dimensional pose dynamic feature vector, and the material interface feature vector composed of three principal components are spliced ​​together to form a 14-dimensional original feature vector. The 14-dimensional original feature vector is normalized so that the numerical range of each dimension is mapped to the interval between 0 and 1. The normalized vector is input into the gating attention mechanism module, which dynamically assigns the importance of each feature dimension through learnable weights and outputs a weighted 14-dimensional real-time representation vector of the welding condition.

7. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, The pre-trained power parameter dynamic control neural network model is a three-layer fully connected neural network with 14 nodes in the input layer, 64 nodes in the hidden layer, and 4 nodes in the output layer. The activation function of the neural network is a modified linear unit. The loss function of the neural network is defined as a weighted multi-objective function, where the arc stability index error weight is 0.5, the spatter rate prediction error weight is 0.3, and the power consumption deviation weight is 0.

2. During the training phase, the neural network is trained end-to-end using 100,000 sets of labeled data collected from real welding experiments. Each set of data contains 14-dimensional input features and corresponding four-dimensional power parameter labels.

8. The intelligent power control and management method for handheld welding equipment according to claim 1, characterized in that, After the power output command is sent to the welding power supply execution unit, a feedback verification step is also included: The actual output current and voltage waveforms are monitored in real time using a current Hall sensor and a voltage divider circuit. The actual waveform is compared with the command waveform point by point, and the mean square error is calculated. When the mean square error exceeds the preset tolerance value of 0.05, the command retransmission mechanism is triggered to regenerate and resend the corrected power output command.

9. A power intelligent control and management system based on handheld welding equipment, characterized in that, include: The multimodal sensing data acquisition unit is used to acquire in real time the physical state signals of the arc, the operating posture signals of the welding torch, and the material interface characteristic signals of the workpiece being welded during the welding process through an arc spectrum sensor, a six-axis inertial measurement unit, and a high-frequency eddy current probe array. The arc stability dynamic index extraction unit is used to perform time-series modeling on the arc physical state signal with a time resolution of milliseconds to extract the arc stability dynamic index. The welding torch posture dynamic feature generation unit is used to analyze the spatial motion trajectory of the welding torch operation posture signal to generate welding torch posture dynamic features. A dissimilar metal interface region type identification unit is used to perform frequency domain and spatial domain joint analysis on the material interface feature signals to identify the dissimilar metal interface region type where the current welding point is located. The welding condition real-time characterization vector construction unit is used to fuse the arc stability dynamic index, the welding torch posture dynamic features and the dissimilar metal interface region type through multi-source feature fusion to construct the welding condition real-time characterization vector. The power parameter dynamic control command generation unit is used to generate a power output command that matches the current millisecond-level welding condition based on the real-time characterization vector of the welding condition and through a pre-trained power parameter dynamic control neural network model. The power supply execution and feedback verification unit is used to send the power output command to the welding power supply execution unit and perform feedback verification of the actual output waveform through a current Hall sensor and a voltage divider circuit.

10. The intelligent power control and management system for handheld welding equipment according to claim 9, characterized in that, In the multimodal sensing data acquisition unit, the photosensitive element of the arc spectral sensor is a back-illuminated complementary metal-oxide-semiconductor image sensor with a pixel size of 3.5 micrometers and a frame rate of 20,000 frames per second; the six-axis inertial measurement unit includes a three-axis microelectromechanical system accelerometer and a three-axis microelectromechanical system gyroscope, with an accelerometer range of ±16 times the gravitational acceleration and a gyroscope zero-bias instability of less than 0.5 degrees per hour; each differential eddy current coil of the high-frequency eddy current probe array is wound with 120 turns of enameled copper wire with a diameter of 0.2 mm, the outer diameter of the coil is 5 mm, and the effective value of the excitation current is 50 mA.

Images