Collaborative control system based on double-sided double-arc root-avoiding welding

Abstract

The application relates to the technical field of automatic welding process control, and discloses a collaborative control system based on double-sided double-arc root-avoiding welding. The system extracts the mel-frequency cepstral coefficient of a welding sound pressure signal through an acoustic acquisition array, and obtains an arc transient short-circuit transition frequency and an energy integral value of an arc burning period through a voltage and current sampling circuit. A collaborative controller inputs the above coefficient and integral value into a long and short time memory network based on physical constraints, solves a time step prediction value of a critical penetration state, and adjusts a dry elongation of a back welding gun and a welding speed before the critical penetration state arrives according to the prediction value. The scheme completes parameter adjustment before penetration occurs, eliminates adjustment lag caused by molten pool thermal inertia and signal transmission delay, and overcomes the back forming burn-through and non-fusion defects in root-avoiding welding.

Description

Technical Field

[0001] This invention relates to the field of automated welding process control technology, and discloses a collaborative control system based on double-sided double-arc root-cleaning welding. Background Technology

[0002] Double-sided, double-arc, root-cleaning-free welding typically involves placing welding torches on opposite sides of the workpiece, relying on the combined heat input from the front and back arcs to melt the metal and achieve full penetration. In existing control schemes, the conventional approach is to place a temperature sensor or vision sensor on the back side to monitor changes in the physical state of the molten pool. When the temperature of the molten pool on the back side exceeds a set threshold or the weld width changes, the control system adjusts the welding current or voltage of the back arc based on the feedback signal. This control logic relies on the sensor's ability to capture physical changes; after acquiring the sensor signal, the control system performs internal logic calculations and then outputs adjustment commands to the back welding power source and actuator.

[0003] Based on the aforementioned feedback control mechanism, due to the thermal inertia of the flow and heat conduction of the liquid metal within the molten pool, when the back-side sensor detects abnormal changes in temperature or weld width, the physical evolution of the penetration state has already occurred and continues to advance. The inherent delays in signal acquisition, line transmission, and command execution in the control system cause the adjustment actions applied to the back-side arc to lag behind the real-time evolution of the molten pool's physical state. In the no-cleaning welding process, this lag prevents the back-side arc from intervening before the molten pool reaches the critical penetration state, directly resulting in physical defects such as burn-through or lack of fusion in the back-side weld formation. Summary of the Invention

[0004] The purpose of this invention is to provide a collaborative control system based on double-sided double-arc root-cleaning welding, which can effectively solve the problems in the background art.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows: A collaborative control system based on double-sided, double-arc, root-cleaning-free welding includes a front welding torch, a back welding torch, an acoustic acquisition array, a voltage and current sampling circuit, and a collaborative controller. The front welding torch and the back welding torch are respectively arranged on both sides of the workpiece to be welded. The acoustic acquisition array and the voltage and current sampling circuit are both connected to the collaborative controller. The acoustic acquisition array acquires the sound pressure signals generated by the front and back welding torches during the welding process and extracts the Mel frequency cepstral coefficients. The voltage and current sampling circuit acquires the arc transient short-circuit transition frequency and the arc-burning energy integral value of the front and back welding torches. The collaborative controller inputs the Mel frequency cepstral coefficients and the arc-burning energy integral value into a physically constrained long short-time memory network and outputs a predicted time step value for the time step before the critical penetration state. Based on the predicted time step value, the collaborative controller adjusts the back welding torch's extension and welding speed before the critical penetration state arrives, thus forming the system.

[0006] Preferably, the acoustic acquisition array includes multiple acoustic sensors, which are equally spaced and arranged around the nozzles of the front welding torch and the back welding torch. The acoustic acquisition array performs high-pass filtering on the acquired raw sound pressure signal, performs frame-by-frame windowing on the filtered sound pressure signal, performs fast Fourier transform on the windowed signal to obtain the power spectrum, inputs the power spectrum into a preset Mel triangular filter bank for nonlinear mapping in the frequency domain, takes the logarithm of the output result after nonlinear mapping and performs discrete cosine transform to extract the Mel frequency cepstral coefficients of a set dimension and transmits them to the cooperative controller.

[0007] Preferably, the voltage and current sampling circuit synchronously acquires the instantaneous voltage and instantaneous current values ​​of the front welding torch and the back welding torch at a preset sampling frequency. When the instantaneous voltage value is less than a preset short-circuit voltage threshold and the instantaneous current value is greater than a preset short-circuit current threshold, it is determined to be a short-circuit transition state. The number of short-circuit transition states within a unit time window is counted to calculate the arc transient short-circuit transition frequency. The time interval corresponding to the short-circuit transition state is removed. The product of the instantaneous voltage value and the instantaneous current value in the remaining time interval is integrated over time to generate the arc-burning period energy integral value and transmit it to the cooperative controller.

[0008] Preferably, the physical constraint-based long short-term memory network includes an input layer, a hidden layer, and an output layer. The hidden layer includes a forget gate, an input gate, and an output gate. The cell state update equation of the hidden layer introduces a physical constraint weight matrix. The diagonal elements of the physical constraint weight matrix are set as the product of the inverse of the thermal diffusivity of the workpiece to be welded and its thickness. The cooperative controller concatenates the Mel frequency cepstral coefficients and the arc-burning energy integral value into a multidimensional feature vector and inputs it into the input layer. The cell state is updated through the forget gate and the input gate. The output layer performs a linear mapping on the cell state.

[0009] Preferably, the output layer includes a fully connected layer and an activation function. The number of neurons in the fully connected layer is equal to the preset maximum prediction time step. The activation function is a normalized exponential function. The collaborative controller extracts the index value of the element with the highest probability density in the output vector of the normalized exponential function, multiplies the index value by the preset physical time interval of a single time step, and generates the predicted time step value of the distance to the critical melting state. The collaborative controller updates the multidimensional feature vector input to the input layer with the physical time interval of the single time step as the sliding step size.

[0010] Preferably, the collaborative controller internally stores a wire extension adjustment mapping table and a welding speed adjustment mapping table. The collaborative controller queries the wire extension adjustment mapping table to obtain the target wire extension adjustment amount corresponding to the predicted time step value, and queries the welding speed adjustment mapping table to obtain the target welding speed adjustment amount corresponding to the predicted time step value. The collaborative controller sends a first pulse sequence to the wire feeding motor of the back welding torch according to the target wire extension adjustment amount, and sends a second pulse sequence to the walking servo motor of the back welding torch according to the target welding speed adjustment amount.

[0011] Preferably, the acoustic acquisition array introduces an adaptive noise cancellation process before performing a fast Fourier transform. The voltage and current sampling circuit extracts the fundamental current signal of the back welding torch as a reference noise signal. The acoustic acquisition array uses the filtered sound pressure signal as the main input signal, calculates the cross-correlation matrix between the main input signal and the reference noise signal, constructs an adaptive filter coefficient matrix based on the cross-correlation matrix, uses the adaptive filter coefficient matrix to perform a weighted summation on the reference noise signal to generate estimated noise, and subtracts the estimated noise from the main input signal to obtain the denoised sound pressure signal.

[0012] Preferably, the voltage and current sampling circuit performs dynamic correction of the arc voltage before calculating the integral value of the arcing period energy. A displacement sensor is provided at the end of the conductive tip of the back welding torch. The displacement sensor measures the actual extension length of the back welding torch. The voltage and current sampling circuit reads the preset resistivity and cross-sectional area of ​​the welding wire. The actual extension length, the resistivity of the welding wire, and the cross-sectional area are substituted into Ohm's law to calculate the extension voltage drop. The collected instantaneous voltage value is added to the extension voltage drop to generate a corrected instantaneous voltage value. The corrected instantaneous voltage value is used to replace the instantaneous voltage value for integral calculation.

[0013] Preferably, the collaborative controller stores an online network parameter update mechanism. The collaborative controller records multiple time step prediction values ​​output by the physical constraint-based long short-term memory network within a historical welding cycle, retrieves the actual penetration state timestamp corresponding to the historical welding cycle, calculates the mean square error loss function between the time step prediction value and the actual penetration state timestamp, and updates the off-diagonal elements in the physical constraint weight matrix along the gradient in the opposite direction of the mean square error loss function using a stochastic gradient descent algorithm, while keeping the values ​​of the diagonal elements fixed.

[0014] Preferably, the collaborative controller is provided with a wire extension limit threshold. When the target wire extension adjustment causes the current wire extension value of the back welding torch to reach the wire extension limit threshold, the collaborative controller stops sending the first pulse sequence to the wire feeding motor, and calculates the lateral offset compensation amount based on the difference between the target wire extension adjustment amount and the wire extension limit threshold. The collaborative controller converts the lateral offset compensation amount into a position offset signal and sends the position offset signal to the cross slide servo driver connected to the front welding torch to adjust the lateral position of the front welding torch.

[0015] Compared with the prior art, the present invention has the following beneficial effects: 1. This invention extracts the Mel-frequency cepstral coefficients of the sound pressure signal using an acoustic acquisition array, and combines this with the energy integral value obtained during the arcing period from a voltage and current sampling circuit. These features are then input into a physically constrained long short-term memory network (LSTM), which outputs a predicted time step value for the distance to the critical penetration state. The co-controller adjusts the back welding torch extension and welding speed in advance based on this predicted value. This scheme completes the issuance and execution of adjustment commands before the penetration state occurs, changing the existing technology's reliance on post-feedback control logic and eliminating interference from molten pool thermal inertia and system signal processing delays on the back welding formation without root cleaning.

[0016] 2. The acoustic acquisition array incorporates adaptive noise cancellation processing, using the fundamental current signal of the back welding torch as a reference noise to construct an adaptive filter coefficient matrix, reducing the interference of welding site noise on acoustic characteristics. The voltage and current sampling circuit obtains the actual extension length through a displacement sensor and calculates the extension voltage drop by combining it with the welding wire resistivity, dynamically correcting the instantaneous voltage value and improving the accuracy of calculating the energy integral value during the arcing period. The collaborative controller uses data from historical welding cycles to calculate the mean square error loss function and updates network parameters online to adapt to changes in working conditions. When the extension adjustment reaches the limit threshold, the system converts the excess difference into a position offset signal to adjust the lateral position of the front welding torch, providing redundant adjustment channels. Attached Figure Description

[0017] Figure 1 This is an overall flowchart of the present invention; Figure 2 This is a flowchart of the acoustic feature extraction and adaptive noise reduction process of the present invention; Figure 3 This is a flowchart of the electrical parameter sampling and arc voltage dynamic correction process of the present invention; Figure 4 This is a flowchart of the Long Short-Term Memory Network prediction process of the present invention; Figure 5 This is a flowchart illustrating the mapping and adjustment process between the welding extension and welding speed of this invention. Figure 6 This is a flowchart of the online network parameter update and redundancy control process of the present invention. Detailed Implementation

[0018] Please refer to the attached document. Figure 1 This embodiment provides a collaborative control system for double-sided, double-arc, no-cleaning welding, including a front welding torch, a back welding torch, an acoustic acquisition array, a voltage and current sampling circuit, and a collaborative controller. The front and back welding torches are respectively positioned on opposite sides of the workpiece to be welded, with their welding axes perpendicular to and coaxially aligned with the workpiece surface. This concentrates the heat input of the front and back arcs at the same location on the weld seam, achieving collaborative heat input from both sides. The workpiece adopts a flat butt joint with an I-groove, suitable for the no-cleaning welding process. The signal output terminals of the acoustic acquisition array and the voltage and current sampling circuit are electrically connected to the signal input terminal of the collaborative controller, and the control output terminal of the collaborative controller is connected to the actuators of the front and back welding torches, respectively.

[0019] The acoustic acquisition array collects the sound pressure signals generated during the welding process of the front and back welding torches. After preprocessing the collected sound pressure signals, Mel-frequency cepstral coefficients are extracted. The acoustic sensors used in the acoustic acquisition array have wideband response characteristics, covering a frequency range of 100Hz-10000Hz, and can completely collect the sound pressure signals generated during the welding arc combustion and molten pool oscillation. The analog signals output by the sensors are transmitted to the analog-to-digital conversion port of the co-controller through shielded cables, reducing the interference of the strong electromagnetic environment at the welding site on signal transmission. The voltage and current sampling circuit synchronously collects the transient short-circuit transition frequency of the arc and the energy integral value during the arc combustion period of the front and back welding torches. The voltage and current sampling circuit uses Hall effect sensors with electrical isolation characteristics to avoid damage to the sampling circuit from the strong electrical signals of the main welding circuit, while ensuring the linearity and accuracy of the sampled signals.

[0020] The collaborative controller concatenates the received Mel frequency cepstral coefficients with the arc-burning energy integral value and the arc transient short-circuit transition frequency to generate a multi-dimensional feature vector. This multi-dimensional feature vector is then input into a physically constrained long short-term memory (LSTM) network. The LSM network is pre-trained using a welding process test dataset. This dataset contains synchronously collected welding process data for workpieces of different thicknesses and materials, along with corresponding timestamps of the critical penetration state. During pre-training, the constraint parameters related to welding physics are fixed, and only the trainable weights and biases are updated. After performing temporal operations on the input multi-dimensional feature vector, the network outputs a predicted time step from the critical penetration state.

[0021] Based on the predicted time step value, the collaborative controller adjusts the wire extension and welding speed of the back welding torch before the critical penetration state arrives. When the predicted time step value is greater than the preset adjustment trigger threshold, the collaborative controller keeps the current operating parameters of the back welding torch unchanged; when the predicted time step value is less than or equal to the preset adjustment trigger threshold, the collaborative controller immediately generates the corresponding control command and sends a pulse sequence to the wire feed motor and travel servo motor of the back welding torch to complete the adjustment of the wire extension and welding speed. The total execution time of the adjustment action is less than the physical time interval corresponding to a single time step, ensuring that the adjustment action is completed before the critical penetration state arrives, forming a complete collaborative control closed loop.

[0022] Table 1. Composition and Physical Meaning of Multidimensional Input Feature Vectors

[0023] This table clearly defines the complete structure of the multidimensional feature vectors input to the physically constrained Long Short-Term Memory (LSTM) network. All feature dimensions correspond to physical quantities that can be directly measured or calculated during the welding process. There are no abstract or uninterpretable feature dimensions, and the computation time windows of each feature are kept consistent, ensuring time synchronization between features. The total dimension of the feature vectors is 16, which can comprehensively characterize the evolution of the penetration state during the welding process, providing comprehensive input information for the network's advanced prediction.

[0024] This embodiment fully constructs the overall architecture and workflow of a collaborative control system based on double-sided double-arc root-cleaning welding. By fusing the Mel frequency cepstral coefficients extracted by the acoustic acquisition array with the arc-period energy integral value and short-circuit transition frequency obtained by the voltage and current sampling circuit, a multi-dimensional characterization of the welding penetration state evolution process is achieved. By introducing a physical constraint-based long short-term memory network, the time step distance to the critical penetration state is predicted in advance. Based on the prediction results, the collaborative controller adjusts the back welding torch extension and welding speed before the critical penetration state arrives. This changes the post-feedback control logic that relies on the already occurred state change in the prior art, eliminates the influence of molten pool thermal inertia and system signal processing and instruction execution delays on the back weld formation, and avoids defects such as burn-through or lack of fusion in the back weld.

[0025] Please refer to the attached document. Figure 2 In a preferred embodiment, the acoustic acquisition array includes multiple acoustic sensors, which are equally spaced and arranged around the nozzles of the front and back welding torches. Specifically, the outer wall of the nozzle of the front welding torch is provided with an annular mounting bracket, which is coaxially arranged with the nozzle. Four acoustic sensors are fixed to the annular mounting bracket by threaded connection. The central angle between two adjacent acoustic sensors is 90 degrees. The pickup surfaces of all acoustic sensors face the arc action area of ​​the welding torch, and the vertical distance between the pickup surface and the nozzle end is 15mm. The nozzle of the back welding torch adopts the same sensor arrangement as the front welding torch. The sampling frequency of all acoustic sensors in both the front and back welding torches is set to 48kHz. The sampling trigger signal is uniformly issued by the co-controller to ensure that the sampling time of all acoustic sensors is completely synchronized.

[0026] The acoustic acquisition array first performs anti-aliasing low-pass filtering on the raw sound pressure signal. The cutoff frequency of the low-pass filter is set to 20kHz to filter out high-frequency signals exceeding half of the sampling frequency, thus avoiding frequency aliasing. The signal after anti-aliasing filtering is converted from analog to digital to a 16-bit digital sound pressure signal, and then subjected to high-pass filtering. The high-pass filter uses an infinite impulse response Butterworth filter with an 8th order and a cutoff frequency of 200Hz to filter out low-frequency environmental vibrations and low-frequency noise generated by mechanical transmission at the welding site.

[0027] The high-pass filtered sound pressure signal was subjected to frame-by-frame windowing. The frame length was set to 20ms, corresponding to 960 sampling points, and the frame shift was set to 10ms, corresponding to 480 sampling points. There was a 50% overlap between adjacent frames to avoid information loss at the frame edges. A Hamming window was used as the window function, which can effectively suppress spectral leakage and improve the frequency resolution of the subsequent fast Fourier transform. The formula for calculating the time-domain sound pressure signal after windowing is as follows:

[0028] in, For the first Time-domain sound pressure signal after frame windowing This is the original time-domain sound pressure signal after high-pass filtering. For Hamming window functions, This is the index of discrete-time sampling points, with a value range of [value range missing]. , This represents the total number of sampling points for a single frame of signal.

[0029] Perform a Fast Fourier Transform (FFT) on the windowed signal for each frame, setting the FFT to 1024 points. Obtain the frequency domain amplitude spectrum of each frame. Square the amplitude spectrum and divide by the number of transform points to obtain the corresponding power spectrum. The formula for calculating the power spectrum is:

[0030] in, For the first Power spectrum of frame signal This is the frequency point index, with a value range of [value range missing]. , This is for Fast Fourier Transform (FFT) operations.

[0031] The calculated power spectrum is input into a preset Mel triangular filter bank for nonlinear frequency domain mapping. The total number of filters in the Mel triangular filter bank is set to 24, covering a frequency range of 200Hz-8000Hz. The center frequencies of the filters are evenly distributed on the Mel frequency scale. The conversion relationship between Mel frequency and linear frequency is as follows: ,in For Mel frequency, The frequency response of each Mel filter is triangular, with an amplitude of 1 at the center frequency and linearly decaying to 0 at the center frequency of the adjacent filter. After the power spectrum passes through each Mel filter, the total energy value of the corresponding filter frequency band is output.

[0032] The natural logarithm of the energy values ​​output by the 24 Mel filters is taken, and the result is input into the Discrete Cosine Transform (DCT) module to perform the DCT. The first 12 dimensions of the transformed coefficients are extracted as the final Mel frequency cepstral coefficients. The calculation formula is as follows:

[0033] in, For the first Vermeer frequency cepstral coefficients The range of values ​​is , For the extracted cepstral coefficient dimension, The total number of filters in the Mel-triangle filter bank. For the first The output energy value of the Mel filter. The extracted Mel frequency cepstral coefficients are transmitted to the feature splicing module via the internal bus of the co-controller, and spliced ​​with the electrical signal features.

[0034] Before performing the Fast Fourier Transform (FFT), the acoustic acquisition array incorporates an adaptive noise cancellation process. The voltage and current sampling circuit extracts the fundamental current signal from the back welding torch as a reference noise signal. This fundamental current signal is extracted by lock-in amplification of the acquired instantaneous current signal. The reference frequency for lock-in amplification is the inverter frequency of the welding power source, which accurately extracts the fundamental current signal related to electromagnetic interference from the welding arc. The acoustic acquisition array uses the high-pass filtered sound pressure signal as the main input signal. This main input signal contains useful acoustic signals generated by the welding arc and molten pool oscillations, as well as coherent noise generated by electromagnetic radiation from the welding power source and cable coupling. The cross-correlation matrix between the main input signal and the reference noise signal is calculated. Based on this cross-correlation matrix, an adaptive filter coefficient matrix is ​​constructed. The adaptive filter employs the least mean square algorithm, and the filter order is set to 32. The update formula for the filter coefficients is as follows:

[0035] in, for The adaptive filter coefficient vector at time t. This is the convergence step size factor of the filter. for Error signal at time, , Main input signal, The reference noise signal is used as a reference. The reference noise signal is weighted and summed using the coefficient matrix of the adaptive filter to generate an estimated noise. The estimated noise is then subtracted from the main input signal to obtain the denoised sound pressure signal. The denoised sound pressure signal then enters the subsequent frame segmentation, windowing, and fast Fourier transform stages.

[0036] Table 2. Comparison of Mel frequency cepstral coefficient penetration state discrimination under different framing parameters.

[0037] This table compares the ability of extracted Mel-frequency cepstral coefficients to distinguish welding penetration states under different framing and windowing parameters. Feature discriminative power is quantified by the ratio of inter-class dispersion to intra-class dispersion; a higher ratio indicates a stronger ability to distinguish different penetration states. The table data shows that when the frame length is set to 20ms and the frame shift to 10ms, the Mel-frequency cepstral coefficients exhibit the highest discriminative power in both the pre-penetration stable region and the critical penetration transition region. This effectively captures the changes in acoustic signal characteristics before the arrival of the critical penetration state, providing highly sensitive input features for subsequent advanced prediction.

[0038] This embodiment refines the sensor arrangement and complete signal processing flow of the acoustic acquisition array. By using an equally spaced, surrounding acoustic sensor array, omnidirectional acquisition of welding sound pressure signals is achieved, avoiding the blind spots of a single sensor. Through a complete process of frame windowing, fast Fourier transform, Mel frequency nonlinear mapping, and discrete cosine transform, Mel frequency cepstral coefficients that can effectively characterize the penetration state are extracted. By introducing an adaptive noise cancellation processing step based on the least mean square algorithm, using the fundamental wave signal of the welding current as reference noise, the interference of coherent electromagnetic noise at the welding site on the sound pressure signal is filtered out, further improving the signal-to-noise ratio of the extracted acoustic features and the ability to characterize the penetration state.

[0039] Please refer to the attached document. Figure 3 In a preferred embodiment, the voltage and current sampling circuit synchronously acquires the instantaneous voltage and current values ​​of the front and back welding torches at a preset sampling frequency. The sampling frequency is set to 100kHz, maintaining an integer multiple synchronization relationship with the sampling frequency of the acoustic acquisition array. The synchronization trigger signal output by the co-controller acts simultaneously on the analog-to-digital converter of the acoustic acquisition array and the analog-to-digital converter of the voltage and current sampling circuit, ensuring that the sampling time starting points of the acoustic signal and the electrical signal are completely consistent, and the time synchronization error does not exceed 10μs, avoiding the impact of time misalignment between different types of features on subsequent feature fusion and prediction results.

[0040] The voltage and current sampling circuits are equipped with independent voltage and current sampling channels for the front and back welding torches, respectively. Each voltage sampling channel uses a Hall voltage sensor with a rated measurement voltage range of 0-100V and a linearity error of no more than 0.2%. Each current sampling channel uses a Hall current sensor with a rated measurement current range of 0-500A and a linearity error of no more than 0.2%. The output terminals of all sampling channels are equipped with differential amplifier circuits and low-pass filter circuits. The gain of the differential amplifier circuit is set to 2, and the cutoff frequency of the low-pass filter circuit is set to 50kHz to filter out high-frequency interference components in the sampling signal. The filtered signal is input to the 16-bit analog-to-digital converter built into the co-controller to complete the conversion from analog signal to digital signal.

[0041] The voltage and current sampling circuit processes the converted digital voltage and current signals in real time. It presets short-circuit voltage and current thresholds. The short-circuit voltage threshold is set to 10V, based on the no-load voltage of the welding power source and the arc voltage characteristics during short-circuit transition. The short-circuit current threshold is set to 250A, based on the steady-state welding current and short-circuit peak current characteristics during welding. When the acquired instantaneous voltage value is less than the preset short-circuit voltage threshold and the instantaneous current value is greater than the preset short-circuit current threshold, it determines that the current sampling moment is in a short-circuit transition state, and records the start and end timestamps of the short-circuit transition state. It counts the number of short-circuit transition states occurring within a unit time window. The duration of the unit time window is consistent with the frame length of the acoustic signal, set to 20ms. The number of short-circuit transitions within the unit time window is divided by the duration of the time window to calculate the arc transient short-circuit transition frequency. This generates the arc transient short-circuit transition frequency values ​​for the front and back welding torches, respectively, and transmits them to the feature splicing module of the collaborative controller.

[0042] The voltage and current sampling circuit, based on the timestamp of the short-circuit transition state, removes the time interval corresponding to the short-circuit transition state within a unit time window, leaving the remaining time interval as the arcing period time interval. The product of the instantaneous voltage and instantaneous current values ​​within the arcing period time interval is integrated over time to generate the arcing period energy integral value. The arcing period energy integral values ​​for both the front and back welding torches are calculated separately and transmitted to the feature splicing module of the collaborative controller. The formula for calculating the arcing period energy integral value is:

[0043] in, This represents the integral value of energy during the arcing period within a unit time window. The arcing period is the time interval within a unit time window after excluding the short-circuit transition state. This is the corrected instantaneous arc voltage value. The instantaneous arc current value was collected synchronously. It is a continuous-time variable.

[0044] Before calculating the energy integral value during the arcing period, the voltage and current sampling circuit performs dynamic correction of the arc voltage to eliminate voltage measurement deviations caused by changes in wire extension. A laser displacement sensor is installed at the end of the contact tip of the back welding torch. The measurement direction of the laser displacement sensor is parallel to the wire feed direction, and the measurement point is located at the end of the wire extending from the contact tip. This allows for real-time measurement of the actual length of the wire extending from the contact tip, i.e., the actual wire extension length. The sampling frequency of the laser displacement sensor is consistent with the sampling frequency of the voltage and current sampling circuit, which is 100kHz, ensuring time synchronization between the wire extension measurement value and the voltage and current sampling value. The storage unit of the voltage and current sampling circuit pre-stores the resistivity of the welding wire material and the cross-sectional area of ​​the welding wire. When the welding wire material is carbon steel, the resistivity is set to... When the welding wire diameter is 1.2mm, the cross-sectional area is set to Substituting the real-time measured actual wire extension length, wire resistivity, wire cross-sectional area, and synchronously acquired instantaneous current value into Ohm's law calculation formula, the real-time wire extension voltage drop is obtained. The calculation formula for the wire extension voltage drop is as follows:

[0045] in, for Voltage drop during the elongation at any given moment. The resistivity of the welding wire material. The actual dry extension length measured by the displacement sensor. This refers to the cross-sectional area of ​​the welding wire. This refers to the instantaneous arc current value acquired synchronously.

[0046] The original instantaneous voltage value is added to the calculated voltage drop due to dry stretching to generate the corrected instantaneous voltage value. The formula for calculating the corrected instantaneous voltage value is as follows:

[0047] in, This is the corrected instantaneous voltage value. The original instantaneous voltage value is acquired by the voltage and current sampling circuit. The corrected instantaneous voltage value is used to replace the original instantaneous voltage value for subsequent arcing period energy integration calculation.

[0048] Table 3 Comparison of Voltage Correction and Arc Energy Integral Deviation under Different Actual Dry Extension Lengths

[0049] This table shows the impact of arc extension voltage drop on the calculation results of arc energy integral under different actual arc extension lengths. The welding test conditions were: carbon steel welding wire diameter 1.2mm, welding current effective value 200A, unit time window length 20ms, and short-circuit transition time percentage 15%. The relative deviation is the proportion of the uncorrected arc energy integral value to the corrected arc energy integral value. As can be seen from the table data, with the increase of actual arc extension length, the arc extension voltage drop gradually increases, and the relative deviation of the uncorrected arc energy integral value also increases linearly. This verifies the necessity of the dynamic correction of arc voltage. Through correction, the energy calculation deviation caused by changes in arc extension can be effectively eliminated, and the accuracy of the calculation of the arc energy integral value can be improved.

[0050] This embodiment refines the synchronous sampling mechanism of the voltage and current sampling circuit, the short-circuit transition state determination process, and the arcing period energy integral calculation method. Through multi-channel synchronous sampling design, the time synchronization of electrical and acoustic signals is ensured. Through the dual-threshold short-circuit transition state determination method, the accurate division between the short-circuit transition state and the arcing period is achieved, ensuring the accuracy of the calculation of the short-circuit transition frequency and the arcing period energy integral value. By introducing a dynamic correction link for arc voltage based on actual dry extension measurement, the voltage measurement deviation caused by dry extension changes is eliminated, avoiding the calculation error of the arcing period energy integral value, and providing accurate electrical signal feature input for subsequent melting penetration state prediction.

[0051] Please refer to the attached document. Figure 4 In a preferred embodiment, the physically constrained long short-term memory network includes an input layer, a hidden layer, and an output layer. The number of neurons in the input layer is consistent with the dimension of the input multidimensional feature vector, which is set to 16 dimensions. The hidden layer is a stacked 2-layer long short-term memory network layer, with 64 hidden units in each layer. Each hidden unit includes a forget gate, an input gate, and an output gate. The cell state update equation of the hidden layer incorporates a physically constrained weight matrix. The output layer includes a fully connected layer and an activation function layer. The number of neurons in the fully connected layer is equal to the preset maximum prediction time step.

[0052] The collaborative controller concatenates the 12-dimensional Mel frequency cepstral coefficients extracted by the acoustic acquisition array, the arc-period energy integral values ​​of the front and back welding torches, and the arc transient short-circuit transition frequencies of the front and back welding torches to generate a 16-dimensional multi-dimensional feature vector. This multi-dimensional feature vector is then input into the input layer of the network. The input layer normalizes the input feature vector using a min-max normalization method, mapping the value of each feature dimension to the [0,1] interval. The normalization parameters are determined through statistical results from the pre-training dataset to ensure a balanced contribution of feature dimensions of different scales to the network.

[0053] The normalized feature vector from the input layer is fed into the first hidden layer. The forget gate in the hidden layer filters the cell state from the previous time step, determining which historical information needs to be retained. The formula for calculating the forget gate is:

[0054] in, for The output value of the forget gate. It is the sigmoid activation function. Here is the weight matrix for the forget gate. for The hidden layer output state at any given time. for The multidimensional feature vectors input into the network at each time step. This is the bias vector for the forget gate.

[0055] The input gate determines the proportion of new information at the current moment that needs to be stored in the cell state, and simultaneously generates candidate cell states. The formulas for calculating the input gate and candidate cell states are as follows:

[0056]

[0057] in, for The output value of the input gate at any time, Here is the weight matrix of the input gate. This is the bias vector for the input gate. for The state of candidate cells at any given time. This is the weight matrix for the candidate cell states. This is the bias vector for the candidate cell state. It is the hyperbolic tangent activation function.

[0058] A physical constraint weight matrix is ​​introduced during the cell state update process to physically constrain the candidate cell states, ensuring that the cell state update conforms to the physical laws of welding heat conduction. The cell state update formula with physical constraints is as follows:

[0059] in, for The constantly updated state of the cell. for Cellular state at any given moment For Hadamard product operations, This is the physical constraint weight matrix. For a diagonally dominant matrix, its diagonal elements satisfy... , The thickness of the workpiece to be welded. The thermal diffusivity of the workpiece to be welded is given, and the off-diagonal elements are trainable parameters.

[0060] The output gate determines the information that the cell state needs to output at the current moment, and generates the hidden layer output state at the current moment. The formulas for calculating the output gate and the hidden state are as follows:

[0061]

[0062] in, for The output value of the output gate at any time. This is the weight matrix of the output gate. This is the bias vector for the output gate. for The output state of the hidden layer at any given time. The output state of the first hidden layer is input to the second hidden layer. The second hidden layer uses the exact same computational logic as the first layer, and its output state is input to the output layer.

[0063] The output layer consists of a fully connected layer and an activation function. The number of neurons in the fully connected layer is equal to the preset maximum prediction time step, set to 20, corresponding to a maximum prediction time length of 200ms. The fully connected layer linearly maps the output state of the second hidden layer, generating a 20-dimensional output vector. Each element of the output vector corresponds to the original output value of a time step category. The activation function uses a normalized exponential function to normalize the output vector of the fully connected layer, generating the prediction probability value for each time step category. The formula for calculating the normalized exponential function is as follows:

[0064] in, For the first The predicted probability value for each time step category. The first output of the fully connected layer The value of each neuron. This is the preset maximum number of prediction time steps. The range of values ​​is .

[0065] The collaborative controller extracts the index value corresponding to the element with the highest probability density in the output vector of the normalized exponential function, multiplies this index value by the preset physical time interval of a single time step, and generates a predicted time step value for the distance from the critical melting state. The physical time interval of a single time step is set to 10ms, and the formula for calculating the predicted time step value is as follows:

[0066] in, This is the predicted time step from the critical melt penetration state. The operation is to retrieve the index corresponding to the maximum value. This refers to the physical time interval corresponding to a preset single time step. The collaborative controller updates the multidimensional feature vector of the input network with the physical time interval of a single time step as the sliding step size, thereby achieving continuous rolling prediction of the critical melt penetration state.

[0067] Please refer to the attached document. Figure 5 The collaborative controller's storage unit pre-stores a weld elongation adjustment mapping table and a welding speed adjustment mapping table. Both tables are generated through prior welding process testing and calibration. During calibration, for workpieces of different thicknesses and materials, the optimal weld elongation adjustment and welding speed adjustment corresponding to different time step prediction values ​​are tested to ensure that the adjusted welding parameters can maintain stable formation of the molten pool at the critical penetration state, avoiding burn-through or incomplete fusion defects. Based on the generated time step prediction value, the collaborative controller queries the weld elongation adjustment mapping table to obtain the target weld elongation adjustment value corresponding to the current time step prediction value, and simultaneously queries the welding speed... The system adjusts the mapping table to obtain the target welding speed adjustment amount corresponding to the predicted value of the current time step. Based on the target wire extension adjustment amount, the collaborative controller sends a first pulse sequence to the wire feeding motor of the back welding torch. The number of pulses in the first pulse sequence is linearly related to the target wire extension adjustment amount. After receiving the pulse sequence, the wire feeding motor drives the welding wire to feed or retract, adjusting the length of the welding wire extending from the contact tip. Based on the target welding speed adjustment amount, the collaborative controller sends a second pulse sequence to the travel servo motor of the back welding torch. The pulse frequency of the second pulse sequence is linearly related to the target welding speed. After receiving the pulse sequence, the travel servo motor adjusts the motor speed, changing the travel speed of the back welding torch.

[0068] The collaborative controller's storage unit contains an online network parameter update mechanism, enabling online updates of network parameters using historical welding data to adapt to changes in different welding conditions. The collaborative controller records multiple time-step prediction values ​​output by the physically constrained long short-term memory network within historical welding cycles, and simultaneously retrieves the actual penetration state timestamps corresponding to the historical welding cycles. These actual penetration state timestamps are calibrated using the metallographic inspection results of the workpiece after welding, or in real-time using the penetration detection results of an online infrared thermal imaging device. The mean square error loss function between the predicted time-step values ​​and the actual penetration state timestamps within the historical welding cycles is calculated. The formula for calculating the loss function is as follows:

[0069] in, The value of the mean squared error loss function. This represents the total number of samples from historical welding cycles. For the first Predicted time step value for each historical welding cycle For the first The timestamp of the actual penetration state corresponding to each historical welding cycle.

[0070] The collaborative controller updates the off-diagonal elements of the physical constraint weight matrix along the inverse direction of the gradient of the mean squared error loss function using the stochastic gradient descent algorithm. Simultaneously, it updates other weight parameters and bias vectors of the network. During the update process, the values ​​of the diagonal elements of the physical constraint weight matrix remain constant to ensure that the network always conforms to the physical laws of welding heat conduction. The update formula for the off-diagonal elements of the physical constraint weight matrix is ​​as follows:

[0071] in, For the first The set of off-diagonal elements of the physical constraint weight matrix at the next iteration. The learning rate for the stochastic gradient descent algorithm is... For the loss function relative to The gradient value is calculated. The online update process uses a mini-batch update method, with a batch size of 32 historical welding samples and a learning rate of 0.001. When the loss function value is less than the preset convergence threshold, the update stops to ensure that the prediction accuracy of the network remains at a stable level.

[0072] The collaborative controller's storage unit is equipped with a wire extension limit threshold, which includes an upper limit threshold and a lower limit threshold. The upper limit threshold is set to 30mm, and the lower limit threshold is set to 8mm. The adjustment range of the wire extension is limited to between 8mm and 30mm to prevent welding instability or wire feeding mechanism malfunction caused by the wire extension exceeding the reasonable range for an extended period. When the target wire extension adjustment causes the current wire extension value of the back welding torch to reach the wire extension limit threshold, the collaborative controller stops sending the first pulse sequence to the wire feeding motor and no longer adjusts the wire extension of the back welding torch. At the same time, based on the difference between the target wire extension adjustment and the wire extension limit threshold, the lateral offset compensation of the front welding torch is calculated. The formula for calculating the lateral offset compensation is as follows:

[0073] in, This is the lateral offset compensation amount for the welding torch on the front side. The preset compensation coefficient,

[0074] This is the difference between the target extension adjustment amount and the extension limit threshold. The co-controller converts the calculated lateral offset compensation amount into a position offset signal, which is then sent to the cross slide servo driver connected to the front welding torch. The cross slide servo driver drives the front welding torch to move laterally along the direction perpendicular to the welding direction by the corresponding compensation distance. By adjusting the lateral position of the front welding torch, the heat input distribution of the front arc is changed, compensating for the insufficient heat input adjustment capability caused by the limited extension adjustment of the back welding torch, and ensuring stable control of the penetration state.

[0075] Table 4. Mapping table of welding torch adjustment parameters corresponding to the predicted time step values ​​from the critical penetration state.

[0076] This table contains the calibration data for the welding extension adjustment mapping table and welding speed adjustment mapping table stored internally by the collaborative controller. The calibration test conditions were: butt welding of a 12mm thick carbon steel plate with an I-groove, a welding current of 280A for the front welding torch, an initial welding current of 220A for the back welding torch, an initial welding extension of 15mm, and an initial welding speed of 300mm / min. In the table, a positive value for the target welding extension adjustment represents an increase in welding extension, and a negative value represents a decrease in welding extension; a negative value for the target welding speed adjustment represents a decrease in welding speed, and a positive value represents an increase in welding speed; the lateral offset compensation coefficient is the proportional coefficient between the lateral offset compensation amount and the difference between the welding extension adjustment amount and the threshold value. This table clarifies the welding torch adjustment parameters corresponding to the predicted values ​​at different time steps, providing a direct execution basis for the collaborative controller's advance adjustment actions and ensuring the matching of adjustment actions with the evolution of the penetration state.

[0077] Please refer to the attached document. Figure 6 This embodiment refines the complete structure and computational logic of the physical constraint-based long short-term memory network. By introducing a physical constraint weight matrix into the cell state update equation, it ensures that the network's prediction results conform to the physical laws of welding heat conduction, thus improving the interpretability and stability of the network prediction. Through the normalized exponential function and index mapping of the output layer, it achieves accurate generation of the predicted time step value from the critical penetration state. Through a pre-calibrated mapping table, it realizes the coordinated advance adjustment of the back welding torch extension and welding speed. Through the online update mechanism of network parameters, it can continuously optimize network parameters using historical welding data to adapt to changes in different welding conditions. By setting a lateral offset redundant adjustment channel under the extension limit threshold, when the back welding torch extension adjustment is limited, it achieves supplementary adjustment of heat input by adjusting the lateral position of the front welding torch, improving the system's adaptability to working conditions and operational stability, and ensuring the back weld formation quality of double-sided double-arc root-cleaning welding.

Claims

1. A collaborative control system based on double-sided double-arc root-cleaning-free welding, characterized in that, It includes a front welding torch, a back welding torch, an acoustic acquisition array, a voltage and current sampling circuit, and a collaborative controller; The front welding gun and the back welding gun are respectively arranged on both sides of the workpiece to be welded; Both the acoustic acquisition array and the voltage and current sampling circuit are connected to the collaborative controller. The acoustic acquisition array acquires the sound pressure signal generated during the welding process of the front welding gun and the back welding gun and extracts the Mel frequency cepstral coefficients; The voltage and current sampling circuit acquires the transient short-circuit transition frequency of the arc and the energy integral value during the arc burning period of the front welding torch and the back welding torch; The collaborative controller inputs the Mel frequency cepstral coefficients and the arc-burning energy integral value into a physical constraint-based long short-term memory network, and outputs a predicted time step value for the distance to the critical melting state. The collaborative controller adjusts the back welding torch extension and welding speed before the critical penetration state arrives based on the predicted time step value. The physical constraint-based long short-term memory network includes an input layer, a hidden layer, and an output layer. The hidden layer contains a forget gate, an input gate, and an output gate. The cell state update equation of the hidden layer introduces a physical constraint weight matrix. The diagonal elements of the physical constraint weight matrix are set as the product of the inverse of the thermal diffusivity of the workpiece to be welded and its thickness. The cooperative controller concatenates the Mel frequency cepstral coefficients and the arc-burning energy integral value into a multidimensional feature vector and inputs it into the input layer. The cell state is updated through the forget gate and the input gate. The output layer performs a linear mapping on the cell state. The collaborative controller stores an online network parameter update mechanism. It records multiple time step prediction values ​​output by the physical constraint-based long short-term memory network within a historical welding cycle, retrieves the actual penetration state timestamp corresponding to the historical welding cycle, calculates the mean square error loss function between the time step prediction value and the actual penetration state timestamp, and updates the off-diagonal elements in the physical constraint weight matrix along the gradient in the opposite direction of the mean square error loss function using a stochastic gradient descent algorithm, while keeping the values ​​of the diagonal elements fixed.

2. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 1, characterized in that, The acoustic acquisition array includes multiple acoustic sensors, which are equally spaced around the nozzles of the front and rear welding torches. The acoustic acquisition array performs high-pass filtering on the acquired raw sound pressure signal, performs frame-by-frame windowing on the filtered sound pressure signal, performs fast Fourier transform on the windowed signal to obtain the power spectrum, inputs the power spectrum into a preset Mel triangular filter bank for nonlinear mapping in the frequency domain, takes the logarithm of the nonlinearly mapped output and performs discrete cosine transform to extract the Mel frequency cepstral coefficients of a set dimension and transmits them to the cooperative controller.

3. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 1, characterized in that, The voltage and current sampling circuit synchronously acquires the instantaneous voltage and instantaneous current values ​​of the front welding torch and the back welding torch at a preset sampling frequency. When the instantaneous voltage value is less than a preset short-circuit voltage threshold and the instantaneous current value is greater than a preset short-circuit current threshold, it is determined to be a short-circuit transition state. The number of short-circuit transition states within a unit time window is counted to calculate the arc transient short-circuit transition frequency. The time interval corresponding to the short-circuit transition state is removed. The product of the instantaneous voltage value and the instantaneous current value in the remaining time interval is integrated over time to generate the arc-burning period energy integral value and transmit it to the cooperative controller.

4. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 3, characterized in that, The output layer includes a fully connected layer and an activation function. The number of neurons in the fully connected layer is equal to the preset maximum prediction time step. The activation function is a normalized exponential function. The collaborative controller extracts the index value of the element with the highest probability density in the output vector of the normalized exponential function, multiplies the index value by the preset physical time interval of a single time step, and generates the predicted time step value of the distance to the critical melting state. The collaborative controller updates the multidimensional feature vector input to the input layer with the physical time interval of the single time step as the sliding step size.

5. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 1, characterized in that, The collaborative controller internally stores a wire extension adjustment mapping table and a welding speed adjustment mapping table. The collaborative controller queries the wire extension adjustment mapping table to obtain the target wire extension adjustment amount corresponding to the predicted time step value, and queries the welding speed adjustment mapping table to obtain the target welding speed adjustment amount corresponding to the predicted time step value. The collaborative controller sends a first pulse sequence to the wire feeding motor of the back welding torch according to the target wire extension adjustment amount, and sends a second pulse sequence to the walking servo motor of the back welding torch according to the target welding speed adjustment amount.

6. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 2, characterized in that, Before performing a fast Fourier transform, the acoustic acquisition array introduces an adaptive noise cancellation process. The voltage and current sampling circuit extracts the fundamental current signal of the back welding torch as a reference noise signal. The acoustic acquisition array uses the filtered sound pressure signal as the main input signal, calculates the cross-correlation matrix between the main input signal and the reference noise signal, constructs an adaptive filter coefficient matrix based on the cross-correlation matrix, uses the adaptive filter coefficient matrix to perform a weighted summation of the reference noise signal to generate estimated noise, and subtracts the estimated noise from the main input signal to obtain the denoised sound pressure signal.

7. The collaborative control system based on double-sided double-arc root-cleaning welding according to claim 3, characterized in that, The voltage and current sampling circuit performs dynamic correction of the arc voltage before calculating the integral value of the arcing period energy. A displacement sensor is provided at the end of the conductive tip of the back welding torch. The displacement sensor measures the actual extension length of the back welding torch. The voltage and current sampling circuit reads the preset resistivity and cross-sectional area of ​​the welding wire. The actual extension length, the resistivity of the welding wire, and the cross-sectional area are substituted into Ohm's law to calculate the extension voltage drop. The collected instantaneous voltage value is added to the extension voltage drop to generate a corrected instantaneous voltage value. The corrected instantaneous voltage value is used to replace the instantaneous voltage value for integral calculation.

8. The collaborative control system based on double-sided double-arc root-cleaning-free welding according to claim 5, characterized in that, The co-controller is equipped with a wire extension limit threshold. When the target wire extension adjustment causes the current wire extension value of the back welding torch to reach the wire extension limit threshold, the co-controller stops sending the first pulse sequence to the wire feeding motor and calculates the lateral offset compensation amount based on the difference between the target wire extension adjustment amount and the wire extension limit threshold. The co-controller converts the lateral offset compensation amount into a position offset signal and sends the position offset signal to the cross slide servo driver connected to the front welding torch.

Images