A corrugated plate self-adaptive welding control method based on weld quality feedback

Abstract

The application discloses a corrugated plate adaptive welding control method based on weld quality feedback, which comprises the following steps: collecting weld quality related parameters and preprocessing; analyzing the weld forming state and determining whether the current weld forming state is stable; while analyzing the weld forming state, the current welding position is identified synchronously; when the weld forming state is unstable, the welding process parameters are adaptively adjusted based on the mapping relationship and the differentiated adjustment strategy combined with the weld forming state analysis result and the welding position state; the welding operation is continuously performed according to the adjusted welding process parameters, and the updated welding state is fed back to form a closed-loop control process. The application can realize real-time sensing of the weld forming state and adaptive adjustment during the corrugated plate welding process, effectively suppresses the weld forming fluctuation, improves the weld consistency and the welding quality stability, reduces the manual parameter adjustment, improves the automatic welding reliability, and is suitable for the welding of corrugated plates of different specifications.

Description

Technical Field

[0001] This invention belongs to the field of welding automation and relates to the welding control technology of corrugated plates, specifically to an adaptive welding control method for corrugated plates based on weld quality feedback. Background Technology

[0002] Corrugated steel sheets are widely used in liquefied natural gas (LNG) storage tanks, pressure vessels, and other fields due to their high structural strength and light weight. The surface of the corrugated steel sheet has a periodic undulating structure of crests and troughs. During the welding process, the relative position of the welding torch and the workpiece, as well as the heat input distribution in the welding area, will constantly change with the changes in crests and troughs, which will easily lead to fluctuations in the state of the weld pool and the quality of the weld formation.

[0003] Existing automated welding equipment typically employs fixed welding process parameters or relies solely on weld seam tracking for torch posture adjustment, primarily focusing on weld seam position tracking and control. However, it neglects real-time feedback on weld formation quality and closed-loop adjustment of process parameters. Particularly in corrugated plate welding, fluctuations in welding heat input due to changes in crest and trough positions often lead to instability in welding quality indicators such as weld width, weld height, and molten pool stability, thus affecting weld consistency and reliability. Furthermore, existing technologies lack differentiated adjustment strategies for welding process parameters at different corrugation positions, making it difficult to adapt to different corrugation structures and welding conditions. Existing welding systems based on single-vision tracking are highly susceptible to blinding interference from strong arc light at corrugated plate crests / troughs; and existing PID control cannot anticipate periodic and drastic changes in spatial heat dissipation conditions, resulting in significant adjustment lag. Summary of the Invention

[0004] Purpose of the invention: In order to solve the problems of large fluctuations in weld formation quality, difficulty in adapting welding process parameters to changes in the position of peaks and troughs, and poor weld consistency in the existing technology of corrugated plate welding, an adaptive welding control method for corrugated plates based on weld quality feedback is provided, thereby improving the automation level and welding quality stability of corrugated plate welding.

[0005] Technical Solution: To achieve the above objectives, this invention provides an adaptive welding control method for corrugated plates based on weld quality feedback, comprising the following steps:

[0006] S1: Collect relevant parameters of weld quality and perform preprocessing;

[0007] S2: Based on the preprocessed data, analyze the weld formation state and determine whether the current weld formation state is stable;

[0008] S3: While analyzing the weld formation state, simultaneously identify whether the current welding position is in the crest or trough area of ​​the corrugated plate.

[0009] S4: When the weld formation state is unstable, the welding process parameters are adaptively adjusted based on the mapping relationship and the identified welding position state, combined with the weld formation state analysis results and the identified welding position state.

[0010] S5: Continue the welding operation according to the adjusted welding process parameters, and at the same time feed back the updated welding status to step S1 to form a continuous closed-loop control process.

[0011] Furthermore, the preprocessing in step S1 includes:

[0012] Transient outlier removal based on Hampel filtering (for splashes and arcs):

[0013] In visual inspection and infrared thermometry, splashing droplets or instantaneous arc bursts can affect the width of the acquired image. or temperature Unrealistic pulse spikes appear; a Hampel filter is used to identify and remove these outliers in real time. Taking a temperature sequence as an example, let the length be... The data sequence within the sliding time window is :

[0014] Calculate the absolute median difference (MAD) of the data within the window:

[0015]

[0016] Outlier detection and replacement: When At that time, determine the current temperature For abnormal artifacts caused by splashing or arcing, the system automatically replaces them with the mid-range values ​​in the window; otherwise, the original values ​​are retained.

[0017] First-order inertial low-pass filter (for high-frequency electromagnetic noise):

[0018] Welding current collected by the arc condition monitoring module and voltage The system contains a large amount of high-frequency ripple generated by switching power supplies, which is then discretized using a first-order inertial low-pass filter.

[0019]

[0020] in, The original sampling current, This is the smoothed current after filtering. This is the filtering result from the previous time step; The filter coefficients (values) ), The value of is related to the system sampling period and cutoff frequency;

[0021] Timestamp alignment of multimodal data (key feature):

[0022] Because the arc sensor has an extremely high sampling rate, while the visual and infrared cameras have lower sampling rates, the control system must align the multi-source data on the time axis to prevent misalignment during EKF matrix calculation.

[0023] Using the low sampling rate of the visual sensor as the base timestamp For high-frequency electrical parameter sequences in Integrate over the interval to find the mean:

[0024]

[0025] This generates observation vectors that are absolutely aligned at the same time. ;

[0026] Dimensionless normalization processing:

[0027] To integrate physical quantities of different dimensions (Celsius, millimeters, amperes) into a single decision model, the relative deviation rate of the smoothed and aligned parameters is calculated:

[0028]

[0029] in, Represents the current parameter. This represents the steady-state theoretical baseline value calibrated for the current corrugated plate specifications. The processed data is directly input into the multi-parameter fusion phase observation model.

[0030] Furthermore, the method for analyzing the weld formation state in step S2 includes:

[0031] A1: At any sampling time For the preprocessed real-time acquisition parameters in the above four dimensions, their phase relative to the current space is calculated respectively. Standardized deviation rate of the ideal benchmark value:

[0032] Temperature deviation rate :

[0033]

[0034] in, This is the actual temperature measured by infrared. The ideal temperature for the current phase;

[0035] Morphological distortion rate :

[0036]

[0037] in, The effective area of ​​the molten pool extracted by thermal imaging;

[0038] Arc variation coefficient :

[0039]

[0040] in, To set the standard deviation of the arc voltage within the time window, This is an average value, used to characterize the intensity of fluctuations in electric arc combustion;

[0041] Geometric dimensional deviation rate :

[0042]

[0043] in, Real-time weld width for visual inspection;

[0044] A2: Calculate the comprehensive instability index at the current moment using a weighted fusion algorithm. :

[0045]

[0046] in, The dynamic weight coefficients for the four dimensions satisfy the following conditions: ;

[0047] A3: Decision Logic

[0048] Preset a tolerance threshold for stable operation :

[0049] when When the current weld formation state is determined to be stable, the control system maintains the current reference parameters and does not trigger adjustments;

[0050] when When the current weld formation state is determined to be unstable, the control system immediately triggers the subsequent adaptive adjustment strategy, and analyzes which sub-deviation term... The direction of parameter adjustment is determined by the factor that contributes the most (e.g., if...). If there is a sudden increase, the arc voltage should be adjusted first; if it is... If there is a sudden increase, the current and speed will be adjusted first.

[0051] Furthermore, in step S3, the current welding position is identified using the constructed nonlinear spatial phase dynamic observer based on extended Kalman filtering, specifically including:

[0052] B1: Establish a nonlinear state-space model

[0053] Define the hidden state vector of the system as ,in, For the actual spatial phase factor, It is the spatial phase change rate (geometrically constrained by welding speed and corrugation pitch). This represents the cumulative transient enthalpy in the molten pool region.

[0054] Establish the nonlinear state transition equation considering the thermo-coupling delay:

[0055]

[0056] Specifically, it can be elaborated as follows:

[0057]

[0058] in, The sampling period of the control system; This represents the current wire welding speed. The spatial pitch of the corrugated plate; This is for noise in the phase estimation process; This refers to process noise related to the phase change rate. The heat retention coefficient; For hot input function, hot input Controlled input The function; To control the input vector; It is a nonlinear heat dissipation function. As The independent variable; This refers to process noise in the thermodynamic model.

[0059] B2: Establishing Multi-Source Fusion Observation Equations

[0060] The highest temperature of the molten pool obtained by the weld quality feedback unit. molten pool width and welding current As observation vector Since these apparent parameters have a nonlinear relationship with internal enthalpy and spatial phase, the observation equations are established as follows:

[0061]

[0062] in, A nonlinear observation function characterizing the mapping relationship between "spatial phase, thermal conduction, and molten pool morphology" Gaussian noise for the sensor;

[0063] B3: EKF Dynamic Iteration and Phase Extraction

[0064] Within one control cycle, the nonlinear function is linearized using Taylor expansion, and the Jacobian matrix is ​​calculated. and ;

[0065] Prediction steps:

[0066]

[0067]

[0068] Update steps (calculate Kalman gain and correct phase):

[0069]

[0070]

[0071] B4: High Confidence Determination of Spatial Regions

[0072] After EKF filtering, the system outputs the optimal estimated spatial phase. and its covariance matrix, when the optimal estimated phase fall into When the Kalman estimation error covariance converges within the interval, a high-confidence assessment identifies it as a peak region; when the optimal estimated phase... fall into When the interval is defined and the covariance converges, the high confidence level indicates that it is a trough region.

[0073] Furthermore, in step S4, the mapping relationship is a mapping relationship between welding quality-related parameters and welding process parameters. The mapping relationship is constructed using a fuzzy adaptive PID control algorithm based on spatial phase. The construction method includes:

[0074] Obtain the deviation between the target weld formation parameters and the real-time acquired parameters. and the rate of change of deviation Based on whether the welding torch assembly is currently located in the crest or trough region of the corrugated plate, a spatial phase factor is introduced. ; Constructing based on deviation and rate of change of deviation A fuzzy controller that takes PID parameter adjustment as input and outputs the PID parameter adjustment value.

[0075] The fuzzy controller sets differentiated fuzzy inference rules for the peak and trough regions, and outputs PID parameter adjustment values ​​in real time to compensate for the reference welding process parameters.

[0076] Furthermore, the mapping relationship in step S4 is expressed as follows:

[0077] The system uses welding current The object to be adjusted; let the target molten pool temperature be... The infrared temperature sensor collects the current actual temperature in real time. ;

[0078] Define temperature deviation and the rate of change of temperature deviation for:

[0079]

[0080]

[0081] Introducing spatial phase factor Used to characterize the current welding position ( or Corresponding to the peak, (Corresponding to the trough); the mapping relationship of welding current is expressed as:

[0082]

[0083]

[0084] in, The reference current varies with the ripple space; , , These are the initial PID parameters; , , This is an adaptive adjustment amount.

[0085] Furthermore, the operation of the fuzzy controller in step S4 includes:

[0086] Temperature deviation and the rate of change of temperature deviation As input, with , , To achieve dynamic updating of parameters as output;

[0087] The fuzzification process involves mapping the input variables to a fuzzy universe of discourse and defining a fuzzy subset as... These represent negative large, negative medium, negative small, zero, positive small, positive medium, and positive large, respectively.

[0088] Combining the adaptive rule of spatial phase, when Located in the peak region (slow heat dissipation, prone to burn-through): If To prevent the thin plate from burning through due to instantaneous heat input overload (i.e., excessively low temperature), the fuzzy rules are configured with smaller values. and larger To suppress overshoot;

[0089] when Located in the trough area (fast heat dissipation, prone to incomplete soldering): If To ensure full melting, the fuzzy rules are configured to be relatively large. and smaller To expedite system response.

[0090] Defuzzy computation: Using the centroid method to output accurate results , , .

[0091] Furthermore, the adaptive adjustment method in step S4 includes:

[0092] C1: Introduce a time-series feature extraction algorithm to transform real-time abnormal parameter streams into structured abnormal feature vectors;

[0093] C2: Based on the extracted anomaly feature vectors, a Deep Deterministic Policy Gradient (DDPG) reinforcement learning network model is adopted to achieve optimal policy generation and adaptive evolution under anomalous conditions;

[0094] C3: Generate the final adjustment action based on the acquired current state vector (including abnormal features and spatial phase).

[0095] Furthermore, in step C1, the time window is... The process of extracting abnormal feature vectors from the data sequence includes:

[0096] C1-1: Drift trend characteristics (mean of first derivative):

[0097] Used to identify whether the heat in the molten pool is continuously accumulating (prone to burn-through) or continuously decaying (prone to incomplete penetration); taking temperature as an example, calculating the heat drift rate. :

[0098]

[0099] when Furthermore, exceeding the threshold indicates an abnormal trend in heat accumulation; conversely, falling below the threshold indicates an abnormal trend in heat decay.

[0100] C1-2: Oscillatory fluctuation characteristics (sliding variance):

[0101] Used to identify whether there is severe arc flash or molten pool tumbling during the welding process (usually caused by shielding gas failure or extreme parameter mismatch).

[0102] With melt width For example, calculate the variance of morphological oscillations. :

[0103]

[0104] in This represents the average melt width within the window;

[0105] C1-3: Spatial phase mismatch:

[0106] Extract the residual between the current actual thermodynamic state and the ideal spatial phase state predicted by the EKF observer:

[0107]

[0108] Ultimately, the system generates structured anomaly feature vectors. This serves as a high-dimensional input source for the subsequent generation of adjustment strategies.

[0109] Further, step C2 includes:

[0110] C2-1: Reinforcement Learning Environment and State Space

[0111] The extracted anomaly feature vector is concatenated with the current device state to define the system's state vector. :

[0112]

[0113] That is, the system has detected the current abnormal situation, its current position on the corrugated plate, and the parameters used in the previous second;

[0114] C2-2: Continuous Action Space

[0115] Define the Actor Action Network The output is a continuous adjustment of process parameters. :

[0116]

[0117] That is, the system determines "how much to adjust the current, how much to adjust the welding speed, and how much to adjust the welding posture angle in the next step";

[0118] C2-3: Reward Function Design

[0119] The DDPG network's ability to optimize based on historical data lies in the design of its reward function. The system calculates immediate rewards based on the weld quality feedback after an action is performed. The reward function comprehensively considers minimizing the forming error and minimizing the motion amplitude:

[0120]

[0121] If the temperature and melt width are close to the ideal values ​​after adjustment, the reward will be high; if the adjustment range is too large... If the excessive size causes system oscillation, a penalty will be imposed;

[0122] C2-4: Network Architecture and Dynamic Update Formula

[0123] The system comprises an Actor network (responsible for outputting the adjustment strategy) and a Critic network (responsible for evaluating the effectiveness of the strategy). Based on a data replay buffer generated from historical welding processes, the Critic network updates its weights by minimizing the temporal difference error (TD Error).

[0124]

[0125] Among them, the target value ;

[0126] The Actor network updates its weights by maximizing the Critic network's rating score, thus achieving a dynamic evolution of the adjustment strategy.

[0127] .

[0128] Further, step C3 includes:

[0129] C3-1: Basic Policy Forward Computation Based on Deep Networks

[0130] The Actor network employs a multi-layer fully connected feedforward neural network; let the state vector be... for The system calculates the basic adjustment coefficient by performing a step-by-step derivation using the following matrix calculation formula for the dimensional column vector:

[0131] First hidden layer calculation:

[0132]

[0133] in, This is the weight matrix. For bias vectors, It is a non-linear activation function;

[0134] Second hidden layer calculation:

[0135]

[0136] Output layer basic coefficient calculation:

[0137] The network output layer uses the hyperbolic tangent function Tanh to strictly map the output values ​​to... Within the interval, the basic motion coefficient vector is obtained. :

[0138]

[0139] C3-2: Physical Boundary Scaling Calculation

[0140] Because the output of a neural network is dimensionless. The coefficients must be calculated by scaling a diagonal matrix, taking into account the physical limits of the current welding equipment.

[0141]

[0142] in, Diagonal matrix for setting the maximum allowable adjustment range of the system:

[0143]

[0144] Calculated The physical adjustment amount suggested by the network;

[0145] C3-3: Expert Rule Action Mask Calculation

[0146] The system uses currently extracted abnormal features (such as heat drift rate) Calculate the mask function:

[0147]

[0148] in, This represents the Hadamard product, which is element-wise multiplication.

[0149] mask The specific conditional discrimination formula is as follows:

[0150] Burn-through prevention rule: When the heat drift rate When this indicates a rapid accumulation of heat and an impending burn-through, the system forcibly prohibits increasing the current and slows down the welding speed.

[0151]

[0152]

[0153] Rules for preventing incomplete penetration: When When (indicating heat loss and impending failure to fuse), the mask rules are inverted to forcibly prohibit reducing current and increasing welding speed;

[0154] Normal rule: When hour, It fully adopts the optimized output of the network.

[0155] Beneficial effects: Compared with existing technologies, this invention can acquire multi-dimensional weld quality feedback information in real time during the corrugated plate welding process. Combined with the weld formation state and the structural characteristics of the corrugated plate, it can precisely and adaptively adjust welding parameters, effectively suppressing weld formation fluctuations and improving the stability and consistency of welding quality. Simultaneously, this invention significantly reduces the need for manual intervention and minimizes manual adjustment errors. It is adaptable to welding scenarios for corrugated plates of different specifications and structural forms, significantly broadening the applicability of the device and method, and possessing outstanding practicality and application value. Attached Figure Description

[0156] Figure 1 This is a schematic diagram of the framework of the method of the present invention;

[0157] Figure 2 This is a schematic flowchart of the method of the present invention;

[0158] Figure 3 This is an overall structural diagram of an automatic welding head used for corrugated plate welding;

[0159] Figure 4 This is a diagram showing the working state of an automatic welding head used for corrugated plate welding. Detailed Implementation

[0160] The present invention will be further illustrated below with reference to the accompanying drawings and specific embodiments. It should be understood that these embodiments are for illustrative purposes only and are not intended to limit the scope of the invention. After reading this invention, any modifications of the invention in various equivalent forms by those skilled in the art will fall within the scope defined by the appended claims.

[0161] Example 1:

[0162] like Figure 1 and Figure 2 As shown, this embodiment provides an adaptive welding control method for corrugated plates based on weld quality feedback, including the following steps:

[0163] S1: Collect relevant parameters of weld quality and perform preprocessing;

[0164] Welding quality-related parameters include one or more of the following: weld pool temperature, weld pool morphology, weld bead width, weld bead height, and welding arc stability.

[0165] Preprocessing includes:

[0166] Transient outlier removal based on Hampel filtering (for splashes and arcs):

[0167] In visual inspection and infrared thermometry, splashing droplets or instantaneous arc bursts can affect the width of the acquired image. or temperature Unrealistic pulse spikes appear; a Hampel filter is used to identify and remove these outliers in real time. Taking a temperature sequence as an example, let the length be... The data sequence within the sliding time window is :

[0168] Calculate the absolute median difference (MAD) of the data within the window:

[0169]

[0170] Outlier detection and replacement: When At that time, determine the current temperature For abnormal artifacts caused by splashing or arcing, the system automatically replaces them with the mid-range values ​​in the window; otherwise, the original values ​​are retained.

[0171] First-order inertial low-pass filter (for high-frequency electromagnetic noise):

[0172] Welding current collected by the arc condition monitoring module and voltage The system contains a large amount of high-frequency ripple generated by switching power supplies, which is then discretized using a first-order inertial low-pass filter.

[0173]

[0174] in, The original sampling current, This is the smoothed current after filtering. This is the filtering result from the previous time step; The filter coefficients (values) ), The value of is related to the system sampling period and cutoff frequency;

[0175] Timestamp alignment of multimodal data (key feature):

[0176] Because the arc sensor has an extremely high sampling rate, while the visual and infrared cameras have lower sampling rates, the control system must align the multi-source data on the time axis to prevent misalignment during EKF matrix calculation.

[0177] Using the low sampling rate of the visual sensor as the base timestamp For high-frequency electrical parameter sequences in Integrate over the interval to find the mean:

[0178]

[0179] This generates observation vectors that are absolutely aligned at the same time. ;

[0180] Dimensionless normalization processing:

[0181] To integrate physical quantities of different dimensions (Celsius, millimeters, amperes) into a single decision model, the relative deviation rate of the smoothed and aligned parameters is calculated:

[0182]

[0183] in, Represents the current parameter. This represents the steady-state theoretical baseline value calibrated for the current corrugated plate specifications. The processed data is directly input into the multi-parameter fusion phase observation model.

[0184] S2: Based on the preprocessed data, analyze the weld formation state and determine whether the current weld formation state is stable;

[0185] In this embodiment, the preprocessed data includes infrared temperature measurement parameters, thermal imaging parameters, arc state parameters, and visual detection parameters;

[0186] Infrared thermometry parameters: monitor the absolute temperature at the center of the molten pool; thermal imaging parameters: extract the temperature gradient at the edge of the molten pool and the morphology of the liquid metal (such as the length and symmetry of the molten pool tail); arc state parameters: monitor the high-frequency arc voltage and current fluctuation rate; visual inspection parameters: obtain the actual weld width or weld reinforcement.

[0187] Methods for analyzing weld formation include:

[0188] A1: At any sampling time For the preprocessed real-time acquisition parameters in the above four dimensions, their phase relative to the current space is calculated respectively. Standardized deviation rate of the ideal benchmark value:

[0189] Temperature deviation rate :

[0190]

[0191] in, This is the actual temperature measured by infrared. The ideal temperature for the current phase;

[0192] Morphological distortion rate :

[0193]

[0194] in, The effective area of ​​the molten pool extracted by thermal imaging;

[0195] Arc variation coefficient :

[0196]

[0197] in, To set the standard deviation of the arc voltage within the time window, This is an average value, used to characterize the intensity of fluctuations in electric arc combustion;

[0198] Geometric dimensional deviation rate :

[0199]

[0200] in, Real-time weld width for visual inspection;

[0201] A2: Calculate the comprehensive instability index at the current moment using a weighted fusion algorithm. :

[0202]

[0203] in, The dynamic weight coefficients for the four dimensions satisfy the following conditions: ;

[0204] A3: Decision Logic

[0205] Preset a tolerance threshold for stable operation :

[0206] when When the current weld formation state is determined to be stable, the control system maintains the current reference parameters and does not trigger adjustments;

[0207] when When the current weld formation state is determined to be unstable, the control system immediately triggers the subsequent adaptive adjustment strategy, and analyzes which sub-deviation term... The direction of parameter adjustment is determined by the factor that contributes the most (e.g., if...). If there is a sudden increase, the arc voltage should be adjusted first; if it is... If there is a sudden increase, the current and speed will be adjusted first.

[0208] S3: While analyzing the weld formation state, simultaneously identify whether the current welding position is in the crest or trough area of ​​the corrugated plate.

[0209] This embodiment identifies the current welding position using a pre-constructed nonlinear spatial phase dynamic observer based on extended Kalman filtering, specifically including:

[0210] B1: Establish a nonlinear state-space model

[0211] Define the hidden state vector of the system as ,in, For the actual spatial phase factor, It is the spatial phase change rate (geometrically constrained by welding speed and corrugation pitch). This represents the cumulative transient enthalpy in the molten pool region.

[0212] Establish the nonlinear state transition equation considering the thermo-coupling delay:

[0213]

[0214] Specifically, it can be elaborated as follows:

[0215]

[0216] in, The sampling period of the control system; This represents the current wire welding speed. The spatial pitch of the corrugated plate; This is for noise in the phase estimation process; This refers to process noise related to the phase change rate. The heat retention coefficient; For hot input function, hot input Controlled input The function; To control the input vector; It is a nonlinear heat dissipation function. As The independent variable; This refers to process noise in the thermodynamic model.

[0217] B2: Establishing Multi-Source Fusion Observation Equations

[0218] The highest temperature of the molten pool obtained by the weld quality feedback unit. molten pool width and welding current As observation vector Since these apparent parameters have a nonlinear relationship with internal enthalpy and spatial phase, the observation equations are established as follows:

[0219]

[0220] in, A nonlinear observation function characterizing the mapping relationship between "spatial phase, thermal conduction, and molten pool morphology" Gaussian noise for the sensor;

[0221] B3: EKF Dynamic Iteration and Phase Extraction

[0222] Within one control cycle, the nonlinear function is linearized using Taylor expansion, and the Jacobian matrix is ​​calculated. and ;

[0223] Prediction steps:

[0224]

[0225]

[0226] Update steps (calculate Kalman gain and correct phase):

[0227]

[0228]

[0229] B4: High Confidence Determination of Spatial Regions

[0230] After EKF filtering, the system outputs the optimal estimated spatial phase. and its covariance matrix, when the optimal estimated phase fall into When the Kalman estimation error covariance converges within the interval, a high-confidence assessment identifies it as a peak region; when the optimal estimated phase... fall into When the interval is defined and the covariance converges, the high confidence level indicates that it is a trough region.

[0231] S4: When the weld formation state is unstable, the welding process parameters are adaptively adjusted based on the mapping relationship and the identified welding position state, combined with the weld formation state analysis results and the identified welding position state.

[0232] The mapping relationship is dynamically updated based on the welding quality data corresponding to different ripple positions in the historical welding process, so as to achieve adaptive optimization for different peak and trough welding conditions;

[0233] The mapping relationship is the mapping relationship between welding quality-related parameters and welding process parameters. A fuzzy adaptive PID control algorithm based on spatial phase is used to construct this mapping relationship. The construction method includes:

[0234] Obtain the deviation between the target weld formation parameters and the real-time acquired parameters. and the rate of change of deviation Based on whether the welding torch assembly is currently located in the crest or trough region of the corrugated plate, a spatial phase factor is introduced. ; Constructing based on deviation and rate of change of deviation A fuzzy controller that takes PID parameter adjustment as input and outputs the PID parameter adjustment value.

[0235] The fuzzy controller sets differentiated fuzzy inference rules for the peak and trough regions, and outputs PID parameter adjustment values ​​in real time to compensate for the reference welding process parameters.

[0236] The mapping relationship is expressed as follows:

[0237] The system uses welding current The object to be adjusted; let the target molten pool temperature be... The infrared temperature sensor collects the current actual temperature in real time. ;

[0238] Define temperature deviation and the rate of change of temperature deviation for:

[0239]

[0240]

[0241] Introducing spatial phase factor Used to characterize the current welding position ( or Corresponding to the peak, (Corresponding to the trough); the mapping relationship of welding current is expressed as:

[0242]

[0243]

[0244] in, The reference current varies with the ripple space; , , These are the initial PID parameters; , , This is an adaptive adjustment amount.

[0245] The operation of the fuzzy controller includes:

[0246] Temperature deviation and the rate of change of temperature deviation As input, with , , To achieve dynamic updating of parameters as output;

[0247] The fuzzification process involves mapping the input variables to a fuzzy universe of discourse and defining a fuzzy subset as... These represent negative large, negative medium, negative small, zero, positive small, positive medium, and positive large, respectively.

[0248] Combining the adaptive rule of spatial phase, when Located in the peak region (slow heat dissipation, prone to burn-through): If To prevent the thin plate from burning through due to instantaneous heat input overload (i.e., excessively low temperature), the fuzzy rules are configured with smaller values. and larger To suppress overshoot;

[0249] when Located in the trough area (fast heat dissipation, prone to incomplete soldering): If To ensure full melting, the fuzzy rules are configured to be relatively large. and smaller To expedite system response.

[0250] Defuzzy computation: Using the centroid method to output accurate results , , .

[0251] Adaptive adjustment methods include:

[0252] C1: Introduce a time-series feature extraction algorithm to transform real-time abnormal parameter streams into structured abnormal feature vectors;

[0253] C2: Based on the extracted anomaly feature vectors, a Deep Deterministic Policy Gradient (DDPG) reinforcement learning network model is adopted to achieve optimal policy generation and adaptive evolution under anomalous conditions;

[0254] C3: Generate the final adjustment action based on the acquired current state vector (including abnormal features and spatial phase).

[0255] Step C1 involves the time window. The process of extracting abnormal feature vectors from the data sequence includes:

[0256] C1-1: Drift trend characteristics (mean of first derivative):

[0257] Used to identify whether the heat in the molten pool is continuously accumulating (prone to burn-through) or continuously decaying (prone to incomplete penetration); taking temperature as an example, calculating the heat drift rate. :

[0258]

[0259] when Furthermore, exceeding the threshold indicates an abnormal trend in heat accumulation; conversely, falling below the threshold indicates an abnormal trend in heat decay.

[0260] C1-2: Oscillatory fluctuation characteristics (sliding variance):

[0261] Used to identify whether there is severe arc flash or molten pool tumbling during the welding process (usually caused by shielding gas failure or extreme parameter mismatch).

[0262] With melt width For example, calculate the variance of morphological oscillations. :

[0263]

[0264] in This represents the average melt width within the window;

[0265] C1-3: Spatial phase mismatch:

[0266] Extract the residual between the current actual thermodynamic state and the ideal spatial phase state predicted by the EKF observer:

[0267]

[0268] Ultimately, the system generates structured anomaly feature vectors. This serves as a high-dimensional input source for the subsequent generation of adjustment strategies.

[0269] Step C2 includes:

[0270] C2-1: Reinforcement Learning Environment and State Space

[0271] The extracted anomaly feature vector is concatenated with the current device state to define the system's state vector. :

[0272]

[0273] That is, the system has detected the current abnormal situation, its current position on the corrugated plate, and the parameters used in the previous second;

[0274] C2-2: Continuous Action Space

[0275] Define the Actor Action Network The output is a continuous adjustment of process parameters. :

[0276]

[0277] That is, the system determines "how much to adjust the current, how much to adjust the welding speed, and how much to adjust the welding posture angle in the next step";

[0278] C2-3: Reward Function Design

[0279] The DDPG network's ability to optimize based on historical data lies in the design of its reward function. The system calculates immediate rewards based on the weld quality feedback after an action is performed. The reward function comprehensively considers minimizing the forming error and minimizing the motion amplitude:

[0280]

[0281] If the temperature and melt width are close to the ideal values ​​after adjustment, the reward will be high; if the adjustment range is too large... If the excessive size causes system oscillation, a penalty will be imposed;

[0282] C2-4: Network Architecture and Dynamic Update Formula

[0283] The system comprises an Actor network (responsible for outputting the adjustment strategy) and a Critic network (responsible for evaluating the effectiveness of the strategy). Based on a data replay buffer generated from historical welding processes, the Critic network updates its weights by minimizing the temporal difference error (TD Error).

[0284]

[0285] Among them, the target value ;

[0286] The Actor network updates its weights by maximizing the Critic network's rating score, thus achieving a dynamic evolution of the adjustment strategy.

[0287] .

[0288] Step C3 includes:

[0289] C3-1: Basic Policy Forward Computation Based on Deep Networks

[0290] The Actor network employs a multi-layer fully connected feedforward neural network; let the state vector be... for The system calculates the basic adjustment coefficient by performing a step-by-step derivation using the following matrix calculation formula for the dimensional column vector:

[0291] First hidden layer calculation:

[0292]

[0293] in, This is the weight matrix. For bias vectors, It is a non-linear activation function;

[0294] Second hidden layer calculation:

[0295]

[0296] Output layer basic coefficient calculation:

[0297] The network output layer uses the hyperbolic tangent function Tanh to strictly map the output values ​​to... Within the interval, the basic motion coefficient vector is obtained. :

[0298]

[0299] C3-2: Physical Boundary Scaling Calculation

[0300] Because the output of a neural network is dimensionless. The coefficients must be calculated by scaling a diagonal matrix, taking into account the physical limits of the current welding equipment.

[0301]

[0302] in, Diagonal matrix for setting the maximum allowable adjustment range of the system:

[0303]

[0304] Calculated The physical adjustment amount suggested by the network;

[0305] C3-3: Expert Rule Action Mask Calculation

[0306] The system uses currently extracted abnormal features (such as heat drift rate) Calculate the mask function:

[0307]

[0308] in, This represents the Hadamard product, which is element-wise multiplication.

[0309] mask The specific conditional discrimination formula is as follows:

[0310] Burn-through prevention rule: When the heat drift rate When this indicates a rapid accumulation of heat and an impending burn-through, the system forcibly prohibits increasing the current and slows down the welding speed.

[0311]

[0312]

[0313] Rules for preventing incomplete penetration: When When (indicating heat loss and impending failure to fuse), the mask rules are inverted to forcibly prohibit reducing current and increasing welding speed;

[0314] Normal rule: When hour, It fully adopts the optimized output of the network.

[0315] S5: Continue the welding operation according to the adjusted welding process parameters, and at the same time feed back the updated welding status to step S1 to form a continuous closed-loop control process.

[0316] Welding process parameters include one or more of the following: welding current, welding voltage, welding speed, wire feed speed, welding torch oscillation parameters, and welding torch posture parameters.

[0317] Example 2:

[0318] Based on the control method of Embodiment 1, this embodiment provides an automatic welding head for corrugated plate welding, such as... Figure 3 As shown, the automatic welding head mainly includes a welding head body 1, a welding head 2, a welding torch assembly 3, a weld quality feedback unit 4, and a control system 5. The components work together to achieve automated welding and quality control of the corrugated plate.

[0319] The welding torch assembly 3 is mounted on the welding head body 1 and is used to perform welding operations on the corrugated plate. The welding torch assembly 3 includes components such as the welding torch 31 and the welding wire guide structure 32. Through the cooperation of these components, the welding wire is stably fed and the arc welding operation is achieved, providing a basic guarantee for the welding operation.

[0320] The weld quality feedback unit 4 is positioned around or near the welding torch assembly 3, corresponding to the welding torch assembly 3. It is used to collect relevant parameters characterizing the weld formation state in real time during the welding process, providing data support for subsequent quality assessment and parameter adjustment. The weld quality feedback unit 4 employs a multi-sensor module combination, including an infrared thermometer for detecting the temperature of the welding area, a thermal imaging sensor for acquiring weld heat distribution information, an arc state monitoring module for monitoring the stability of the arc combustion, and a weld visual inspection module for acquiring weld appearance morphology information.

[0321] The control system 5, as the core control component, includes a quality analysis module 51 and a parameter adjustment module 52. It establishes electrical or signal communication connections with the weld quality feedback unit 4 and the welding torch assembly 3, respectively. It can receive multi-dimensional welding quality parameters collected by the weld quality feedback unit 4 and implement real-time closed-loop control of the welding process of the welding torch assembly 3 based on these parameters, thereby realizing dynamic control of welding quality.

[0322] like Figure 4As shown, the sensor modules of the weld quality feedback unit 4 are arranged around the axis of the welding torch 31, and the detection end of each sensor module is precisely pointed to the weld formation area 7. Regardless of whether the welding torch 31 travels to the crest 61 or trough 62 of the corrugated plate, this arrangement can effectively overcome the detection blind spot caused by the sudden change in spatial height, ensuring that the detection range fully covers the core welding area and guaranteeing the continuity and effectiveness of parameter acquisition.

[0323] During the welding process, infrared temperature sensor 41, thermal imaging sensor 42, arc condition monitoring module 43, and weld visual inspection module 44 simultaneously collect welding quality-related parameters from four different dimensions: temperature, heat distribution, arc characteristics, and appearance. Especially during the continuous crossing of wave crests 61 and troughs 62 by the welding torch, these multi-source sensors can keenly capture the thermodynamic abrupt changes caused by the periodic alteration of spatial heat dissipation conditions. These parameters collectively constitute the state characterization system of the weld formation zone, comprehensively reflecting the dynamic differences in temperature field distribution, molten pool stability, arc combustion state, and weld appearance quality at different wave positions.

[0324] After receiving the aforementioned multi-dimensional parameters, the control system 5 performs a comprehensive analysis and evaluation of the weld formation state to determine whether the current weld formation is stable. Through the fusion processing of multi-source feedback information, compared to single-parameter detection, the accuracy, reliability, and real-time performance of welding state judgment are significantly improved, providing precise data for subsequent welding parameter adjustments.

[0325] Example 3:

[0326] To verify the effectiveness and efficacy of the present invention, the following experiments and data analyses were conducted in this embodiment:

[0327] 1. Experimental conditions and scheme setup

[0328] This experiment used a 1.2mm thick SUS304 stainless steel corrugated sheet as the welding object, with a corrugation pitch of 100mm and a crest height of 36mm. Automated continuous welding was performed using gas metal arc welding (GMAW).

[0329] The experiment included two control groups:

[0330] Control group: A conventional visual gap tracking system was used in conjunction with standard fixed-parameter PID control.

[0331] Experimental group: The integrated control method of this invention is used to start the EKF-based nonlinear spatial phase dynamic observer and activate the DDPG reinforcement learning parameter adjustment strategy with expert rule mask.

[0332] 2. Comparison of weld dimensional accuracy and consistency

[0333] Three-dimensional morphological scanning was performed on two sets of 2000mm continuous welds to extract the weld width and reinforcement height data.

[0334] Control group data: Due to the drastic changes in heat dissipation conditions at the peaks and troughs, the PID control exhibited significant lag. Tests showed that the weld width in the control group fluctuated by ±1.8mm, with significant widening at the peaks and narrowing at the troughs. The weld reinforcement height fluctuated by ±1.2mm, resulting in poor overall weld uniformity.

[0335] Experimental data: Thanks to the MIMO spatial modulation matrix constructed in this invention, the system can adjust parameters in advance based on the estimated spatial phase. Tests show that the weld width fluctuation range in the experimental group was strictly controlled within ±0.3mm, the weld reinforcement height fluctuation range was reduced to ±0.2mm, and the forming was extremely uniform and smooth. Compared with traditional methods, the dimensional accuracy was improved by approximately 83%.

[0336] 3. Comparison of welding defect rate and stability

[0337] X-ray flaw detection and macroscopic metallographic statistics were used to address the "wave crest burn-through" and "wave trough incomplete penetration" problems that are very common in thin corrugated plates.

[0338] Control group data: When climbing to the peak area, due to the sudden change in dry extension and heat accumulation, the probability of local burn-through defects is 6.5%; in the trough area, the probability of root incomplete penetration is 8.2%; the overall weld defect rate is as high as 14.7%.

[0339] Experimental group data: After introducing the "expert rule action mask", when the heat drift rate When approaching the threshold, the system forcibly cut off the parameter components that were causing the abnormality to worsen. Experimental results showed that the peak burn-through rate was 0%, the trough non-fusion rate dropped to 0.5% (occurring occasionally at the arc initiation end), and the overall defect rate dropped to 0.5%, achieving a qualitative leap in welding stability.

[0340] 4. Dynamic response and anti-arc interference capability analysis

[0341] A pulsed, strong airflow disturbance was artificially introduced during the welding process, and the coefficient of variation of the system's arc voltage was observed. And recovery time.

[0342] Control group data: Under intense arc flash and airflow disturbance, the visual sensor experienced momentary blindness and lacked effective feedback, causing the traditional system to lose the position target, the arc voltage to fluctuate violently, and the system required about 450ms to stabilize again.

[0343] Experimental data: This invention employs multi-source data fusion based on Hampel filtering and EKF dynamic iteration. Even if the vision module fails instantaneously, the system can still predict the current phase based on the "inertia" of the thermodynamic state transition equation. Tests show that, under the same disturbance, the arc variation coefficient... The peak value decreased by 70%, and the system's response time to return to steady state was only 40ms, demonstrating significantly better anti-interference capabilities than existing technologies.

[0344] in conclusion:

[0345] Experimental data fully demonstrate that the method provided in this invention, through multi-source data fusion, EKF phase prediction, and masked deep reinforcement learning adjustment, completely solves the problem of uneven heat input caused by abrupt changes in the spatial attitude of corrugated plates. It exhibits extremely significant improvements in three core industrial indicators: weld formation consistency, defect suppression rate, and anti-interference response speed.

Claims

1. An adaptive welding control method for corrugated plates based on weld quality feedback, characterized in that, Includes the following steps: S1: Collect relevant parameters of weld quality and perform preprocessing; S2: Based on the preprocessed data, analyze the weld formation state and determine whether the current weld formation state is stable; S3: While analyzing the weld formation state, simultaneously identify whether the current welding position is in the crest or trough area of ​​the corrugated plate. S4: When the weld formation state is unstable, the welding process parameters are adaptively adjusted based on the mapping relationship and the identified welding position state, combined with the weld formation state analysis results and the identified welding position state. S5: Continue the welding operation according to the adjusted welding process parameters, and at the same time feed back the updated welding status to step S1 to form a continuous closed-loop control process.

2. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 1, characterized in that, The preprocessing in step S1 includes: Transient outlier removal based on Hampel filtering: In visual inspection and infrared thermometry, splashing droplets or instantaneous arc bursts can affect the width of the acquired image. or temperature If non-real pulse spikes appear, a Hampel filter is used to identify and remove these outliers in real time. First-order inertial low-pass filter: Welding current collected by the arc condition monitoring module and voltage The system contains a large amount of high-frequency ripple generated by switching power supplies, which is then discretized using a first-order inertial low-pass filter. ； in, The original sampling current, This is the smoothed current after filtering. This is the filtering result from the previous time step; These are the filter coefficients; Timestamp alignment for multimodal data: Using the low sampling rate of the visual sensor as the base timestamp For high-frequency electrical parameter sequences in Integrate over the interval to find the mean: ； This generates observation vectors that are absolutely aligned at the same time. ; Dimensionless normalization processing: In order to integrate physical quantities with different dimensions into a single judgment model, the relative deviation rate of the smoothed and aligned parameters is calculated: ； in, Represents the current parameter. This represents the steady-state theoretical benchmark value calibrated for the current corrugated sheet specifications.

3. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 1, characterized in that, The method for analyzing the weld formation state in step S2 includes: A1: At any sampling time For the preprocessed real-time acquired parameters, their phase relative to the current space is calculated. Standardized deviation rate of the ideal benchmark value: Temperature deviation rate : ； in, This is the actual temperature measured by infrared. The ideal temperature for the current phase; Morphological distortion rate : ； in, The effective area of ​​the molten pool extracted by thermal imaging; Arc variation coefficient : ； in, To set the standard deviation of the arc voltage within the time window, This is an average value, used to characterize the intensity of fluctuations in electric arc combustion; Geometric dimensional deviation rate : ； in, Real-time weld width for visual inspection; A2: Calculate the comprehensive instability index at the current moment using a weighted fusion algorithm. : ； in, The dynamic weight coefficients for the four dimensions satisfy the following conditions: ; A3: Decision Logic Preset a tolerance threshold for stable operation : when When the current weld formation state is determined to be stable, the control system maintains the current reference parameters and does not trigger adjustments; when If the current weld formation state is determined to be unstable, the control system will immediately trigger the subsequent adaptive adjustment strategy.

4. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 3, characterized in that, In step S3, the current welding position is identified using the constructed nonlinear spatial phase dynamic observer based on extended Kalman filtering, specifically including: B1: Establish a nonlinear state-space model Define the hidden state vector of the system as ,in, For the actual spatial phase factor, The rate of change of spatial phase. This represents the cumulative transient enthalpy in the molten pool region. Establish the nonlinear state transition equation considering the thermo-coupling delay: ； Specifically, it can be elaborated as follows: ； in, The sampling period of the control system; This represents the current wire welding speed. The spatial pitch of the corrugated plate; This is for noise in the phase estimation process; This refers to process noise related to the phase change rate. The heat retention coefficient; For hot input function, hot input Controlled input The function; To control the input vector; It is a nonlinear heat dissipation function. As The independent variable; This refers to process noise in the thermodynamic model. B2: Establishing Multi-Source Fusion Observation Equations The highest temperature of the molten pool obtained by the weld quality feedback unit. molten pool width and welding current As observation vector Since these apparent parameters have a nonlinear relationship with internal enthalpy and spatial phase, the observation equations are established as follows: ； in, A nonlinear observation function characterizing the mapping relationship between "spatial phase, thermal conduction, and molten pool morphology" Gaussian noise for the sensor; B3: EKF Dynamic Iteration and Phase Extraction Within one control cycle, the nonlinear function is linearized using Taylor expansion, and the Jacobian matrix is ​​calculated. and ; Prediction steps: ； ； Update steps: ； ； B4: High Confidence Determination of Spatial Regions After EKF filtering, the system outputs the optimal estimated spatial phase. and its covariance matrix, when the optimal estimated phase fall into When the Kalman estimation error covariance converges within the interval, a high-confidence assessment identifies it as a peak region; when the optimal estimated phase... fall into When the interval is defined and the covariance converges, the high confidence level indicates that it is a trough region.

5. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 4, characterized in that, In step S4, the mapping relationship is the mapping relationship between welding quality-related parameters and welding process parameters. The mapping relationship is constructed using a fuzzy adaptive PID control algorithm based on spatial phase. The construction method includes: Obtain the deviation between the target weld formation parameters and the real-time acquired parameters. and the rate of change of deviation Based on whether the welding torch assembly is currently located in the crest or trough region of the corrugated plate, a spatial phase factor is introduced. ; Constructing based on deviation and rate of change of deviation A fuzzy controller that takes PID parameter adjustment as input and outputs the PID parameter adjustment value. The fuzzy controller sets differentiated fuzzy inference rules for the peak and trough regions, and outputs PID parameter adjustment values ​​in real time to compensate for the reference welding process parameters.

6. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 5, characterized in that, The mapping relationship in step S4 is expressed as follows: The system uses welding current The object to be adjusted; let the target molten pool temperature be... The infrared temperature sensor collects the current actual temperature in real time. ; Define temperature deviation and the rate of change of temperature deviation for: ； ； Introducing spatial phase factor This is used to characterize the current welding position; The mapping relationship of welding current is expressed as: ； ； in, The reference current varies with the ripple space; , , These are the initial PID parameters; , , This is an adaptive adjustment amount.

7. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 6, characterized in that, The adaptive adjustment method in step S4 includes: C1: Introduce a time-series feature extraction algorithm to transform real-time abnormal parameter streams into structured abnormal feature vectors; C2: Based on the extracted abnormal feature vectors, a deep deterministic policy gradient reinforcement learning network model is adopted to achieve optimal policy generation and adaptive evolution under abnormal conditions; C3: Generate the final adjustment action based on the acquired current state vector.

8. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 7, characterized in that, In step C1, the time window is... The process of extracting abnormal feature vectors from the data sequence includes: C1-1: Drift trend characteristics: Calculate the heat drift rate : ； when Furthermore, exceeding the threshold indicates an abnormal trend in heat accumulation; conversely, falling below the threshold indicates an abnormal trend in heat decay. C1-2: Oscillating Fluctuation Characteristics: Calculate the variance of morphological oscillations : ； in This represents the average melt width within the window; C1-3: Spatial phase mismatch: Extract the residual between the current actual thermodynamic state and the ideal spatial phase state predicted by the EKF observer: ； Ultimately, the system generates structured anomaly feature vectors. .

9. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 8, characterized in that, Step C2 includes: C2-1: Reinforcement Learning Environment and State Space The extracted anomaly feature vector is concatenated with the current device state to define the system's state vector. : ； C2-2: Continuous Action Space Define the Actor Action Network The output is a continuous adjustment of process parameters. : ； C2-3: Reward Function Design The system calculates an immediate reward based on the weld quality feedback after the action is performed. : ； C2-4: Network Architecture and Dynamic Update Formula The system includes an Actor network and a Critic network. Based on a data playback buffer generated from historical welding processes, the Critic network updates its weights by minimizing the temporal difference error. ； Among them, the target value ; The Actor network updates its weights by maximizing the Critic network's rating score, thus achieving a dynamic evolution of the adjustment strategy. 。 10. The adaptive welding control method for corrugated plates based on weld quality feedback according to claim 9, characterized in that, Step C3 includes: C3-1: Basic Policy Forward Computation Based on Deep Networks The Actor network employs a multi-layer fully connected feedforward neural network; let the state vector be... for The system calculates the basic adjustment coefficient by performing a step-by-step derivation using the following matrix calculation formula for the dimensional column vector: First hidden layer calculation: ； in, This is the weight matrix. For bias vectors, It is a non-linear activation function; Second hidden layer calculation: ； Output layer basic coefficient calculation: The network output layer uses the hyperbolic tangent function Tanh to strictly map the output values ​​to... Within the interval, the basic motion coefficient vector is obtained. : ； C3-2: Physical Boundary Scaling Calculation Because the output of a neural network is dimensionless. The coefficients must be calculated by scaling a diagonal matrix, taking into account the physical limits of the current welding equipment. ； in, Diagonal matrix for setting the maximum allowable adjustment range of the system: ； Calculated The physical adjustment amount suggested by the network; C3-3: Expert Rule Action Mask Calculation The system calculates a mask function based on the currently extracted anomaly features: ； in, This represents the Hadamard product, which is element-wise multiplication. mask The specific conditional discrimination formula is as follows: Burn-through prevention rule: When the heat drift rate At this time, the system will forcibly prohibit increasing the current and slowing down the welding speed: ； ； Rules for preventing incomplete penetration: When At this time, the mask rules are inverted to forcibly prohibit reducing the current and speeding up the welding process; Normal rule: When hour, It fully adopts the optimized output of the network.

Images