An intelligent control method for welding process parameters of automobile parts

Abstract

The application discloses a kind of welding process parameter intelligent control method for automobile parts, belong to intelligent manufacturing technical field.The method includes: the multi-physical field collaborative perception system of four kinds of sensors of integrated vision, acoustics, electrical signal, thermal field is constructed;Double closed-loop adaptive control of inner ring molten pool dynamic control and outer ring weld quality prediction is implemented;Three-level active defense mechanism containing defect embryo identification, cause tracing, targeted compensation is executed;Through dynamic knowledge graph and reinforcement learning, realize knowledge accumulation and strategy evolution;Multi-modal information deep integration is realized by adopting feature layer and decision layer two-level fusion.The present application can identify abnormal signs and implement precise intervention during the incubation period before defect formation, and change passive after-repair into active prevention in advance.Double closed-loop collaborative control realizes the organic combination of millisecond-level fast response and quality stability.The system has self-learning ability and can continuously optimize control strategy.The method can be widely applied to intelligent welding production line of automobile manufacturing industry.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing technology, specifically relating to an intelligent control method for welding process parameters of automotive parts. Background Technology

[0002] In the automotive manufacturing industry, welding is a core process for connecting key components such as body structural parts and chassis components. Welding quality directly affects the safety, reliability, and service life of a vehicle. With the development of lightweight automotive trends and the widespread application of new materials such as high-strength steel and aluminum alloys, higher requirements are placed on the control of welding process parameters.

[0003] Existing welding parameter control technologies have the following main shortcomings: 1. Traditional welding parameter control often relies on preset parameters or simple feedback adjustment, lacking the ability to predict dynamic changes in the welding process. This results in lag in parameter adjustment and makes it difficult to effectively prevent welding defects.

[0004] 2. Most existing intelligent control methods are based on single sensor information or simple superposition of multi-sensor data, failing to deeply explore the coupling relationship between multi-physical field information, resulting in low information utilization and limited control accuracy.

[0005] 3. Current welding quality control mainly relies on post-production inspection and passive adjustment, lacking in-depth understanding of the defect formation mechanism and proactive defense mechanism, resulting in high scrap rate and high rework cost.

[0006] 4. Existing control systems lack adaptability and require extensive testing and manual debugging when facing new materials and new working conditions. The lack of knowledge accumulation and inheritance mechanisms makes it difficult to achieve true intelligence.

[0007] Therefore, there is an urgent need for an intelligent control method for welding process parameters that can achieve deep fusion of multi-source information, possess proactive defect defense capabilities, and support autonomous knowledge evolution. Summary of the Invention

[0008] The purpose of this invention is to overcome the above-mentioned defects of the prior art and provide an intelligent control method for welding process parameters of automotive parts. By constructing a multi-level collaborative control architecture, it can achieve precise control of the welding process and proactive defense against defects.

[0009] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent control of welding process parameters for automotive parts, characterized by comprising the following steps: S1: Construct a multi-physics collaborative sensing system. This system integrates visual sensors, acoustic sensors, electrical signal acquisition units, and infrared thermal imagers to collect information on the molten pool morphology, arc acoustic characteristics, welding electrical parameter waveforms, and temperature field distribution during the welding process. Furthermore, the collected multi-source heterogeneous data undergoes time synchronization and spatial registration to ensure consistency of data from different sensors across time and space. Time synchronization employs the IEEE 1588 precise time protocol, with a synchronization accuracy better than 10 microseconds to ensure a frame interval of less than 5%. Spatial registration utilizes the Zhang Zhengyou calibration method to unify the coordinate systems of each sensor.

[0010] S2: Implement dual-loop adaptive control. The dual-loop control includes an inner loop for dynamic control of the molten pool and an outer loop for predictive control of weld quality. The inner loop, based on the geometric and dynamic characteristics of the molten pool, employs an improved fuzzy control strategy combined with Kalman filter prediction compensation to achieve millisecond-level rapid parameter adjustment. The outer loop constructs a time-series prediction model based on a long short-term memory network, analyzes historical weld formation trends, predicts the weld quality state within future time windows, and generates feedforward control signals. The inner and outer loops interact and fuse information through an intelligent decision fusion module, dynamically adjusting control weights according to the welding stage.

[0011] S3: Implement proactive vulnerability defense controls. This step includes a three-tiered defense mechanism: The first level is defect embryo identification. Micro-perturbation signals in the welding process are extracted through empirical mode decomposition, a 128-dimensional perturbation feature spectrum is constructed, and a defect entropy index is defined to quantify the disorder of the system. When the defect entropy exceeds the adaptive threshold, defect embryos in the gestation period are identified.

[0012] The second level is defect cause tracing, which constructs a three-layer Bayesian network containing parameter nodes, process state nodes, and defect nodes. The posterior probability of each parameter anomaly is calculated through Bayesian inference, and combined with Granger causality test, the causal chain and propagation path leading to defect formation are identified.

[0013] The third level is targeted compensation intervention, which generates a targeted parameter compensation plan based on the source analysis results. This includes main compensation (60% intensity) acting on the main cause, auxiliary compensation (30% intensity) acting on related parameters, and protective compensation (10% intensity) to maintain system stability. The compensation process follows an S-shaped curve for a smooth transition.

[0014] S4: Knowledge Accumulation and Strategy Evolution. A dynamic knowledge graph based on an attribute graph model is constructed, with each case node containing six attribute domains: material properties, geometric properties, process properties, parameter sequences, quality assessment, and defect information. Case matching is achieved through multi-dimensional similarity calculation and kd-tree retrieval. Reinforcement learning is implemented using a proximal policy optimization algorithm. The state space contains 12 continuous variables, and the action space is a discretized set of parameter adjustments. The control strategy is continuously optimized by designing a comprehensive reward function.

[0015] S5: Achieve deep fusion of multimodal information. Feature layer fusion uses the minimum variance criterion to weightedly combine similar features, with weights dynamically adjusted through online variance estimation. Decision layer fusion employs Dempster-Shafer evidence theory, combining judgments from different modalities as evidence to calculate confidence and similarity. A cross-modal attention mechanism is introduced, determining the importance of each modality through mutual information calculation. Anomaly detection utilizes an integrated approach of isolated forest, local anomaly factors, and autoencoders to improve detection reliability.

[0016] Beneficial effects

[0017] 1. Significantly enhanced preventive control capabilities: Through defective embryo identification technology, abnormal signs can be detected during the gestation period before defects are formed, enabling early intervention and transforming passive post-event remediation into proactive pre-event prevention, effectively reducing the incidence of defects.

[0018] 2. Significantly improved control accuracy and response speed: The dual-loop collaborative control architecture achieves an organic combination of fast response and stable control. The inner loop ensures millisecond-level dynamic adjustment, while the outer loop provides forward-looking quality assurance. The overall control accuracy and stability are significantly better than traditional methods.

[0019] 3. Significantly enhanced information utilization efficiency: Multi-physics collaborative sensing and multi-modal deep fusion technology fully taps into the complementarity and correlation between different sensor information, improving the accuracy and reliability of welding status identification.

[0020] 4. Outstanding adaptive and self-learning capabilities: The knowledge accumulation and strategy evolution mechanism enables the system to continuously learn and optimize autonomously, and can continuously summarize experience from practice, improve control performance, and reduce dependence on human experience.

[0021] 5. Good versatility and scalability: The method framework has good versatility and can adapt to the control requirements of different materials, thicknesses and welding positions, making it easy to promote and apply in automobile manufacturing enterprises. Attached Figure Description

[0022] Figure 1 A flowchart illustrating the steps of the method described in this invention is shown. Detailed Implementation

[0023] Exemplary embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0024] Combination Figure 1 This invention provides an intelligent control method for welding process parameters of automotive parts, comprising the following steps: S1: Constructing a multi-physics collaborative sensing system This step aims to establish a comprehensive and multi-dimensional information acquisition system for the welding process, achieving a comprehensive perception of the welding physical field by integrating multiple sensors.

[0025] Specifically, the multi-physics collaborative sensing system integrates four types of sensing devices. The visual sensor uses a CMOS high-speed camera with a frame rate of 5000 frames per second, equipped with a narrow-band filter with a center wavelength of 650nm and a bandwidth of 10nm to filter out welding arc light interference. The camera's optical axis is mounted at a 45-degree angle to the welding torch axis, with the lens center 200mm vertically and 150mm horizontally from the welding point. This configuration clearly captures the molten pool morphology while avoiding spatter damage. The acoustic sensor uses an array of three omnidirectional condenser microphones with a frequency response range of 20Hz to 100kHz and a dynamic range of 120dB. The three microphones are arranged in an equilateral triangle with sides of 100mm, and the array center is located 300mm directly above the welding torch. Beamforming technology is used to achieve noise suppression and sound source localization.

[0026] The electrical signal acquisition unit is directly connected to the welding power supply output. A Hall effect current sensor and a differential voltage probe are used to measure the welding current and voltage, respectively. The current sensor has a range of 0-500A, an accuracy of 0.1%, and a response time of less than 1 microsecond; the voltage probe has a range of 0-100V and an input impedance greater than 10MΩ. The data acquisition card has a sampling frequency of 20kHz and a resolution of 16 bits. The infrared thermal imager uses an uncooled microbolometer detector with a working wavelength of 8-14μm, an upper limit of the temperature measurement range of 1500℃, a temperature resolution better than 0.05℃, and a spatial resolution of 640×480 pixels.

[0027] Furthermore, data synchronization and registration are crucial for ensuring the spatiotemporal consistency of multi-source information. Time synchronization employs the IEEE 1588 Precision Time Protocol, using an industrial switch supporting the PTP protocol as the master clock and each sensor as a slave clock. Through a delay request-response mechanism and clock skew compensation, timestamp synchronization accuracy is achieved better than 10 microseconds, ensuring it is less than 5% of the frame interval. Spatial registration uses a checkerboard calibration board with 9×7 corner points and a checkerboard size of 20mm×20mm. The intrinsic and extrinsic parameter matrices of each sensor are calculated using the Zhang Zhengyou calibration method to establish a unified world coordinate system.

[0028] Data preprocessing includes four aspects. First, the visual image is filtered using a 5×5 median filter to remove salt-and-pepper noise. Then, contrast-limited adaptive histogram equalization is applied for enhancement, with a block size of 8×8 and a contrast limit factor of 3.0. Melt pool contour extraction uses Canny edge detection with a low threshold of 50 and a high threshold of 150.

[0029] Secondly, acoustic signal preprocessing includes pre-emphasis, framing, and windowing. The pre-emphasis transfer function is: in: The z-domain transfer function of the pre-emphasis filter; The pre-weighting factor is 0.97. This is the unit delay operator, specifically the unit delay operator in the Z-transform, which represents a signal delay of one sampling period.

[0030] The time-domain implementation is as follows: in: The original signal; This is the signal after pre-emphasis.

[0031] The frame length is 25 milliseconds, corresponding to 500 sampling points; the frame shift is 10 milliseconds, corresponding to 200 sampling points, achieving 60% overlap. A Hamming window is used as the window function. in: This is the window function value for the nth sampling point; The frame length is N = 500 here.

[0032] Next, the electrical signal is processed by a fourth-order Butterworth low-pass filter with a cutoff frequency of 1kHz. The normalized transfer function is: in: These are normalized complex frequency domain variables; frequency transformation is required in practical applications. ,in rad / s.

[0033] Finally, the thermal image is processed using bilateral filtering to remove noise while preserving the edges of the temperature gradient: in: These are the filtered pixel values; These are the original pixel values; These are the normalization coefficients; It is an 11×11 neighborhood window.

[0034] Gaussian kernel function in the spatial domain: in: Spatial distance; Pixel.

[0035] Gaussian kernel function in grayscale region: in: This represents the grayscale difference. Grayscale levels.

[0036] Through the above preprocessing, the raw data collected by each sensor is converted into a standardized format suitable for subsequent analysis, laying a reliable foundation for feature extraction, state recognition, and control decisions.

[0037] S2: Implement dual closed-loop adaptive control This step achieves precise adjustment of the welding process through the coordinated control of the inner and outer rings. The inner ring is responsible for the dynamic and rapid response of the molten pool, while the outer ring is responsible for the predictive control of weld quality.

[0038] Specifically, the inner-loop control uses the dynamic features of the molten pool as input to achieve millisecond-level parameter adjustment. First, molten pool features are extracted from the preprocessed image. Geometric features include the molten pool area A, length L, width W, aspect ratio R, centroid coordinates (xc, yc), and trailing angle θ. The molten pool area is obtained through pixel counting. in: The area of ​​the molten pool is (mm²). This represents the number of pixels in the molten pool region. The pixel equivalent is obtained through calibration; in this embodiment, it is 0.05 mm / pixel.

[0039] The length and width are calculated using the minimum bounding rectangle algorithm. Aspect Ratio During normal welding, the radius (R) ranges from 1.5 to 3.0. The centroid coordinates are calculated using moments. in: This is the molten pool region; This is a binarized image.

[0040] Dynamic characteristics are obtained by analyzing consecutive frames. The molten pool oscillation frequency is obtained by performing an FFT transform on the area time series and taking the frequency corresponding to the maximum amplitude, which is within the normal range of 2-20Hz. The center movement speed is calculated by the centroid displacement of adjacent frames.

[0041] The inner-loop controller employs an improved fuzzy control strategy. The input variable is the molten pool area deviation. and rate of change The input universes of discourse are [-50, 50] mm² and [-100, 100] mm² / s, respectively, and are divided into seven fuzzy sets. The membership functions use a combination of triangular and trapezoidal shapes. in: For the triangle parameter; for the zero set, .

[0042] Fuzzy inference employs the Mamdani method, with the rule base constructed based on collected expert experience and experimental data. Typical rules include: when the molten pool area is too large and continues to increase, the output current adjustment is significantly reduced; when the molten pool area is too small and continues to decrease, the output current adjustment is significantly increased; when the molten pool area is close to the target and the rate of change is small, the current is maintained. Defuzzification uses the centroid method. in: Let i be the activation degree of the i-th output fuzzy set; The center value.

[0043] Furthermore, Kalman filtering is introduced to predict and compensate for the molten pool area, improving the foresight of the control. By establishing a dynamic state-space model of the molten pool, the area of ​​the molten pool at future moments is estimated, and the predicted value is added as a feedforward signal to the control loop.

[0044] The outer loop control employs a two-layer LSTM network to predict weld quality. The network input includes a sequence of welding parameters (100×4D) from the past 10 seconds, a sequence of molten pool features (100×4D), and static environmental factors (8D). The first LSTM layer contains 128 hidden units, and the second layer contains 64. The network output is the predicted weld width, weld height, and penetration depth for the next 5 seconds. Training uses a mean squared error loss function and the Adam optimizer with an initial learning rate of 0.001, which decays to 0.9 times the original rate every 1000 training epochs. The training data consists of 10,000 historical welding records.

[0045] Coordination between the inner and outer loops is achieved through an intelligent decision fusion module. Control weights are dynamically adjusted based on the welding stage. The welding process is divided into four stages: arc initiation (0-2 seconds), transition (2-5 seconds), stabilization (5 seconds to 3 seconds before arc termination), and arc termination (the last 3 seconds). During the arc initiation stage, the inner loop weight is set to 0.8, and the outer loop weight to 0.2, prioritizing the establishment of a stable molten pool. During the stabilization stage, both the inner and outer loop weights are 0.5, balancing rapid response and quality control. During the arc termination stage, the outer loop weight is increased to 0.7 to ensure the quality of the weld finishing.

[0046] Weight switching employs a smooth transition to avoid abrupt changes in control: in: The weight at time t; The weight of the previous stage; As the weight for the next stage; The transition rate parameter is set to 0.1 s. −1 That is, 90% of the weight transition is completed per second; This marks the point of phase transition.

[0047] The final control output is obtained through weighted fusion: in: This is the final control command; This is the current regulation amount of the inner loop output; This is the feedforward compensation amount for the outer loop output; and For the corresponding weights, satisfying ; This is the bias compensation term, used to eliminate steady-state error, and is adjusted online through integral control.

[0048] Through dual closed-loop collaboration, the inner loop ensures millisecond-level rapid response to dynamic changes in the molten pool, while the outer loop provides forward-looking control based on quality prediction, achieving an optimized balance between control speed and weld quality stability.

[0049] S3: Execution Defect Proactive Defense Control This step is the core innovation of this invention, achieving proactive prevention of welding defects through a three-tiered defense mechanism, including defect embryo identification, cause tracing, and targeted compensation. The defect embryo refers to early microscopic abnormalities before defect formation.

[0050] Level 1: Defective Embryo Identification Defective embryo identification is based on micro-perturbation signal analysis, detecting abnormal signs during the gestation period before defects form. Micro-perturbations are defined as tiny fluctuations superimposed on a steady-state signal, with an amplitude less than 1% of the steady-state value.

[0051] For welding current signals, empirical mode decomposition (EMD) is used to extract micro-perturbation components. EMD decomposes the original signal into intrinsic mode functions and residual terms: in: This is the original welding current signal, in amperes. Let i be the i-th intrinsic mode function; For the residual term; To determine the number of layers, this embodiment uses 8 layers.

[0052] The micro-perturbation signal is defined as the sum of the first three high-frequency intrinsic mode functions: in: It represents the current micro-perturbation signal; the first three intrinsic mode functions contain the main high-frequency perturbation information.

[0053] Micro-perturbations in acoustic signals are obtained through short-time energy fluctuation analysis. First, the short-time energy is calculated: in: Let m be the short-time energy of the m-th frame; These are the sampled values ​​of the sound signal; For Hamming window functions; Given the frame length, 500 sampling points are used. For frame shift, 200 sampling points are taken; The frame number; This is the sequence number of the sampling point within the frame.

[0054] The standard deviation of energy movement serves as an indicator of acoustic micro-perturbations: in: The energy standard deviation of the m-th frame; The moving window length is set to 21 frames. The average energy within the window; For the frame index within the window.

[0055] The perturbation of the molten pool area is defined as the difference between the observed value and the filtered estimate: in: For the area perturbation at time k; This represents the actual observed area of ​​the molten pool; This is the steady-state estimate obtained through Kalman filtering.

[0056] Based on the extracted micro-perturbation signals, a 128-dimensional feature vector is constructed. The feature vector contains three types of features: 40-dimensional time-domain features, 48-dimensional frequency-domain features, and 40-dimensional time-frequency-domain features. Time-domain features include mean, variance, skewness, kurtosis, maximum value, minimum value, zero-crossing rate, and autocorrelation coefficient. Eight features are extracted for each type of micro-perturbation signal, totaling 24 features across the three signals, plus 16 combined features. Frequency-domain features include spectral centroid, dominant frequency, spectral width, and number of spectral peaks, obtained through Fast Fourier Transform. The time-frequency-domain features are extracted through three-level wavelet packet decomposition, yielding the energy distribution across eight frequency bands.

[0057] The 40-dimensional time-domain features include: 8 statistical features of current micro-perturbations: mean, variance, skewness, kurtosis, maximum / minimum value, zero-crossing rate, and autocorrelation coefficient; 8 features of acoustic micro-perturbations; 8 features of molten pool area micro-perturbations; and 16 cross-signal combination features, including current-acoustic cross-correlation and current-area covariance.

[0058] K-means clustering was used to classify historical perturbation patterns into 8 categories. The Euclidean distance between the current feature and each cluster center was calculated: in: It is the distance to the i-th cluster center; Let j be the j-th component of the current feature vector; is the j-th dimension coordinate of the i-th cluster center; 128 is the feature dimension.

[0059] The distance is converted into probability using the softmax function: in: Let be the probability of belonging to the i-th class; For the temperature parameter, take 0.5; The clustering index ranges from 1 to 8. Specifically, T is the probability distribution softening coefficient, which is dimensionless and set to 0.5. It is used to adjust the sharpness of the probability distribution; the smaller T is, the more precise the classification.

[0060] Defining defect entropy quantifies the degree of disorder in a system: in: Let be the defect entropy at time t; when When, define .

[0061] The dynamic threshold is adjusted according to welding conditions: in: The threshold at time t; The baseline threshold is set at 2.5. The material adjustment factor is set to 0.3. The thickness adjustment factor is set to 0.2. The position adjustment factor is set to 0.15. For material factors, carbon steel is 1.0, stainless steel is 1.2, and aluminum alloy is 1.5. The thickness factor is calculated using the following formula: Where t is the plate thickness in millimeters. Use dimensionless values: for example, take 1.1 when the plate thickness is 3mm, and take 1.3 when the plate thickness is 5mm; The welding position factor is 1.0 for flat welding, 1.1 for horizontal welding, 1.2 for vertical welding, and 1.3 for overhead welding.

[0062] When the defect entropy exceeds the dynamic threshold, a defective embryo is determined to exist.

[0063] Level 2: Defect Cause Tracing A three-layer Bayesian network is constructed for causal reasoning. The bottom layer parameter nodes include five nodes: welding current I, voltage U, wire feed speed Vf, shielding gas flow rate Qg, and welding speed Vw. The middle layer state nodes include four nodes: weld pool stability PS, heat input level HI, arc stability AS, and shielding effect PE. The top layer defect nodes include four nodes: porosity PO, lack of fusion LF, crack CR, and undercut UC.

[0064] The conditional probabilities between nodes are learned from historical data and smoothed using Laplace. in: For conditional probability; The number of times Y and X occur simultaneously; Let X be the number of times it occurs; The smoothing parameter is set to 1. Let Y be the number of states.

[0065] Given the feature E of a defective embryo, calculate the posterior probability of parameter abnormalities: in: This is the i-th parameter node; The prior probability is set to 0.05; Let the likelihood probability be denoted as '('). The marginal probability of the evidence.

[0066] The three causal paths with the highest probability were identified, and their path contributions were: in: The contribution of the k-th path; Let be the probability of the k-th path.

[0067] Level 3: Targeted Compensatory Intervention Define the defect loss function: in: Loss due to defects; This is the current state vector; , , , These are the conditional probabilities of porosity, lack of fusion, cracks, and undercut, respectively. , , , These are the corresponding weighting coefficients.

[0068] Parameter compensation vector calculation: in: For parameter compensation vector; The diagonal compensation gain matrix has diagonal elements of 2.0 (current), 1.5 (voltage), 1.8 (wire feed speed), 1.0 (gas flow rate), and 1.2 (welding speed). This represents the gradient of the loss function with respect to the parameters.

[0069] The gradient is calculated using finite differences: in: For the i-th parameter; Let be the disturbance quantity, take .

[0070] The compensation is implemented in three layers. The main compensation intensity is 60% of the calculated value, the auxiliary compensation is 30%, and the protective compensation is 10%. The parameter time evolution uses the sigmoid function. in: The parameter value at time t; These are the initial parameter values; For compensation amount; The transition rate coefficient is set to 0.5. To compensate for the start time; This refers to the current moment.

[0071] Indicators for evaluating the effectiveness of compensation: in: To ensure the effectiveness of compensation; To compensate for the entropy of the previous defects; The current defect entropy; The defect entropy under normal conditions is set to 1.5.

[0072] When the effectiveness index falls below 0.5, the compensation strategy is adjusted or the backup plan is switched. Through the synergistic effect of the three-tiered defense, proactive prevention of welding defects is achieved.

[0073] S4: Accumulate knowledge and evolve strategies This step enables the system to continuously optimize and evolve autonomously by constructing a dynamic knowledge graph and using reinforcement learning mechanisms.

[0074] Specifically, the knowledge graph is constructed using an attribute graph model, where nodes represent welding cases and edges represent similarity relationships between cases. Each case node contains six attribute domains: in: For case nodes; It is a material property vector, which includes the material type code and the content of 8 chemical elements (mass percentage of C, Si, Mn, P, S, Cr, Ni, Mo); It is a geometric attribute vector, containing plate thickness, bevel angle, blunt edge height, assembly clearance, and welding position encoding; This is a process attribute vector, containing welding method code, shielding gas type, and welding wire diameter; The parameter sequence matrix has a dimension of T×5, where T is the number of time steps and 5 is the number of parameters; This is a quality assessment vector, which includes weld width, weld reinforcement height, penetration depth, and surface quality score. This is a defect information vector, containing the defect type, location, and severity.

[0075] Case similarity calculation employs a multi-dimensional comprehensive evaluation. Material similarity is based on a weighted Euclidean distance of chemical composition: in: Material similarity; Let k be the mass percentage of the k-th element in case i. The mass percentage of the k-th element in case j; Let be the weight of the k-th element, where (carbon), (silicon), (manganese), (phosphorus), (sulfur), (chromium), (nickel), (molybdenum).

[0076] Geometric similarity primarily considers differences in plate thickness: in: Geometric similarity; and The plate thicknesses for cases i and j are in millimeters, respectively. For the plate thickness and bandwidth parameters, take 2 mm; Assuming a plate thickness weight, take 0.7; The bevel weight is set to 0.3; This is the bevel similarity function, which is 1 when the bevel forms are the same and 0.5 when they are different.

[0077] The parameter sequence similarity uses a dynamic time warping algorithm. The local distance between two sequences X and Y is defined as follows: in: Let X be the distance between sequence X at time i and sequence Y at time j. Let X be the k-th parameter value of sequence X at time i; 5 represents the k-th parameter value of sequence Y at time j; 5 represents the number of parameters.

[0078] Dynamic time warping solves for the cumulative distance matrix recursively: in: Cumulative distance; boundary conditions , , .

[0079] The parameter sequence similarity is defined as: in: For dynamic time-normalized distance; and These represent the sequence length; divided by Used for normalization.

[0080] Quality assessment similarity: in: Let m be the quality index of case i; and These are the maximum and minimum values ​​of the m-th indicator; For weighted index, (Weld width) (Yugao), (penetration depth) (Surface quality).

[0081] Overall similarity is calculated using a weighted combination: Case retrieval is accelerated using kd-trees. During construction, the dimension with the highest variance is selected for splitting. in: For the segmentation dimension; Let be the variance of the d-th dimension. The median is chosen as the split point, and the process is recursively built until the number of cases in the leaf nodes is less than 10.

[0082] Policy evolution employs a proximal policy optimization algorithm. The state space is defined to contain 12 continuous variables: in: This refers to the welding current. Voltage; This refers to the wire feeding speed; For welding speed; The area of ​​the molten pool; Aspect ratio; The oscillation frequency; Defect entropy; For temperature gradient; Acoustic characteristic index; It is the arc stability index; To protect gas pressure.

[0083] The action space is a discrete parameter adjustment, with each parameter having 5 actions: decrease by 10%, decrease by 5%, hold, increase by 5%, and increase by 10%.

[0084] The policy network is a three-layer fully connected network with 256 neurons in the first layer and 128 neurons in the second layer. The output layer generates the action probability distribution. The value network has the same structure and outputs a state value estimate.

[0085] Reward function design: Quality rewards are based on a Gaussian function. in: , , These are weld width, reinforcement height, and penetration depth, respectively. mm mm (Plate thickness) is the target value; mm mm The standard deviation is denoted as .

[0086] Stability Bonus: in: This represents the standard deviation of the molten pool area. This represents the average area of ​​the molten pool.

[0087] Defect Penalty: in: Defect entropy; The threshold value is used.

[0088] Energy penalty: in: This refers to the welding current. The time step is set to 0.1 seconds.

[0089] Change penalty: in: This is the action value for the i-th parameter.

[0090] PPO objective function: in: For policy network parameters; Importance sampling ratio; For the estimation of the advantage function; The trimming parameter is set to 0.2; It indicates the expectation of the time step.

[0091] Training uses a batch size of 64 and a learning rate of 3× The strategy is updated every 2048 steps. Through knowledge accumulation and reinforcement learning, the system continuously optimizes its control strategy.

[0092] S5: Achieving deep fusion of multimodal information This step achieves effective integration of information from different sensors through two levels: feature layer fusion and decision layer fusion, thereby improving the accuracy and reliability of welding status identification.

[0093] Specifically, feature layer fusion optimizes the combination of physical quantities that can be measured simultaneously by multiple sensors. Taking molten pool area as an example, both visual sensors and infrared thermal imagers can provide measurement values. Visual measurement values... The error variance is: in: The total variance of visual measurements; The variance of the segmentation error is taken as 0.5 mm². To determine the calibration error variance, a value of 0.2 mm² is used. The image noise variance is set to 0.3 mm².

[0094] Thermal imaging measurements The error variance is: in: The total variance of thermal imaging measurements; The variance of the temperature measurement error is taken as 0.4 mm². The error variance is defined for the gradient region and is set to 0.6 mm². The variance caused by the change in emissivity is taken as 0.3 mm².

[0095] Using the minimum variance criterion for fusion, the fusion estimate is: in: The area of ​​the molten pool after fusion; and For the weighting coefficients, satisfying .

[0096] The optimal weights are obtained by minimizing the fusion variance: in: ; .

[0097] Online variance estimation uses an exponentially weighted moving average: in: Let be the variance estimate of the i-th sensor at time t; The forgetting factor is set to 0.95. The value measured at time t; This is a mean estimate.

[0098] The mean estimate also uses an exponentially weighted moving average: in: The mean forgetting factor is set to 0.98.

[0099] The decision-making level fusion adopts the Dempster-Shafer evidence theory. The identification framework is defined as follows: in: This is the normal state; Stomatal tendency; It has a tendency to not fuse; It is prone to cracking.

[0100] Each sensing modality provides a basic probability assignment function. The visual modality, based on the molten pool morphology features, has the following basic probability assignment: , , , The acoustic modes are based on the characteristics of electric arc sound, and their basic probability assignments are: , , , The electrical signal modes are based on current-voltage stability, and their fundamental probability assignments are: , , , The thermal field mode is based on the temperature distribution, and its basic probability is assigned as follows: , , , .

[0101] Dempster's combination rules are: in: Assign values ​​to the basic probabilities of the combined set A; and The focal points are from two different sources of evidence; This represents the conflict coefficient.

[0102] The conflict coefficient is calculated as follows: in: This represents all focal element combinations whose intersection is empty.

[0103] Four modal recursive combinations: in: This represents Dempster's combination operation; This is the final fusion result.

[0104] The trust function and the plausibility function are as follows: in: Let A be the trust level of set A; Let A be the similarity score of set A; For all focal elements.

[0105] Decision-making rules are based on maximum trust level: in: This is the final state to be determined.

[0106] The cross-modal attention mechanism determines the importance of each mode through mutual information. The mutual information between mode i and welding state S is: in: The mutual information between the i-th mode and the state; For joint probability; and This represents the marginal probability.

[0107] Attention weights are normalized using softmax: in: Let be the attention weights for the i-th modality; For the temperature parameter, take 0.5; For modal indexes, from 1 to 4.

[0108] The fused feature vector is: in: The final fused feature has 64 dimensions. Let be the eigenvector of the i-th mode.

[0109] Anomaly detection employs an ensemble of three algorithms. Isolation forest anomaly score: in: The isolated forest score for sample x; This represents the average path length. The normalization constant is ; Let i be the harmonic number; This represents the number of samples.

[0110] Local anomalous factors: in: , which is a local anomaly factor for sample x; Let x be the set of its k nearest neighbors, where k is 20; This represents the locally accessible density.

[0111] Autoencoder anomaly score is based on reconstruction error: in: The score for anomalies in the autoencoder; The input vector; This is the reconstructed vector.

[0112] Integrated anomaly score: when When this occurs, the security control mode is triggered: in: For safety control parameters; These are nominal parameters; This is the current parameter.

[0113] When an anomaly is detected, if the current parameter Greater than the nominal value ,but = -0.5∣ - |; if less than, then = +0.5∣ - | means always adjusting in a conservative direction. By integrating multi-level information, the system makes full use of the complementary information from various sensors, thereby improving the accuracy of welding status identification and anomaly detection.

[0114] In summary, the intelligent control method for welding process parameters provided by this invention achieves intelligent control of the welding process through the organic integration of five core technologies: multi-physics field collaborative sensing, dual closed-loop adaptive control, proactive defect defense, knowledge evolution, and multimodal fusion. This method elevates defect prevention from passive response to proactive prevention, significantly improving welding quality and production efficiency, and providing strong technical support for the intelligent transformation of the automotive manufacturing industry.

[0115] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent control of welding process parameters for automotive parts, characterized in that, Includes the following steps: Construct a multi-physics collaborative sensing system that integrates at least four types of sensors to collect information on the molten pool morphology, acoustics, electrical parameters, and temperature field during the welding process; Implement dual closed-loop adaptive control: the inner loop performs fast response control based on the dynamic characteristics of the molten pool, and the outer loop performs quality look-ahead control based on a time-series prediction model. The active defense control for defects includes a three-level defense mechanism: the first level identifies defective embryos by extracting and analyzing micro-perturbation signals; the second level traces the cause of defects through a probabilistic inference network; and the third level implements hierarchical targeted compensation based on the traceability results. Construct a dynamic knowledge graph and continuously optimize the control strategy using reinforcement learning algorithms; Multimodal sensor information is integrated through a multi-level information fusion mechanism.

2. The method according to claim 1, characterized in that, The first-level defect embryo identification includes: Micro-perturbation components are extracted from multi-source sensor signals, wherein the amplitude of the micro-perturbation components is less than a preset ratio of the steady-state signal; Construct a multidimensional perturbation feature vector that includes time domain, frequency domain, and time-frequency domain; The disorder index of the probability distribution calculation system based on the perturbation mode; When the disorder index exceeds the threshold dynamically adjusted according to welding conditions, a defective embryo is determined to exist.

3. The method according to claim 1, characterized in that, The second level of defect cause tracing includes: Construct a multi-layer probabilistic network comprising a parameter layer, a state layer, and a defect layer; Based on Bayesian inference, the probability of parameter anomalies is inferred from the defect symptoms; Verify the causal relationships between parameters using time-series causal analysis; Identify the main propagation path from abnormal parameters to defect formation.

4. The method according to claim 1, characterized in that, The third-level targeted compensation includes: Based on the probability of parameter anomalies and the degree of causal contribution, the compensation is divided into three levels: primary compensation, secondary compensation, and protective compensation. The main compensation acts on the main inducing parameters, the auxiliary compensation acts on the related parameters, and the protective compensation is used to maintain system stability. The compensation process uses a smooth transition function to avoid sudden parameter changes; Real-time assessment of compensation effectiveness and dynamic adjustment of compensation strategies.

5. The method according to claim 1, characterized in that, In the aforementioned dual-closed-loop adaptive control: The inner-loop controller employs fuzzy inference combined with predictive compensation, achieving a response time in the millisecond range. The outer loop controller uses a deep temporal network to predict the quality trend of future time windows; The control effects of the inner and outer loops are coordinated through a variable weighting mechanism during the welding stage.

6. The method according to claim 1, characterized in that, The dynamic knowledge graph includes: A graph structure with welding cases as nodes and similarity relationships as edges; Each case node contains multi-dimensional attributes including material, geometry, process, parameters, quality, and defects; Rapid case retrieval and matching are achieved through multi-dimensional similarity calculation and efficient index structure.

7. The method according to claim 1, characterized in that, The reinforcement learning optimization includes: Define a continuous state space that includes welding parameters and process characteristics; The motion space is designed as a set of discrete parameter adjustments; Construct a composite reward function that comprehensively considers quality, stability, energy consumption, and defects; The policy gradient algorithm is used to achieve online optimization of the control policy.

8. The method according to claim 1, characterized in that, The multi-level information fusion includes: The feature layer employs adaptive weighted fusion of multi-sensor measurements of similar physical quantities; The decision-making body adopts evidence theory to integrate state judgments from different modalities; Dynamically assign cross-modal attention weights using information theory metrics.

9. The method according to any one of claims 1 to 8, characterized in that, It also includes anomaly detection and security control: An integrated approach using multiple anomaly detection algorithms is employed to identify abnormal states. When an anomaly is detected, it automatically switches to the preset security control mode; The safety control mode gradually adjusts the parameters to a safe range.

Images