Adaptive Control Method for Welding of High-Pressure Common Rail Solenoid Valves Based on Multimodal Feature Fusion
By generating adversarial networks through multimodal feature fusion and thermal conduction constraints at the interface of dissimilar materials, the problem of reduced welding status identification accuracy caused by sensor interference during the welding process of high-pressure common rail solenoid valves was solved, achieving stability and consistency of welding quality and improving the adaptive capability of welding control.
Patent Information
- Application Number
- CN202611175994.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-08-04
- Publication Date
- 2026-11-17
AI Technical Summary
Existing multimodal fusion welding control methods are susceptible to sensor interference under complex working conditions, which leads to reduced welding status identification accuracy and makes it difficult to meet the requirements of high-pressure common rail solenoid valve welding for penetration control accuracy and product qualification rate.
An adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion is adopted. The welding process is divided into four stages by the molten pool image and welding electrical signal features. The modal comprehensive confidence is calculated by combining the three-dimensional geometric features of the keyhole and Kalman filtering. A generative adversarial network with heat conduction constraints at the interface of dissimilar materials is used with a dual encoder-single decoder structure to generate missing modal compensation features, which are then weighted and spliced to output the final fused feature vector. The method predicts the weld depth, weld width and defect probability in real time and dynamically adjusts the laser power and defocusing amount.
It improves the anti-interference ability and robustness of the welding process, ensures the consistency and stability of welding quality, and enhances the reliability of the adaptive control system for welding high-pressure common rail solenoid valves.
Smart Images

Figure CN123103178A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of welding control technology for high-pressure common rail solenoid valves, specifically to an adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion. Background Technology
[0002] The high-pressure common rail solenoid valve is a core precision component of the diesel engine fuel injection system. The consistency of its weld penetration and sealing performance directly determine the engine's operational reliability and service life. Currently, the industry commonly uses pulsed laser welding to achieve precise connections for solenoid valves. To improve the controllability of the welding process, multimodal sensor fusion technology is gradually being introduced. This technology simultaneously acquires multi-dimensional data, including visual, thermal infrared, acoustic, and electrical signals, to characterize the welding state and thus achieve adaptive adjustment of process parameters.
[0003] However, existing multimodal fusion welding control methods generally adopt fusion strategies with fixed weights or simple confidence weights, which do not fully consider the strong dynamic characteristics of the welding process and the nonlinear time-varying correlation between different modes. Under complex working conditions in industrial sites, sensors are susceptible to spatter, arc interference or temporary failure, resulting in distorted fusion features and a sharp drop in welding status identification accuracy, making it difficult to meet the stringent requirements of high-pressure common rail solenoid valve welding for penetration control accuracy and product qualification rate. Summary of the Invention
[0004] The purpose of this invention is to provide an adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion, so as to solve the problems mentioned in the background art.
[0005] To address the aforementioned technical problems, this invention provides the following technical solution: an adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion, applied to the micro-seam pulsed laser welding process of dissimilar materials in high-pressure common rail solenoid valves, characterized in that: The keyhole's three-dimensional geometric features were extracted from the molten pool image, and the single-pass welding process was divided into four stages by combining the welding electrical signal features. A modal sensitive defect library corresponding to each stage was established offline. The four stages include the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage. Based on the three fundamental dimensions of the data, a keyhole fluctuation correction factor and a Kalman filter smoothing mechanism are introduced to calculate the comprehensive confidence score of the four original modes, and each original mode is divided into high-confidence mode, low-confidence mode, and missing mode. The three fundamental dimensions of the data include completeness, temporal stability, and anomaly deviation. The four original modes include visual, thermal infrared, acoustic, and electrical signals. The original features of the four original modalities are extracted. The original features classified as high-confidence modalities are used as high-confidence modal features, and the original features classified as low-confidence modalities are used as low-confidence modal features. For missing modes, a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is adopted. It integrates high-confidence mode features with interface thermal conduction features calculated based on the thermal conduction equation between heterogeneous materials to generate compensation features for missing modes. Among them, the heterogeneous material interface thermal conduction constraint generative adversarial network is pre-trained offline using a total loss function that includes interface thermal conduction constraint loss, and real-time adaptation is achieved through an online incremental update mechanism that only updates the bias parameters of the last layer of the decoder. Based on the current stage's defect risk probability, stage baseline weight, and comprehensive confidence score of each modality, calculate the comprehensive weight of each modality, and weight and concatenate the high-confidence modality features, low-confidence modality features, and compensation features to output the final fused feature vector; The fused feature vector is input into the welding state identification model to predict the weld depth, weld width and defect probability in real time. The model predictive control algorithm is used to dynamically adjust the laser power and defocusing amount, and the control effect is fed back to the modal sensitive defect library to update the defect risk probability.
[0006] Specifically, the fused feature vector is input into a lightweight MLP and then passed through three hidden layers to output the final welding state identification result. The first hidden layer has an output dimension of 64 dimensions and uses the ReLU activation function; the second hidden layer has an output dimension of 32 dimensions, uses the ReLU activation function, and uses L2 regularization to constrain the weights to avoid overfitting; the third hidden layer has an output dimension of 16 dimensions and uses the ReLU activation function; finally, an output layer outputs a 3-dimensional result, corresponding to the predicted weld penetration depth, predicted weld width, and defect probability value. The defect probability value is normalized using the Sigmoid function to obtain the probability of welding defects occurring. Preferably, the training set of the welding state identification model is a fusion feature sample containing dissimilar material pulsed laser welding of high-pressure common rail solenoid valves. The mean square error of weld depth and weld width combined with the cross-entropy loss of defect classification is used as the total loss function. The Adam optimizer is used, with the initial learning rate set to 0.001 and the number of training rounds set to 100. Every 20 rounds, the learning rate is decayed to 0.5.
[0007] Furthermore, the specific method for establishing the modal-sensitive defect library corresponding to each stage is as follows: Multiple sets of high-pressure common rail solenoid valve single-pass welding samples with grading and defects were collected. The molten pool images, welding electrical signals and final welding defect detection results were acquired simultaneously during the welding process. The molten pool images were denoised and the welding electrical signals were filtered and preprocessed to achieve time synchronization between the molten pool images and the welding electrical signals. The three-dimensional geometric features of the keyhole, including keyhole area, depth and opening width, were extracted from the preprocessed molten pool images. Based on the time-domain variation characteristics of welding electrical signals, the single-pass welding process is divided into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage. The defect type and the time point of defect occurrence for each welding sample are determined by metallographic inspection, sealing test, and X-ray flaw detection. The time point of defect occurrence is associated with the divided welding stages, and the welding stage to which each defect belongs is marked. The frequency of occurrence of various welding defects in each welding stage is statistically analyzed in all samples, and the occurrence rate of each type of defect in each welding stage is calculated. The occurrence rate is the proportion of the number of occurrences of a certain type of defect in a certain welding stage to the total number of occurrences of that type of defect. Defect types with an occurrence rate greater than a preset threshold are identified as high-incidence defect types in the corresponding welding stage. For each welding stage, the recognition accuracy of the high-incidence defect type is statistically analyzed using four modal features: visual, thermal infrared, acoustic, and electrical signals. Modal features with recognition accuracy greater than a preset threshold are identified as sensitive modal features of the defect type. Welding stages, corresponding high-incidence defect types, and defect-sensitive modal features are associated and stored to establish a modal sensitive defect library.
[0008] Furthermore, the specific method for dividing the single-pass welding process into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage is as follows: At the same time as welding starts, the high-speed vision sensor and the electrical signal acquisition module are triggered to work synchronously to acquire high-speed images of the welding area and welding current signals in real time, ensuring that the timestamps of the two are perfectly aligned. Based on the three-dimensional geometric features of the keyhole, the contour features of the keyhole are extracted, and the collected welding current signals are analyzed in real time to identify the rising, peak plateau and falling phases of the current. From the moment welding starts, the pre-melting stage is defined as the moment when the high-speed vision first detects the complete keyhole outline. During this stage, the welding current is rising and the material surface gradually melts but does not form a keyhole. The period from the moment when the high-speed vision system first detects a complete keyhole outline until the moment when the keyhole area change rate fluctuation does not exceed the keyhole transient stability threshold within three consecutive sampling cycles is defined as the keyhole transient formation stage. During this stage, the welding current approaches its peak value, the keyhole appears from nothing and expands rapidly, and the area changes drastically. The keyhole transient stability threshold is calculated by multiplying the maximum absolute value of the keyhole area change rate between two adjacent sampling cycles in all samples during the stable penetration stage by a safety factor. The period from the moment when the rate of change of keyhole area does not exceed the stable threshold of the penetration stage to the moment when the welding current begins to decrease is determined to be the stable penetration stage. During this stage, the welding current is maintained at the peak plateau, the keyhole shape and area remain relatively stable, and the penetration depth meets the design requirements. The stable threshold of the penetration stage is determined by calculating the standard deviation of the relative rate of change of keyhole area within a continuous sampling period to quantify the degree of fluctuation and determine the stable moment. The period from the moment the welding current begins to decrease until the welding ends and the weld is completely solidified is defined as the solidification and shrinkage stage. During this stage, the welding current gradually decreases to zero, the keyhole gradually collapses and disappears, and the weld metal cools and solidifies.
[0009] Furthermore, the specific method for calculating the overall confidence score of the four original modes is as follows: For the four original modes—visual, thermal infrared, acoustic, and electrical signals—normalized scores were calculated for three basic dimensions: data integrity, temporal stability, and anomaly deviation, with scores ranging from 0 to 1. The original features for each of the four modes were as follows: visual mode extracted molten pool contour, weld texture, and spatter area morphology; thermal infrared mode extracted the mean temperature field, temperature gradient, and high-temperature region proportion of the molten pool; acoustic mode extracted the spectral energy of the welding frequency band and peak impact noise; and electrical signal mode extracted the mean welding current and voltage, fluctuation variance, and temporal distortion. Extract the keyhole area change rate of the current frame, and calculate the keyhole fluctuation correction factor based on the current welding stage and the keyhole area change rate; Kalman filtering is used to predict the overall confidence score of the previous frame to obtain the predicted confidence value of the current frame. The comprehensive confidence score for each original modality in the current frame is calculated using the formula Ci(k) = α × βi(k) × [ω1Di(k) + ω2Si(k) + ω3Ai(k)] + (1 - α) × Ki(k), where α represents the weighting coefficient of the basic dimension, (1-α) is the weighting coefficient of the Kalman prediction; βi(k) represents the keyhole fluctuation correction factor of the i-th mode in the k-th frame, Ki(k) represents the Kalman filter prediction confidence of the i-th mode in the k-th frame; ω1, ω2, and ω3 are the weighting coefficients of the three basic dimensions, and the sum of the three is 1. Di(k) = 1 - (frame drop rate in the current sliding window + data interruption duration percentage in the current sliding window), which represents the data integrity score of the i-th modality in the k-th frame. Si(k) = max[0, 1 - min (variance of the feature vector within the current sliding window / variance threshold under normal operating conditions, 1)], which represents the temporal stability score of the i-th modality in the k-th frame; Ai(k) = max[0, 1 - (Euclidean distance between the current frame feature and the mean of historical normal features / the abnormal distance threshold of this modality)], which represents the abnormal deviation score of the i-th modality in the k-th frame; Based on the overall confidence score, each original mode is divided into three levels: high confidence mode, low confidence mode, and missing mode. Specifically, a high confidence mode is defined as a comprehensive confidence score ≥ 0.7, a low confidence mode is defined as a comprehensive confidence score ≥ 0.3 and < 0.7, and a missing mode is defined as a comprehensive confidence score < 0.3.
[0010] Furthermore, a specific method for generating compensation features for missing modes using a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is as follows: A generative adversarial network (GAN) with thermal conduction constraints at heterogeneous material interfaces using a dual encoder-single decoder structure is used to generate compensated features; the dual encoder includes a first encoder and a second encoder. The fusion features of all high-confidence modes in the current welding process are input into the first encoder to extract high-dimensional modal characterization features; based on the electrical signal parameters of the current welding stage and the thermal property parameters of the dissimilar materials, the interface thermal conduction features are calculated through the interface thermal conduction equation of the dissimilar materials, and the interface thermal conduction features are input into the second encoder to extract high-dimensional physical characterization features. High-dimensional modal representation features and high-dimensional physical representation features are concatenated and fused at the decoder input, and then input into the decoder to generate compensation features for the corresponding missing modalities. The generative adversarial network (GAN) is pre-trained offline using a total loss function that includes adversarial loss, feature reconstruction loss, and interface thermal conduction constraint loss. The interface thermal conduction constraint loss is used to ensure that the generated compensation features conform to the physical laws of thermal conduction in dissimilar material welding. After offline pre-training, the GAN is adapted in real time during the welding process using an online incremental update mechanism. The online incremental update mechanism only updates the bias parameters of the last layer of the decoder. The update method is as follows: the true features of the high-confidence mode in the current welding process are used as labels, the mean square error between the decoder output and the true features is used as the loss function, and the bias parameters of the last layer of the decoder are updated using gradient descent. The update time for a single update does not exceed 1ms.
[0011] Furthermore, the specific method for weighted concatenation of high-confidence modal features, low-confidence modal features, and compensation features to output the final fused feature vector is as follows: Based on historical data statistics, the defect risk probability Rj of the current stage is calculated in real time. For each mode, the comprehensive weight Wi(k) of the i-th mode is calculated as Wi(k) = Ci(k) × wi(k) × [1 + Rj × (w,i)], where wi(k) is the baseline weight of the i-th mode in the current welding stage, (w,i) is the identification weight of the i-th mode for the j-th type of defect in the current stage, and Ci(k) is the confidence level of the i-th mode in the k-th frame. For the i-th mode that is determined to be a missing mode, its comprehensive weight is multiplied by a preset discount coefficient to obtain the corrected compensation feature weight. The preset discount coefficient is used to reflect the uncertainty of the compensation feature. For high-confidence and low-confidence modes, their original feature vectors are multiplied by the corresponding comprehensive weights; for missing modes, their compensation feature vectors are multiplied by the corrected compensation feature weights; weighted concatenation and output: all weighted modal feature vectors are concatenated end to end in a preset order to form a one-dimensional final fused feature vector and output.
[0012] Furthermore, the specific method for dynamically adjusting laser power and defocusing amount using model predictive control algorithms is as follows: The welding state identification model inputs the real-time generated fusion feature vector and outputs the current moment's predicted penetration depth, weld width, and defect probability. Set the prediction time domain to 3 steps and the control time domain to 1 step; set the target values for melt depth and melt width; set the laser power adjustment range, the defocusing amount adjustment range, and the maximum rate of change limit for both. Based on the current predicted values of weld penetration and weld width, as well as the current values of laser power and defocusing amount, the trend of weld penetration and weld width changes in the next three moments is predicted through a pre-established linearized state equation of the welding process. Construct an objective function, which is the sum of squared errors between the predicted and target values of weld depth and weld width at the next three time points, plus the sum of squared changes in laser power and defocusing amount. Under the constraints of laser power, defocusing amount, and their rate of change, solve for the optimal control increment sequence that minimizes the objective function. Take the first element of the optimal control increment sequence and add it to the current laser power and defocusing amount respectively to obtain the laser power setpoint and defocusing amount setpoint for the next time point. Send the setpoints to the welding equipment actuator to complete one process parameter adjustment. After welding is completed, the weld penetration, weld width and defect detection are performed. The actual detection results are compared with the defect probability prediction results of each stage in the welding process. Based on the comparison results, the number of defect occurrences and the total number of samples in the modal sensitive defect library corresponding to the welding stage are updated, and the defect risk probability of that stage is recalculated for use in the three-dimensional dynamic fusion weight calculation of the subsequent welding process.
[0013] Compared with the prior art, the beneficial effects achieved by the present invention are as follows: The present invention adopts a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure, which integrates high-confidence modal features and interface thermal conduction features, and can make full use of the physical prior knowledge of heterogeneous material welding, thereby improving the authenticity and rationality of missing modal compensation features; the introduction of interface thermal conduction constraint loss ensures that the generated compensation features conform to the physical laws of welding, and avoids welding state identification deviation caused by invalid completion. This invention can effectively adapt to sensor interference and modal loss problems under complex welding conditions, enhance the anti-interference capability and robustness of the high-pressure common rail solenoid valve welding adaptive control system, and help ensure the consistency and stability of welding quality. Attached Figure Description
[0014] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of the adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to the present invention. Detailed Implementation
[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0016] Please see Figure 1 This invention provides a technical solution: an adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion, applied to the micro-seam pulsed laser welding process of dissimilar materials in high-pressure common rail solenoid valves, characterized in that: The keyhole's three-dimensional geometric features were extracted from the molten pool image, and the single-pass welding process was divided into four stages by combining the welding electrical signal features. A modal sensitive defect library corresponding to each stage was established offline. The four stages include the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage. Based on the three fundamental dimensions of the data, a keyhole fluctuation correction factor and a Kalman filter smoothing mechanism are introduced to calculate the comprehensive confidence score of the four original modes, and each original mode is divided into high-confidence mode, low-confidence mode, and missing mode. The three fundamental dimensions of the data include completeness, temporal stability, and anomaly deviation. The four original modes include visual, thermal infrared, acoustic, and electrical signals. The original features of the four original modalities are extracted. The original features classified as high-confidence modalities are used as high-confidence modal features, and the original features classified as low-confidence modalities are used as low-confidence modal features. For missing modes, a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is adopted. It integrates high-confidence mode features with interface thermal conduction features calculated based on the thermal conduction equation between heterogeneous materials to generate compensation features for missing modes. Among them, the heterogeneous material interface thermal conduction constraint generative adversarial network is pre-trained offline using a total loss function that includes interface thermal conduction constraint loss, and real-time adaptation is achieved through an online incremental update mechanism that only updates the bias parameters of the last layer of the decoder. Based on the current stage's defect risk probability, stage baseline weight, and comprehensive confidence score of each modality, calculate the comprehensive weight of each modality, and weight and concatenate the high-confidence modality features, low-confidence modality features, and compensation features to output the final fused feature vector; The fused feature vector is input into the welding state identification model to predict the weld depth, weld width and defect probability in real time. The model predictive control algorithm is used to dynamically adjust the laser power and defocusing amount, and the control effect is fed back to the modal sensitive defect library to update the defect risk probability.
[0017] Furthermore, the specific method for establishing the modal-sensitive defect library corresponding to each stage is as follows: Multiple sets of high-pressure common rail solenoid valve single-pass welding samples with grading and defects were collected. The molten pool images, welding electrical signals and final welding defect detection results were acquired simultaneously during the welding process. The molten pool images were denoised and the welding electrical signals were filtered and preprocessed to achieve time synchronization between the molten pool images and the welding electrical signals. The three-dimensional geometric features of the keyhole, including keyhole area, depth and opening width, were extracted from the preprocessed molten pool images. Based on the time-domain variation characteristics of welding electrical signals, the single-pass welding process is divided into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage. The defect type and the time point of defect occurrence for each welding sample are determined by metallographic inspection, sealing test, and X-ray flaw detection. The time point of defect occurrence is associated with the divided welding stages, and the welding stage to which each defect belongs is marked. The frequency of occurrence of various welding defects in each welding stage is statistically analyzed in all samples, and the occurrence rate of each type of defect in each welding stage is calculated. The occurrence rate is the proportion of the number of occurrences of a certain type of defect in a certain welding stage to the total number of occurrences of that type of defect. Defect types with an occurrence rate greater than a preset threshold are identified as high-incidence defect types in the corresponding welding stage. For each welding stage, the recognition accuracy of the high-incidence defect type is statistically analyzed using four modal features: visual, thermal infrared, acoustic, and electrical signals. Modal features with recognition accuracy greater than a preset threshold are identified as sensitive modal features of the defect type. Welding stages, corresponding high-incidence defect types, and defect-sensitive modal features are associated and stored to establish a modal sensitive defect library.
[0018] Furthermore, the specific method for dividing the single-pass welding process into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage is as follows: At the same time as welding starts, the high-speed vision sensor and the electrical signal acquisition module are triggered to work synchronously to acquire high-speed images of the welding area and welding current signals in real time, ensuring that the timestamps of the two are perfectly aligned. Based on the three-dimensional geometric features of the keyhole, the contour features of the keyhole are extracted, and the collected welding current signals are analyzed in real time to identify the rising, peak plateau and falling phases of the current. From the moment welding starts, the pre-melting stage is defined as the moment when the high-speed vision first detects the complete keyhole outline. During this stage, the welding current is rising and the material surface gradually melts but does not form a keyhole. The period from the moment when the high-speed vision system first detects a complete keyhole outline until the moment when the keyhole area change rate fluctuation does not exceed the keyhole transient stability threshold within three consecutive sampling cycles is defined as the keyhole transient formation stage. During this stage, the welding current approaches its peak value, the keyhole appears from nothing and expands rapidly, and the area changes drastically. The keyhole transient stability threshold is calculated by multiplying the maximum absolute value of the keyhole area change rate between two adjacent sampling cycles in all samples during the stable penetration stage by a safety factor. The period from when the rate of change of keyhole area does not exceed the stable threshold of the penetration stage to when the welding current begins to decrease is considered the stable penetration stage. During this stage, the welding current remains at its peak plateau, the keyhole shape and area remain relatively stable, and the penetration depth meets the design requirements. The stable threshold of the penetration stage is determined by calculating the standard deviation of the relative rate of change of keyhole area within a continuous sampling period to quantify the degree of fluctuation and determine the stable moment. The specific steps are as follows: For each sampling moment, the difference between the keyhole area of the current frame and the keyhole area of the previous frame is calculated. The difference is divided by the keyhole area of the previous frame to obtain the relative change rate of the keyhole area at the current moment. Taking three consecutive sampling periods as a sliding window, the standard deviation of the relative change rate of the keyhole area within the window is calculated as the fluctuation value corresponding to the window. Based on the statistical data of more than 1,000 historical qualified welding samples, the standard deviation of the relative change rate of the keyhole area in the stable penetration stage of all samples is extracted, and its 95th percentile is taken as the stability threshold of the penetration stage. When the fluctuation value of a certain sliding window is less than or equal to the stability threshold of the penetration stage, the last sampling moment of the window is determined to be the moment when the fluctuation of the keyhole area change rate is stable.
[0019] The period from the moment the welding current begins to decrease until the welding ends and the weld is completely solidified is defined as the solidification and shrinkage stage. During this stage, the welding current gradually decreases to zero, the keyhole gradually collapses and disappears, and the weld metal cools and solidifies.
[0020] Furthermore, the specific method for calculating the overall confidence score of the four original modes is as follows: For the four original modes—visual, thermal infrared, acoustic, and electrical signals—normalized scores were calculated for three basic dimensions: data integrity, temporal stability, and anomaly deviation, with scores ranging from 0 to 1. The original features for each of the four modes were as follows: visual mode extracted molten pool contour, weld texture, and spatter area morphology; thermal infrared mode extracted the mean temperature field, temperature gradient, and high-temperature region proportion of the molten pool; acoustic mode extracted the spectral energy of the welding frequency band and peak impact noise; and electrical signal mode extracted the mean welding current and voltage, fluctuation variance, and temporal distortion. Extract the keyhole area change rate of the current frame, and calculate the keyhole fluctuation correction factor based on the current welding stage and the keyhole area change rate. Specifically, determine the welding stage to which the current frame belongs through the aforementioned welding stage division method, including pre-melting, keyhole transient formation, stable penetration and solidification shrinkage. Calculate the rate of change of the keyhole area in the current frame: ΔQ = |Q k -Q k-1 | / Q k-1 ×100%, where Q k Q is the keyhole area extracted in the current frame. k-1 This represents the area of the keyhole extracted in the previous frame. If no keyhole was detected in the previous frame, then ΔQ is set to 100%. Currently in the pre-melting stage or solidification shrinkage stage: there is no normal keyhole fluctuation, and the keyhole fluctuation correction factor is always set to 1; If the current stage is stable penetration: the keyhole should remain stable, and any sudden change in area is considered an interference; when ΔQ < 20%, the keyhole fluctuation correction factor is 1; when 20% ≤ ΔQ < 50%, the keyhole fluctuation correction factor is 0.5; when ΔQ ≥ 50%, the keyhole fluctuation correction factor is 0.
[0021] If the current keyhole is in the transient formation stage: allow the keyhole to expand rapidly. When ΔQ < 300%, it is judged as normal keyhole fluctuation, and the keyhole fluctuation correction factor is 1; when 300% ≤ ΔQ < 500%, it is judged as mild interference, and the keyhole fluctuation correction factor is 0.5; when ΔQ ≥ 500%, it is judged as severe interference, and the keyhole fluctuation correction factor is 0.
[0022] Kalman filtering is used to predict the overall confidence score of the previous frame to obtain the predicted confidence value of the current frame. The comprehensive confidence score for each original modality in the current frame is calculated using the formula Ci(k) = α × βi(k) × [ω1Di(k) + ω2Si(k) + ω3Ai(k)] + (1 - α) × Ki(k), where α represents the weighting coefficient of the basic dimension, (1-α) is the weighting coefficient of the Kalman prediction; βi(k) represents the keyhole fluctuation correction factor of the i-th mode in the k-th frame, Ki(k) represents the Kalman filter prediction confidence of the i-th mode in the k-th frame; ω1, ω2, and ω3 are the weighting coefficients of the three basic dimensions, and the sum of the three is 1. Di(k) = 1 - (frame drop rate in the current sliding window + data interruption duration percentage in the current sliding window), which represents the data integrity score of the i-th modality in the k-th frame. Si(k) = max[0, 1 - min (variance of the feature vector within the current sliding window / variance threshold under normal operating conditions, 1)], which represents the temporal stability score of the i-th modality in the k-th frame; Ai(k) = max[0, 1 - (Euclidean distance between the current frame feature and the mean of historical normal features / the abnormal distance threshold of this modality)], which represents the abnormal deviation score of the i-th modality in the k-th frame; Based on the overall confidence score, each original mode is divided into three levels: high-confidence mode, low-confidence mode, and missing mode.
[0023] Furthermore, a specific method for generating compensation features for missing modes using a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is as follows: A generative adversarial network (GAN) with thermal conduction constraints at heterogeneous material interfaces using a dual encoder-single decoder structure is used to generate compensated features; the dual encoder includes a first encoder and a second encoder. The fusion features of all high-confidence modes in the current welding process are input into the first encoder to extract high-dimensional modal characterization features; based on the electrical signal parameters of the current welding stage and the thermal property parameters of the dissimilar materials, the interface thermal conduction features are calculated through the interface thermal conduction equation of the dissimilar materials, and the interface thermal conduction features are input into the second encoder to extract high-dimensional physical characterization features. Specifically, the welding current and welding voltage at the current moment are collected as electrical signal input parameters; the thermal conductivity, specific heat capacity, and density of stainless steel and copper, two dissimilar materials, are pre-stored as thermal property input parameters; a one-dimensional transient heat conduction model of the dissimilar material interface is established, simplifying the welding area along the thickness direction into a three-layer structure of stainless steel layer, interface layer and copper layer, and establishing one-dimensional transient heat conduction equations for each layer, where the interface layer considers the contact thermal resistance between the two materials; Initial and boundary conditions are set. The initial condition is the ambient temperature before welding begins. The effective heat flux density of the laser heat source is calculated based on the welding current and welding voltage at the current moment, and used as the boundary condition for the upper surface of the stainless steel layer. The lower surface of the copper layer is set as the adiabatic boundary condition. The temperature distribution is obtained by solving the heat conduction equation. The above one-dimensional transient heat conduction equation is discretized and solved using the finite difference method to calculate the temperature distribution of the stainless steel layer, interface layer and copper layer along the thickness direction at the current moment. The interface heat conduction characteristics are extracted from the calculated temperature distribution, including the temperature value at the interface, the interface heat flux density, the temperature gradient of the stainless steel side interface, and the temperature gradient of the copper side interface. These characteristics together constitute the interface heat conduction characteristics.
[0024] High-dimensional modal representation features and high-dimensional physical representation features are concatenated and fused at the decoder input, and then input into the decoder to generate compensation features for the corresponding missing modalities. The generative adversarial network (GAN) is pre-trained offline using a total loss function that includes adversarial loss, feature reconstruction loss, and interface thermal conduction constraint loss. The interface thermal conduction constraint loss is used to ensure that the generated compensation features conform to the physical laws of thermal conduction in dissimilar material welding. After offline pre-training, the GAN is adapted in real time during the welding process using an online incremental update mechanism. The online incremental update mechanism only updates the bias parameters of the last layer of the decoder. The update method is as follows: the true features of the high-confidence mode in the current welding process are used as labels, the mean square error between the decoder output and the true features is used as the loss function, and the bias parameters of the last layer of the decoder are updated using gradient descent. The update time for a single update does not exceed 1ms.
[0025] Specifically, multiple sets of complete multimodal data were collected without interference during the pulsed laser welding process of stainless steel-copper dissimilar materials in high-pressure common rail solenoid valves, and the corresponding welding stages and interface heat conduction parameters were labeled. An ITCM-GAN network structure was constructed, including a generator and a discriminator. The generator adopted a dual-encoder-single-decoder structure. The first encoder was a high-confidence modality feature encoder, which input high-confidence modality concatenation features and output high-dimensional semantic encoded features through three hidden layers. The output dimension of the first hidden layer was 128 dimensions, using the ReLU activation function; the output dimension of the second hidden layer was 64 dimensions, using the ReLU activation function and employing L2 regularization to constrain the weights to avoid overfitting; the output dimension of the third hidden layer was 32 dimensions, using the ReLU activation function; the second... The encoder is an interface heat conduction feature encoder. After inputting the interface heat conduction features, it passes through two hidden layers to output physical constraint encoded features. The first hidden layer has a 64-dimensional output dimension and uses the ReLU activation function; the second hidden layer has a 16-dimensional output dimension, uses the ReLU activation function, and utilizes L2 regularization to constrain the weights to avoid overfitting. The decoder takes the concatenation result of the high-dimensional semantic encoded features and the physical constraint encoded features as input, passes through three hidden layers to output initial compensation features. The first hidden layer has a 64-dimensional output dimension and uses the ReLU activation function; the second hidden layer has a 16-dimensional output dimension, uses the ReLU activation function, and utilizes L2 regularization to constrain the weights to avoid overfitting. The first hidden layer has a 32-dimensional output, uses the ReLU activation function, and leverages L2 regularization to constrain the weights and prevent overfitting. The third hidden layer also has a 32-dimensional output, uses a linear activation function, and outputs initial compensation features with the same dimension as the original modality. The discriminator contains three parallel branches: a authenticity discrimination branch, a stage consistency discrimination branch, and a physical consistency discrimination branch. The authenticity discrimination branch passes the input features through two hidden layers to output the authenticity judgment result. The first hidden layer has a 32-dimensional output and uses the ReLU activation function; the second hidden layer has a 32-dimensional output. The input features are processed through two hidden layers, with the first hidden layer having a 32-dimensional output using the ReLU activation function and the second hidden layer having a 4-dimensional output using the Softmax activation function. The physical consistency branch processes the input features through two hidden layers, outputting the physical consistency result. The first hidden layer has a 32-dimensional output using the ReLU activation function, and the second hidden layer has a 1-dimensional output using the Sigmoid activation function. The total loss function is set as L=Lb+λ1×Lt+λ2×Ls, where Lb is the basic loss, including adversarial loss and feature reconstruction loss; Lt is the interface thermal conduction constraint loss; Ls is the stage consistency loss; and λ1 and λ2 are weight coefficients pre-set according to the dissimilar material welding process. The ITCM-GAN model is iteratively trained using the training dataset until the loss function converges, resulting in a pre-trained ITCM-GAN model. Preferably, the training set for ITCM-GAN consists of 10,000 sets of interference-free, complete multimodal samples of stainless steel-copper dissimilar material pulsed laser welding of high-pressure common rail solenoid valves, labeled with the corresponding welding stages and interface heat conduction parameters. The total loss function is used as the training objective, with the Adam optimizer, an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. The learning rate is decayed to 0.5 every 20 epochs.
[0026] Furthermore, the specific method for weighted concatenation of high-confidence modal features, low-confidence modal features, and compensation features to output the final fused feature vector is as follows: Based on historical data statistics, the defect risk probability Rj of the current stage is calculated in real time. For each mode, the comprehensive weight Wi(k) of the i-th mode is calculated as Wi(k) = Ci(k) × wi(k) × [1 + Rj × (w,i)], where wi(k) is the baseline weight of the i-th mode in the current welding stage, (w,i) is the identification weight of the i-th mode for the j-th type of defect in the current stage, and Ci(k) is the confidence level of the i-th mode in the k-th frame. For the i-th mode that is determined to be a missing mode, its comprehensive weight is multiplied by a preset discount coefficient to obtain the corrected compensation feature weight. The preset discount coefficient is used to reflect the uncertainty of the compensation feature. For high-confidence and low-confidence modes, their original feature vectors are multiplied by the corresponding comprehensive weights; for missing modes, their compensation feature vectors are multiplied by the corrected compensation feature weights; weighted concatenation and output: all weighted modal feature vectors are concatenated end to end in a preset order to form a one-dimensional final fused feature vector and output.
[0027] Furthermore, the specific method for dynamically adjusting laser power and defocusing amount using model predictive control algorithms is as follows: The welding state identification model inputs the real-time generated fusion feature vector and outputs the current moment's predicted penetration depth, weld width, and defect probability. Set the prediction time domain to 3 steps and the control time domain to 1 step; set the target values for melt depth and melt width; set the laser power adjustment range, the defocusing amount adjustment range, and the maximum rate of change limit for both. Based on the current predicted values of weld penetration and weld width, as well as the current values of laser power and defocusing amount, the trend of weld penetration and weld width changes in the next three moments is predicted through a pre-established linearized state equation of the welding process. Construct an objective function, which is the sum of squared errors between the predicted and target values of weld depth and weld width at the next three time points, plus the sum of squared changes in laser power and defocusing amount. Under the constraints of laser power, defocusing amount, and their rate of change, solve for the optimal control increment sequence that minimizes the objective function. Take the first element of the optimal control increment sequence and add it to the current laser power and defocusing amount respectively to obtain the laser power setpoint and defocusing amount setpoint for the next time point. Send the setpoints to the welding equipment actuator to complete one process parameter adjustment. After welding is completed, the weld penetration, weld width and defect detection are performed. The actual detection results are compared with the defect probability prediction results of each stage in the welding process. Based on the comparison results, the number of defect occurrences and the total number of samples in the modal sensitive defect library corresponding to the welding stage are updated, and the defect risk probability of that stage is recalculated for use in the three-dimensional dynamic fusion weight calculation of the subsequent welding process.
[0028] Specifically, a lightweight multilayer perceptron is used as the welding condition identification model to achieve real-time prediction of weld penetration, weld width, and defect probability. The specific steps are as follows: A multilayer perceptron model consisting of an input layer, two hidden layers, and an output layer is constructed. The number of nodes in the input layer is consistent with the dimension of the fused feature vector. The hidden layer uses the ReLU activation function. The output layer contains three nodes, corresponding to the predicted melt depth, predicted melt width, and defect probability value, respectively.
[0029] Multiple sets of welding process fusion feature vectors and corresponding actual weld depth, weld width and defect detection results are collected to construct a training dataset; the model is iteratively trained using the sum of the mean square error of weld depth and weld width and the cross-entropy loss of defect classification as the total loss function until the loss function converges, thus obtaining a pre-trained welding state identification model.
[0030] The real-time generated fusion feature vector is input into the pre-trained welding state identification model, which outputs the current moment's predicted weld depth, weld width, and defect probability.
[0031] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.
[0032] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.
Claims
1. An adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion, applied to the micro-seam pulsed laser welding process of dissimilar materials in high-pressure common rail solenoid valves, characterized in that: The keyhole three-dimensional geometric features are extracted from the molten pool image, and the single-pass welding process is divided into four stages by combining the welding electrical signal features. A modal sensitive defect library corresponding to each stage is established offline. The comprehensive confidence score of four original modes is calculated based on three fundamental dimensions of the data, distinguishing them into high-confidence modes, low-confidence modes, and missing modes. The three fundamental dimensions of the data include completeness, temporal stability, and anomaly deviation. The four original modes include visual, thermal infrared, acoustic, and electrical signals. For the missing modes, a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is used to generate compensation features for the missing modes. Based on the current stage's defect risk probability, stage baseline weight, and comprehensive confidence score of each modality, calculate the comprehensive weight of each modality, and weight and concatenate the high-confidence modality features, low-confidence modality features, and compensation features to output the final fused feature vector; The fused feature vector is input into the welding state identification model to predict the weld depth, weld width and defect probability in real time. The model predictive control algorithm is used to dynamically adjust the laser power and defocusing amount, and the control effect is fed back to the modal sensitive defect library to update the defect risk probability.
2. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 1, characterized in that: The specific methods for establishing the modality-sensitive defect library corresponding to each stage are as follows: Collect qualified and defective single-pass welding samples of high-pressure common rail solenoid valves, and simultaneously acquire images of the molten pool, welding electrical signals, and final welding defect detection results during the welding process; extract the three-dimensional geometric features of the keyhole from the molten pool images, including keyhole area, depth, and opening width features; Based on the time-domain variation characteristics of welding electrical signals, the single-pass welding process is divided into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage. The defect type and the time point of defect occurrence for each welding sample are determined, and the time point of defect occurrence is associated with the divided welding stages to mark the welding stage to which each defect belongs. The frequency of occurrence of various welding defects in each welding stage is statistically analyzed in all samples. The occurrence rate of each type of defect in each welding stage is calculated, and the defect types with an occurrence rate greater than a preset threshold are identified as the high-incidence defect types in the corresponding welding stage. For the common defect types in each welding stage, the recognition accuracy of four modal features for the defect type is statistically analyzed, and the modal with a recognition accuracy greater than a preset threshold is identified as the sensitive mode of the defect type. The welding stage, the corresponding high-incidence defect type, and the defect-sensitive mode are associated and stored to establish a modal-sensitive defect library.
3. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 2, characterized in that: The specific method for dividing the single-pass welding process into the pre-melting stage, the keyhole transient formation stage, the stable penetration stage, and the solidification shrinkage stage is as follows: Simultaneously with the start of welding, high-speed images of the welding area and welding current signals are acquired in real time; based on the three-dimensional geometric features of the keyhole, the keyhole contour features are extracted, and the acquired welding current signals are analyzed in real time to identify the rising, peak plateau, and falling phases of the current. From the moment welding begins, the pre-melting stage is defined as the moment when the high-speed vision system first detects the complete keyhole outline. The transient keyhole formation stage is defined as the moment when the keyhole area change rate fluctuates within three consecutive sampling periods and does not exceed the keyhole transient stability threshold. The stable penetration stage is defined as the moment when the keyhole area change rate fluctuates within three consecutive sampling periods and does not exceed the penetration stage stability threshold. The penetration stage stability threshold is determined by calculating the standard deviation of the relative change rate of the keyhole area within consecutive sampling periods to quantify the degree of fluctuation and determine the stable moment. The solidification shrinkage stage is defined as the moment when the welding current begins to decrease and welding ends and the weld is completely solidified.
4. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 3, characterized in that: The specific method for calculating the overall confidence score of the four original modes is as follows: For the four original modes of vision, thermal infrared, acoustic and electrical signals, normalized scores of three basic dimensions, namely data integrity, temporal stability and abnormal deviation, are calculated respectively; the keyhole area change rate of the current frame is extracted, and the keyhole fluctuation correction factor is calculated based on the current welding stage and the keyhole area change rate. Kalman filtering is used to predict the overall confidence score of the previous frame to obtain the predicted confidence value of the current frame; the overall confidence score of each original modality of the current frame is calculated according to the formula Ci(k)=α×βi(k)×[ω1Di(k)+ω2Si(k)+ω3Ai(k)]+(1-α)×Ki(k), where α represents the basic dimension weighting coefficient and (1-α) is the Kalman prediction weighting coefficient; βi(k) represents the keyhole fluctuation correction factor for the i-th mode in the k-th frame, Ki(k) represents the Kalman filter prediction confidence for the i-th mode in the k-th frame; ω1, ω2, and ω3 are the weight coefficients of the three basic dimensions, and the sum of the three is 1; Di(k) represents the data integrity score for the i-th mode in the k-th frame, Si(k) represents the temporal stability score for the i-th mode in the k-th frame, and Ai(k) represents the anomaly deviation score for the i-th mode in the k-th frame. Based on the overall confidence score, each original mode is divided into three levels: high-confidence mode, low-confidence mode, and missing mode.
5. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 4, characterized in that: The specific method for generating compensation features for missing modes using a heterogeneous material interface thermal conduction constraint generative adversarial network with a dual encoder-single decoder structure is as follows: A generative adversarial network (GAN) with thermal conduction constraints at heterogeneous material interfaces using a dual encoder-single decoder structure is used to generate compensated features; the dual encoder includes a first encoder and a second encoder. The fusion features of all high-confidence modes in the current welding process are input into the first encoder to extract high-dimensional modal representation features; Based on the electrical signal parameters of the current welding stage and the thermal properties of the dissimilar materials, the interface thermal conduction characteristics are calculated through the interface thermal conduction equation of the dissimilar materials. The interface thermal conduction characteristics are then input into the second encoder to extract high-dimensional physical characterization features. High-dimensional modal representation features and high-dimensional physical representation features are concatenated and fused at the decoder input, and then input into the decoder to generate compensation features for the corresponding missing modalities.
6. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 5, characterized in that: The specific method for weighted concatenation of high-confidence modal features, low-confidence modal features, and compensation features to output the final fused feature vector is as follows: Based on historical data statistics, the defect risk probability Rj of the current stage is calculated in real time. For each mode, the comprehensive weight Wi(k) of the i-th mode is calculated as Wi(k) = Ci(k) × wi(k) × [1 + Rj × (w,i)], where wi(k) is the baseline weight of the i-th mode in the current welding stage, (w,i) is the identification weight of the i-th mode for the j-th type of defect in the current stage, and Ci(k) is the confidence level of the i-th mode in the k-th frame. For the i-th mode that is determined to be a missing mode, its comprehensive weight is multiplied by a preset discount coefficient to obtain the corrected compensation feature weight. The preset discount coefficient is used to reflect the uncertainty of the compensation feature. For high-confidence and low-confidence modes, their original feature vectors are multiplied by the corresponding comprehensive weights; for missing modes, their compensation feature vectors are multiplied by the corrected compensation feature weights; weighted concatenation and output: all weighted modal feature vectors are concatenated end to end in a preset order to form a one-dimensional final fused feature vector and output.
7. The adaptive control method for welding high-pressure common rail solenoid valves based on multimodal feature fusion according to claim 6, characterized in that: The specific method for dynamically adjusting laser power and defocusing amount using a model predictive control algorithm is as follows: The welding state identification model inputs the real-time generated fusion feature vector and outputs the current moment's predicted penetration depth, weld width, and defect probability. Set the prediction time domain and control time domain; set the target values for melt depth and melt width; set the laser power adjustment range, defocusing amount adjustment range, and the maximum rate of change limit for both; Based on the current predicted values of weld penetration and weld width, as well as the current values of laser power and defocusing amount, the trend of weld penetration and weld width changes in the next three moments is predicted through a pre-established linearized state equation of the welding process. Construct an objective function and, under the constraints of laser power, defocusing amount, and their rate of change, find the optimal control increment sequence that minimizes the objective function. Take the first element of the optimal control increment sequence and add it to the current laser power and defocusing amount respectively to obtain the laser power setpoint and defocusing amount setpoint for the next moment. Send the setpoints to the welding equipment actuator to complete one process parameter adjustment.
8. A computer device, characterized in that: It includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.
9. A computer-readable storage medium, characterized in that: The device stores a computer program that, when executed by a processor, causes the processor to perform the steps of the method as described in any one of claims 1 to 7.
