Multi-modal sensing and intelligent quality control method for friction stir welding of spaceflight structure

Abstract

The invention relates to the technical field of aerospace manufacturing and intelligent welding control, and solves the technical problems of insufficient single-mode perception, limited quality prediction precision, lagging parameter regulation and control and weak model adaptive capacity in traditional welding process monitoring. In particular to a spaceflight structure friction stir welding multi-modal sensing and intelligent quality control method, which combines multi-modal real-time sensing and reinforcement learning by introducing a body intelligent body thought, so that a system can autonomously understand a welding state, predict a quality trend and actively adjust process parameters in a welding process; and adaptive optimization control with quality as constraint is realized. According to the method, the limitation of traditional off-line modeling and manual adjustment is broken through, an intelligent control framework with self-learning and self-evolution capacity is provided for spaceflight-level friction stir welding, and a new technical approach is provided for achieving high-consistency and high-reliability welding seams. The intelligent level of friction stir welding of the spaceflight structural part and the welding seam consistency level are remarkably improved.

Description

Technical Field

[0001] This invention relates to the fields of aerospace manufacturing and intelligent welding control technology, and in particular to a multimodal sensing and intelligent quality control method for friction stir welding of aerospace structures. Background Technology

[0002] Friction stir welding (FSW) is a solid-state welding technique that joins high-strength aluminum alloys, magnesium alloys, titanium alloys, and dissimilar metals using a non-consumable method. Due to its advantages such as low defect rate, high weld strength, minimal deformation, and no pollution, it has been widely used in the manufacturing of aerospace structural components. However, the welding of complex aerospace modules, tanks, and skin-like components typically involves complex conditions such as multi-curved surfaces, varying thicknesses, and non-uniform heat dissipation, resulting in a highly nonlinear thermo-mechanical coupling process. Traditional control methods based on fixed process parameters struggle to guarantee weld consistency and internal quality stability.

[0003] Current welding process control largely relies on manual experience or offline experimental parameter settings, making it difficult to adapt to dynamic changes in working conditions. For example, slight variations in feed rate, spindle speed, and pressure under different materials, fixture stiffness, and ambient temperature can lead to thermal input deviations, causing quality problems such as voids, incomplete welds, tunnel defects, or grain coarsening. Although quality prediction models based on machine learning or deep neural networks have emerged in recent years, most of these models are based on single-modal data (such as force or temperature signals), lacking cross-modal fusion and real-time feedback capabilities, making it difficult to respond quickly to sudden disturbances. Furthermore, traditional parameter optimization methods are mainly based on multi-objective iterative calculations or heuristic algorithms (such as genetic algorithms and particle swarm optimization) based on static data, which have limited search efficiency and real-time performance, failing to meet the dynamic control requirements of high-cycle, continuous aerospace welding production. The lack of an intelligent control structure that unifies perception, cognition, and decision-making processes within a single system is the core bottleneck currently facing quality control in high-end friction stir welding. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a multimodal sensing and intelligent quality control method for friction stir welding of aerospace structures, which solves the technical problems of insufficient single-modal sensing, limited quality prediction accuracy, lag in parameter adjustment, and weak model adaptive capability in traditional welding process monitoring.

[0005] To address the aforementioned technical problems, this invention provides the following technical solution: a multimodal sensing and intelligent quality control method for friction stir welding of aerospace structures, comprising the following steps: S1. Synchronously acquire multi-modal signals during the welding process to comprehensively characterize the dynamic behavior and potential defect features of the welding process, including multi-source physical data generated during the welding process, including mechanical information, acoustic information, thermal field changes and surface image data; S2. Input the multimodal signal into the welding quality prediction model, extract the feature variables that reflect the welding thermal stability and material coupling state, and combine them with the training network to predict and evaluate tensile properties and defect risks, and output an index that comprehensively reflects the welding quality. S3. Construct an embodied intelligent agent for welding control tasks, and automatically form a strategy update mechanism based on historical data and online feedback to achieve adaptive decision optimization of welding process parameters including feed rate, spindle speed and pressure. S4. Under the premise that the tensile properties of the weld are not lower than the set threshold, the optimal combination of welding parameters is solved by combining energy consumption and process stability, using reinforcement learning or gradient-based search algorithm, and then sent to the welding equipment in real time to realize the dynamic control of process parameters. S5. After completing a section of weld, the welding quality prediction model and the embodied intelligent agent are incrementally updated based on the actual quality inspection results and sensor process records to form a closed-loop feedback path, thereby realizing self-learning ability and quality robustness in complex manufacturing tasks.

[0006] Furthermore, in step S1, the specific process includes the following steps: S11. Deploy multi-dimensional force sensors in key parts of the friction stir welding equipment to obtain mechanical information on the workpiece applied by the stirring head during the welding process, including axial pressure, lateral thrust and instantaneous change rate, so as to characterize the material flow resistance and contact strength in the weld area, and serve as the basic indicators for judging the viscoplastic state and the degree of stirring in the weld area. S12. Set up a high-frequency response acoustic emission sensor to collect acoustic information of elastic sound waves excited by microcrack propagation, interface discontinuity or material internal delamination during the welding process. By performing short-time Fourier transform analysis on the original acoustic information, extract the time-frequency energy spectrum characteristics and main frequency band distribution characteristics for capturing microscopic signs before defect initiation. S13. Use an infrared thermal imager or thermocouple array to monitor the temperature of the welding area throughout the entire process, capture the thermal field changes in the temperature field distribution of the weld center and heat-affected zone, and calculate the average temperature, maximum gradient, and temperature fluctuation coefficient. This serves as the basis for evaluating the uniformity of welding heat input and the stability of the thermal cycle; S14. Real-time welding images are acquired using a high-resolution industrial camera. Visual texture analysis algorithms are used to extract image features including weld width, surface smoothness, and forming symmetry. Meanwhile, optical flow is used to track the dynamic deformation of the weld area to reflect the strain recovery behavior of the metal material in the post-weld stage. S15. Perform time synchronization processing and anomaly cleaning on the modal data streams collected in steps S11-S14 to obtain multi-source physical data.

[0007] Furthermore, the welding quality prediction model includes a front-end, which uses a causal convolutional temporal convolutional network to capture short-term process disturbances and local periodic noise, and a back-end. The backend uses a multi-head self-attention Transformer encoder to aggregate cross-window dependencies, enabling long-range information exchange between different welding stages; And a fusion layer is used to perform channel splicing and layer normalization on the local robust representation of the temporal convolutional network output and the global dependency representation of the Transformer encoder output, so as to suppress the bias of the single structure under sudden perturbations or trend drift. In addition, a top-level branch for the multi-task system is set up, corresponding to mechanical property regression, thermal stability estimation, and defect rate discrimination, including: The return branch outputs the expected tensile properties. And its proportional stability to the reference strength is constrained by the relative error loss; Thermally stable branch, direct regression temperature fluctuation coefficient To continue the thermal mode's characterization of energy uniformity; Defective branches, combining visual morphology and acoustic band characteristics to output defect rate The probability estimation is used to improve sensitivity to rare defects by focusing loss.

[0008] Furthermore, in step S2, the specific process includes the following steps: S21. Unify the encoding of multimodal signals and extract features to form a multidimensional feature vector for input to the welding quality prediction model, including: The force signal corresponding to the mechanical information is extracted using sliding window statistics. The main energy frequency band is obtained by wavelet packet decomposition of the acoustic signal corresponding to the acoustic information. Construct temperature trend and fluctuation indicators from the thermal sensing signals corresponding to changes in the thermal field. Visual parameters, including weld contour, defect area, and geometric deviation, are extracted from surface image data. S22. Utilize the welding quality prediction model to model the multidimensional feature vector and output the predicted tensile properties of the weld at the current moment. Temperature fluctuation coefficient And the defect rate inferred from image / acoustic analysis. Each indicator represents a key dimension of welding quality; S23. Construct a normalized welding quality scoring function to quantitatively evaluate whether the current welding condition meets the process standards for aerospace structural components. The expression is as follows: ; in, For comprehensive quality indicators; , where is the weighting coefficient, reflecting the intensity of the influence of each quality indicator on the overall quality score; These are standard tensile property reference values; The score of the stability factor for temperature fluctuations after normalization. S24. Assess the overall quality index of the current welding section. With set quality threshold When comparing, If the quality is deemed satisfactory, adjustment is not initiated; when... If the problem persists, proceed to the next optimization step to find feasible parameter adjustment paths.

[0009] Furthermore, the embodied intelligent agent includes a perception layer, a strategy layer, and a feedback update module; The perception layer is used to perform normalization, noise reduction, and temporal compression. The policy layer is used to process the state vector within a millisecond time limit. Perform forward reasoning to generate next parameter action suggestions to meet the real-time and stability requirements of the edge side; The feedback update module is used to inject the execution results returned by the device into a new round of perception-prediction information closed loop, so as to continuously correct strategy deviations and absorb new operating condition knowledge.

[0010] Furthermore, in step S3, the specific process includes the following steps: S31. The multi-dimensional feature vectors used as multimodal sensing data and the prediction output of the welding quality prediction model are used as a joint state description, and are uniformly encoded with historical welding behavior into a state vector. Control strategies are generated by embodied intelligent agents; S32. The policy layer of the embodied agent is trained using a proximal policy optimization algorithm, and expected actions are generated based on historical interaction experience to optimize the predicted quality Q value in the current state. ,Right now: ; in, It is the feed rate; It is the spindle speed; This refers to the downward pressure. S33. Use the prediction quality Q-value as a reward signal. The parameters and actions executed are paired and stored in the experience replay cache, and the long-term returns are calculated using a discounted reward function. ,Right now: ; in, This is a discount factor used to balance the effects of short-term and long-term control. S34. The policy update mechanism is formed by iteratively updating the network parameters of the policy layer using the gradient descent method in batches.

[0011] Furthermore, the historical welding behavior includes multi-dimensional feature vectors and comprehensive indicators from the most recent control cycles. Decomposition amount And the executed parameter trajectory and its rate of change.

[0012] Furthermore, in step S4, the specific process includes the following steps: S41. A multi-objective optimization function for the welding process is established by combining weld defect risk, energy consumption, and comprehensive indicators. The expression is: ; in, It is the feed rate; It is the spindle speed; This refers to the downward pressure. The defect risk item is output by a defect regression model learned from historical data. The energy consumption per unit length of welding is estimated based on the heat input model. This is a comprehensive quality scoring function; The target balancing weights are either manually set or learned. S42. Under the constraint conditions, the comprehensive quality scoring function is satisfied. and Under the condition that the process parameters are within a legal range, the process parameters are searched using the strategy gradient method or black-box optimization method to find a multi-objective optimization function. The minimum optimal parameter combination ; S43. Combine the optimal parameters The control parameters are sent to the friction welding equipment control system, and the welding control parameters are updated in real time through the set interface, and a smooth switching is achieved in the form of a gradually changing control curve. S44. Continuously monitor the multimodal sensing signals during the updated process. If abnormal deviations, sudden changes, or failure to achieve the expected quality improvement trend occur, immediately stop the current strategy or revert to the previous stable control parameter state to ensure production safety and process stability.

[0013] Furthermore, in step S5, the specific process includes the following steps: S51. After welding, perform tensile property testing and internal defect analysis on the sample. By comparing the actual results with the prediction results output by the welding quality prediction model, evaluate the generalization accuracy and local error range of the welding quality prediction model. S52. Construct the actual error term as a loss function and backpropagate it to the model parameters of the welding quality prediction model, and use a fine-tuning strategy to update the incremental model. S53. The parameters, actions, welding feedback results, and environmental states of each execution are combined into a triplet to continuously update the policy network through reinforcement learning. This enables the policy layer of the embodied agent to absorb experience from short-term memory, realizing a complete closed-loop mechanism from raw signal acquisition to decision control and result feedback, forming a "perception-reasoning-execution-correction" cycle.

[0014] By employing the above technical solution, the present invention provides a multimodal sensing and intelligent quality control method for friction stir welding of aerospace structures, which has at least the following beneficial effects: 1. This invention integrates multimodal data, introduces an embodied intelligent agent decision-making mechanism and reinforcement learning optimization strategy to construct a welding quality control system that can autonomously perceive, learn in real time and dynamically optimize, thereby fundamentally improving the intelligence and stability level of friction stir welding of aerospace structural components.

[0015] 2. The method proposed in this invention breaks through the limitations of traditional offline modeling and manual adjustment, and provides an intelligent control framework with self-learning and self-evolution capabilities for aerospace-grade friction stir welding, providing a new technical approach to achieve highly consistent and highly reliable welds. Attached Figure Description

[0016] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 This is a flowchart of the method for predicting and optimizing the quality of friction stir welding in this invention. Detailed Implementation

[0017] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. This will allow for a full understanding of how the present application uses technical means to solve technical problems and achieve technical effects, and to facilitate its implementation.

[0018] This embodiment proposes a multimodal perception and intelligent quality control method for friction stir welding of aerospace structures. It introduces the concept of an embodied agent, combining real-time multimodal perception with reinforcement learning. This enables the system to autonomously understand the welding state, predict quality trends, and proactively adjust process parameters during the welding process, achieving adaptive optimization control constrained by quality. This method overcomes the limitations of traditional offline modeling and manual adjustment, providing an intelligent control framework with self-learning and self-evolution capabilities for aerospace-grade friction stir welding, and offering a new technical approach to achieving highly consistent and reliable welds. Figure 1 As shown, the method includes the following steps: S1. Synchronously acquire multimodal signals during the welding process to comprehensively characterize the dynamic behavior and potential defect features of the welding process, including multi-source physical data generated during welding, including mechanical information, acoustic information, thermal field changes, and surface image data. The specific process includes the following steps: S11. A multi-dimensional force sensor module is arranged on the main shaft loading assembly and support base of the friction stir welding equipment to collect the axial pressure between the stirring head and the workpiece in real time using a high-frequency sampling method. ), lateral thrust ( The force signal and its instantaneous rate of change are analyzed. By performing moving average and spectral analysis on the force signal, the mechanical disturbance characteristics caused by plastic flow of material, wear of the stirring pin, or instability of the contact surface during welding can be identified. This sensing module forms a direct feedback channel of "force-flow-mass" in welding. The continuity of the force signal fluctuation amplitude reflects the flow resistance of the metal in the weld zone, while its stability characterizes the contact strength between the stirring pin and the workpiece.

[0019] When force signals exhibit periodic high-frequency fluctuations, it often indicates that the material has not been sufficiently plasticized locally or has experienced slippage and delamination. Based on this mechanical information, the viscoplastic state and mixing level of the weld zone can be dynamically determined, providing a physical constraint basis for subsequent process optimization and early defect warning. This multi-dimensional force sensing design significantly improves the analytical capability for complex three-dimensional welding force fields compared to traditional single-axis monitoring methods, making force signals not only a process quantity but also one of the key input channels for quality prediction models.

[0020] S12. A high-sensitivity acoustic emission sensor array is deployed near the weld boundary in the welding area to collect the elastic acoustic waves released by the metal material under the coupling effect of frictional heat and plastic deformation during the welding process. The acoustic emission signal contains high-frequency characteristic components of hidden defects such as microcrack initiation, porosity formation, and interface non-bonding within the weld. Its time response is usually completed within milliseconds, so it needs to be captured without distortion through a high sampling rate (≥200kHz) and a wideband front-end amplification circuit. After short-time Fourier transform (STFT) or wavelet transform, the acoustic energy spectrum distribution, the peak frequency, the center frequency of energy density, and their variation trends can be extracted.

[0021] A sudden increase in acoustic spectrum energy often corresponds to crack propagation, while a shift in the dominant frequency to higher frequencies reflects localized material hardening. By establishing a statistical mapping relationship between acoustic emission characteristics and weld defects, potential quality anomalies can be predicted seconds in advance during the welding process, enabling a shift from "post-production detection" to "real-time diagnosis." This acoustic modality complements the limitations of force and temperature signals, giving the system the ability to detect micro-perturbations in the internal structure of the weld and macroscopic stress release, laying the foundation for multi-modal collaborative quality identification.

[0022] S13. Infrared thermal imagers and multi-point thermocouple arrays are deployed on both sides of the welding channel and behind the stirring pin, respectively, to synchronously monitor the weld center and heat-affected zone, ensuring complete capture of the thermal field distribution in both time and space. The infrared thermal imagers are responsible for acquiring macroscopic temperature field distribution cloud maps, while the multi-point thermocouple arrays are used to calibrate local point temperature differences and gradient changes, thereby achieving bidirectional comparison between surface radiation temperature and actual internal temperature. By performing temperature fitting and curve smoothing on continuous frame images, the weld center temperature is extracted. Boundary temperature With time derivative Calculate the average temperature Maximum temperature gradient and temperature fluctuation coefficient The calculation formula is: ; in, This represents the standard deviation of the welding temperature series. A larger value indicates uneven heat input, which may lead to insufficient metal flow or grain coarsening.

[0023] The stability of temperature distribution not only reflects the uniformity of energy input in the weld but also directly affects the recrystallization behavior and microstructure density of the material. By introducing thermal modes, real-time determination of welding thermal cycle stability and process window can be achieved, providing physically interpretable process characteristics for subsequent optimization control. Compared with traditional single-point temperature monitoring, this method has the advantages of area coverage and multi-layer data fusion, greatly enhancing the accuracy and robustness of thermal process visualization.

[0024] S14. Real-time welding images are acquired using a high-resolution industrial camera. Visual texture analysis algorithms are used to extract image features including weld width, surface smoothness, and forming symmetry. At the same time, optical flow is used to track the dynamic deformation of the weld area to reflect the strain recovery behavior of the metal material in the post-weld stage.

[0025] S15. Install a high-resolution industrial camera system at a suitable angle behind the stirring head to continuously image the weld formation area, acquiring the geometric contour and fusion zone boundary features of the weld surface in real time. After grayscale conversion, edge detection, and texture enhancement, visual feature parameters such as weld width, surface roughness, reinforcement height, symmetry, and boundary straightness are extracted. Simultaneously, optical flow is used to track the dynamic deformation of the weld area to reflect the strain recovery behavior of the metal material in the post-weld stage.

[0026] S16. Time synchronization processing and outlier cleaning are performed on the collected modal data streams to obtain multi-source physical data, ensuring the consistency of data used in subsequent modeling in terms of timing, accuracy, and completeness. In this embodiment, the visual frames are time-stamped with force, sound, and heat modal signals through a synchronous trigger signal to ensure the temporal consistency of multi-source data. During data cleaning, outlier frames (such as spatter occlusion and brightness abrupt changes) are removed or repaired, and interpolation smoothing is performed to maintain continuity. By introducing the visual modality, phenomena such as weld surface defects, misalignment, or vibration marks can be transformed into quantifiable indicators, providing intuitive evidence of appearance dimensions for the quality prediction model. The combined use of this modality with force, sound, and heat signals enables a unified description of the surface and inner states of the welding process, achieving dual-domain perception fusion from "physical response" to "visual shaping."

[0027] S2. Input the multimodal signals into the welding quality prediction model, extract feature variables reflecting the welding thermodynamic stability and material coupling state, and combine them with the trained network to predict and evaluate tensile properties and defect risks, outputting a comprehensive index reflecting welding quality. The specific process includes the following steps: S21. After performing unified timestamp alignment and modal normalization on the time-series data from the four channels of force, sound, heat, and vision, a hierarchical encoding strategy is adopted to generate multi-dimensional feature vectors that can be directly consumed by the learner, namely: The force signals corresponding to the mechanical channel are statistically summarized using a fixed-length sliding window to construct quantitative quantities that characterize the stability of the process, such as window mean, variance, peak value, peak-valley amplitude, spectral center, and spectral kurtosis.

[0028] The acoustic signal corresponding to the acoustic channel is first subjected to wavelet packet decomposition to separate multi-scale detail components, and then the energy ratio of each scale, the center frequency of the main energy band, the instantaneous bandwidth and the entropy value are extracted to capture the precursor signs of microcracks and interface discontinuities.

[0029] The thermal signal corresponding to the thermal sensing channel is constructed based on the fusion sequence of thermal images and thermocouples to determine the temperature trend slope, local extremum density, and temperature fluctuation coefficient. Indices such as cross-interval gradient directly map the uniformity of heat input and the stability of thermal cycling.

[0030] The surface image data corresponding to the visual channel is used to extract weld contours, surface roughness, excess height and forming symmetry through edge detection and texture descriptors. Suspicious defect areas and geometric deviations are marked by connected component analysis and converted into scale-independent morphological features at the pixel level.

[0031] The aforementioned features are concatenated at the encoding end in the form of a combination of "local statistics + frequency domain energy + trend term + morphological parameters", which not only preserves the intrinsic differences of multi-physics fields, but also achieves cross-modal comparability on the numerical scale, providing a stable, sufficient and interpretable input representation for subsequent sequence modeling.

[0032] S22. Input the multidimensional feature vector formed in step S21 into the welding quality prediction model of the fusion regression network, and output the expected tensile properties of the weld at the current moment. Temperature fluctuation coefficient And the defect rate inferred from image / acoustic analysis. Each indicator represents a key dimension of welding quality.

[0033] In this embodiment, the network front-end uses a causal convolutional temporal convolutional network (TCN) to capture short-term process disturbances and local periodic noise. The back-end uses a multi-head self-attention Transformer encoder to aggregate cross-window dependencies, completing long-range information interaction between different welding stages. The fusion layer performs channel splicing and layer normalization on the local robust representation of the TCN output and the global dependency representation of the Transformer output to suppress the bias of a single structure under sudden disturbances or trend drift.

[0034] The model's top-level design incorporates multiple task branches, corresponding to mechanical property regression, thermal stability estimation, and defect rate determination, respectively: Expected tensile properties of the return branch output Its proportional stability to the reference strength is constrained by the relative error loss. The thermally stable branch directly regresses the temperature fluctuation coefficient. To continue the characterization of energy uniformity using the thermal mode. The defect rate is output by combining the visual morphology and acoustic band characteristics of the defective branch. The probability estimation is used to improve sensitivity to rare defects by focusing loss.

[0035] Through a multi-task design that shares a main trunk and dedicated branches, the three indicators are mutually regularized and constrained, ensuring that the output remains stable and transferable under non-ideal conditions such as welding load fluctuations, material batch differences, and equipment stiffness variations.

[0036] S23. Construct a normalized welding quality scoring function to quantitatively evaluate whether the current welding condition meets the process standards for aerospace structural components. The expression is as follows: ; in, For comprehensive quality indicators; , where is the weighting coefficient, reflecting the intensity of the influence of each quality indicator on the overall quality score; These are standard tensile property reference values; The score represents the normalized value of the stability factor for temperature fluctuations.

[0037] Comprehensive quality indicators Numerical consistency with upper bounds and constraints with lower bounds makes it easy to use as a target or constraint for subsequent optimization. At the same time, the three physically interpretable sub-indices allow error sources to be located by dimension, facilitating targeted intervention and traceability analysis on the process side.

[0038] S24. Comprehensive quality index obtained for each control cycle Threshold determination is performed on continuous sequences, using a quality threshold. To determine the pass / fail criteria and to avoid frequent jitter by considering the hysteresis band and minimum hold time, the following applies: when Furthermore, by maintaining stability within the set time window, the current parameter trajectory can be preserved to ensure forming continuity and consistency of heat input; when If the value continuously exceeds the lower hysteresis threshold or shows a significant downward trend, the parameter optimization routine is triggered, which submits the current state, historical window, and prediction signal to the decision-making layer for constraint search of feasible parameter sets.

[0039] This judgment mechanism is based on the interpretable decomposition of scores. It can intervene in advance under boundary conditions and avoid over-adjustment during short-term fluctuations, thereby achieving a balance between quality stability and process stability. It provides clear, unified and traceable entry conditions for the generation and execution of strategies in subsequent steps S3-S4.

[0040] S3. Construct an embodied intelligent agent oriented towards welding control tasks, and automatically form a strategy update mechanism based on historical data and online feedback to achieve adaptive decision optimization of welding process parameters, including feed rate, spindle speed, and pressure. The specific process includes the following steps: S31. Using the time-aligned multi-dimensional feature vectors of the multimodal sensing data and the prediction output of the welding quality prediction model as a joint state description, the multi-dimensional feature vectors and comprehensive indicators from the most recent control cycles are combined. Decomposition amount The executed parameter trajectories and their rates of change are uniformly encoded into state vectors. .

[0041] The embodied intelligent agent includes a perception layer, a policy layer, and a feedback update module, and its design meets the requirements of edge deployment, real-time computing, and low-latency decision-making.

[0042] Then, the perception layer of the embodied agent performs normalization, noise reduction, and temporal compression; the policy layer processes the state vector within a millisecond time limit. Forward reasoning is performed to generate next step parameter action suggestions to meet the real-time and stability requirements of the edge side; the feedback update module injects the execution results returned by the device and a new round of perception-prediction information into the closed loop to continuously correct strategy deviations and absorb new operating condition knowledge.

[0043] The overall structure is robust to communication latency and control jitter. It uses a fixed-step asynchronous buffer and priority queue to ensure that critical states are processed first, so that strategy decisions can maintain low latency and high availability under different equipment stiffness, material batches and thermal environment disturbances.

[0044] S32. The policy layer of the embodied agent is trained using a proximal policy optimization algorithm, and expected actions are generated based on historical interaction experience to optimize the predicted quality Q value in the current state. ,Right now: ; in, It is the feed rate; It is the spindle speed; This is the downward pressure, i.e., the desired action. It is the desired feed rate, spindle speed, and depth of cut adjustment scheme to optimize the predicted quality Q value under the current state.

[0045] In this embodiment, the policy layer uses PPO as the core training paradigm and adopts a shared trunk + dual-head structure (policy head and value head) to directly output the desired action within the continuous action space. The Gaussian distribution mean and variance are used, and unified control of different dimensions is achieved through learnable action scales. Before transmission, the action is controlled by safety projection and rate of change limiting (…). Constraints are used to avoid abrupt changes in the thermo-mechanical field.

[0046] in, It is the absolute value of the change in feed rate, representing the adjustment amount of the rate between two adjacent control cycles. If it is too large, it will cause a sudden change in the material flow rate, resulting in local voids or insufficient mixing. It is the absolute value of the spindle speed change, reflecting the intensity of the adjustment of the stirring needle rotation power. Excessive speed change may cause abnormal instantaneous heat input, resulting in uneven weld nugget structure or thermal cracks. It is the absolute value of the variation in the depth of pressure (or axial pressure), representing the stability of the vertical pressure control. Excessive variation can easily cause accelerated wear of the stirring pin or collapse of the weld surface.

[0047] The comprehensive indicators given in step S2 during the training process The primary optimization objective is to incorporate energy consumption and stability penalties into the reward shaping process, thereby improving the overall performance of the strategy. It simultaneously considers energy efficiency and process smoothness. Based on the sampling-update rhythm of the rolling horizon, the strategy utilizes historical interaction fragments to adapt to different speed ranges, plate thicknesses, and clamping stiffnesses, thereby maintaining reliable parameter adjustment capabilities under unspecified nonlinear and time-varying characteristics.

[0048] S33. Use the prediction quality Q-value as a reward signal. The parameters and actions executed are paired and stored in the experience replay cache, and the long-term returns are calculated using a discounted reward function. ,Right now: ; in, This is a discount factor that adjusts the trade-off between long-term quality improvement and short-term stability. A larger discount factor is used when the weld crosses a material joint or heat dissipation boundary, causing a lag in the quality response. This encourages the strategy to learn a gradual adjustment of "stability first, then optimization"; when takt time requirements are strict and operating conditions are stable, a smaller discount factor is used. This helps improve sensitivity to immediate quality fluctuations. Combined with piecewise normalization and dominance function standardization, the reward value range is compressed into a stable interval, facilitating subsequent numerical optimization and variance control of the PPO target.

[0049] S34. The network parameters of the policy layer are iteratively updated using the gradient descent method with batch updates to form a policy update mechanism, which improves the generalization ability and response speed of the embodied agent in similar task scenarios, and enhances policy diversity by combining entropy regularization terms to prevent getting trapped in local optima.

[0050] Within each update cycle, the policy and value head are jointly optimized using multiple mini-batches. A PPO truncation target is employed to limit the deviation of the ratio between the old and new policies, and an adaptive KL divergence threshold is used as an early stopping criterion to suppress overfitting and policy collapse. An entropy regularization term is introduced to maintain the degree of exploration, allowing for the attempt of diverse actions even with changes in new materials, thicknesses, or clamping boundary conditions, thus avoiding getting trapped in local optima.

[0051] The value function is jointly trained using a smooth regression loss and a biased estimation of the advantage function, combined with a hierarchical learning rate and weight decay to balance convergence speed and stability. The trained policy is exported to the edge with lightweight weights. Combined with tensor inference and pipelined parallelism, it ensures that forward computation and safe projection are completed within the control cycle, thereby obtaining a stable, fast, and transferable decision response on the actual production line.

[0052] S4. Under the premise that the tensile properties of the weld are not lower than a set threshold, and considering energy consumption and process stability, the optimal combination of welding parameters is solved using reinforcement learning or gradient-based search algorithms, and then sent to the welding equipment in real time to achieve dynamic control of process parameters. The specific process includes the following steps: S41. A multi-objective optimization function for the welding process is established by combining weld defect risk, energy consumption, and comprehensive indicators. The expression is: ; in, It is provided by the defect probability regressor, and the input includes the defect rate predicted in step S2. Rather than being sensitive to parameter combinations, it penalizes combinations that may result in incomplete welding, porosity, or tunneling defects. This is the energy consumption per unit length of welding, estimated based on the heat input model; It is a comprehensive quality scoring function, provided by the quality scoring function, reflecting both strength compliance and thermal stability; The target balance weights, whether manually set or learned, can be calibrated offline or updated adaptively online according to the task focus (such as quality priority or energy efficiency priority).

[0053] The multi-objective optimization function incorporates three physically interpretable indices into a unified optimizable framework, ensuring computability while retaining adjustable space for different production strategies. This allows the parameter search to converge toward a compromise solution of "low defects, low energy consumption, and high quality," facilitating efficient solution-solving within constraints and safety boundaries.

[0054] S42. Under the constraint conditions, the comprehensive quality scoring function is satisfied. and Under the condition that the process parameters are within a legal range, the process parameters are searched using the strategy gradient method or black-box optimization method to find a multi-objective optimization function. The minimum optimal parameter combination ; The comprehensive quality scoring function satisfies the constraints. and Under the condition of a legally defined process range, two complementary solution strategies are adopted: First, continuous control search based on policy gradients can directly reuse the policy layer trained in step S3 for "hot start," performing local fine-grained descent within the current working condition neighborhood to improve convergence speed; Second, black-box global optimization (genetic algorithm / Bayesian optimization) is used for global exploration across materials and thicknesses, relying on surrogate models (Gaussian process / tree structure Parzen estimation) for multi-objective optimization functions. Perform high-value sampling driven by uncertainty characterization and acquisition function.

[0055] The two strategies work together in a hierarchical manner: first, a global coarse search is performed to determine the cluster of potential optimal solutions, and then local fine-tuning is used to approximate the feasible optimum, thereby ensuring the overall quality scoring function. Under the premise of finding a multi-objective optimization function Minimum optimal parameter combination It balances the breadth of exploration with the accuracy of convergence, and improves the adaptability to non-convex and time-varying conditions.

[0056] S43. Combine the optimal parameters The parameters are sent to the friction welding equipment control system, and the welding control parameters are updated in real time through the set interface. The switching is achieved smoothly in the form of a gradually changing control curve to avoid system overshoot and control oscillation.

[0057] The optimal parameter combination obtained After safety projection and consistency verification, the parameters are sent to the servo and spindle circuits via the equipment's numerical control interface (NC / PLC) and switched using piecewise linear or S-shaped gradual curves. The update process interpolates to the target value at a fixed sampling period, constrains the parameter increment per unit time, and introduces feedforward compensation at key turning points to suppress transient disturbances under thermo-mechanical field coupling.

[0058] At the same time, a short-window quality score monitoring function is enabled to monitor the overall quality indicators. The system quickly checks the upward trend and fluctuation range, and automatically slows down the interpolation step size if overshoot signs are detected. This smooth switching strategy accelerates the arrival at the optimal operating point and reduces control oscillations while ensuring the continuity of weld formation. This allows parameter adjustments to be seamlessly integrated into the production cycle, avoiding secondary disturbances to process stability.

[0059] S44. Continuously monitor the multimodal sensing signals during the updated process. If abnormal deviations, sudden changes, or failure to achieve the expected quality improvement trend occur, immediately stop the current strategy or revert to the previous stable control parameter state to ensure production safety and process stability.

[0060] Within several control cycles following parameter updates, key statistics and comprehensive quality indicators for the four modes of force, acoustics, heat, and visual performance were analyzed. The trajectory is monitored online, and three criteria are set: offset threshold, abrupt change detection, and trend consistency. When mechanical fluctuations exceed the steady-state confidence interval, acoustic emission band density rises abnormally, and temperature fluctuation coefficients... Continuously increasing or comprehensive quality indicators If the expected improvement slope is not observed, a safety rollback is immediately triggered to restore the parameters to the previous stable operating condition; at the same time, the state and actions at the time of triggering are recorded to be incorporated into the experience sample update strategy network and agent model.

[0061] In this embodiment, the monitoring process is based on minimum lag detection and rapid rollback, forming a closed-loop control logic of "early detection, rapid rollback, and re-optimization". This not only prevents potential defects from evolving into irreversible stages, but also ensures the predictability and consistency of production under complex boundary conditions.

[0062] S5. After completing a section of weld, the welding quality prediction model and the embodied intelligent agent are incrementally updated based on the actual quality inspection results and sensor process records to form a closed-loop feedback path, thereby achieving self-learning capability and quality robustness in complex manufacturing tasks. Step S5 specifically includes the following steps: S51. After welding, tensile property tests and internal defect analyses are performed on the samples. The generalization accuracy and local error range of the welding quality prediction model are evaluated by comparing the actual results with the predicted results output by the welding quality prediction model. Specifically, after each welding segment, representative samples are obtained according to standardized sampling and testing procedures, and uniaxial tensile tests are conducted to obtain the actual tensile strength. Non-destructive testing (ultrasound / DR / X-ray / phased array) and necessary metallographic sections were used to quantitatively evaluate internal defects, and indicators such as defect area ratio, void volume fraction and defect location distribution were obtained.

[0063] Compare the above measured results with the output of step S2. , , The model is aligned segment by segment, and statistical quantities such as relative error of tensile properties, consistency deviation of temperature fluctuation, and consistency of defect detection are calculated. Generalization accuracy curves and local error ranges are provided under stratified conditions such as material batch, plate thickness, clamping stiffness, and environmental temperature and humidity. When systematic deviations are found in specific working conditions (such as heat dissipation boundaries, thickness transitions, and lap start / stop ends), they are marked as high-risk sample domains for priority absorption in subsequent incremental learning. This provides data support for the robustness assessment and correction of the welding quality prediction model while ensuring traceability of detection.

[0064] S52. Construct the actual error term as a loss function and backpropagate it to the model parameters of the welding quality prediction model. Use a fine-tuning strategy to update the model incrementally to ensure that the model maintains the performance of the old scenario under the new data.

[0065] In this embodiment, the incremental learning loss is constructed using the performance deviation, thermal stability deviation, and defect consistency deviation calculated in step S51. The regression error, stability difference metric, and classification / probability calibration error are weighted by multiple tasks, and online / quasi-online updates are performed through a parameter-efficient strategy without changing the backbone structure of the network.

[0066] Specifically, Elastic Weight Consolidation uses the Fisher information matrix to approximate the second-order sensitivity of important parameters, introducing stabilization regularization to key weights to reduce the risk of forgetting old scenarios; LoRA fine-tuning only updates a small number of incremental parameters on low-rank adaptation branches, while keeping the main branch frozen or with a small learning rate, allowing the model to absorb new operating condition knowledge with lower computational and storage costs; during incremental updates, a hold set and validation gate are set to continuously monitor the performance regression of the old distribution, and if necessary, roll back to the previous stability checkpoint and adjust the loss weights, thereby balancing new data adaptation with the preservation of historical capabilities and ensuring prediction consistency across batches, materials, and devices.

[0067] S53. The parameters, actions, welding feedback results, and environmental conditions of each execution are combined into a triplet to continuously update the policy network through reinforcement learning. This allows the policy layer of the embodied agent to absorb experience from short-term memory and continuously improve its quality assurance capabilities.

[0068] In this embodiment, the parameters and actions executed each time, the welding feedback results, and the environmental conditions are combined to form a triplet. The data is written to the replay cache in chronological order and then sampled hierarchically. Based on this experience pool, policy updates are triggered at a fixed rhythm. Advantage function estimation and policy entropy regularization are combined to improve the balance between exploration and utilization. Under multiple environmental conditions (different materials, thicknesses, clamping and heat dissipation boundaries), distribution bias is corrected by domain randomization and importance sampling, so that the policy network has higher update weights for high-value fragments in short-term memory.

[0069] When a specific state-action-result combination is detected to be strongly correlated with quality degradation, its sampling probability is automatically reduced or a safety shielding constraint is introduced, forming a dual-channel update logic of "experience absorption - risk suppression". This continuously improves the ability to suppress quality fluctuations and the robustness of parameter adjustments while maintaining the production cycle.

[0070] S54. Achieve a complete closed-loop mechanism from raw signal acquisition to decision control and result feedback, forming a "perception-reasoning-execution-correction" cycle, continuously reducing the need for manual intervention, and enhancing the system's adaptability and self-evolution capabilities under different material types, structural forms, thickness differences and environmental fluctuations.

[0071] This embodiment uses a unified timeline to connect the data and control links of "multimodal acquisition - feature encoding - quality prediction - parameter optimization - equipment execution - detection feedback - incremental learning - strategy update". Bidirectional verification points and health thresholds are set at the equipment layer, model layer and strategy layer to ensure that each closed loop produces a replayable, auditable and traceable full process trajectory.

[0072] When the production object, structural form, or environmental boundary changes, knowledge transfer is completed within several control cycles through online domain adaptation and rapid fine-tuning. Simultaneously, ineffective or overly aggressive adjustments are suppressed by relying on safety rollback and threshold hysteresis. This is achieved through the integration of statistical samples and empirical triples. Through accumulation, the welding quality prediction model and strategy network gradually form a stable evolutionary pattern of "stability first, optimization later, and learning by doing" in a wider feasible domain. Human intervention is only required when the strategy is abnormal or the equipment alarms. The overall control shifts from passive correction to active protection. The system's adaptability to material type, plate thickness difference and environmental disturbance and its noise resistance continue to expand with iteration.

[0073] This embodiment constructs a multimodal real-time sensing system to synchronously collect and fuse multi-source physical data such as mechanical signals, acoustic information, thermal field changes, and surface images generated during the welding process, comprehensively depicting the dynamic behavior and potential defect characteristics of the welding process. Based on the extracted multimodal features, a welding quality prediction model is established to identify key variables reflecting the welding thermodynamic stability and material coupling state, predicting weld tensile properties and defect risks. An embodied intelligent agent integrating sensing, prediction, and decision-making modules is constructed, combining historical data and real-time feedback to form a strategy update mechanism, achieving adaptive optimization control of feed rate, spindle speed, and pressure. Under the constraint of ensuring that the weld tensile properties are not lower than a preset threshold, reinforcement learning or gradient optimization algorithms are used to solve for the optimal parameter combination for energy consumption and quality balance, and the results are sent to the welding equipment in real time to achieve dynamic adjustment of process parameters. After welding is completed, the prediction model and control strategy are incrementally updated based on actual detection results and sensor data, forming a closed-loop system of "sensing-prediction-optimization-feedback". This method can significantly improve the intelligence of the friction stir welding process and the weld consistency level of aerospace structural components.

[0074] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Moreover, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0075] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. Since the above embodiments are substantially similar to the method embodiments, their descriptions are relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0076] The above embodiments provide a detailed description of the present invention. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A multimodal sensing and intelligent quality control method for friction stir welding of aerospace structures, characterized in that, The method includes the following steps: S1. Synchronously acquire multi-modal signals during the welding process to comprehensively characterize the dynamic behavior and potential defect features of the welding process, including multi-source physical data generated during the welding process, including mechanical information, acoustic information, thermal field changes and surface image data; S2. Input the multimodal signal into the welding quality prediction model, extract the feature variables that reflect the welding thermal stability and material coupling state, and combine them with the training network to predict and evaluate tensile properties and defect risks, and output an index that comprehensively reflects the welding quality. S3. Construct an embodied intelligent agent for welding control tasks, and automatically form a strategy update mechanism based on historical data and online feedback to achieve adaptive decision optimization of welding process parameters including feed rate, spindle speed and pressure. S4. Under the premise that the tensile properties of the weld are not lower than the set threshold, the optimal combination of welding parameters is solved by combining energy consumption and process stability, using reinforcement learning or gradient-based search algorithm, and then sent to the welding equipment in real time to realize the dynamic control of process parameters. S5. After completing a section of weld, the welding quality prediction model and the embodied intelligent agent are incrementally updated based on the actual quality inspection results and sensor process records to form a closed-loop feedback path, thereby realizing self-learning ability and quality robustness in complex manufacturing tasks.

2. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 1, characterized in that, In step S1, the specific process includes the following steps: S11. Deploy multi-dimensional force sensors in key parts of the friction stir welding equipment to obtain mechanical information on the workpiece applied by the stirring head during the welding process, including axial pressure, lateral thrust and instantaneous change rate, so as to characterize the material flow resistance and contact strength in the weld area, and serve as the basic indicators for judging the viscoplastic state and the degree of stirring in the weld area. S12. Set up a high-frequency response acoustic emission sensor to collect acoustic information of elastic sound waves excited by microcrack propagation, interface discontinuity or material internal delamination during the welding process. By performing short-time Fourier transform analysis on the original acoustic information, extract the time-frequency energy spectrum characteristics and main frequency band distribution characteristics for capturing microscopic signs before defect initiation. S13. Use an infrared thermal imager or thermocouple array to monitor the temperature of the welding area throughout the entire process, capture the thermal field changes in the temperature field distribution of the weld center and heat-affected zone, and calculate the average temperature, maximum gradient, and temperature fluctuation coefficient. This serves as the basis for evaluating the uniformity of welding heat input and the stability of the thermal cycle; S14. Real-time welding images are acquired using a high-resolution industrial camera. Visual texture analysis algorithms are used to extract image features including weld width, surface smoothness, and forming symmetry. Meanwhile, optical flow is used to track the dynamic deformation of the weld area to reflect the strain recovery behavior of the metal material in the post-weld stage. S15. Perform time synchronization processing and anomaly cleaning on the modal data streams collected in steps S11-S14 to obtain multi-source physical data.

3. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 1, characterized in that, The welding quality prediction model includes a front-end, which uses a causal convolutional temporal convolutional network to capture short-term process disturbances and local periodic noise, and a back-end. The backend uses a multi-head self-attention Transformer encoder to aggregate cross-window dependencies, enabling long-range information exchange between different welding stages; And a fusion layer is used to perform channel splicing and layer normalization on the local robust representation of the temporal convolutional network output and the global dependency representation of the Transformer encoder output, so as to suppress the bias of the single structure under sudden perturbations or trend drift. In addition, a top-level branch for the multi-task system is set up, corresponding to mechanical property regression, thermal stability estimation, and defect rate discrimination, including: The return branch outputs the expected tensile properties. And its proportional stability to the reference strength is constrained by the relative error loss; Thermally stable branch, direct regression temperature fluctuation coefficient To continue the thermal mode's characterization of energy uniformity; Defective branches, combining visual morphology and acoustic band characteristics to output defect rate The probability estimation is used to improve sensitivity to rare defects by focusing loss.

4. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 3, characterized in that, In step S2, the specific process includes the following steps: S21. Unify the encoding of multimodal signals and extract features to form a multidimensional feature vector for input to the welding quality prediction model, including: The force signal corresponding to the mechanical information is extracted using sliding window statistics. The main energy frequency band is obtained by wavelet packet decomposition of the acoustic signal corresponding to the acoustic information. Construct temperature trend and fluctuation indicators from the thermal sensing signals corresponding to changes in the thermal field. Visual parameters, including weld contour, defect area, and geometric deviation, are extracted from surface image data. S22. Utilize the welding quality prediction model to model the multidimensional feature vector and output the predicted tensile properties of the weld at the current moment. Temperature fluctuation coefficient And the defect rate inferred from image / acoustic analysis. Each indicator represents a key dimension of welding quality; S23. Construct a normalized welding quality scoring function to quantitatively evaluate whether the current welding condition meets the process standards for aerospace structural components. The expression is as follows: ； in, For comprehensive quality indicators; , where is the weighting coefficient, reflecting the intensity of the influence of each quality indicator on the overall quality score; These are standard tensile property reference values; The score of the stability factor for temperature fluctuations after normalization. S24. Assess the overall quality index of the current welding section. With set quality threshold When comparing, If the quality is deemed satisfactory, adjustment is not initiated; when... If the problem persists, proceed to the next optimization step to find feasible parameter adjustment paths.

5. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 1, characterized in that, The embodied intelligent agent includes a perception layer, a strategy layer, and a feedback update module; The perception layer is used to perform normalization, noise reduction, and temporal compression. The policy layer is used to process the state vector within a millisecond time limit. Perform forward reasoning to generate next parameter action suggestions to meet the real-time and stability requirements of the edge side; The feedback update module is used to inject the execution results returned by the device into a new round of perception-prediction information closed loop, so as to continuously correct strategy deviations and absorb new operating condition knowledge.

6. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 5, characterized in that, In step S3, the specific process includes the following steps: S31. The multi-dimensional feature vectors used as multimodal sensing data and the prediction output of the welding quality prediction model are used as a joint state description, and are uniformly encoded with historical welding behavior into a state vector. Control strategies are generated by embodied intelligent agents; S32. The policy layer of the embodied agent is trained using a proximal policy optimization algorithm, and expected actions are generated based on historical interaction experience to optimize the predicted quality Q value in the current state. ,Right now: ； in, It is the feed rate; It is the spindle speed; This refers to the downward pressure. S33. Use the prediction quality Q-value as a reward signal. The parameters and actions executed are paired and stored in the experience replay cache, and the long-term returns are calculated using a discounted reward function. ,Right now: ； in, This is a discount factor used to balance the effects of short-term and long-term control. S34. The policy update mechanism is formed by iteratively updating the network parameters of the policy layer using the gradient descent method in batches.

7. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 6, characterized in that, The historical welding behavior includes multi-dimensional feature vectors and comprehensive indicators from the most recent control cycles. Decomposition amount And the executed parameter trajectory and its rate of change.

8. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 1, characterized in that, In step S4, the specific process includes the following steps: S41. A multi-objective optimization function for the welding process is established by combining weld defect risk, energy consumption, and comprehensive indicators. The expression is: ； in, It is the feed rate; It is the spindle speed; This refers to the downward pressure. The defect risk item is output by a defect regression model learned from historical data. The energy consumption per unit length of welding is estimated based on the heat input model. This is a comprehensive quality scoring function; The target balancing weights are either manually set or learned. S42. Under the constraint conditions, the comprehensive quality scoring function is satisfied. and Under the condition that the process parameters are within a legal range, the process parameters are searched using the strategy gradient method or black-box optimization method to find a multi-objective optimization function. The minimum optimal parameter combination ; S43. Combine the optimal parameters The control parameters are sent to the friction welding equipment control system, and the welding control parameters are updated in real time through the set interface, and a smooth switching is achieved in the form of a gradually changing control curve. S44. Continuously monitor the multimodal sensing signals during the updated process. If abnormal deviations, sudden changes, or failure to achieve the expected quality improvement trend occur, immediately stop the current strategy or revert to the previous stable control parameter state to ensure production safety and process stability.

9. The multimodal sensing and intelligent quality control method for aerospace structure friction stir welding according to claim 1, characterized in that, In step S5, the specific process includes the following steps: S51. After welding, perform tensile property testing and internal defect analysis on the sample. By comparing the actual results with the prediction results output by the welding quality prediction model, evaluate the generalization accuracy and local error range of the welding quality prediction model. S52. Construct the actual error term as a loss function and backpropagate it to the model parameters of the welding quality prediction model, and use a fine-tuning strategy to update the incremental model. S53. The parameters, actions, welding feedback results, and environmental states of each execution are combined into a triplet to continuously update the policy network through reinforcement learning. This enables the policy layer of the embodied agent to absorb experience from short-term memory, realizing a complete closed-loop mechanism from raw signal acquisition to decision control and result feedback, forming a "perception-reasoning-execution-correction" cycle.

Images