A laser intelligent welding weld quality detection method and system

Abstract

This invention discloses a method and system for weld quality inspection in laser intelligent welding, relating to the fields of welding technology and quality inspection technology. The method includes the following steps: acquiring multimodal signal data from the weld surface, defining a multimodal feature extraction network, and generating an initial feature representation based on the multimodal feature extraction network and the multimodal signal data. The invention utilizes a parameter dynamic adjustment module constructed through a transfer learning framework and evolutionary algorithm, enabling efficient conversion between different materials and improving the flexibility and generalization ability of the inspection system. The application of a variational Gaussian mixture model combined with Bayesian inference allows the inspection system to extract valuable trends from the complex data distribution of the hierarchical feedback adjustment unit, dynamically adjusting core parameters and further enhancing the adaptability of the inspection system to different material properties.

Description

Technical Field

[0001] This invention relates to the field of welding technology and quality inspection technology, specifically to a method and system for inspecting the quality of laser intelligent welding welds. Background Technology

[0002] In the industrial manufacturing sector, laser intelligent welding, with its advantages of high precision and high efficiency, is widely used in industries such as automotive manufacturing and aerospace. Weld quality directly affects the safety, durability, and service life of products; therefore, weld quality inspection is of paramount importance. Traditional manual visual inspection or conventional non-destructive testing techniques, such as ultrasonic and X-ray inspection, suffer from low efficiency, high cost, reliance on experience, and difficulties in data retention, making them insufficient to fully meet the demands of modern industry for high-precision, intelligent inspection.

[0003] While various laser weld inspection technologies have been developed, such as inspection systems integrating 3D laser scanning and AI analysis, enabling millimeter-level 3D modeling, dual-mode inspection, and intelligent defect classification, practical applications still present significant challenges in adapting to different metal materials. Different metals, such as copper, aluminum, and steel, exhibit substantial differences in reflectivity and thermal properties. For example, copper's high reflectivity leads to significant laser signal reflection, weakening the effective signal received by the sensor and affecting the accurate acquisition of weld information. Aluminum, on the other hand, has high thermal conductivity, resulting in rapid heat dissipation during welding, and its temperature field and thermal stress distribution at the weld differ significantly from other metals. Existing detection thresholds and algorithms have limited adaptability to these differences, leading to decreased detection accuracy and stability, thus impacting the consistent assessment of weld quality and the achievement of high-precision inspection results for various metal materials. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a method and system for inspecting weld quality in laser intelligent welding, thereby solving the problems of insufficient adaptability to multiple materials and decreased detection accuracy and stability in existing technologies.

[0005] To achieve the above objectives, the present invention provides a method for inspecting weld quality in laser intelligent welding, comprising the following steps:

[0006] Acquire multimodal signal data of the weld surface, define a multimodal feature extraction network, and generate an initial feature representation based on the multimodal feature extraction network and the multimodal signal data;

[0007] By utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm to enhance the correlation between optical signals and thermal field signals in the initial feature representation, a multimodal sensing module is determined.

[0008] Based on the multimodal sensing module, a hierarchical feedback adjustment unit is constructed. The hierarchical feedback adjustment unit is used to process feedback signals from different levels to generate an optimized detection path.

[0009] Based on the hierarchical feedback adjustment unit, a parameter dynamic adjustment module is constructed using a transfer learning framework and an evolutionary algorithm. The core parameters are adjusted based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy.

[0010] Based on the optimized detection path and the adaptive detection strategy, an incremental feature extension module is constructed using graph embedding-based structured analysis technology. A weld quality inspection system is then constructed based on the incremental feature extension module, the multimodal feature extraction network, the multimodal perception module, the hierarchical feedback adjustment unit, and the parameter dynamic adjustment module.

[0011] Furthermore, the step of constructing a parameter dynamic adjustment module based on the hierarchical feedback adjustment unit using a transfer learning framework and evolutionary algorithm, and adjusting the core parameters based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy, includes the following steps:

[0012] Using a transfer learning framework and evolutionary algorithm, an initial parameter dynamic adjustment module is constructed based on the hierarchical feedback adjustment unit. The adaptability of the initial parameter dynamic adjustment module among different materials is optimized by differential evolution algorithm and genetic algorithm.

[0013] Based on the optimized detection path, the variational Gaussian mixture model is used to analyze the variational pattern of the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit, and the core parameters in the initial parameter dynamic adjustment module are dynamically adjusted by Bayesian inference to generate the target parameter dynamic adjustment module.

[0014] By utilizing a context-sensitive mechanism, the target parameter dynamic adjustment module adjusts the detection strategy according to the characteristics of different materials, thereby generating an adaptive detection strategy.

[0015] Furthermore, the module that uses a context-sensitive mechanism to dynamically adjust the target parameters adjusts the detection strategy according to the characteristics of different materials to generate an adaptive detection strategy includes the following steps:

[0016] By using a context-sensitive mechanism combined with graph embedding technology and attention allocation mechanism, key properties are analyzed from the characteristics of different materials in the hierarchical feedback adjustment unit. These key properties include reflectivity distribution, thermal conductivity difference, and thermal stress distribution characteristics.

[0017] Based on the key attributes, the detection threshold in the target parameter dynamic adjustment module is dynamically adjusted by combining Bayesian inference and Gaussian process modeling to generate a preliminary adaptive detection strategy.

[0018] The initial adaptive detection strategy is optimized using a time series decomposition model through reinforcement learning to generate an optimized detection strategy.

[0019] The optimal detection strategy is searched from the optimized detection strategy by combining the particle swarm optimization algorithm and the simulated annealing algorithm;

[0020] An active sampling mechanism is introduced to select target samples from the optimal detection strategy to generate an adaptive detection strategy.

[0021] Furthermore, the method of using a context-sensitive mechanism combined with graph embedding technology and attention allocation mechanism to analyze key properties from the characteristics of different materials in the hierarchical feedback adjustment unit includes the following steps:

[0022] The characteristics of different materials in the hierarchical feedback control unit are collected and cleaned using signal preprocessing technology to obtain the material characteristic data of the hierarchical feedback control unit;

[0023] A context-sensitive mechanism is introduced to perform preliminary screening and dimensionality reduction on the material property data by combining principal component analysis and independent component analysis, so as to obtain the target material property information.

[0024] By applying graph embedding technology and complex network modeling, the correlation between materials is analyzed from the target material property information to construct a material relationship graph;

[0025] An attention allocation mechanism combined with deep reinforcement learning is applied to highlight important attributes in the material relationship graph, resulting in the key attribute weight allocation results.

[0026] The key attributes are extracted by analyzing the material relationship diagram and the key attribute weight allocation results.

[0027] Furthermore, based on the optimized detection path, the variational Gaussian mixture model is used to analyze the changing patterns from the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit, and Bayesian inference is used to dynamically adjust the core parameters in the initial parameter dynamic adjustment module to generate the target parameter dynamic adjustment module, including the following steps:

[0028] Based on the optimized detection path, the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit is modeled using variational Gaussian mixture model and statistical modeling tools, and potential patterns and trends are analyzed from the feedback signals to generate data distribution change patterns.

[0029] Based on the data distribution variation pattern, the core parameters in the initial parameter dynamic adjustment module are dynamically adjusted through Bayesian inference, Gaussian process modeling, and random forest to generate the target parameter dynamic adjustment module.

[0030] Furthermore, the step of enhancing the correlation between the optical signal and the thermal field signal in the initial feature representation by utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm to determine the multimodal sensing module includes the following steps:

[0031] An initial multimodal feature extraction network is defined using a hierarchical graph embedding network, and the initial feature representation is analyzed from the intrinsic relationship between the optical signal and the thermal field signal of the initial multimodal feature extraction network.

[0032] Based on the initial feature representation, a cross-modal information fusion mechanism is introduced, combined with a bidirectional gated loop unit, to jointly encode the optical signal and thermal field signal in the initial feature representation, thereby obtaining a multimodal information representation;

[0033] Based on the multimodal information representation, a generative adversarial network is constructed, and key data is analyzed from the adversarial training process between the generator and the discriminator in the generative adversarial network.

[0034] The initial multimodal feature extraction network is optimized based on the key data and the variational Gaussian mixture model to obtain the target multimodal feature extraction network, and the multimodal perception module is determined based on the target multimodal feature extraction network.

[0035] Furthermore, the incremental feature expansion module, constructed using graph embedding-based structured analysis technology based on the optimized detection path and the adaptive detection strategy, includes the following steps:

[0036] The optimized detection path and the adaptive detection strategy are combined with time series modeling and Bayesian inference to determine the incremental update rules.

[0037] Based on the incremental update rule, the multimodal data in the multimodal perception module is modeled and processed using graph embedding-based structured analysis technology combined with hierarchical clustering algorithm to obtain a structured representation of the multimodal data.

[0038] The data obtained from the hierarchical feedback adjustment unit is encoded based on the multimodal data structured representation combined with a dynamic graph embedding network and a variational Gaussian mixture model, and the data is input into the multimodal data structured representation to output an updated feature representation;

[0039] By applying an attention allocation mechanism combined with deep reinforcement learning, key attributes in the updated feature representation are highlighted to obtain an optimized feature representation.

[0040] The parameters in the incremental update process of the parameter dynamic adjustment module are optimized and adjusted by combining reinforcement learning and particle swarm optimization algorithms to generate optimized parameter configuration;

[0041] An incremental feature expansion module is generated based on the optimized feature representation and the optimized parameter configuration.

[0042] This invention also provides a laser intelligent welding weld quality inspection system, comprising:

[0043] The acquisition module is used to acquire multimodal signal data of the weld surface and define a multimodal feature extraction network to generate an initial feature representation based on the multimodal signal data and the multimodal feature extraction network.

[0044] An enhancement module is used to enhance the correlation between the optical signal and the thermal field signal in the initial feature representation by utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm, thereby determining the multimodal sensing module;

[0045] The module is used to construct a hierarchical feedback adjustment unit based on the multimodal sensing module, and to process the feedback signals from different levels using the hierarchical feedback adjustment unit to generate an optimized detection path;

[0046] The adjustment module is used to construct a parameter dynamic adjustment module based on the hierarchical feedback adjustment unit using a transfer learning framework and an evolutionary algorithm, and to adjust the core parameters based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy.

[0047] The module is used to construct an incremental feature expansion module based on graph embedding-based structured analysis technology, based on the optimized detection path and the adaptive detection strategy. The weld quality inspection system is constructed based on the incremental feature expansion module, the multimodal feature extraction network, the multimodal perception module, the hierarchical feedback adjustment unit, and the parameter dynamic adjustment module.

[0048] Furthermore, the acquisition module includes a high-resolution optical camera and an infrared thermal imager, which are used to capture the optical reflection characteristics and thermal field distribution characteristics of the weld surface, respectively.

[0049] The present invention also provides a computing device, including a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform the method described in any one of the preceding descriptions.

[0050] Beneficial effects

[0051] 1. This invention utilizes a parameter dynamic adjustment module constructed through a transfer learning framework and evolutionary algorithm, enabling efficient conversion between different materials and improving the flexibility and generalization ability of the detection system. The application of a variational Gaussian mixture model combined with Bayesian inference allows the detection system to extract valuable trends from the complex data distribution of the hierarchical feedback adjustment unit, dynamically adjusting core parameters and further enhancing the system's adaptability to different material properties. Furthermore, the application of a context-sensitive mechanism allows the detection system to flexibly adjust its detection strategy based on specific material characteristics, generating the optimal strategy and ensuring high performance and rapid convergence in various application scenarios. In summary, this method effectively solves the problems of insufficient multi-material adaptability, static parameter configuration, and imperfect feedback mechanisms in existing solutions, providing a more intelligent and efficient weld quality inspection solution. Attached Figure Description

[0052] Figure 1 A schematic diagram of the module structure of the laser intelligent welding weld quality inspection system in this embodiment of the invention;

[0053] Figure 2 This is a schematic diagram illustrating the working principle of the multimodal sensing module in an embodiment of the present invention. Detailed Implementation

[0054] 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.

[0055] This invention provides a method and system for inspecting weld quality in laser intelligent welding, the specific implementation of which is described in conjunction with the appendix. Figure 1 and attached Figure 2 Please provide a detailed explanation.

[0056] Figure 1 The diagram illustrates the modular structure of a weld quality inspection system, including a data acquisition module, an enhancement module, a construction module, an adjustment module, and a utilization module. These modules are logically connected to implement the specific workflow of weld quality inspection. Figure 2 This diagram illustrates the working principle of the multimodal sensing module, and details the joint encoding process of optical signals and thermal field signals under the cross-modal information fusion mechanism, as well as the role of generative adversarial networks in optimizing feature representation.

[0057] In practical applications, the acquisition module first obtains multimodal signal data from the weld surface, including optical and thermal signals. The acquisition module monitors the weld surface in real time using a sensor array, which includes a high-resolution optical camera and an infrared thermal imager, to capture the optical reflection characteristics and thermal distribution characteristics of the weld surface, respectively. The acquired multimodal signal data is then transmitted to the enhancement module for processing. The acquisition module and the enhancement module are connected via a gigabit Ethernet interface (transmission rate ≥ 1Gbps, latency ≤ 10ms), and a CRC check mechanism is used to ensure that the signal data is transmitted without loss or errors, guaranteeing real-time performance and integrity.

[0058] The enhancement module processes the acquired multimodal signal data based on a multimodal feature extraction network to generate an initial feature representation. The multimodal feature extraction network employs a hierarchical design: a convolutional neural network (CNN) with three convolutional layers (kernel sizes of 3×3, 5×5, and 3×3, all with a stride of 1) and two pooling layers (max pooling, 2×2) are used to extract spatial features of the optical signal (such as crack edges and pore morphology); an autoencoder with an input layer (dimension = number of thermal field signal sampling points × number of time steps), three hidden layers (dimensions of 512, 256, and 128, respectively), and an output layer are used to extract temporal series features of the thermal field signal (such as the rate of change of temperature gradient). The enhancement module further introduces a cross-modal information fusion mechanism, combining a bidirectional gated recurrent unit to jointly encode the optical and thermal field signals in the initial feature representation, obtaining a multimodal information representation. This cross-modal information fusion mechanism highlights key attributes through an attention allocation mechanism, ensuring that the correlation between the optical and thermal field signals is fully explored. Attention weight calculation (based on scaled dot product attention): , ,in, For query vectors (such as optical signal features). For example, the key vector (such as the characteristics of thermal field signals). The dimension of the key vector. Let be the attention weight of the i-th query on the j-th key.

[0059] Subsequently, a generative adversarial network (GAN) is introduced to optimize the initial multimodal feature extraction network, generate the target multimodal feature extraction network, and determine the multimodal perception module. In the GAN, the generator is responsible for generating latent feature representations, while the discriminator evaluates the generated feature representations, continuously optimizing the quality of the feature representations through an adversarial training process.

[0060] The module constructs a hierarchical feedback adjustment unit based on the multimodal perception module. This unit processes feedback signals from different levels to generate an optimized detection path. The hierarchical feedback adjustment unit employs a three-layer tree structure: the first layer (root node) receives the aggregated data of the original multimodal signals; the second layer (6 child nodes) corresponds to optical reflection intensity, optical texture distribution, peak thermal temperature, thermal diffusion rate, stress concentration factor, and stress distribution uniformity, respectively; the third layer (12 child nodes) represents the subdivisions of each indicator in the second layer (e.g., thermal diffusion rate is subdivided into lateral diffusion rate and vertical diffusion rate). Each node aggregates lower-layer signals using a weighted average algorithm (weights are dynamically allocated based on feature importance) to generate upper-layer feedback results. The hierarchical feedback adjustment unit analyzes the data distribution of the feedback signals using statistical modeling tools, identifies potential patterns and trends, and generates an optimized detection path based on these patterns and trends. The lower bound of evidence (ELBO) objective function for the variational Gaussian mixture model is: Where X is the observed data (feedback signal) and Z is the latent variable (cluster label). These are the model parameters (mean, covariance). Given a variational distribution, the objective is to maximize... To approximate the true posterior The building module and the adjustment module are connected via a data bus to ensure that the output of the hierarchical feedback adjustment unit can be transmitted to the parameter dynamic adjustment module in real time.

[0061] The adjustment module, based on a hierarchical feedback adjustment unit, utilizes a transfer learning framework and evolutionary algorithms to construct a dynamic parameter adjustment module. This module, combined with the optimized detection path, adjusts core parameters to generate an adaptive detection strategy. The construction process of the dynamic parameter adjustment module consists of three steps.

[0062] The first step involves constructing a module for dynamically adjusting initial parameters based on a hierarchical feedback adjustment unit using a transfer learning framework and evolutionary algorithms. Differential evolutionary algorithm and genetic algorithm are used to optimize the module's adaptability across different materials. The mutation operation formula for the differential evolutionary algorithm is: ,in, For the i-th mutant individual in the t-th generation, Three distinct individuals are randomly selected, and F∈[0,2] is the scaling factor (here, F=0.8). Crossover operation formula: Where CR∈[0,1] is the crossover probability (here CR=0.6). For random dimensions.

[0063] The second step involves using a variational Gaussian mixture model to extract the variational patterns from the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit, based on the optimized detection path. Then, the core parameters in the dynamic adjustment module are dynamically adjusted by Bayesian inference, thereby generating the target parameter dynamic adjustment module.

[0064] The third step utilizes a context-sensitive mechanism to dynamically adjust the target parameter adjustment module based on the characteristics of different materials, generating an adaptive detection strategy. This context-sensitive mechanism, combined with graph embedding technology and attention allocation, analyzes key attributes from the characteristics of different materials in the hierarchical feedback adjustment unit, including reflectivity distribution, thermal conductivity differences, and thermal stress distribution characteristics. These key attributes are used to dynamically adjust the detection threshold through Bayesian inference and Gaussian process modeling, generating a preliminary adaptive detection strategy. This preliminary adaptive detection strategy is further optimized using a time-series decomposition model and reinforcement learning, and finally, a particle swarm optimization algorithm and simulated annealing algorithm are combined to search for the optimal detection strategy. The velocity and position update formulas for the particle swarm optimization algorithm are as follows:

[0065]

[0066] Where w is the inertia weight (here w=0.729), and c_1=c_2=1.494 are the acceleration coefficients. It is a random number. For the individual's optimal position, The position is the global optimum, and k is the number of iterations.

[0067] An incremental feature expansion module is constructed using graph embedding-based structured analysis techniques based on an optimized detection path and adaptive detection strategy. The construction process of the incremental feature expansion module includes the following steps: First, using the optimized detection path and adaptive detection strategy, combined with time series modeling (using an ARIMA model, order p=3, q=2) and Bayesian inference, an incremental update rule is determined: incremental updates are triggered when the distribution difference (KL divergence) between newly acquired data and historical data is ≥0.1, with an update frequency of once every 50ms. Each update retains only the core features of the first 1000 historical data points (filtered using L1 regularization) to reduce redundancy. Second, based on the incremental update rule, graph embedding-based structured analysis techniques combined with a hierarchical clustering algorithm are used to model and process the multimodal data in the multimodal perception module, obtaining a structured representation of the multimodal data. Third, based on the structured representation of the multimodal data, a dynamic graph embedding network and a variational Gaussian mixture model are used to encode the data obtained from the hierarchical feedback adjustment unit, and the data is input into the structured representation of the multimodal data to output the updated feature representation. Subsequently, an attention allocation mechanism combined with deep reinforcement learning is applied to highlight key attributes in the updated feature representation, resulting in an optimized feature representation. Finally, reinforcement learning and particle swarm optimization algorithms are used to optimize the parameters in the incremental update process of the parameter dynamic adjustment module, generating an optimized parameter configuration. Based on the optimized feature representation and the optimized parameter configuration, an incremental feature expansion module is generated.

[0068] In practical applications, the collaborative workflow of the above modules is as follows: After the acquisition module acquires the multimodal signal data of the weld surface, the data is first transmitted to the enhancement module for feature extraction and cross-modal information fusion processing to generate a multimodal perception module. Subsequently, the construction module builds a hierarchical feedback adjustment unit based on the multimodal perception module, generates an optimized detection path, and passes it to the adjustment module. The adjustment module uses a transfer learning framework and evolutionary algorithm to build a parameter dynamic adjustment module, and generates an adaptive detection strategy in conjunction with the optimized detection path. Finally, the module uses the optimized detection path and adaptive detection strategy to build an incremental feature expansion module, completing the construction of the entire weld quality inspection system.

[0069] In terms of hardware implementation, this system is executed by a computing device, which includes a processor and a memory. The memory stores the computer program, and the processor runs the computer program to execute the aforementioned weld quality inspection method. The computing device is connected to the acquisition module via a communication interface, receiving multimodal signal data from the sensor array and transmitting the data to various functional modules for processing via an internal bus. The memory also stores key algorithm modules such as pre-trained multimodal feature extraction network models, generative adversarial network models, and variational Gaussian mixture models, ensuring that the system can operate efficiently under different material conditions.

[0070] In this embodiment, the connections and positions between modules are rationally designed to ensure efficient signal data transmission and processing accuracy. For example, a high-speed data transmission interface is used between the acquisition module and the enhancement module to avoid data loss and latency issues. The enhancement module and the construction module are connected via a dedicated data channel to ensure that the output of the multimodal sensing module can be quickly transmitted to the hierarchical feedback adjustment unit. The adjustment module and the utilization module are connected via a dual-channel data bus, supporting bidirectional transmission of optimized detection paths and adaptive detection strategies. Furthermore, the coordination between modules is optimized to ensure the system maintains high performance in complex environments. For example, the cross-modal information fusion mechanism in the enhancement module works in conjunction with the generative adversarial network to ensure effective mining of deep correlations in multimodal signal data. The hierarchical feedback adjustment unit in the construction module works closely with the parameter dynamic adjustment module in the adjustment module to ensure that the dynamic adjustment of the detection path and core parameters can respond to environmental changes in real time.

[0071] To enable those skilled in the art to fully understand and implement this invention, the specific implementation principles of this invention will be further explained below in conjunction with specific application scenarios.

[0072] In real-world industrial manufacturing scenarios, taking the welding of aluminum alloy vehicle frames in the automotive industry as an example, weld quality inspection systems are applied to real-time monitoring on the production line. After a welding robot completes the laser intelligent welding of a section of aluminum alloy vehicle frame, defects such as micro-cracks, porosity, or lack of fusion may exist on the weld surface. At this point, the weld quality inspection system is activated, and it gradually completes operations such as multi-modal signal acquisition, feature extraction and fusion, dynamic path optimization, and adaptive strategy generation according to a preset process.

[0073] First, the acquisition module monitors the weld surface in real time using a sensor array. A high-resolution optical camera captures the optical reflection characteristics of the weld surface, while an infrared thermal imager records the thermal field distribution characteristics of the weld area. Due to the high thermal conductivity of aluminum alloy, heat dissipates rapidly during welding, resulting in a significant difference in temperature field distribution in the weld area compared to other metal materials. Therefore, the acquired thermal field signals need to be processed in conjunction with the optical signals to comprehensively reflect the weld's condition. The acquisition module transmits these multimodal signal data to the enhancement module in real time via a high-speed data transmission interface, ensuring data integrity and timeliness.

[0074] Subsequently, the enhancement module processes the acquired data based on a multimodal feature extraction network. A convolutional neural network extracts spatial features of the weld surface from the optical signal, such as the geometry and distribution of cracks; an autoencoder extracts time-series features from the thermal field signal, such as the temperature change trend in the weld area. A cross-modal information fusion mechanism jointly encodes these two signals using bidirectional gated recurrent units to highlight key attributes. For example, for the common thermal stress concentration phenomenon in aluminum alloy welding, the cross-modal information fusion mechanism uses an attention allocation mechanism to prioritize abnormal temperature gradient regions in the thermal field signal and associates them with surface crack features in the optical signal. A generative adversarial network further optimizes this process. The generator is responsible for generating latent feature representations, which are evaluated by the discriminator. Through adversarial training, the quality of the feature representations is continuously improved, ultimately determining the multimodal perception module.

[0075] The module constructs a hierarchical feedback adjustment unit based on the multimodal sensing module. This hierarchical feedback adjustment unit adopts a tree-like structure, with each node corresponding to a different feedback signal source. For example, the first-layer node receives the optical reflection characteristics of the weld surface, the second-layer node analyzes the thermal field distribution characteristics, and the third-layer node focuses on the stress distribution characteristics within the material. Statistical modeling tools analyze the data distribution of the feedback signals to identify potential patterns and trends. For instance, by analyzing the thermal field signal in the aluminum alloy weld area, a significant temperature gradient change was found during the welding cooling process, which may be related to the residual stress distribution within the weld. Based on this analysis, the module generates an optimized detection path to ensure that subsequent detection can cover the critical areas of the weld.

[0076] The adjustment module is built upon a hierarchical feedback adjustment unit to dynamically adjust parameters. First, a dynamic parameter adjustment module is constructed using a transfer learning framework and evolutionary algorithms. Differential evolution and genetic algorithms are used to optimize the module's adaptability across different materials. For example, in the aluminum alloy welding scenario, the initial parameter dynamic adjustment module adjusts the weight coefficients of the thermal field signal based on the high thermal conductivity of aluminum alloy, giving it a more significant role in the detection process. Next, based on the optimized detection path, a variational Gaussian mixture model is used to extract the variation patterns from the data distribution of feedback signals at different levels within the hierarchical feedback adjustment unit. For instance, core parameters are dynamically adjusted through Bayesian inference, enabling the detection system to adapt to the complex thermal field distribution in the aluminum alloy weld area. A context-sensitive mechanism, combined with graph embedding technology and attention allocation mechanism, further analyzes the reflectivity distribution, thermal conductivity differences, and thermal stress distribution characteristics of the aluminum alloy, generating a preliminary adaptive detection strategy. Subsequently, this strategy is optimized using a time series decomposition model and reinforcement learning, and finally, a particle swarm optimization algorithm and simulated annealing algorithm are combined to search for the optimal detection strategy.

[0077] An incremental feature extension module is constructed based on the optimized detection path and adaptive detection strategy. First, time series modeling and Bayesian inference are used to determine the incremental update rules. For example, a dynamically updated time window is set for the thermal field signal in the aluminum alloy weld area to ensure the system can capture the temperature field change trend in real time. Second, structured analysis techniques based on graph embedding combined with hierarchical clustering algorithms are used to model and process the multimodal data in the multimodal perception module, resulting in a structured representation of the multimodal data. For example, surface crack features in the optical signal are associated with temperature gradient anomaly regions in the thermal field signal to form a unified feature representation. Third, a dynamic graph embedding network and a variational Gaussian mixture model encode the data obtained from the hierarchical feedback adjustment unit, and the encoded data is input into the structured representation of the multimodal data, outputting an updated feature representation. Subsequently, an attention allocation mechanism combined with deep reinforcement learning is applied to highlight key attributes in the updated feature representation, such as prioritizing features related to weld cracks. Finally, reinforcement learning and particle swarm optimization algorithms are combined to optimize the parameters in the incremental update process of the parameter dynamic adjustment module, generate the optimized parameter configuration, and generate an incremental feature expansion module based on the optimized feature representation and the optimized parameter configuration.

[0078] Through the above steps, the weld quality inspection system achieves efficient inspection of welds on aluminum alloy vehicle frames. During the inspection process, the system can accurately identify minute cracks on the weld surface and, in conjunction with thermal field signals, determine whether these cracks are accompanied by internal thermal stress concentration. Furthermore, the system can dynamically adjust the inspection strategy based on the material properties of the aluminum alloy, ensuring high accuracy and stability under various welding conditions. Finally, the entire inspection process is completed with the support of computing equipment; the processor runs the computer program in memory, executing the above methods to ensure the system can operate efficiently in complex industrial environments.

[0079] 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.

[0080] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for inspecting weld quality in laser intelligent welding, characterized in that, Includes the following steps: Acquire multimodal signal data of the weld surface, define a multimodal feature extraction network, and generate an initial feature representation based on the multimodal feature extraction network and the multimodal signal data; By utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm to enhance the correlation between optical signals and thermal field signals in the initial feature representation, a multimodal sensing module is determined. Based on the multimodal sensing module, a hierarchical feedback adjustment unit is constructed. The hierarchical feedback adjustment unit is used to process feedback signals from different levels to generate an optimized detection path. Based on the hierarchical feedback adjustment unit, a parameter dynamic adjustment module is constructed using a transfer learning framework and an evolutionary algorithm. The core parameters are adjusted based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy. Based on the optimized detection path and the adaptive detection strategy, an incremental feature expansion module is constructed using graph embedding-based structured analysis technology. A weld quality inspection system is then constructed based on the incremental feature expansion module, the multimodal feature extraction network, the multimodal perception module, the hierarchical feedback adjustment unit, and the parameter dynamic adjustment module. The process of constructing a parameter dynamic adjustment module based on the hierarchical feedback adjustment unit using a transfer learning framework and evolutionary algorithm, and adjusting core parameters based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy includes the following steps: Using a transfer learning framework and evolutionary algorithm, an initial parameter dynamic adjustment module is constructed based on the hierarchical feedback adjustment unit. The adaptability of the initial parameter dynamic adjustment module among different materials is optimized by differential evolution algorithm and genetic algorithm. Based on the optimized detection path, the variational Gaussian mixture model is used to analyze the variational pattern of the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit, and the core parameters in the initial parameter dynamic adjustment module are dynamically adjusted by Bayesian inference to generate the target parameter dynamic adjustment module. The target parameter dynamic adjustment module utilizes a context-sensitive mechanism to adjust the detection strategy according to the characteristics of different materials, generating an adaptive detection strategy. This specifically includes the following steps: By using a context-sensitive mechanism combined with graph embedding technology and attention allocation mechanism, key properties are analyzed from the characteristics of different materials in the hierarchical feedback adjustment unit. These key properties include reflectivity distribution, thermal conductivity difference, and thermal stress distribution characteristics. Based on the key attributes, the detection threshold in the target parameter dynamic adjustment module is dynamically adjusted by combining Bayesian inference and Gaussian process modeling to generate a preliminary adaptive detection strategy. The initial adaptive detection strategy is optimized using a time series decomposition model through reinforcement learning to generate an optimized detection strategy. The optimal detection strategy is searched from the optimized detection strategy by combining the particle swarm optimization algorithm and the simulated annealing algorithm; An active sampling mechanism is introduced to select target samples from the optimal detection strategy to generate an adaptive detection strategy; The process of enhancing the correlation between optical signals and thermal field signals in the initial feature representation by utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm to determine the multimodal sensing module includes the following steps: An initial multimodal feature extraction network is defined using a hierarchical graph embedding network, and the initial feature representation is analyzed from the intrinsic relationship between the optical signal and the thermal field signal of the initial multimodal feature extraction network. Based on the initial feature representation, a cross-modal information fusion mechanism is introduced, combined with a bidirectional gated loop unit, to jointly encode the optical signal and thermal field signal in the initial feature representation, thereby obtaining a multimodal information representation; Based on the multimodal information representation, a generative adversarial network is constructed, and key data is analyzed from the adversarial training process between the generator and the discriminator in the generative adversarial network. The initial multimodal feature extraction network is optimized based on the key data and the variational Gaussian mixture model to obtain the target multimodal feature extraction network, and the multimodal perception module is determined based on the target multimodal feature extraction network. The incremental feature expansion module, constructed based on the optimized detection path and the adaptive detection strategy using graph embedding-based structured analysis technology, includes the following steps: The optimized detection path and the adaptive detection strategy are combined with time series modeling and Bayesian inference to determine the incremental update rules. Based on the incremental update rule, the multimodal data in the multimodal perception module is modeled and processed using graph embedding-based structured analysis technology combined with hierarchical clustering algorithm to obtain a structured representation of the multimodal data. The data obtained from the hierarchical feedback adjustment unit is encoded based on the multimodal data structured representation combined with a dynamic graph embedding network and a variational Gaussian mixture model, and the data is input into the multimodal data structured representation to output an updated feature representation; By applying an attention allocation mechanism combined with deep reinforcement learning, key attributes in the updated feature representation are highlighted to obtain an optimized feature representation. The parameters in the incremental update process of the parameter dynamic adjustment module are optimized and adjusted by combining reinforcement learning and particle swarm optimization algorithms to generate optimized parameter configuration; An incremental feature expansion module is generated based on the optimized feature representation and the optimized parameter configuration.

2. The method for inspecting weld quality in laser intelligent welding according to claim 1, characterized in that, The method of using a context-sensitive mechanism combined with graph embedding technology and attention allocation mechanism to analyze the key properties of different materials in the hierarchical feedback adjustment unit includes the following steps: The characteristics of different materials in the hierarchical feedback control unit are collected and cleaned using signal preprocessing technology to obtain the material characteristic data of the hierarchical feedback control unit; A context-sensitive mechanism is introduced to perform preliminary screening and dimensionality reduction on the material property data by combining principal component analysis and independent component analysis, so as to obtain the target material property information. By applying graph embedding technology and complex network modeling, the correlation between materials is analyzed from the target material property information to construct a material relationship graph; An attention allocation mechanism combined with deep reinforcement learning is applied to highlight important attributes in the material relationship graph, resulting in the key attribute weight allocation results. The key attributes are extracted by analyzing the material relationship diagram and the key attribute weight allocation results.

3. The method for inspecting weld quality in laser intelligent welding according to claim 1, characterized in that, Based on the optimized detection path, the variational Gaussian mixture model is used to analyze the variational patterns of the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit, and Bayesian inference is used to dynamically adjust the core parameters in the initial parameter dynamic adjustment module to generate the target parameter dynamic adjustment module, including the following steps: Based on the optimized detection path, the data distribution of feedback signals at different levels in the hierarchical feedback adjustment unit is modeled using variational Gaussian mixture model and statistical modeling tools, and potential patterns and trends are analyzed from the feedback signals to generate data distribution change patterns. Based on the data distribution variation pattern, the core parameters in the initial parameter dynamic adjustment module are dynamically adjusted through Bayesian inference, Gaussian process modeling, and random forest to generate the target parameter dynamic adjustment module.

4. A laser intelligent welding weld quality inspection system, used to execute the laser intelligent welding weld quality inspection method according to any one of claims 1-3, characterized in that, include: The acquisition module is used to acquire multimodal signal data of the weld surface and define a multimodal feature extraction network to generate an initial feature representation based on the multimodal signal data and the multimodal feature extraction network. An enhancement module is used to enhance the correlation between the optical signal and the thermal field signal in the initial feature representation by utilizing a cross-modal information fusion mechanism and an adaptive compensation algorithm, thereby determining the multimodal sensing module; The module is used to construct a hierarchical feedback adjustment unit based on the multimodal sensing module, and to process the feedback signals from different levels using the hierarchical feedback adjustment unit to generate an optimized detection path; The adjustment module is used to construct a parameter dynamic adjustment module based on the hierarchical feedback adjustment unit using a transfer learning framework and an evolutionary algorithm, and to adjust the core parameters based on the parameter dynamic adjustment module and the optimized detection path to generate an adaptive detection strategy. The module is used to construct an incremental feature expansion module based on graph embedding-based structured analysis technology, based on the optimized detection path and the adaptive detection strategy. The weld quality inspection system is constructed based on the incremental feature expansion module, the multimodal feature extraction network, the multimodal perception module, the hierarchical feedback adjustment unit, and the parameter dynamic adjustment module.

5. The laser intelligent welding weld quality inspection system according to claim 4, characterized in that, The acquisition module includes a high-resolution optical camera and an infrared thermal imager, which are used to capture the optical reflection characteristics and thermal field distribution characteristics of the weld surface, respectively.

6. A computing device, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program, and the processor is configured to run the computer program to perform a laser intelligent welding weld quality inspection method according to any one of claims 1 to 3.

Images