A high-adaptability laser welding parameter adjustment method, device, medium and equipment

Abstract

The application belongs to the field of industrial control automation, and specifically discloses a high-adaptability laser welding parameter adjustment method, device, medium and equipment. The method comprises the following steps: real-time monitoring a laser welding process and collecting welding data, wherein the welding data comprises a welding speed, a welding power and a welding focus shifting distance; pre-processing the welding data; inputting the pre-processed welding data into a trained welding quality monitoring model for analysis, and automatically adjusting laser welding parameters according to an output result of the welding quality monitoring model. The application can improve welding quality and welding efficiency.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of industrial control automation, and specifically relates to a high-adaptability laser welding parameter adjustment method, device, medium and equipment. BACKGROUND

[0002] In modern industrial manufacturing, laser welding is widely used as an efficient and precise connection technology. However, traditional laser welding programs usually rely on pre-set fixed parameters and path planning. Although this mode can meet the needs in specific application scenarios, its adaptability and flexibility are particularly insufficient when facing different workpieces, materials and changing environmental conditions.

[0003] Specifically, existing laser welding programs often need to be customized for each specific workpiece and material, including but not limited to welding speed, power size, focal length adjustment and other key parameters. Once these parameters are determined, they are difficult to adjust in real time during actual operation, resulting in low efficiency and unstable welding quality when dealing with multiple materials or complex-shaped workpieces. In addition, traditional methods require operators to have high levels of professional knowledge to manually adjust or optimize based on experience, which not only increases labor costs but also makes it difficult to ensure consistency and stability of each welding.

[0004] Therefore, how to improve the intelligent level of the laser welding system so that it can automatically identify different welding tasks and adaptively adjust related parameters to achieve an efficient and stable welding process has become an important direction of current research. SUMMARY

[0005] In view of the deficiencies in the prior art, the purpose of the present application is to provide a high-adaptability laser welding parameter adjustment method, device, medium and equipment, which aims to overcome the limitations of traditional laser welding technology by monitoring and dynamically adjusting welding parameters in real time, thereby significantly improving welding quality and efficiency.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] A high-adaptability laser welding parameter adjustment method, the method comprising: monitoring a laser welding process in real time and collecting welding data, wherein the welding data includes welding speed, welding power and welding focus shifting distance; preprocessing the welding data; inputting the preprocessed welding data into a trained welding quality monitoring model for analysis, and dynamically adjusting laser welding parameters according to the output results of the welding quality monitoring model.

[0008] Optionally, the welding data is pre-processed, including: cleaning the welding data; dimension reduction on the cleaned welding data; and normalization processing on the dimension-reduced welding data.

[0009] Optionally, the welding quality monitoring model comprises: an input layer, a multi-channel feature extraction module, a time sequence feature fusion module, a dynamic attention mechanism, an adaptive decision module, and an output layer, wherein the input layer is configured to input the pre-processed welding data; the multi-channel feature extraction module is configured to extract multi-dimensional features in the pre-processed welding data; the time sequence feature fusion module is configured to fuse the output features of the multi-channel feature extraction module to obtain a fusion feature vector; the dynamic attention mechanism is configured to dynamically adjust and optimize feature weight distribution; the adaptive decision module is configured to predict welding quality and dynamically adjust welding parameters based on the welding quality; and the output layer is configured to output a welding quality score and an adjustment instruction of the welding parameters.

[0010] Optionally, the multi-channel feature extraction module adopts a three-channel parallel architecture, comprising a 1D-CNN time sequence channel, a Bi-LSTM time sequence channel, and a statistical feature channel.

[0011] Optionally, the time sequence feature fusion module adopts a cross-attention guided feature enhancement architecture, configured to fuse the outputs of the 1D-CNN time sequence channel, the Bi-LSTM time sequence channel, and the statistical feature channel.

[0012] Optionally, the adaptive decision module adopts a double-branch output structure, comprising: a quality prediction branch and a parameter adjustment branch, wherein the quality prediction branch is configured to score the welding quality based on welding feature data; and the parameter adjustment branch is configured to adjust the welding parameters based on the welding quality score and generate a parameter adjustment vector.

[0013] The application also provides a high-adaptability laser welding parameter adjustment device, comprising: a monitoring and collecting module configured to monitor a laser welding process in real time and collect welding data, the welding data comprising welding speed, welding power, and welding focus shifting distance; a pre-processing module configured to pre-process the welding data; and an adjustment module configured to input the pre-processed welding data into a trained welding quality monitoring model for analysis, and automatically adjust laser welding parameters according to the output result of the welding quality monitoring model.

[0014] Optionally, the pre-processing module comprises: a cleaning sub-module configured to clean the welding data; a dimension reduction sub-module configured to perform dimension reduction on the cleaned welding data; and a normalization sub-module configured to perform normalization processing on the dimension-reduced welding data.

[0015] The application also provides a storage medium comprising instructions which, when executed on a computer, cause the computer to perform the method of any one of the preceding.

[0016] The application also provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of the preceding when executing the program.

[0017] Compared with the prior art, the application has the beneficial effects of:

[0018] The technical solution can realize efficient welding process optimization by constructing a welding quality monitoring model, scoring the welding quality and adjusting the welding parameters in real time during the welding process, thereby improving the welding quality and welding efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0019] Figure 1 is a flowchart of a high-adaptability laser welding parameter adjustment method provided by an embodiment of the application;

[0020] Figure 2 is a structural diagram of a welding quality monitoring model provided by another embodiment of the application;

[0021] Figure 3 is a structural diagram of a high-adaptability laser welding parameter adjustment device provided by another embodiment of the application. DETAILED DESCRIPTION

[0022] The specific embodiments of the application will be described in detail below with reference to the accompanying drawings. Although specific embodiments of the application are shown in the drawings, it should be understood that the application can be implemented in various forms and should not be limited by the embodiments described herein. On the contrary, these embodiments are provided to enable a more thorough understanding of the application and to fully convey the scope of the application to those skilled in the art.

[0023] It should be noted that some terms are used in the specification and claims to refer to specific components. Those skilled in the art should understand that the same component can be referred to by different terms. The specification and claims of this specification do not distinguish components based on differences in terms, but rather on differences in function. As mentioned throughout the specification and claims, "including" or "comprising" is an open term, and should be interpreted as "including but not limited to". The subsequent description in the specification is a preferred embodiment of implementing the application, and is for the purpose of illustrating the general principles of the specification, and not to limit the scope of the application. The scope of protection of the application is defined by the appended claims.

[0024] For the convenience of understanding the embodiments of the present application, the following will be further explained and described with specific embodiments as examples in conjunction with the accompanying drawings, and each of the accompanying drawings does not constitute a limitation to the embodiments of the present application.

[0025] In one example embodiment, as shown in Figure 1 The present application provides a high-adaptability laser welding parameter adjustment method, which comprises the following steps:

[0026] S100: Real-time monitoring of the laser welding process and collecting welding data, wherein the welding data comprises welding speed, welding power and welding defocusing distance;

[0027] S200: Preprocessing of the welding data;

[0028] S300: Inputting the preprocessed welding data into a trained welding quality monitoring model for analysis, and automatically adjusting the laser welding parameters according to the output results of the welding quality monitoring model.

[0029] In another example embodiment, in step S200, the preprocessing of the welding data comprises the following steps:

[0030] S201: Cleaning the welding data by principal component analysis (PCA);

[0031] In this step, the welding data is cleaned by principal component analysis, which can remove noise and useless information caused by measurement errors or system interference, thereby effectively reducing the data dimension while retaining the most representative features.

[0032] S202: Dimensionality reduction of the cleaned welding data;

[0033] In this step, the cleaned welding data can be reduced in dimension by using feature selection techniques, i.e. according to the correlation or importance between each feature and the target variable, the most relevant features are selected for retention, and redundant or irrelevant features are removed. In addition, linear discriminant analysis (LDA) can also be selected, which projects the cleaned welding data into a low-dimensional space based on class labels to maximize the inter-class distance and minimize the intra-class distance, thereby improving the classification effect.

[0034] S203: Normalization of the dimensionality-reduced welding data.

[0035] In this step, z-score standardization can be used for normalization, i.e. the value of each feature is subtracted from the mean of the feature and then divided by its standard deviation, which converts the data to a distribution with a mean of 0 and a standard deviation of 1. By normalization, the welding data can be converted to a unified scale, thereby eliminating the influence of dimensional differences between different features on the model.

[0036] In another exemplary embodiment, in step S300, as Figure 2 As shown, the welding quality monitoring model includes an input layer, a multi-channel feature extraction module, a temporal feature fusion module, a dynamic attention mechanism, an adaptive decision module, and an output layer. The input layer is used to input preprocessed welding data. The multi-channel feature extraction module is used to extract multi-dimensional features from the preprocessed welding data. The multi-channel feature extraction module adopts a three-channel parallel architecture to collaboratively process three parameters: welding rate, welding power, and welding focus shift distance. Specifically, the three channels include a 1D-CNN temporal channel, a Bi-LSTM temporal channel, and a statistical feature channel. The 1D-CNN temporal channel deploys three sets of convolutional kernels of different scales (e.g., 3×3, 5×5, 7×7) to capture instantaneous fluctuation features during the welding process (such as power pulses and focus shift abrupt changes). Each set of convolutional kernels contains 64 units, which are activated by ReLU and output as feature maps. These feature maps are then compressed using max pooling to obtain key feature vectors (output dimension 192). The Bi-LSTM temporal channel employs a bidirectional LSTM structure (128 units each for forward and backward directions) to deeply model the long-range dependencies of welding parameters (e.g., the impact of welding rate variations on subsequent weld quality) and outputs a 256-dimensional state vector to preserve the complete temporal context. The statistical feature channel calculates engineering statistics in real-time within a 50ms sliding window, including dynamic mean (for monitoring power trends), variance (for detecting rate stability), and linear fitting slope (for tracking the rate of change in focus distance), and encodes these into structured features through a fully connected layer.

[0037] In summary, the multi-channel feature extraction module can comprehensively capture the instantaneous changes, time-dependent characteristics and overall trends in the welding process through multi-channel parallel processing, which is conducive to improving the accuracy and diversity of feature extraction, thereby providing strong support for the prediction of welding quality and the adjustment of welding parameters. The time sequence feature fusion module is used to fuse the output features of the multi-channel feature extraction module to obtain a fusion feature vector. The time sequence feature fusion module adopts a cross-attention guided feature enhancement architecture to effectively fuse multi-source time sequence features. First, the module inputs the local transient features (192 dimensions) from the 1D-CNN channel output as the key-value vector (Key-Value) and the global time sequence features (256 dimensions) from the Bi-LSTM channel output as the query vector (Query) into the multi-head cross-attention layer (including 4 heads, each head has a dimension of 120), and dynamically filters the most relevant local features with the current welding stage through the attention weight; then, the 32-dimensional engineering parameters of the statistical feature channel are spliced with the output of the multi-head cross-attention layer by using the gated residual mechanism, and the contribution of each feature in the final output is adjusted by using the learnable gating weight; finally, the feature dimension is compressed by the time sequence compression unit (including two layers of dilated causal convolution, the dilation rates are 2 and 4 respectively, the convolution kernel size is 3, and the output channel is 128) while preserving the key time-dependent characteristics, and finally outputs a 256-dimensional spatio-temporal enhanced feature vector.

[0038] The dynamic attention mechanism is used to dynamically adjust and optimize the feature weight distribution. The dynamic attention mechanism adopts a two-level attention structure, specifically including feature-level attention and time-level attention. The feature-level attention focuses on the feature weight distribution of the three channels. Specifically, since the welding process is a complex and time-varying process, the influence of feature information at different stages on the welding quality may be different. The feature-level attention dynamically adjusts the weight of each channel, so that the model can pay more attention to the features of a specific channel according to the actual needs in different welding stages. For example, in the arc striking stage, the welding speed changes greatly, which can easily cause the molten pool to be unstable or the weld to be uneven. In this case, the change of welding speed has a greater impact on the welding quality than the change of power. At this time, the dynamic attention mechanism enhances the feature weight related to the welding speed and reduces the attention to the power features, so that the model can more sensitively capture the welding speed fluctuations and adjust the welding speed in time to stabilize the molten pool, thereby ensuring the quality of the welded joint.

[0039] The time-level attention mechanism dynamically adjusts the weights of historical features based on different stages of the welding process. Throughout the welding process, the influence of historical data may change over time. Through this mechanism, the model can dynamically determine how to utilize previous time-series data based on the current welding stage's state and characteristics. For example, in the initial stage of laser welding, the process is extremely sensitive to changes in temperature, rate, and power. At this stage, the model relies more heavily on historical information, particularly the trend of welding rate changes, to predict the stability of the molten pool and the formation of the weld joint. In this phase, historical data helps the model identify the potential impact of welding rate fluctuations on joint quality, allowing for timely adjustments. In the stable stage, the welding process gradually becomes more stable. At this point, more attention is paid to current welding parameters (such as real-time power control), and the reliance on historical data is relatively low because welding rate and power changes are more stable and their impact on quality is minimal. Therefore, the model dynamically adjusts the weights of historical information according to the needs of different stages, thereby optimizing welding quality.

[0040] In summary, by introducing a dynamic attention mechanism, the model can flexibly respond to the needs of different stages in the welding process, thereby helping to improve welding quality and process adaptability.

[0041] The adaptive decision module is used to predict welding quality and dynamically adjust welding parameters based on the welding quality. The adaptive decision module adopts a dual-branch output structure, specifically including a quality prediction branch and a parameter adjustment branch. The quality prediction branch scores the welding quality based on welding feature data. Specifically, the welding quality can be represented by continuous real numbers, where a high score corresponds to a low welding quality. The welding quality score can be calculated using the following formula:

[0042]

[0043] in, Indicates the welding quality score; Indicates at time The collected welding feature vectors; Represents the weight matrix; This indicates the bias term.

[0044] The parameter adjustment branch adjusts welding parameters based on welding quality scores and ultimately generates a parameter adjustment vector, which indicates how to adjust each welding parameter to optimize welding quality. Furthermore, the adaptive decision module incorporates a feedback control mechanism. Through this mechanism, the parameter adjustment branch continuously adjusts various parameters during the welding process, ensuring optimal quality for each welding operation. Specifically, the parameter adjustment branch adjusts welding parameters based on the following formula:

[0045]

[0046] wherein, denotes a parameter adjustment vector, denotes an adjustment coefficient, denotes a target welding quality score, denotes a welding quality score, denotes a current welding parameter.

[0047] In summary, the adaptive decision module adopts a double-branch output structure, enabling the model to simultaneously perform quality prediction and welding parameter adjustment during the welding process. The quality prediction branch outputs a welding quality score using welding feature data, while the parameter adjustment branch adjusts the welding parameters based on the quality score, ultimately generating a parameter adjustment vector to optimize the welding process. By adopting such a dynamic adjustment mechanism, the stability and efficiency of the welding quality can be ensured under different welding stages and conditions. The output layer outputs the welding quality score and welding parameter adjustment instructions to help welding operators adjust welding parameters in real time during the welding process, thereby continuously improving welding quality.

[0048] In another exemplary embodiment, the welding quality monitoring model is trained by the following steps in step S300:

[0049] S301: Collect welding data under different conditions, including welding speed, welding power, and welding focus shifting distance, etc. The collection frequency is set to 100 ms to ensure real-time and sufficient granularity of the data.

[0050] S302: Preprocess the collected welding data, wherein the preprocessing steps are as described above.

[0051] S303: Construct a feature vector.

[0052] In this step, two feature vectors need to be constructed, including a welding process parameter feature vector and a parameter adjustment action feature vector wherein, denotes the welding speed (unit: m / min), denotes the welding power (unit: W), denotes the focus shifting distance (unit: mm); denotes the power adjustment step size, is the focus shifting direction. It should be noted that, since the welding speed directly affects the heat input during laser welding, frequent adjustment of the welding speed may cause fluctuations in welding temperature, leading to inconsistencies in penetration and width, etc. Therefore, the welding speed Keep constant to reduce the complexity of the control process.

[0053] S304: Constructing the Actor-Critic model and solving the policy function in the Actor network;

[0054] In this step, first, the Actor-Critic model receives the welding process parameter feature vector as input, and outputs the welding action adjustment change vector , which includes the power adjustment step and the focus adjustment amount, as part of the decision strategy, which can be represented by the following function:

[0055]

[0056] wherein, represents the decision strategy function; , both represent weight matrices; , both represent bias terms; and RELU represents an activation function.

[0057] Then, the Actor-Critic model receives the state-action feature vector [] and outputs the value function for evaluating the quality of the current state and action. Specifically, the value function is represented as follows:

[0058]

[0059] wherein, represents the value function, and represents the expected value when performing action in state ; , represents the weight matrix, , represents the bias term, and RELU represents the activation function.

[0060] S305: Defining the loss function;

[0061] In this step, during the model training process, the loss function is used to measure the gap between the model output and the target value, and the loss function is specifically shown as follows:

[0062]

[0063] wherein, represents the loss function; represents the batch size, and represents the number of samples used each time the update is performed; represents the loss function value in the first the error calculated in the sample; represents the predicted value of the Critic network at a given state and action ; ; represents the target value; represents when the absolute value of the error is less than or equal to 1, the loss function uses the square error multiplied by 0.5 to obtain the loss; represents when the absolute value of the error is greater than 1, the loss function uses a linear penalty, i.e., the absolute value of the error minus 0.5.

[0064] S306: Calculate the target value;

[0065] In this step, the calculation of the target value is based on the discount reward mechanism in reinforcement learning, and the specific formula is as follows:

[0066] wherein, represents the immediate reward, and represents the direct reward value returned by the environment after performing action in state ; γ represents the discount factor, and the initial value is set to 0.99, which is used to balance the weight of current and future rewards; represents that this is the predicted value of the Critic network for the next state and the action taken in this state.

[0067] Further, the calculation formula of is as follows:

[0068]

[0069] wherein, is the current welding defect rate, and is used to neutralize the influence of welding defects on the immediate reward, and is usually taken =0.5.

[0070] S307: Update the network parameters, which can be specifically calculated by the following formula:

[0071]

[0072]

[0073] wherein, represents the target network update coefficient, and the initial value is 0.005; , These represent the online network parameters of the Actor and Critic networks, respectively. , This represents the target network parameters for the Actor and Critic networks.

[0074] S308: Set the model stopping condition.

[0075] In this step, when the loss function value When the model stops training, at this time The value prediction is stable, and the optimization results of the model and parameters have converged. It can accurately adjust the welding process parameters to ensure the stability and consistency of welding quality.

[0076] Below, this application provides a complete example of the technical solution using specific data.

[0077] First, this application obtained welding data for a welding process in real time using sensors under different working conditions, as shown in Table 1 (sampling frequency of 100ms).

[0078] Table 1

[0079]

[0080] Furthermore, the welding data shown in Table 1 was preprocessed, and the preprocessed welding data is shown in Table 2:

[0081] Table 2

[0082]

[0083] Furthermore, based on the data shown in Table 2, welding feature vectors are constructed as shown in Table 3:

[0084] Table 3

[0085]

[0086] Finally, based on the quality scoring formula and parameter adjustment formula shown above, the quality score at each time point and the total score are calculated, where it is assumed that the weight matrix... The quality score for a time of 100ms is:

[0087]

[0088] Similarly, the quality scores for times of 200ms, 300ms, 400ms, 500ms, and 600ms are as follows: , , , , The overall quality score is: .

[0089] Based on the parameter adjustment formula as described above, we can get .

[0090] Based on the above score, the score is proportionally distributed to the adjustment of welding power, welding speed and focus shifting distance. Assuming that 60% of the score is used for welding power adjustment, 30% for welding speed adjustment and 10% for focus shifting distance adjustment, the specific adjustment data is as follows: first, the welding power adjustment amplitude is 7.9368x0.6=4.722 scores, assuming that the quality score increases by 0.1 for every 1 watt of power increase, so the power adjustment amplitude is 47.22W; second, the welding speed adjustment amplitude is 7.9368x0.3=2.380 scores, assuming that the quality score increases by 0.05 for every 1 meter / minute speed increase, so the speed adjustment amplitude is 47.6 m / min; finally, the focus shifting distance adjustment amplitude is 7.9368x0.1=0.7937 scores, assuming that the quality score increases by 0.02 for every 1 millimeter of focus shifting distance increase, so the focus adjustment amplitude is 39.68 mm. Through these specific adjustments, the welding process can be optimized to ensure stable welding quality and achieve the target score.

[0091] In another exemplary embodiment, as shown in Figure 3 , the application also provides a high adaptability laser welding parameter adjustment device, which comprises: a monitoring and collecting module 100 for real-time monitoring of the laser welding process and collecting welding data, the welding data including welding speed, welding power and welding focus shifting distance; a preprocessing module 200 for preprocessing the welding data; an adjustment module 300 for inputting the preprocessed welding data into a trained welding quality monitoring model for analysis, and automatically adjusting the laser welding parameters according to the output results of the welding quality monitoring model.

[0092] Optionally, the preprocessing module 200 comprises: a cleaning sub-module for cleaning the welding data using a PCA algorithm; a dimension reduction sub-module for dimension reduction of the cleaned welding data; a normalization processing sub-module for normalization processing of the dimension-reduced welding data.

[0093] In another exemplary embodiment, the application also provides a storage medium comprising instructions which, when executed on a computer, cause the computer to perform the method of any one of the preceding embodiments.

[0094] In another exemplary embodiment, the application also provides an electronic device comprising: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of the preceding embodiments when executing the program.

[0095] The above merely provides the optional examples of the present application but are not intended to limit the present application. Even though the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art will be able to make modifications or equivalent replacements to the technical solutions described in the foregoing embodiments or make equivalent replacements to some of the technical features. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A highly adaptive laser welding parameter adjustment method, characterized by, The method comprises: real-time monitoring of a laser welding process and collecting welding data, wherein the welding data includes welding speed, welding power and welding focus shifting distance; preprocessing the welding data; inputting the preprocessed welding data into a trained welding quality monitoring model for analysis, and dynamically adjusting laser welding parameters according to the output results of the welding quality monitoring model; The welding quality monitoring model comprises: input layer, multi-channel feature extraction module, time sequence feature fusion module, dynamic attention mechanism, adaptive decision module and output layer, wherein, the input layer is used for inputting the preprocessed welding data; the multi-channel feature extraction module is used for extracting multi-dimensional features in the preprocessed welding data; the time sequence feature fusion module is used for fusing the output features of the multi-channel feature extraction module to obtain a fusion feature vector; the dynamic attention mechanism is used for dynamically adjusting and optimizing feature weight distribution; the adaptive decision module is used for predicting welding quality and dynamically adjusting welding parameters based on welding quality; the output layer is used for outputting welding quality score and welding parameter adjustment instruction; The multi-channel feature extraction module adopts a three-channel parallel architecture to cooperatively process three parameters of welding speed, welding power and welding focus shifting distance, and the three channels include a 1D-CNN time sequence channel, a Bi-LSTM time sequence channel and a statistical feature channel, wherein the 1D-CNN time sequence channel deploys three groups of convolution kernels with different scales to capture instantaneous fluctuation features in the welding process, and outputs feature maps after ReLU activation, and then obtains key feature vectors through maximum pooling operation; the Bi-LSTM time sequence channel adopts a bidirectional LSTM structure to deeply model long-range dependence of welding parameters, and outputs a 256-dimensional state vector to retain complete time sequence context; the statistical feature channel calculates engineering statistics in a 50ms sliding window in real time, including dynamic mean, variance and linear fitting slope, and encodes them into structured features through a fully connected layer; The time sequence feature fusion module adopts a cross-attention guided feature enhancement architecture, which inputs local transient features output from the 1D-CNN channel as key vectors, and global time sequence features output from the Bi-LSTM channel as query vectors into a multi-head cross-attention layer, and dynamically filters the most relevant local features for the current welding stage through attention weights; subsequently, a gated residual mechanism is used to concatenate 32-dimensional engineering parameters of the statistical feature channel with the output of the multi-head cross-attention layer, and adjust the contribution of each feature in the final output using learnable gating weights; finally, a time sequence compression unit composed of two layers of dilated causal convolution is used to compress the feature dimension while preserving the key time sequence dependence, and finally output a 256-dimensional spatio-temporal enhanced feature vector; The dynamic attention mechanism includes feature-level attention and time-level attention, wherein the feature-level attention focuses on feature weight distribution of the three channels, and the time-level attention dynamically adjusts the weight of historical features based on different stages of the welding process.

2. The method of claim 1, wherein, The preprocessing of the welding data comprises: cleaning the welding data; dimensionality reduction on the cleaned welding data; normalization processing on the dimensionality reduced welding data.

3. The method of claim 1, wherein, The adaptive decision module adopts a double-branch output structure, including: a quality prediction branch and a parameter adjustment branch, wherein the quality prediction branch is used for scoring the welding quality based on the welding feature data; the parameter adjustment branch is used for adjusting the welding parameters based on the welding quality score and generating a parameter adjustment vector.

4. A high adaptability laser welding parameter adjustment device, characterized in that, The device includes: a monitoring and collecting module for real-time monitoring of the laser welding process and collecting welding data, including welding speed, welding power and welding focus shifting distance; a preprocessing module for preprocessing the welding data; an adjustment module for inputting the preprocessed welding data into a trained welding quality monitoring model for analysis, and automatically adjusting the laser welding parameters according to the output results of the welding quality monitoring model; The welding quality monitoring model includes: an input layer, a multi-channel feature extraction module, a time series feature fusion module, a dynamic attention mechanism, an adaptive decision module and an output layer, wherein the input layer is used to input the preprocessed welding data; the multi-channel feature extraction module is used to extract multi-dimensional features in the preprocessed welding data; the time series feature fusion module is used to fuse the output features of the multi-channel feature extraction module to obtain a fusion feature vector; the dynamic attention mechanism is used to dynamically adjust and optimize feature weight distribution; the adaptive decision module is used to predict welding quality and dynamically adjust welding parameters based on welding quality; the output layer is used to output welding quality score and welding parameter adjustment instructions; The multi-channel feature extraction module adopts a three-channel parallel architecture to cooperatively process welding speed, welding power and welding focus shifting distance, the three channels including a 1D-CNN time series channel, a Bi-LSTM time series channel and a statistical feature channel, wherein the 1D-CNN time series channel deploys three groups of convolution kernels of different scales to capture instantaneous fluctuation features in the welding process, outputs feature maps after ReLU activation, and obtains key feature vectors after maximum pooling operation; the Bi-LSTM time series channel adopts a bidirectional LSTM structure to model the long-range dependence relationship of welding parameters in depth, and outputs a 256-dimensional state vector to retain the complete time series context; the statistical feature channel calculates engineering statistics within a 50ms sliding window in real time, including dynamic mean, variance and linear fitting slope, and encodes them into structured features through a fully connected layer. The timing feature fusion module adopts a cross-attention guided feature enhancement architecture, which inputs the local transient features from the 1D-CNN channel output as key vectors and the global timing features from the Bi-LSTM channel output as query vectors into the multi-head cross-attention layer, and dynamically screens the most relevant local features for the current welding stage through the attention weight; subsequently, the 32-dimensional engineering parameters of the statistical feature channel are spliced with the output of the multi-head cross-attention layer by using the gated residual mechanism, and the contribution of each feature in the final output is adjusted by using the learnable gating weight; finally, the feature dimension is compressed by the timing compression unit composed of two layers of dilated causal convolution under the premise of retaining the key timing dependency, and a 256-dimensional spatiotemporal enhanced feature vector is finally output. The dynamic attention mechanism includes feature-level attention and time-level attention, wherein the feature-level attention focuses on the feature weight distribution of the three channels, and the time-level attention dynamically adjusts the weight of the historical features based on different stages of the welding process.

5. The apparatus of claim 4, wherein, The preprocessing module comprises: a cleaning submodule for cleaning the welding data; a dimension reduction submodule for dimension reduction of the cleaned welding data; a normalization submodule for normalizing the dimension-reduced welding data.

6. A storage medium, characterized by The computer program comprises instructions which, when executed on a computer, cause the computer to carry out the method of any one of claims 1-3.

7. An electronic device, comprising: The electronic device comprises: a memory, a processor and a computer program stored on the memory and executable on the processor, wherein the processor implements the method of any one of claims 1-3 when executing the program.

Images