Complex microstructure ultrafast laser processing quality prediction method based on multi-task learning

Abstract

The application discloses a kind of complex microstructure ultrafast laser multi-stage processing process quality prediction method based on multi-task deep learning, and relates to laser processing quality prediction technical field.The method is first to the sparsely sampled historical processing data data enhancement, constructs sufficient training sample set.On this basis, construct quality prediction model based on multi-stage multi-task learning, the model is by introducing the feature extraction module of structured state space Explicit modeling interstage physical effect transmission, and utilize task clustering and unique characteristic modulation to realize the collaborative prediction of roughness, depth and verticality and multiple strong coupling quality indexes.In training, the adaptive loss function based on uncertainty is used to automatically balance multi-task learning.When applied, input each stage process parameter, and the model can output the quality index prediction value of each stage in whole process.The application solves the problem of quality accurate prediction under the condition of multi-stage cumulative effect and multi-index coupling relationship.

Description

Technical Field

[0001] This invention belongs to the field of laser processing quality prediction technology, specifically relating to a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. Background Technology

[0002] With the increasing demand for precision manufacturing, ultrafast laser processing has become a core technology in micro-nano manufacturing and is widely used in aerospace, biomedicine and other fields. This technology achieves high-precision processing through layer-by-layer scanning, but the processing quality is significantly affected by the coupling of multi-layer process parameters and the interlayer transmission of physical effects such as heat accumulation and residual stress. At the same time, there are strong coupling relationships between multiple quality indicators such as roughness, depth and perpendicularity, which makes accurate prediction of the quality of the entire process a serious challenge.

[0003] Currently, laser processing quality prediction mainly employs two methods: finite element simulation and data-driven approaches. Finite element simulation is computationally too time-consuming, making it difficult to meet real-time requirements. Meanwhile, data-driven methods based on machine learning mostly predict only a single quality index of the final state independently, or simplify a multi-stage process into a single-stage model. These methods fail to model the cumulative transmission of physical effects between stages and ignore the inherent correlations between multiple quality indices, resulting in limited prediction accuracy and insufficient generalization ability in real multi-stage processing scenarios.

[0004] Existing technologies have two prominent limitations. First, the predictions of multiple quality indicators are isolated, failing to reflect the objective trade-offs and synergistic relationships between the indicators, and the prediction results may deviate from physical reality. Second, they generally neglect the stage characteristics of processing and fail to explicitly model the evolution and transmission of key states such as heat and stress between layers, resulting in the model's inability to capture the error propagation process and limiting its practical application value in complex multi-stage processing. Therefore, there is an urgent need to develop a new quality prediction method that can unify the modeling of stage transmission and the coupling of multiple indicators. Summary of the Invention

[0005] To address the aforementioned problems in existing technologies, this invention provides a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. The technical problem to be solved by this invention is achieved through the following technical solution: This invention provides a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning, comprising: Step 1: Obtain the historical dataset of ultrafast laser multi-stage processing. The historical dataset includes multiple processing samples, each containing process parameter data and corresponding multivariate quality index data for multiple processing stages. Step 2: Perform data augmentation on the historical dataset to increase the number of samples and construct an augmented dataset for model training; Step 3: Based on the augmented dataset, construct and train a multivariate quality prediction model based on multi-stage multi-task learning; Step 4: Input the process parameters of each processing stage of the workpiece to be predicted into the trained multivariate quality prediction model, and output the predicted values ​​of multivariate quality indicators for each processing stage.

[0006] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The present invention provides a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. Employing a multi-task learning framework, it achieves parallel and joint prediction of multiple strongly coupled quality indicators through shared feature extraction and task-specific feature modulation, and collaborative prediction based on a data-driven task clustering strategy. This method can automatically learn the inherent trade-offs between indicators, making the prediction results more consistent with physical constraints and avoiding the indicator conflicts that may arise from independent prediction by a single task. By modeling the sequence dependencies and physical effects between each processing stage, such as heat accumulation and stress transfer mechanisms, it overcomes the shortcomings of traditional single-stage models that ignore process accumulation errors. It can continuously and dynamically predict the quality of the entire process from the initial layer to the final layer, significantly improving the systematicity and accuracy of the prediction.

[0007] 2. The present invention provides a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. Addressing the problems of high experimental costs, limited sample sizes, and imbalanced data distribution in laser processing, a data augmentation strategy is designed. This strategy not only effectively expands the training samples but also preserves and enhances the physical information contained in outlier patterns caused by specific process parameters, thereby improving the model's ability to utilize small samples and its generalization ability to complex operating conditions.

[0008] The above description is merely an overview of the technical solution of the present invention. In order to better understand the technical means of the present invention and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of the present invention more apparent and understandable, preferred embodiments are described in detail below with reference to the accompanying drawings. Attached Figure Description

[0009] Figure 1 This is a flowchart of a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning, provided in an embodiment of the present invention. Figure 2 This is a flowchart of a data augmentation technology provided in an embodiment of the present invention; Figure 3 This is a framework diagram of a multivariate quality prediction model based on multi-stage, multi-task learning provided in an embodiment of the present invention; Figure 4 This is a block diagram of a multi-stage, multi-task feature extraction network provided in an embodiment of the present invention; Figure 5 This is a block diagram of a multi-stage, multi-task quality prediction network provided in an embodiment of the present invention; Figure 6 This is a process diagram of laser processing provided in an embodiment of the present invention; Figure 7 This is a structural diagram of a multivariate quality prediction model based on multi-stage multi-task learning provided in an embodiment of the present invention; Figure 8 This is a mean graph of all stages of an ablation experiment provided by an embodiment of the present invention; Figure 9 This is a mean plot comparing a multi-task model with a given embodiment of the present invention. Figure 10 This is a fitting graph of the original real data and the prediction results of multiple models provided in the embodiments of the present invention; Detailed Implementation To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following describes in detail, with reference to the accompanying drawings and specific embodiments, a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning proposed in accordance with the present invention.

[0010] The foregoing and other technical contents, features, and effects of the present invention will be clearly presented in the following detailed description of specific embodiments in conjunction with the accompanying drawings. Through the description of the specific embodiments, a more in-depth and concrete understanding can be gained of the technical means and effects adopted by the present invention to achieve its intended purpose. However, the accompanying drawings are for reference and illustration only and are not intended to limit the technical solutions of the present invention.

[0011] This invention provides a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. Please refer to... Figure 1 , Figure 1 This is a flowchart illustrating a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning, provided in an embodiment of the present invention. Figure 1 As shown in this embodiment, the method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning includes the following steps: Step 1: Obtain the historical dataset of ultrafast laser multi-stage processing. The historical dataset includes multiple processing samples, each containing process parameter data and corresponding multivariate quality index data for multiple processing stages.

[0012] In this embodiment, the multi-dimensional quality indicators include at least two of the following: surface roughness, machining depth, and sidewall perpendicularity.

[0013] Step 2: Perform data augmentation on the historical dataset to increase the number of samples and build an augmented dataset for model training.

[0014] In laser processing, it is often necessary to know the roughness, depth, and perpendicularity values ​​of a specific layer. However, during data acquisition, considering the significant manpower and resources required for layer-by-layer data collection, and the relatively long experimental cycle, and given that depth changes within a single layer do not produce significant variations, a more efficient and economical approach is to scan five layers and measure the part data once. Furthermore, variations in laser processing parameters can cause significant changes in laser energy, leading to substantial differences in the quality indicators of parts processed with certain parameters compared to other quality indicators. This is mainly manifested in a few outliers in roughness calculations, although these are not considered outliers. Directly processing these outlier data using only outlier removal methods would result in an incomplete dataset. Additionally, while small sample datasets can be directly used for model training, models trained on small sample datasets often struggle to guarantee predictive performance and stability. Even using general sample augmentation methods, the model still struggles to learn the information contained in the outliers during training.

[0015] To address the above issues, this embodiment proposes a data augmentation method that includes phase expansion processing of the scanning layer and outlier sample processing. This data processing method can simultaneously expand both the phase of the data and the number of samples, while focusing on a few outlier data points, enabling the trained model to learn outlier information.

[0016] Specifically, the scanning layer stage expansion processing includes: based on the sparsely sampled processing stage data in the historical dataset, using Hermite interpolation to interpolate the unsampled processing stages, generating stage expansion samples containing process parameter data and multivariate quality index data for each processing stage.

[0017] Outlier sample processing includes: based on The principle is to identify outliers in the multivariate quality index data of the historical dataset and mark the stage-expanded samples containing outliers as outlier samples. The CTGAN network is used to expand the outlier samples, and the samples generated by the CTGAN network are filtered based on Euclidean distance. The filtered samples are combined with the outlier samples to form expanded outlier samples. The expanded outlier samples are mixed with normal samples in the historical dataset after the scan-layer stage expansion process, and then the CTGAN network is used to expand the samples again. The samples generated by the CTGAN network are filtered based on Euclidean distance. The filtered samples are combined with the historical dataset after the scan-layer stage expansion process to form an augmented dataset.

[0018] In this embodiment, expanding the scanning layers meets the requirement of obtaining predicted values ​​of quality indicators for each layer in a real processing environment. Furthermore, extrapolating the actual first-layer data adds learnable features and improves the predictive ability of shallow scanning indicators. The expanded data contains complete processing parameters and quality indicators for each layer. Outlier samples are then extracted based on one-hot encoding, and a two-step Conditional Tabular Generative Adversarial Network (CTGAN) data augmentation method based on Euclidean distance is used to expand the multi-stage data. The dataset after layer expansion and imbalanced sample augmentation possesses complete processing parameters and quality indicators, and contains a sufficient number of training samples to meet the training requirements of a deep learning quality prediction model.

[0019] Please refer to the above. Figure 2 The data augmentation process in this embodiment will be explained using the example of expanding from 9 stages to 45 stages. Figure 2 This is a data augmentation processing flowchart provided in an embodiment of the present invention. Let the original dataset be... , No. The process parameters for the sample are as follows: The corresponding processing indicators are ,in This represents the original stage number for each sample. The original number of samples is [number], and the augmented dataset is [number]. The data augmentation process mainly includes the following steps: 1) First, expand the scanning layer stage. For the same laser-scanned part sample data, based on the scan layers of the acquired data, fill in the processing parameters and quality indicators of all scan layers without acquired data. Hermite interpolation, as a polynomial interpolation method, ensures the high-order smoothness of the interpolation function by constraining the function value, derivative value, and even higher-order derivative to be the same at a specified point. The specific calculation process is as follows: Let the parameter values ​​of the acquired scanning data be... The processing index value at the location is The derivative value is Then Hermite interpolation polynomial Satisfying equation (2-22): (2-22).

[0020] According to the above formula One condition can determine that there is only one condition that does not exceed [a certain value]. Hermitian interpolation polynomial of degree It can be known that and Equations (2-23) and (2-24) must be satisfied respectively: (2-23); (2-24); The constructed Hermite interpolation polynomial can be determined based on equations (2-22), (2-23), and (2-24), thereby enabling the expansion of the laser scanning layer.

[0021] 2) Use one-hot encoding to obtain outlier samples. Outliers are mainly found in roughness data. For roughness data that has not undergone stage augmentation, use... Outlier data is labeled layer by layer, with outliers marked as 1 and others as 0. All layer labels for each sample are concatenated to obtain a one-hot code for each sample. When all bits in a sample's one-hot code are 1, it indicates the presence of a roughness value that is an outlier. Such samples are separated from other normal samples and designated as outlier samples. Samples with all bits in their one-hot codes set to 0 are designated as normal samples.

[0022] 3) Two-step CTGAN data augmentation based on Euclidean distance. For the obtained outlier samples, CTGAN is used for augmentation. The objective function of CTGAN is shown in equation (2-25): (2-25); in Represents a generator. Indicates the discriminator, Condition variables are used to control specific attributes of the generated data. Indicates data from real data distribution The sample, Indicates the distribution of noise The sample, Represents mathematical expectation, generator The goal is to minimize Discriminator The goal is to maximize The generator is used to generate new data, and its training objective is to make the generated data as indistinguishable from the discriminator as possible. The discriminator is used to distinguish between the generated data and the real data, and its training objective is to identify the source of the data as accurately as possible. By having the generator and discriminator compete against each other during training, the generator and discriminator can be trained to generate data similar to the real data, thus achieving the purpose of data augmentation.

[0023] The outlier markers in section 2) are used on the data generated by CTGAN. The values ​​are then re-encoded and labeled with one-hot encoding. Outlier samples from the augmented dataset are filtered, and the Euclidean distance is used to measure the distance between each augmented outlier sample and the original outlier class samples. The samples closest to the original outlier class samples are retained, ensuring that the number of outlier class samples after augmentation is the same as the number of normal class samples. The samples in the outlier class set... Samples in augmented data The Euclidean distance is calculated using the formula shown in equation (2-26): (2-26); in This indicates the sample dimensions (including process parameter dimensions and quality indicator dimensions). , , The total number of outlier samples. To expand the total number of samples.

[0024] The expanded outlier sample set is then mixed with normal class samples, and CTGAN is used for further sample augmentation. Euclidean distance is then used for sample selection. Since an excessively large expanded dataset may cause the distribution of the expanded data to deviate from the original data distribution, and an excessively small expanded dataset may not achieve the purpose of data augmentation, the few data points closest to the original dataset are retained to form a dataset of 1000 samples, thereby achieving the purpose of data sample augmentation.

[0025] Step 3: Based on the augmented dataset, construct and train a multivariate quality prediction model based on multi-stage multi-task learning.

[0026] To address the challenge of parallel prediction of multiple quality indicators across multiple scanning stages in ultrafast laser processing, this embodiment proposes a structured state space sequence improved multi-stage multi-task attention network (S4-MSMTAN) based on multi-task prediction methods. This model comprises a shared feature extraction module, a task-specific feature generation module, a clustering tower network prediction module, and an adaptive loss adjustment module, all connected sequentially.

[0027] The shared feature extraction module extracts shared features based on the process parameters of the current processing stage and the predicted values ​​of multivariate quality indicators from the previous processing stage. The task-specific feature generation module generates corresponding weights for the tasks that predict each quality indicator and uses these weights to weight and modulate the shared features to obtain the task-specific features. The clustered tower network prediction module predicts all quality indicators of the current processing stage based on the task-specific features through the tower network of each cluster. The adaptive loss adjustment module adaptively adjusts the weight of each task's loss in the total loss based on the uncertainty of each task during the training phase.

[0028] Please refer to the above. Figure 3 The structure of the multivariate quality prediction model based on multi-stage multi-task learning in this embodiment will be described in detail. Figure 3 This is a framework diagram of a multivariate quality prediction model based on multi-stage, multi-task learning provided in an embodiment of the present invention.

[0029] like Figure 3 As shown, the shared feature extraction module in this embodiment employs a deep neural network based on the S4 structured state space. Ultrafast laser processing quality prediction is a multi-stage, multi-task prediction problem. Constructing a shared feature extraction module based on the S4 structured state space deep neural network enables the extraction of shared features in laser processing. Based on the extracted shared features, weights are calculated through the neural network, and the shared features are weighted to obtain the unique features of each cluster of tasks. During shared feature extraction at a certain stage, the process parameters of this stage and the prediction results of quality indicators from previous stages are simultaneously used as input, allowing the model to acquire quality information from previous scanning stages during feature extraction, increasing feature diversity. The neural network then calculates weights and weights the shared features of the current stage, gradually acquiring the unique features of each cluster of tasks through multiple iterations. During the first stage of multi-task quality prediction, no information from other stages is required.

[0030] In ultrafast laser processing, multiple quality indicators exist, such as roughness, depth, and perpendicularity. Based on the correlations among these indicators, a clustered tower network is constructed to improve the efficiency and accuracy of the model's quality prediction. A comprehensive evaluation index of multiple correlations among various quality indicators in a historical dataset is calculated. Based on the correlation calculation results, the quality indicators are divided into multiple clusters. The features extracted by the feature extraction module are used as input to the clustered tower network for quality prediction. The tower model constructed through a deep neural network predicts the values ​​of each task indicator within each cluster.

[0031] The adaptive loss adjustment module in this embodiment performs adaptive weight adjustment based on uncertain multi-task loss. The loss can be calculated based on the predicted value of each task in each stage and the actual value in the training dataset. The loss of multiple tasks in all stages is adjusted by adaptive adjustment of loss weights based on uncertainty to obtain the final total loss of the model.

[0032] Understandably, during the model training phase, the backpropagation and parameter update module calculates the gradient descent direction based on the total loss of the calculated model and updates the parameters of each module of the model until the model converges.

[0033] Specifically, in the design process of the multivariate quality prediction model framework based on multi-stage, multi-task learning, the following is defined: This is a dataset of ultrafast laser processing with multivariate quality labels, in which The total number of samples, Indicates the first The feature space of each sample Indicates the first The label space of each sample Number of stages This represents the total number of features in a single stage. The total number of labels for a single stage, i.e. Indicates the first The first sample Phase 1 feature, Indicates the first The first sample Phase 1 In the context of ultrafast laser processing, the features include multiple adjustable process parameters and the number of scanning layers, while the labels are quality label values ​​such as roughness, depth, and perpendicularity. When the process parameters and the number of scanning layers are set differently, the quality label values ​​such as roughness will also change. Therefore, it is necessary to learn the potential correlation between the number of scanning layers and process parameters in ultrafast laser processing and multiple quality label values, and construct a quality prediction model. This model should be trained to achieve a comprehensive understanding of the ultrafast laser processing feature space. To tag space The mapping enables multi-stage quality prediction and provides support for subsequent laser processing technology optimization. The constructed quality prediction model is shown in Equation (3-1).

[0034] (3-1); in This represents the quality prediction model obtained through training. Represents the predicted label space. and These represent the shared layer parameter set and the unique layer parameter set of the model, respectively. This represents the parameters of the weight adjustment module. For the first... Stage prediction compares the current stage's process parameters with those of the previous stage. The model is represented as Equation (3-2), with the predictive indicators for each stage as input.

[0035] (3-2); in Indicates the first The model built in the stage, Indicates the first Predicted labels for each stage Indicates the first The process parameter characteristics of each stage, and They represent the first The model parameter sets of the shared layer and the specific layer of the stage. These are the parameters for the model weight adjustment module.

[0036] Furthermore, the structure and data processing flow of the feature extraction module and the quality prediction module in the multivariate quality prediction model based on multi-stage multi-task learning in this embodiment are described in detail.

[0037] In ultrafast laser processing of precision parts, laser scanning is required layer by layer until quality requirements are met. To characterize the transmission relationship of this serial data generation process, a serial transmission relationship model needs to be constructed when predicting the layer-by-layer quality of ultrafast laser processing. Therefore, this embodiment constructs a multi-stage, multi-task parallel quality prediction model to solve the prediction problem involving multiple serial processing stages, each of which is a multi-task prediction process. To improve the prediction accuracy and efficiency of the multi-stage, multi-task quality prediction model, the collaborative extraction of shared and unique features becomes a core issue in model design. A reasonable design of the feature extraction module can significantly improve the model's generalization, robustness, accuracy, and other excellent performance characteristics.

[0038] The feature extraction network of the multi-stage, multi-task quality prediction model designed in this embodiment is as follows: Figure 4 As shown, Figure 4 This is a block diagram of a multi-stage, multi-task feature extraction network provided in an embodiment of the present invention. For example... Figure 4 As shown, the current process data and the predicted labels from previous stages are used as inputs to the feature extraction network. The S4 structured state space model is introduced into the shared feature extraction module, which improves the feature extraction capability and model efficiency of high-dimensional input data. At the same time, the weights calculated by the task-specific feature generation module are used to weight the shared features to obtain the unique features of each task.

[0039] In this embodiment, the input to the feature extraction network is the process parameters of the current stage and the predicted labels from multiple previous stages. The input of the process parameters of the current stage ensures that the model obtains the parameter information for this stage, while the input of the predicted quality labels from multiple previous stages allows the model to obtain information from previous processing stages, increasing the diversity of input features and avoiding poor model performance due to a single processing parameter. Since information is transferred between stages, it is impossible to extract shared features across all stages. Therefore, feature extraction and quality prediction are performed separately for each stage based on the stage of the data. The feature extraction for each stage mainly includes two parts: shared feature extraction based on the shared feature extraction module and task-specific feature extraction based on the task-specific feature generation module.

[0040] For the Feature extraction for each stage, with the input process parameters for the current stage as features. and its predecessor Multivariate predictive labels for each stage ,in Indicates the number of stages in the transmission. Indicates the number of input samples. This represents the number of process parameter features for each sample. This represents the number of tasks in the multi-task prediction for each sample, i.e., the number of labels to predict. Indicates its preceding The total number of tasks for predicting labels in each stage.

[0041] In this embodiment, the shared feature extraction module specifically includes: a series of cascaded linear layers, a ReLU activation function layer, an S4 structured state-space model loop, a batch normalization layer, and a Dropout layer.

[0042] Specifically, the linear-activation layer structure achieves efficient linear transformation and introduces nonlinearity to overcome linear limitations; the batch normalization layer standardizes the input distribution of the neural network, making network training more stable and faster; and the S4 structured state space solves the problem of high-dimensional input features caused by inter-stage information transfer. It uses the entire input sequence for training, learns each dimension through the structured state space, and then concatenates the multiple dimensions, making it suitable for addressing the high-dimensional input features in this embodiment. Before the input data enters each layer of the network, its dimensions are expanded to facilitate subsequent data shape transformation. Let the input of each stage after passing through the linear layer and activation function be... The calculation formula is shown in equation (3-3): (3-3); in Indicates the number of samples. Indicates the expanded dimension. This indicates the parameters set for the hidden layer. Represents a linear neural network layer. This represents the ReLU activation function layer.

[0043] Then it is input into the S4 structured state-space loop body. The S4 structured state space is an extension of the discrete state-space model, so the input of the loop body is first... Perform mapping to obtain the discretization step size and average step size , ,in For the Sigmoid function. Set the state-space parameters to... And initialize four learnable parameters: state matrix Input projection matrix Output projection matrix residual term and state space Then, the state matrix and input matrix are discretized to ensure the stability of the state transition, obtaining the following results: and The calculation formulas are shown in equations (3-4) and (3-5): (3-4); (3-5); in The function for calculating the matrix exponent. This involves element-wise matrix multiplication. After the discretization step, the state space is calculated for the nth hidden layer input. Dimensional Data Its corresponding state space and output The calculation formulas are shown in equations (3-6) and (3-7): (3-6); (3-7); in This is a residual connection, which mitigates the vanishing gradient phenomenon by preserving input information.

[0044] After traversing all After processing the dimensional data, the output of the S4 structured state space can be obtained. Finally, Input batch normalization and regularization layers, for Standardization is performed in batches, and Dropout is used to randomly "drop" a portion of neurons to avoid overfitting of the model. The output of the shared feature extraction module is shown in (3-8).

[0045] (3-8); in For batch normalization function, The regularization function is obtained as follows: For the first Shared features extracted during the phase.

[0046] In this embodiment, the task-specific feature generation module is a trainable weight calculation module. After obtaining the first... After the shared features of each stage are extracted, a trainable weight calculation module is used to calculate the weights of all tasks, thus weighting the extracted shared features. Therefore, the number of these weight calculation modules is the same as the number of tasks that need to be predicted in each stage. Therefore, the first The weights of all tasks in the phase are: To improve the accuracy of weight calculation for each task, this weight calculation module can be built based on a neural network.

[0047] In this embodiment, the weight calculation module consists of a linear layer and an activation layer, as shown in equation (3-9). The extracted shared features are used as input to the weight calculation module. Weight coefficients are used to weight the extracted shared features to obtain the unique features of each task. , No. Phase 1 Task-specific characteristics The calculation formula is shown in equation (3-10). Then, based on the multi-task correlation clustering results, the cluster-specific features are spliced ​​together to obtain the input-specific features of each quality prediction tower network.

[0048] (3-9); (3-10); in Indicates the first Phase 1 The weighting coefficient of the task. This represents the weighting function.

[0049] when Since there are no prediction results from previous stages as input, only the process parameter features of this stage are used as input to extract important features.

[0050] For the In this stage, the features obtained through the feature extraction network are input into the quality prediction network, and the multivariate quality index prediction results for this stage can be obtained, as shown in Equation (3-11): (3-11); in .

[0051] Input data can obtain unique features after passing through a multi-stage, multi-task feature extraction network. As The unique characteristics of each task are identified, and then task clustering is performed. The model prediction quality index and parameter update direction are obtained through the quality prediction tower network and loss calculation module.

[0052] It should be noted that before constructing and training the multivariate quality prediction model, task clustering is performed in advance based on task relevance. The same network cluster can predict labels for one or more tasks. During model training, after obtaining the predicted labels for all tasks, these labels are passed to subsequent stages as input. Simultaneously, the total loss for each task in each stage is calculated based on the predicted labels. Furthermore, the total loss for each task in all stages is adaptively weighted based on uncertainty, ultimately obtaining the total loss of the multi-stage, multi-task model. After obtaining the total model loss, the network parameters are updated, thus progressively updating the model parameters and improving model performance.

[0053] Specifically, the task clustering process in this embodiment includes: calculating multiple correlation evaluation indicators between prediction tasks of each pair of quality indicators based on historical datasets to form a comprehensive evaluation score; setting a threshold based on the average comprehensive evaluation score of all task pairs, and classifying tasks with comprehensive evaluation scores higher than the threshold into the same cluster, thereby determining the final task clustering scheme; and fixing the number of tower networks in the clustering tower network prediction module of the multivariate quality prediction model and the set of tasks that each cluster tower network is responsible for predicting according to the clustering scheme.

[0054] In this embodiment, the Pearson correlation coefficient is used. (Pearson correlation coefficient) and Spearman rank correlation coefficient Spearman correlation, Kendall's rank correlation coefficient (Kendall RankCorrelation Coefficient), Cosine Similarity (Cosine Similarity) and the coefficient of determination (Coefficient of Determination) These correlation evaluation indicators assess the linear correlation, monotonicity, ordinal correlation, and directional similarity among multiple tasks, and the comprehensive evaluation results are based on the above indicators. Perform task clustering, based on task relative tasks Taking the correlation as an example, the comprehensive evaluation results The calculation formula is shown in equation (3-12): (3-12).

[0055] Since the closer the first four correlation evaluation indicators are to 0, the lower the correlation between tasks, the absolute value is used to reflect the magnitude of the correlation, and tasks with a correlation greater than the average correlation of all tasks are grouped into the same cluster.

[0056] Furthermore, see also Figure 5 The quality prediction process of this embodiment will be described. Figure 5 This is a block diagram of a multi-stage, multi-task quality prediction network provided in an embodiment of the present invention. Figure 5 As shown, for the first Quality prediction at each stage, the feature extraction stage obtains Based on the unique characteristics of each task, multiple tasks are clustered according to the clustering method described above. Assume all tasks can be divided into... The first cluster is then assembled based on the clustering results, using the unique features of each cluster. Phase 1 The unique characteristics of clusters are The unique characteristics of quality prediction clustering tower networks are: By inputting multiple sets of unique features into several clustered tower networks consisting of linear layers, ReLU activation function layers, and Dropout regularization layers, prediction values ​​for multiple tasks belonging to that cluster can be obtained. After passing through all the quality prediction tower networks, prediction values ​​for all tasks in that stage can be obtained. .

[0057] Based on the predicted values ​​for this stage and the true value of the training set data Able to calculate the MSE loss for each task separately. ,Task The calculation formula is shown in equation (3-13): (3-13).

[0058] After obtaining the losses for all tasks across all stages, calculate the sum of losses for each task across all stages, assuming the task... The total loss is Then, based on the losses of all tasks... An adaptive weight adjustment for multi-task loss based on uncertainty is implemented. This adaptive weight adjustment method reduces the negative impact of noisy tasks (with high uncertainty), focuses loss calculation on low-noise tasks (with high determinism), suppresses interference from high-loss tasks while enhancing the learning of low-loss tasks, and ensures that all tasks participate in model training through a learnable log-variance regularization term. First, define... Learnable log-variance for each task And initialize it, using the reciprocal of the variance as the task weight. ,in The formula for calculating the total loss is shown in equation (3-13).

[0059] (3-13); In each task's loss, the first term is a weighted loss term, and the second term is a learnable log-variance regularization term. This module shares gradients with the main model, and its parameters are updated along with the model training, thus enabling efficient and adaptive end-to-end training of the multi-task balancing strategy. In the optimization of the multi-stage, multi-task quality prediction model, Adam is chosen as the optimizer, and gradient descent is used to optimize all parameters of the quality prediction model network. The model parameters are updated as shown in equation (3-15). When the total model loss... Save the model parameters upon convergence.

[0060] (3-15); in It is the learning rate.

[0061] The pseudocode of the multivariate quality prediction model based on multi-stage multi-task learning in this embodiment is shown in Table 1.

[0062] Table 1

[0063] Step 4: Input the process parameters of each processing stage of the workpiece to be predicted into the trained multivariate quality prediction model, and output the predicted values ​​of multivariate quality indicators for each processing stage.

[0064] In step 4 of this embodiment, for the prediction of the first processing stage, the model input is the process parameters of the first processing stage; for the prediction of the second processing stage... Processing stage prediction, model input includes the first Process parameters and preceding steps in the processing stage Predicted values ​​of multi-source quality indicators for the first processing stage; for the first... Processing stage prediction, model input includes the first Process parameters of the processing stage and the first Pre-processing stage Predicted values ​​of multiple quality indicators for each processing stage, among which The number of information transmission stages is an integer greater than or equal to 1.

[0065] This invention presents a method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning. This method is applicable to the parallel prediction of multi-stage, multi-task, and multi-dimensional quality indicators in ultrafast laser processing. By mining the correlation between process parameters and quality indicators at each stage, it predicts multi-stage, multi-quality labels in ultrafast laser processing, overcoming the problem of low prediction accuracy caused by insufficient data learning in traditional single-stage and single-objective predictions. The proposed method breaks through the limitations of traditional single-stage modeling, achieving unified prediction of quality transfer and multi-indicator coupling relationships in multi-stage ultrafast laser processing. This has significant theoretical and applied value for improving the consistency of finished products in the field of multi-stage precision machining.

[0066] Furthermore, the performance of the multivariate quality prediction model based on multi-stage multi-task learning proposed in this invention is compared with other quality prediction models through simulation experiments, so as to illustrate the effectiveness of the quality prediction method for ultrafast laser processing of complex microstructures based on multi-task learning in this invention.

[0067] This simulation experiment studies multi-stage data from ultrafast laser processing, including the laser processing process as follows: Figure 6 As shown, Figure 6 This is a process diagram of laser processing provided by an embodiment of the present invention. The laser processing system consists of a laser source, a scanning galvanometer, an XY-axis motion platform for moving the workpiece, a Z-axis motion platform for moving the galvanometer to adjust the focal length, and a series of optical lenses. The laser source is a Pharos 6-600-PP femtosecond laser from Light Conversion, with a processing wavelength of 1030nm, a maximum output power of 8W, a pulse width of 191fs, a repetition rate of 1kHz-1MHz, a beam quality M^2<1.3, and a beam diameter of 6.67mm. The scanning galvanometer is mainly used to achieve rapid beam deflection; here, a Scancube III 10 2D galvanometer from Scanlab is selected, equipped with a Sill S4LFT4147-328 field lens (focal length 48.1mm).

[0068] During the processing, two problems were found in collecting processing data layer by layer: firstly, the noise data from laser processing was much greater than the data difference between adjacent layers; secondly, the measurement process was cumbersome, requiring a significant amount of time and manpower to process each layer and measure the parts. To address these issues, when conducting the orthogonal experiment for the data used in this embodiment, a method of scanning five layers and measuring once was adopted until all 45 layers were processed, obtaining data from nine stages: layers 5, 10, 15, 20, 25, 30, 35, 40, and 45.

[0069] Since laser processing is a type of precision machining, a laser confocal scanning microscope is needed to scan the surface of the processed parts when measuring their quality data. The data obtained from the scan includes three-dimensional point cloud data and surface image data. After processing these two types of data, three quality indicators—roughness, perpendicularity, and depth—can be extracted. These three quality indicators are integrated with the process parameters used in the orthogonal experiment to form structured data that facilitates subsequent data expansion and model training. Some of the extracted raw data is shown in Table 2.

[0070] Table 2

[0071] To achieve quality prediction and parameter recommendation for each scan layer, this invention designs a data augmentation method based on CTGAN, which realizes the expansion of stages and the expansion of sample size. Specifically, the processing data that was originally acquired once every 5 layers is expanded to have corresponding processing data for each layer. At the same time, the original 106 data points are expanded to 1000 data points. Some of the expanded data are shown in Table 3.

[0072] Table 3

[0073] To verify the effectiveness of data augmentation techniques, two sets of comparative experiments were set up to compare the model prediction performance of shallow data before and after stage augmentation and the prediction results obtained by different data augmentation methods.

[0074] Experiments revealed that stage expansion significantly improves the prediction performance of shallow data. To quantitatively demonstrate the effectiveness of stage expansion, the prediction accuracy of each indicator before and after stage expansion for the fifth-layer data was compared. Detailed comparison results are shown in the table and figures. The average improvement rate for MSE of the six quality indicators was 45.4%, the average improvement rate for MAE was 32.4%, and the average improvement rate for R² was 14.3%. This is because without stage expansion, the input for the fifth-layer quality data prediction only included the process parameter data of the fifth layer, resulting in a relatively simple input and poor prediction performance for shallow data. For the expanded data, the input for the fifth-layer data prediction included not only the corresponding process parameter data but also the prediction outputs from previous layers, increasing the diversity of input features, improving the situation of single features, and enhancing the model's prediction performance for shallow data.

[0075] The multi-stage data from laser processing was obtained through data augmentation, with five layers separating each pair of actual stage data. Therefore, the chosen number of layers for information transmission was [number missing]. The model employs a multi-stage, multi-task feature extraction network to extract information from the preceding real processing data and the extended layer data between these two real data layers. This improves model performance while avoiding excessive processing layer information that could slow down the model. The processed multi-stage process parameters and transmitted information are input into the multi-stage, multi-task feature extraction network to extract important features for each stage. These important features, carrying transmitted information, are then input into the clustering tower network prediction module for multi-stage, multi-task quality prediction. The specific structure of the proposed multivariate quality prediction model based on multi-stage, multi-task learning is as follows: Figure 7 As shown, Figure 7 This is a structural diagram of a multivariate quality prediction model based on multi-stage multi-task learning provided in an embodiment of the present invention. The specific parameters of each module in the model are shown in Table 4.

[0076] Table 4 (3 represents the number of prediction modules)

[0077] To measure model performance, mean squared error (MSE), mean absolute error (MAE), and coefficient of determination (CDE) are used. These three evaluation metrics are used to quantitatively assess the model's performance. For the first... The first stage The calculation principles for each indicator of each task are shown in the formula.

[0078] ; ; ; in For the number of samples, Indicates the first Phase 1 The average of each task.

[0079] In model training, different hyperparameters affect the model's predictive performance and computational cost. Comparison of experimental results for multiple model parameters revealed that the learning rate and batch size have the greatest impact on model performance. Additionally, the L2 regularization coefficient also has some influence on model performance. Therefore, these three hyperparameters were tuned. During parameter tuning, a random search was first used to determine the approximate parameter range shown in Table 5. Then, within this range, a grid search was used to determine the optimal parameter values. Based on the experimental comparison results, the learning rate was ultimately determined to be... The L2 regularization coefficient is The batch size is 16.

[0080] Table 5

[0081] To verify the effectiveness of the proposed model, two sets of comparative experiments were designed, including an ablation experiment and a multi-stage model comparison experiment. The experimental analysis results were all based on data obtained after data processing, and the average was calculated by repeating the experiment ten times to reduce the impact of randomness in the experimental results. The mean and standard deviation of the evaluation metrics from the ten experiments were analyzed to determine the model's predictive performance. The learning rate, L2 regularization coefficient, and batch size of all comparative models were implemented using random search and grid search. All algorithm models were written in Python 3.8.0 and implemented using torch 2.0.1+cu117.

[0082] 1. Ablation test To verify the contribution of each part of the proposed model to its performance, ablation experiments were conducted to compare the mean values ​​of all tasks across all stages. The clustering design, weight adjustment mechanism, S4 state space, multi-layer transit, and multi-stage prediction methods were removed sequentially to verify the effectiveness of adding each module to the model. The comparison results are shown in Table 6. Figure 8 As shown, Figure 8 This is a mean graph of all stages of an ablation experiment provided by an embodiment of the present invention. The experimental results show that the model performance gradually decreases as components are removed. Converting the multi-stage model to a single-stage model results in a significant performance drop, demonstrating the effectiveness of multi-stage models in this application scenario.

[0083] Table 6

[0084] 2. Multi-stage model comparison experiment The proposed model S4-MSMTAN was experimentally compared with multi-stage single-task models (MS-DNN and MS-CNN) and multi-stage multi-task models (MS-MMoE, MS-PLE, MS-DUPN, MS-MTAN, and MSM-MMOE-S4W) to verify the effectiveness of the proposed model in multi-stage processing scenarios. The comparison results are shown in Table 7 and... Figure 9 As shown, Figure 9 This is a mean plot comparing the proposed model with a multi-task model, provided by an embodiment of the present invention. Among all multi-stage comparison models, the MS-MTAN model has the best prediction performance; therefore, the improvement rate of the proposed model relative to the MS-MTAN model is calculated. For all stages, the average improvement rate of MSE for each task is 33.3%, the average improvement rate of MAE for each task is 20.0%, and the average improvement rate of R2 for each task is 7.5%. By comparing the experimental results of MS-MMoE with MSM-MMOE-S4W and MS-MTAN with S4-MSMTAN, it can be demonstrated that the improvement method of the present invention can also improve the model performance of other multi-task models. Compared with other multi-stage models, the multi-stage model of the present invention uses multi-layer transmission, enriching the input features of the current layer by transmitting the quality prediction data of the previous five layers. The use of the S4 state space improves the model's ability to process high-dimensional feature data, and the combination of an uncertainty-weighted dynamic adjustment method solves the problem of uneven performance of multiple tasks in multi-task learning. These improvements effectively improve the performance of the proposed model.

[0085] Table 7

[0086] 3. Significance test To further verify the effectiveness of the method proposed in this invention, a systematic analysis method using a one-tailed paired t-test was employed for significance testing. The test indices were MSE, MAE, and R², and a null test was established. , , ,in This is the method proposed in this invention. The selected comparison models include the Multi-Stage Single-Task Learning Model (MS-DNN) and the Multi-Stage Multi-Task Learning Model (MSM-MMoE-S4W). The performance of the proposed method is evaluated compared to other comparison methods. For each stage of quality prediction, there are six quality indicators. The mean values ​​of the quality indicators for nine stages were calculated. Therefore, the significance test of the mean results is equivalent to performing a significance test. The cumulative error probability is caused by multiple tests, and for a single stage (stage 15 is selected for comparison in this invention), it is equivalent to performing 6 tests, resulting in a cumulative error probability. Therefore, the Bonferroni method is used for correction, and the significance level of the mean is set to [value missing]. The significance level of the 15-stage test alone was [missing information]. Table 8 shows the results of the one-tailed paired t-test for the selected comparison models. The table reveals that the results are below the significance level for all stages and for stage 15 alone; therefore, the null hypothesis is rejected. , and This strongly demonstrates that the method proposed in this invention is significantly superior to other methods.

[0087] Table 8

[0088] The above experiments were all conducted on data-augmented datasets. To verify the model's performance on the original dataset, this invention compares the proposed model MSM_MTAN_S4W with the five best-performing comparison models on the original dataset. The selected comparison models include MS-MTAN, MSM-MMOE-S4W, MS-DUPN, and MS-PLE. The trained model was then tested on the original real dataset, and fitting plots were generated between the original real data and the prediction results of multiple models, as shown in the figure. Figure 10 As shown, Figure 10 This is a fitting graph of the original real data and the prediction results of multiple models provided in the embodiments of the present invention. It can be clearly seen from the graph that the accuracy of the proposed model is much higher than that of other comparative models, which verifies the superiority of the proposed model S4-MSMTAN.

[0089] It should be noted that, in this document, relational terms such as "first" and "second" are used merely 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 are intended to cover non-exclusive inclusion, such that an article or device comprising a list of elements includes not only those elements but also other elements not expressly listed. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device comprising said element. Terms such as "connected" or "linked" are not limited to physical or mechanical connections but can include electrical connections, whether direct or indirect. The orientations or positional relationships indicated by terms such as "upper," "lower," "left," and "right" are based on the orientations or positional relationships shown in the accompanying drawings and are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be construed as limiting the invention.

[0090] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features or characteristics described may be combined in any suitable manner in one or more embodiments or examples. In addition, those skilled in the art can combine and integrate the different embodiments or examples described in this specification.

[0091] The above description, in conjunction with specific preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. For those skilled in the art, various simple deductions or substitutions can be made without departing from the concept of the present invention, and all such modifications and substitutions should be considered within the scope of protection of the present invention.

Claims

1. A method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning, characterized in that, include: Step 1: Obtain the historical dataset of ultrafast laser multi-stage processing. The historical dataset includes multiple processing samples, each containing process parameter data and corresponding multivariate quality index data for multiple processing stages. Step 2: Perform data augmentation on the historical dataset to increase the number of samples and construct an augmented dataset for model training; Step 3: Based on the augmented dataset, construct and train a multivariate quality prediction model based on multi-stage multi-task learning; Step 4: Input the process parameters of each processing stage of the workpiece to be predicted into the trained multivariate quality prediction model, and output the predicted values ​​of multivariate quality indicators for each processing stage. The multivariate quality prediction model based on multi-stage multi-task learning includes: a shared feature extraction module, a task-specific feature generation module, a clustering tower network prediction module, and an adaptive loss adjustment module connected in sequence. The shared feature extraction module is used to extract shared features based on the process parameters of the current processing stage and the predicted values ​​of multiple quality indicators of the previous processing stage. The task-specific feature generation module is used to generate corresponding weights for the corresponding tasks that predict each quality index, and to use the weights to perform weighted modulation on the shared features to obtain the task-specific features. The clustered tower network prediction module is used to predict all quality indicators of the current processing stage based on the unique characteristics of each task through the tower network of each cluster. The adaptive loss adjustment module is used to adaptively adjust the weight of each task's loss in the total loss based on the uncertainty of each task during the training phase. Before constructing and training the multivariate quality prediction model, task clustering is performed based on task relevance. The task clustering process includes: Based on the historical dataset, multiple correlation evaluation indicators are calculated between the prediction tasks of each pair of quality indicators to form a comprehensive evaluation score; A threshold is set based on the average comprehensive evaluation score among all task pairs. Tasks with comprehensive evaluation scores higher than the threshold are grouped into the same cluster, thereby determining the final task clustering scheme. The clustering scheme is used to fix the number of tower networks in the clustering tower network prediction module of the multivariate quality prediction model and the set of tasks that each cluster tower network is responsible for predicting.

2. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 1, characterized in that, The multi-dimensional quality indicators include at least two of the following: surface roughness, machining depth, and sidewall perpendicularity.

3. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 1, characterized in that, The data augmentation process includes: scan layer stage augmentation and outlier sample processing.

4. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 3, characterized in that, The scan layer stage augmentation process includes: Based on sparsely sampled processing stage data from historical datasets, Hermite interpolation is used to interpolate unsampled processing stages, generating stage-enlarged samples containing process parameter data and multivariate quality index data for each processing stage.

5. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 3, characterized in that, The outlier sample processing includes: based on The principle is to identify outliers in the multivariate quality index data of the historical dataset and mark the stage-expanded samples containing outliers as outlier samples. The outlier samples are augmented using the CTGAN network. The samples generated by the CTGAN network are then filtered based on Euclidean distance. The filtered samples are combined with the outlier samples to form augmented outlier samples. The augmented outlier samples are mixed with the normal samples in the historical dataset after the scanning layer augmentation process. The CTGAN network is then used to augment the samples, and the samples generated by the CTGAN network are filtered based on Euclidean distance. The filtered samples are then combined with the historical dataset after the scanning layer augmentation process to form the augmented dataset.

6. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 1, characterized in that, The shared feature extraction module includes: a series of cascaded linear layers, a ReLU activation function layer, an S4 structured state-space model loop, a batch normalization layer, and a Dropout layer.

7. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 1, characterized in that, The adaptive loss adjustment module sets a learnable log-variance parameter for each quality metric prediction task; the total loss of model training is calculated using the following formula: ； In the formula, This represents the total loss during model training. Indicates the total number of tasks. Indicates the first The total loss across all stages of the task. Indicates the first The learnable log-variance parameter corresponding to each task Indicates the first The weight of each task, .

8. The method for predicting the quality of ultrafast laser processing of complex microstructures based on multi-task learning according to claim 1, characterized in that, In step 4, For the prediction of the first processing stage, the model input is the process parameters of the first processing stage; For the Processing stage prediction, when At that time, the model input includes the first Process parameters and preceding steps in the processing stage Predicted values ​​of multi-source quality indicators for each processing stage; when At that time, the model input includes the first Process parameters of the processing stage and the first Pre-processing stage Predicted values ​​of multiple quality indicators for each processing stage, among which The number of information transmission stages, which is an integer greater than or equal to 1.

Images