Laser residual life estimation method based on adaptive fusion mechanism and dynamic weighting

Abstract

The invention discloses a laser residual life estimation method based on an adaptive fusion mechanism and dynamic weighting. According to the method, a two-stage migration strategy is adopted. The method comprises the following steps: firstly, pre-training a model by utilizing complete degradation data of a source domain to obtain universal degradation characteristic representation; then, bottom-layer feature extraction network parameters of the pre-trained model are frozen, and a small amount of early degradation data of a target domain is utilized to perform fine adjustment on an adaptive fusion module and a top-layer prediction network of the model based on a dynamic weighting strategy; and finally predicting the residual life of the target domain. Aiming at the problem of low prediction precision of a gallium nitride laser under multiple working conditions, efficient migration of a model from a single source domain to multiple target domains is realized through adaptive fusion and a dynamic weighting mechanism, the problems that a traditional migration learning method is weak in generalization ability and depends on a large amount of target domain data are effectively solved, and the prediction accuracy of the gallium nitride laser under multiple working conditions is improved. And the estimation precision and robustness in a cross-working-condition scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of laser technology, and in particular to a laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting. Background Technology

[0002] Gallium nitride (GaN) materials have shown broad application prospects in optoelectronics, energy, and communications due to their excellent physical and chemical properties. Gallium nitride-based lasers, as a key representative, have become a hot topic in the field of optoelectronic devices due to their high efficiency, high power density, and long lifetime.

[0003] High-power lasers are typical degrading systems, with their output optical power gradually decreasing over time. When the optical power falls below a set threshold, the laser is considered to be unable to meet the preset requirements and is in a degraded and failed state. Accurately predicting its optical power degradation is crucial for lifetime prediction.

[0004] The remaining useful life (RUL) estimate of high-power gallium nitride lasers plays a crucial role in the reliable and safe operation of these lasers. Laser lifetime is defined as the operating time during which its output optical power decreases to a failure threshold.

[0005] Traditional methods involve applying high operating stress without altering the failure mechanism, testing the failure time, and then extrapolating the lifetime under normal stress using an accelerated model. This method cannot achieve online lifetime prediction.

[0006] Another type of approach is based on data-driven methods. These methods train a prediction model using labeled, degraded data to predict the laser's output optical power and thus its remaining lifetime. With the continuous development of deep learning technology, data-driven methods have become widely used, among which transfer learning has been extensively researched and applied due to its ability to transfer domain knowledge.

[0007] However, the mainstream paradigm of existing transfer learning methods uses offline training based on labeled source domain data and unlabeled static target domain data. The model parameters exhibit fixed, static characteristics during deployment, making it difficult to effectively adapt to the working conditions of the target application and limiting the model's prediction accuracy. Secondly, these methods merely transfer a model trained under one working condition to another; if the target working condition changes, retraining is required. This paradigm restricts the application of transfer learning models.

[0008] This invention employs an adaptive fusion and dynamic weighting method to achieve effective prediction of laser degradation patterns by the source domain model under different actual operating conditions. Summary of the Invention

[0009] To address the problems existing in the prior art, the purpose of this invention is to provide a laser remaining lifetime estimation method based on an adaptive fusion mechanism and dynamic weighting.

[0010] To solve the above problems, the present invention adopts the following technical solution.

[0011] A laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire source domain degradation data and target domain degradation data of the laser; Step S2, Model Construction: Construct a lifetime prediction model, which includes a low-level feature extraction network, an adaptive fusion module, and a high-level prediction network; Step S3, Pre-training: Use existing source domain degradation data to pre-train the model, enabling the model to extract complete degradation features; Step S4, parameter fine-tuning: Freeze the parameters of the low-level feature extraction network of the pre-trained model, and use a small amount of early degradation data of the target domain to fine-tune the parameters of the adaptive fusion module and the top-level prediction network of the model based on the dynamic weighting strategy. Step S5, Testing and Evaluation: Test the fine-tuned model using the remaining life cycle data of the target domain operating conditions, and output the remaining life estimation results.

[0012] Furthermore, the specific steps of pre-training in step S3 include: 1. Load the preprocessed source domain training set and initialize the model parameters; 2. Set the hyperparameters required for model training, mainly including the maximum number of training epochs and the learning rate; 3. Input the source domain training data into the model and start iterative training. In one training round, the model calculates the predicted value through forward propagation and compares it with the true label to calculate the loss. Then, it updates the model parameters through backpropagation. 4. After each training round, determine whether the current number of training rounds has exceeded the preset maximum number of training rounds; If "No" is answered, it means that the training is not yet complete. The process returns to step 3 to continue the next round of model training. If "yes" indicates that all preset training rounds have been completed, the training process will terminate, and proceed to step 5. 5. After the training loop ends, save the current model weights and parameters to finally obtain a source domain pre-trained model that has been fully trained on the source domain data. Furthermore, the specific steps for parameter fine-tuning in step S4 include: S41: Data Loading Load a small amount of early degradation data from N target domains as the training set, and the corresponding remaining full lifecycle data as the test set; S42: Model Loading and Configuration Load the model weights pre-trained on the source domain data. In order to preserve the general degradation features learned by the pre-trained model in the source domain, freeze the backbone network of the model so that its parameters will not change during the subsequent fine-tuning process. At the same time, set the training parameters required for the fine-tuning stage, including the maximum number of training epochs and the learning rate. S43: Iterative training of the target domain in turn Within each training epoch, the model will perform batch traversal of these N target domain data, for the currently selected target domain... The model will use its corresponding adaptive module Set to trainable state; simultaneously, all other adaptive modules corresponding to the target domain will be frozen to ensure that updates are performed correctly. At that time, other modules are unaffected. S44: Forward Propagation, Loss Calculation, and Dynamic Weighting From the currently selected target domain A batch of data is taken out and input into the model. The model uses a dynamic weighting strategy to weight and fuse the outputs of each sub-estimater to obtain the final model output. Then, the root mean square error (RMSE) between the predicted value of the final model output and the true label is calculated, and this RMSE value is used as the total loss of the batch of data. S45: Model Parameter Update Based on the calculated total loss, the gradient is computed using the backpropagation algorithm. Then, the model parameters are updated using the stochastic gradient descent algorithm. In this step, only the adaptive modules that are currently in a trainable state are updated. Only then will the parameters be updated; S46: Target domain traversal judgment After completing the training and parameter update of a batch of data, it is necessary to determine whether all N target domains have been traversed within the current training cycle. If "No", the process returns to step S43 to select the next target domain. And continue with dynamic configuration and training. If "yes", it means the current training cycle has been completed. Proceed to step S47; S47: Training Round Determination After completing a training cycle, it is necessary to determine whether the current total number of training rounds has exceeded the preset maximum number of training rounds. If "No", it means that the training has not yet reached the preset termination condition. The process returns to step S43 and starts the next training round, restarting the iterative training of all N target domains in turn. If "yes" is answered, it means that the entire training process has reached the maximum number of rounds, and the training will terminate.

[0013] Furthermore, the specific steps of testing and evaluation in step S5 include: (1) Predicted output: The test data for each target working condition is input into the model after fine-tuning in step S4. The model outputs the estimated remaining lifetime of the test samples under each target working condition through forward propagation. (2) Calculation of indicators: Based on the estimated remaining life and the corresponding true value, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are calculated as evaluation indicators, as shown in equations (8) and (9).

[0014] (8) (9) in, This represents the number of samples for the k-th target operating condition; and Let RUL represent the estimated and actual values ​​of the i-th sample for the k-th target working condition, respectively.

[0015] The adaptive module is a lightweight incremental structure designed for RUL estimation under specific target conditions. It can take data from unknown conditions as input and adapt it to a pre-trained Informer model by adding a small number of parameters. It includes two fully connected layers. The first fully connected layer takes the output of the Informer's encoder as input and reduces the original input dimension. Projected to By controlling The size of the parameter is used to limit the number of parameters in the adaptive module, typically... Then, the middle of the module is through a non-linear layer, which is implemented by an activation function, as shown in Equation (10).

[0016] (10) In the output phase, the module restores the input dimension through a second FCN, Remapping back As the output of the module, a residual connection is used to add the input of the adaptive module to the final output, as shown in Equation (11). This ensures the effectiveness of training even if the parameters of the adaptive module are initially initialized close to 0, through the setting of an approximate identity mapping.

[0017] (11) in It is the output of the encoder. This is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer.

[0018] As can be seen from the above formula, FCN is implemented by a single linear layer and does not have a bias term, which further reduces the number of model parameters and improves the model's efficiency. After training multiple adaptive modules, an adaptive fusion mechanism is used to integrate their outputs to achieve a better feature representation. This mechanism first calculates the attention score of each adaptive module with the input feature, then uses the Softmax function to obtain weights in the range [0-1]. Finally, the final feature representation is obtained by weighting and summing the weights and the outputs of the adaptive modules, as shown in Equation (12). (12) in ( () represents the output of the nth adaptive module. This indicates a splicing operation. This represents the output of the M integrated adaptive modules. Indicates the output of the distillation layer. This indicates a splicing operation. This represents the concatenated feature representation. This represents the fused feature representation after using the adaptive fusion mechanism. It is the weight matrix of the adaptive fusion mechanism, which performs a linear transformation on the output of the fusion mechanism. It is a bias term in linear transformation.

[0019] The RUL estimator consists of three parts. The first part is the projection layer, which uses the final output of the decoder as input to reduce the risk of model overfitting. The second part is the sub-estimator layer, the number of which corresponds to the number M of target conditions. It can generate M RUL estimates simultaneously. Each sub-estimator includes a Dense layer and two activation function layers. The Tanh activation function is located before the Dense layer and learns the non-linear relationship between features and labels. The Sigmoid activation function is located after the Dense layer and maps the output value to the range [0,1]. The third part is dynamic weighting, which calculates the weights of the M estimates based on the mean absolute error (MAE) of the prediction results, assigns weights to the estimates, and calculates the final RUL estimate. The calculation process is as follows: in This represents the mean absolute error between the estimate and the true value of the k-th sub-estimator; The set of MAEs is the result of the M sub-estimators sorted in ascending order; Pos is the initial index set after sorting; p represents the index of the initial error at the k-th position after sorting. It is the weight of the output of the k-th sub-estimator. According to the calculation process, the sub-estimator with smaller error is assigned a higher weight, and the sum of all weights is 1. The final RUL estimate of the model can be expressed as Equation (13).

[0020] (13) in, and These are the weights and estimates of the i-th estimator, respectively. This represents the model's final predicted value.

[0021] Beneficial effects of the present invention Compared with the prior art, the advantages of this invention are: This invention addresses the issue of low generalization ability and prediction accuracy of gallium nitride laser models under various operating conditions in traditional transfer learning methods. The new model can transfer a model trained under one operating condition to multiple operating conditions, overcoming the limitation of traditional methods that only apply to a single operating condition, thus further enhancing the generalization ability of the estimation model. This method employs a two-stage training strategy based on pre-training and parameter fine-tuning to share and optimize model parameters. Attached Figure Description

[0022] Figure 1 A flowchart illustrating an overall method for estimating the remaining lifetime of a laser based on an adaptive fusion mechanism and dynamic weighting, provided in this embodiment of the invention. Figure 2 This is a flowchart illustrating the specific process of source domain pre-training in step S3 of this embodiment of the invention.

[0023] Figure 3 This is a flowchart illustrating the parameter fine-tuning process in step S4 of this embodiment of the invention.

[0024] Figure 4 The diagram shows the model structure of a laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting, provided in an embodiment of the present invention. Detailed Implementation

[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0026] Please see Figures 1 to 4 This paper assumes the existence of degradation data under various operating conditions, with corresponding data labels. This represents the existing complete degraded dataset in the source domain, where Indicates the number of samples in the source domain. and Indicates the first in the source domain The degradation data of each sample and its corresponding capacity label. This represents incomplete degraded data in the target domain under M different operating conditions, where Indicates the number of target domains. Indicates the first Degradation data of the laser in each target domain Indicates the first Number of samples in each target domain and Indicates the first In the target domain, the first The study focuses on degraded data of individual samples and their corresponding capacity labels. It primarily employs transfer learning to transfer the model from one source domain to multiple target domains, enabling RUL prediction under various operating conditions.

[0027] Method Description The proposed structure To achieve laser RUL prediction under various operating conditions, this paper proposes a multi-objective domain transfer learning method based on adaptive fusion mechanism and dynamic weighting, the structure of which is as follows: Figure 4 As shown, the core of this structure is an Informer encoder-decoder architecture. We added M adaptive modules corresponding to different target conditions to each encoder and decoder, resulting in M ​​adaptive modules. An adaptive fusion mechanism integrates the outputs of all adaptive modules to adaptively extract degradation features under various conditions. Finally, the decoder output is input to the RUL estimator to obtain the RUL estimate for the current loop.

[0028] Degeneracy feature extraction based on Informer Informer's network structure mainly consists of two parts: an encoder and a decoder. Figure 4 (a)(d).

[0029] Sparse autonomy mechanism The sparse self-attention mechanism is based on sparsity measure, which retains only the u keys that contribute the most to the query, while other positions are ignored, thereby reducing the computational scale. The sparsity measure is calculated as shown in Equation (14).

[0030] (14) in, It is a query vector. It is a key vector. It is the dimension of the embedded vector. is the total number of keys, and d represents the hidden dimension of the attention space. The sparsity measure evaluates the overall strength of the dot product values ​​of all keys by using logarithmic and exponential terms, and calculates the average of all dot products to compare the concentration of the attention distribution. If a query has a large sparsity measure value, it means that its attention distribution is more diverse, and there are significant dot products that contribute more; conversely, if the measure value is small, the distribution is closer to uniform, and the contribution is weaker. To further reduce computational complexity, a fast approximation method is used to calculate the measure: randomly sampling from the key set K. The dot product is calculated sparsityally, where c is a sampling factor. The approximate formula is shown in Equation (15), which uses a combination of maximum and mean values ​​to avoid a complete calculation of each dot product.

[0031] (15) The results of the sparsity metric calculation are used to filter the query points that contribute the most, retaining the top [number]. A subset is selected after a sparse query. After filtering by the above sparsity measures, the sparse self-attention mechanism can be expressed as in equation (16).

[0032] (16) in, It is a sparse subset selected from query Q. It is a hidden dimension of the attention space.

[0033] Distillation mechanism In the Informer, a CNN-based distillation mechanism is used to prioritize dominant features, aiming to generate a concentrated self-attention feature map in the next layer. This mechanism can significantly reduce the length of the model's input features, thereby improving the model's robustness and reducing its memory footprint. The calculation from layer j to layer (j+1) is as shown in Equation (17).

[0034] (17) in, It is the first The input features of the layer; Use convolutional kernels to extract local features along the time axis; As an activation function, it enhances the nonlinear expressive power; Pooling is performed on the time dimension with a fixed step size, reducing the sequence length by half.

[0035] The encoder forms a "pyramid structure" by stacking multiple downsampling layers and fuses the output features of all downsampling layers, as shown in Equation (18): (18) in, It is the first Downsampling output of the layer; This represents the feature concatenation operation. By fusing features at different scales, the model can simultaneously focus on short-term and long-term dependencies. The distillation mechanism is integrated into the encoder and works in conjunction with the sparse self-attention mechanism. The initial input is processed by the sparse self-attention mechanism to extract short-term dependencies. After reducing the time dimension using downsampling, long-term dependencies are further captured. Finally, the downsampled outputs of each layer are concatenated and fused to form the final encoder output, providing multi-scale features for the decoder.

[0036] Adaptive modules and adaptive fusion mechanisms The adaptive module is a lightweight incremental structure designed for RUL estimation under specific target working conditions. It can take data from unknown working conditions as input and adapt it to a pre-trained Informer model by adding a small number of parameters. Its structure is as follows: Figure 4 As shown in (b), it mainly consists of two fully connected layers (FCN). The first FCN takes the output of the Informer's encoder as input and converts the original input dimension... Projected to By controlling The size of the parameter is used to limit the number of parameters in the adaptive module. Typically... Then, the middle of the module passes through a non-linear layer, which is implemented by an activation function, as shown in equation (19).

[0037] (19) In the output phase, the module restores the input dimension through a second FCN, Remapping back The input of the adaptive module is used as the output of the module. Simultaneously, a residual connection is used to add the input of the adaptive module to the final output, as shown in Equation (20). This ensures the effectiveness of training even if the parameters of the adaptive module are initially initialized close to 0, through an approximate identity mapping.

[0038] (20) in It is the output of the encoder. This is the weight matrix of the first fully connected layer. It is the weight matrix of the second fully connected layer.

[0039] As can be seen from the above formula, FCN is implemented by a linear layer and no bias term is set, which further reduces the number of model parameters and improves the efficiency of the model.

[0040] After training multiple adaptive modules, an adaptive fusion mechanism is used to integrate their outputs to achieve better feature representation, as shown in the following structure. Figure 4 (c). The mechanism first calculates the attention score of each adaptive module with the input feature, then uses the Softmax function to obtain the weights in the range of [0-1], and finally obtains the final feature representation by weighted summation of the weights and the output of the adaptive module, as shown in Equation (21).

[0041] (twenty one) in ( () represents the output of the nth adaptive module. This indicates a splicing operation. This represents the output of the M integrated adaptive modules. Indicates the output of the distillation layer. This indicates a splicing operation. This represents the concatenated feature representation. This represents the fused feature representation after using the adaptive fusion mechanism. It is the weight matrix of the adaptive fusion mechanism, which performs a linear transformation on the output of the fusion mechanism. It is a bias term in linear transformation.

[0042] SOH estimation based on dynamic weighting like Figure 4 (e) and (f), the RUL estimator consists of three parts. The first part is the projection layer, which uses the final output of the decoder as input to reduce the risk of model overfitting. The second part is the sub-estimator layer, the number of which corresponds to the number M of target load conditions, and can generate M RUL estimates simultaneously, such as... Figure 4(f) Each sub-estimator includes a Dense layer and two activation function layers. The Tanh activation function is located before the Dense layer and learns the non-linear relationship between features and labels; the Sigmoid activation function is located after the Dense layer and maps the output value in the range [0,1]; the third part is dynamic weighting, which calculates the weights of M estimates based on the mean absolute error (MAE) of the prediction results, assigns weights to the estimates, and calculates the final RUL estimate. The calculation process is as shown in Equation (22).

[0043] (twenty two) in This represents the mean absolute error between the estimate and the true value of the k-th sub-estimator, where n is the number of data samples in the current batch. and Let represent the estimated value and the corresponding true value of the i-th sample data in the k-th sub-estimator, respectively; The set of MAEs is the result of the M sub-estimators sorted in ascending order; Pos is the initial index set after sorting; p represents the index of the initial error at the k-th position after sorting. This is the weight of the output of the k-th sub-estimator. According to the calculation process, the sub-estimator with the smaller error receives a higher weight, and the sum of all weights is 1. The final RUL estimate of the model is calculated as shown in equation (23).

[0044] (twenty three) in, and These are the weights and estimates of the i-th estimator, respectively. This represents the final estimated value of the model.

[0045] Model training This invention uses a transfer learning method based on pre-training fine-tuning to train the model. For example... Figure 1 As shown, the main process includes two stages: the first is the pre-training stage in step S3, which uses existing complete degradation data to pre-train the model; the second is the parameter fine-tuning stage in step S4, which freezes the parameters of the bottom-level network of the pre-trained model and uses a small amount of early degradation data under the target condition to fine-tune the parameters of the top-level network layer of the model. Implementation method: Two-stage training strategy 3.1 Pre-training phase: (Refer to...) Figure 2 The pre-training phase of this method includes the following steps: 1. Load the preprocessed source domain training dataset and initialize the model parameters; 2. Set the hyperparameters required for model training, mainly including the maximum number of training epochs (max_epoch) and the learning rate. 3. Input the preprocessed source domain training data into the model and begin iterative training. In one training epoch, the model calculates the predicted values ​​through forward propagation and compares them with the true labels to calculate the loss. The calculation is shown in Equation (24), and then the model parameters are updated through backpropagation; (twenty four) in, and Let represent the estimated and actual remaining lifetime of the i-th sample in the source domain data, respectively. This indicates the total number of samples in the source domain data.

[0046] 4. After each training round, determine whether the current training round number has exceeded the preset maximum training round number (max_epoch). If "No" (current epoch number ≤ max_epoch): This means that training has not yet been completed, and the process returns to step 3 to continue the next round of model training; If "yes" (current epoch > max_epoch): This means that all preset training epochs have been completed, the training process terminates, and proceeds to step 5; 5. After the training loop ends, save the current model weights and parameters to finally obtain a source domain pre-trained model that has been fully trained on the source domain data. 3.2 Parameter fine-tuning stage: Reference Figure 3 The fine-tuning phase involves adjusting the parameters of the adaptive module and the RUL estimator. This includes the following steps: S41: Data Loading Load the parameters of the pre-trained model in the source domain, load a small amount of early degradation data from N target domains as the training set, and load the corresponding remaining full lifecycle data as the test set. S42: Initialize and fine-tune the configuration Freeze the parameters of the model's underlying backbone network (Informer structure) to prevent them from changing during subsequent fine-tuning. Simultaneously, set the training parameters required for the fine-tuning phase, including the maximum number of training epochs (max_epoch) and the learning rate, and initialize a global epoch counter. S43: Iterative training of the target domain in turn Within each training epoch, the model will perform batch processing on these N target domain datasets. For the currently selected target domain... The model will use its corresponding adaptive module Set to a trainable state. Meanwhile, all other target domains ( ,in The corresponding adaptive module will be frozen to ensure that updates are performed correctly. At this time, other modules are unaffected; S44: Forward Propagation, Loss Calculation, and Dynamic Weighting From the currently selected target domain A batch of data is extracted and input into the model. The model uses a dynamic weighting strategy to weight and fuse the outputs of its internal sub-estimators to obtain the final model output. Next, the root mean square error (RMSE) between the predicted value of the final model output and the true label is calculated, and this RMSE value is used as the total loss for that batch of data. S45: Model Parameter Update Based on the calculated total loss, the gradient is computed using the backpropagation algorithm. Then, the model parameters are updated using stochastic gradient descent. In this step, only the adaptive module currently in a trainable state ( Only parameters that are specified will be updated; S46: Target domain traversal judgment After completing the training and parameter update of a batch of data, it is necessary to determine whether all N target domains have been traversed within the current training epoch. If "No" (i.e., not all N target domains have been cycled), the process returns to step S43 to select the next target domain. ), and continue dynamic configuration and training; If "yes", it means the current training epoch has been completed. Proceed to step S47; S47: Training Round Determination After completing a training cycle (epoch), it is necessary to determine whether the current total number of training epochs has exceeded the preset maximum number of training epochs (max_epoch).

[0047] If "No" is selected (i.e., the current epoch number is ≤ max_epoch), it means that the training has not yet reached the preset termination condition. The process returns to step S43 to start the next training epoch and restart the iterative training of all N target domains in turn.

[0048] If "yes" (i.e., current epoch > max_epoch), it means that the entire training process has reached the maximum number of epochs, and training terminates.

[0049] This paper proposes a multi-target domain laser RUL estimation method based on adaptive fusion mechanism and dynamic weighting, which improves the low generalization ability and prediction accuracy of gallium nitride lasers under various operating conditions in traditional transfer learning methods. This model can transfer a model trained under one operating condition to multiple operating conditions, solving the problem that traditional methods can only transfer to a single operating condition, and further enhancing the generalization ability of the estimation model. The method employs a two-stage training strategy based on pre-training and parameter fine-tuning to share and optimize model parameters.

[0050] First, the model is pre-trained using complete degradation data to enable it to learn the complete degradation trajectory of the laser. Then, early degradation data collected under various unknown operating conditions are input into the model to fine-tune the corresponding module parameters, thereby improving the prediction accuracy under the target operating conditions.

[0051] This paper designs a network called the Adaptive Module for extracting laser degradation features. During training, a corresponding adaptive module is assigned to different operating conditions. During fine-tuning, the parameters are optimized by alternately inputting corresponding target data. Furthermore, while optimizing the module for one operating condition, modules for other operating conditions are frozen to avoid catastrophic forgetting of parameters. Simultaneously, an adaptive fusion mechanism based on multi-head attention is proposed to fuse the outputs of these multiple adaptive modules, further obtaining the feature representation of the laser's RUL under various operating conditions. Finally, the prediction results of all sub-estimators in the RUL estimator are weighted and calculated to obtain the RUL estimate for the target operating condition. The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.

Claims

1. A laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting, characterized in that, Includes the following steps: Step S1, Data Acquisition and Preprocessing: Acquire the degradation data of the laser and the degradation data of the target domain; Step S2, Model Construction: Construct a lifetime prediction model, which includes a bottom-level feature extraction network Informer, a top-level prediction network adaptive fusion module, and a RUL estimator based on a dynamic weighting strategy; Step S3, Pre-training: The model is pre-trained using the source domain degradation data to enable the model to extract complete degradation features; Step S4, parameter fine-tuning: Freeze the parameters of the low-level feature extraction network of the pre-trained model, and fine-tune the parameters of the top-level prediction network of the model using a small amount of early degradation data of the target domain. Step S5, Testing and Evaluation: Test the fine-tuned model using the remaining life cycle data of the target domain operating conditions, and output the remaining life estimation results and accuracy indicators.

2. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The specific steps of data acquisition and preprocessing in step S1 include: (1) Collect monitoring parameters of the laser throughout its entire life cycle under source domain and target domain conditions; (2) Linear interpolation and maximum-minimum normalization methods are used to preprocess the monitoring parameters; (3) Using the sliding time window technique, the normalized time series data is divided into multiple fixed-length subsequences to construct source domain datasets and target domain datasets for model training and testing.

3. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The underlying feature extraction network Informer in step S2 includes an encoder and a decoder: (1) The encoder uses a multi-head probabilistic sparse self-attention mechanism to extract the long-range temporal dependencies of the feature sequence and outputs the encoded feature vector; (2) The decoder receives the encoded feature vector and outputs the predicted hidden layer feature representation in a generative manner.

4. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The adaptive fusion module in step S2 includes an adaptive module and an adaptive fusion mechanism: (1) The adaptive module is a lightweight incremental structure designed for RUL estimation under specific target working conditions. It can take the data of unknown working conditions as input and adapt it to the Informer model by adding a small number of parameters. Its structure includes two fully connected layers (FCN). The first fully connected layer takes the output of the Informer encoder as input and changes the original input dimension. Projected to By controlling The size of the parameter is used to limit the number of parameters in the adaptive module, typically... ≫ Then, between the two fully connected layers of the module is a non-linear layer implemented by an activation function, as shown in equation (1): （1） In the output phase, the module restores the data to the input dimension through a second fully connected layer. Remapping back As the output of the module, a residual connection is used to add the input of the adaptive module to the final output, as shown in Equation (2). This ensures the effectiveness of training even if the parameters of the adaptive module are initially close to 0, through the setting of an approximate identity mapping. （2） in It is the output of the encoder. This is the weight matrix of the first fully connected layer. This is the weight matrix of the second fully connected layer. As can be seen from the above formula, FCN is implemented by a single linear layer and does not have a bias term, which further reduces the number of model parameters and improves model efficiency. (2) The adaptive fusion mechanism integrates the outputs of multiple adaptive modules to achieve better feature representation. The mechanism first calculates the attention score of each adaptive module with the input feature, then uses the Softmax function to obtain the weights in the range of [0-1]. Finally, the final feature representation is obtained by weighted summation of the weights and the outputs of the adaptive modules, as shown in Equation (3): (3) in ( () represents the output of the nth adaptive module. This indicates a splicing operation. This represents the output of n integrated adaptive modules. This represents the output of the encoder in the Informer network. This indicates a splicing operation. This represents the concatenated output representation, where F represents the fused feature after using the adaptive fusion mechanism. It is the weight matrix of the adaptive fusion mechanism, which performs a linear transformation on the output of the fusion mechanism. It is a bias term in linear transformation.

5. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The RUL estimator in step S2 consists of three parts: a projection layer, a sub-estimator layer, and a dynamic weighting strategy. (1) The projection layer uses the output of the decoder as input to reduce the risk of model overfitting; (2) The number of sub-estimator layers corresponds to the number of target working conditions M, and can generate M RUL estimates simultaneously. Each sub-estimator includes a Dense layer and two activation function layers. The Tanh activation function is located before the Dense layer and learns the non-linear relationship between features and labels; the Sigmoid activation function is located after the Dense layer and maps the output value in the range [0,1]. (3) The dynamic weighting strategy calculates the weights of M estimates based on the mean absolute error (MAE) of the prediction results, assigns weights to the estimates, and calculates the final RUL estimate. The calculation process is as shown in equation (4): （4） in The MAE represents the sum of the estimated and true values ​​of the k-th sub-estimator. and represent the estimated value and the true value of the i-th sample in the k-th sub-estimator of the input dataset, respectively, where n is the total number of samples in the current batch of data; It is the set of results from the M sub-estimators sorted in ascending order; Pos is the set of indices of the initial position and the current position after sorting. This represents the index of the initial error at the k-th position after sorting. This represents the weight of the output of the k-th sub-estimator. As the calculation process shows, the sub-estimator with the smaller error receives a higher weight, and the sum of all weights is ultimately 1. The final RUL estimate of the model can be expressed as: in, and These are the weights and estimates of the i-th estimator, respectively. It is the model's final predicted value.

6. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The specific pre-training process in step S3 includes the following steps: (1) Data loading and initialization: Directly call the source domain dataset processed in step S1 and initialize the network parameters of the model; (2) Parameter settings: Set the hyperparameters required for model pre-training, mainly including the maximum number of training rounds and the learning rate; (3) Iterative training: Input the source domain training data into the model and start iterative training. In one training round, the model calculates the predicted value through forward propagation and compares it with the true label to calculate the loss. The calculation is performed as shown in equation (5), and then the model parameters are updated through backpropagation: （5） in and Let represent the estimated and true remaining lifetime of the i-th sample in the source domain training dataset, respectively. This represents the number of samples in the source domain training dataset; (4) Termination judgment: After each training round, determine whether the current number of training rounds has exceeded the preset maximum number of training rounds; If "No" indicates that the training is not yet complete, the process returns to step (3) to continue the next round of model training; If "yes" indicates that all preset training rounds have been completed, the training process ends, and proceeds to step (5). (5) Model saving: After the training cycle ends, the current model weights and parameters are saved, and finally a source domain pre-trained model that has been fully trained on the source domain data is obtained.

7. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The specific process for parameter fine-tuning in step S4 includes the following steps: (1) Data loading: Directly call the target domain dataset processed in step S1; (2) Model loading and configuration: Load the pre-trained model weights on the source domain data, freeze the underlying Informer network of the model, and set the training parameters required for the fine-tuning stage, including the maximum number of training rounds and the learning rate. (3) Alternating training of target domains: Start the training round loop, and in each training round, traverse N target domain data in turn; for the currently selected target domain, the model will set its corresponding adaptive module to a trainable state. At the same time, the adaptive modules corresponding to all other target domains will be frozen to ensure that other modules are not affected when updating. (4) Forward propagation and weighted fusion: Take a batch of data from the currently selected target domain and input it into the model. Each sub-estimater in the model calculates the initial prediction value. The model calculates the fusion weight of each sub-estimater according to the dynamic weighting strategy and performs a weighted summation on the initial prediction value to obtain the final remaining lifetime prediction value. (5) Loss calculation and parameter update: Calculate the root mean square error (RMSE) between the final remaining lifetime prediction value and the true label, and use the RMSE value as the total loss function for the current batch; in this step, only the network parameters of the top-level prediction network of the currently selected target domain are updated; (6) Target domain traversal judgment: After completing the training and parameter update of a batch of data, it is necessary to determine whether all M target domains have been traversed in the current training cycle; If "No", the process returns to step (3), selects the next target domain, and continues dynamic configuration and training; If "yes", the process proceeds to step (7); (7) Training round judgment: After completing a training cycle, it is necessary to judge whether the current total number of training rounds has reached the preset maximum number of training rounds; If "no", the process returns to step (3) and begins the next training round, restarting the iterative training of all M target domains in turn; If "yes", the fine-tuning process is terminated, and the final model containing M target domain-specific adaptive fusion modules is output.

8. The laser remaining lifetime estimation method based on adaptive fusion mechanism and dynamic weighting according to claim 1, characterized in that: The specific process for testing and evaluation in step S5 includes the following steps: (1) Predicted output: The test data for each target working condition is input into the model after fine-tuning in step S4. The model outputs the estimated remaining lifetime of the test samples under each target working condition through forward propagation. (2) Calculation of indicators: Based on the estimated remaining life and the corresponding true value, the root mean square error (RMSE) and mean absolute percentage error (MAPE) are calculated as evaluation indicators, as shown in equations (6) and (7): （6） （7） in, This represents the number of samples for the k-th target operating condition; and Let RUL represent the estimated and actual values ​​of the i-th sample for the k-th target working condition, respectively.

Images