LED spectrum prediction method based on one-dimensional convolutional neural network

Abstract

This application relates to the field of lighting technology, and in particular to an LED spectral prediction method based on a one-dimensional convolutional neural network. The method includes: acquiring data from a monochrome LED chip, which includes multi-dimensional parameters and corresponding spectral data; preprocessing the monochrome LED chip data, including spectral wavelength axis alignment, smoothing filtering, feature standardization, feature selection, and output data decoupling, to obtain a training dataset; constructing a one-dimensional convolutional neural network model, which includes a residual block-based shared convolutional feature extractor and three independent prediction branches, outputting normalized spectral shape, spectral intensity, and photoelectric conversion efficiency prediction results, respectively; and training the model using decoupling loss functions, including spectral shape loss, spectral smoothness constraint loss, spectral intensity loss, and photoelectric conversion efficiency loss. This method solves the problems of feature redundancy, failure to fully exploit local correlation information of input features, and inability to simultaneously constrain multiple prediction targets.

Description

Technical Field

[0001] This application relates to the field of lighting technology, and in particular to an LED spectrum prediction method based on a one-dimensional convolutional neural network. Background Technology

[0002] As a core component of multi-color lighting systems, the spectral power distribution (SPD) and photoelectric conversion efficiency (WPE) of monochrome LEDs directly determine the downstream LED chip combinations and the light distribution effect of the luminaires. Currently, the mainstream method for obtaining monochrome LED spectral data is experimental measurement, which involves collecting spectral data point by point using a spectrometer for specific types and packages of chips under different drive currents.

[0003] To address the limitations of experimental measurements, the industry has recently attempted to introduce deep learning methods into LED spectrum prediction. However, existing models mostly employ a single fully connected network structure, failing to fully exploit the local correlations between input features (such as dominant wavelength, chip size, and operating current). Furthermore, their loss functions are often simplistic, unable to simultaneously constrain the waveform consistency of the spectrum shape and the numerical accuracy of the spectral intensity, resulting in significant deviations in both waveform matching and absolute intensity of the predicted spectrum. In addition, existing methods do not consider the redundancy of input features; directly inputting all features into the model may introduce noise, affecting prediction accuracy and model generalization ability. Summary of the Invention

[0004] To address the aforementioned technical problems, this application provides an LED spectrum prediction method based on a one-dimensional convolutional neural network, comprising: Acquire data from a monochrome LED chip, which includes multi-dimensional parameters and corresponding spectral data. The data from the monochrome LED chip is preprocessed, including spectral wavelength axis alignment, smoothing filtering, feature standardization, feature selection, and output data decoupling, to obtain the training dataset. A one-dimensional convolutional neural network model is constructed, which includes a shared convolutional feature extractor based on residual blocks and three independent prediction branches, which output three prediction results: normalized spectral shape, spectral intensity and photoelectric conversion efficiency, respectively. The model is trained using a decoupling loss function, which includes spectral shape loss, spectral smoothness constraint loss, spectral intensity loss, and photoelectric conversion efficiency loss.

[0005] In one embodiment, spectral wavelength axis alignment aligns all spectral data to a unified band; smoothing filtering is used to eliminate measurement noise, and the smoothing filtering employs at least one of Savitzky-Golay filter, moving average filter, Gaussian filter, median filter, or low-pass filter; feature selection is used to remove redundant features, and the feature selection employs at least one of Copula entropy-based feature selection method, mutual information method, analysis of variance method, LASSO regression method, random forest feature importance method, or principal component analysis method; output data decoupling processing includes: normalizing, standardizing, or logarithmic transformation of spectral shape, spectral intensity, and photoelectric conversion efficiency, respectively.

[0006] In one embodiment, the Copula entropy-based feature selection method measures the independence and redundancy of features by calculating the Copula entropy between input features, quantifies the information contribution of features to the prediction target, and identifies redundant features by calculating the Copula entropy between features in order to eliminate redundant features.

[0007] In one embodiment, the shared convolutional feature extractor includes an ingress convolutional layer and multiple residual blocks. Each residual block contains two layers of one-dimensional convolution, normalization, activation function, and regularization, and includes skip connections to implement residual learning. When the number of input and output channels of the residual block is inconsistent, the dimension is adjusted by convolution.

[0008] In one embodiment, the three prediction branches are as follows: The spectral shape prediction branch contains multiple fully connected hidden layers and outputs normalized spectral shape data through linear mapping. The spectral intensity prediction branch outputs the spectral intensity value. The photoelectric conversion efficiency prediction branch outputs the photoelectric conversion efficiency value.

[0009] In one embodiment, the decoupling loss function includes root mean square error loss of spectral shape, cosine similarity loss of spectral smoothness constraint, and root mean square error loss of spectral intensity and photoelectric conversion efficiency, and each loss term is weighted and summed by weighting coefficients; In the decoupling loss function, the weighting coefficients of the spectral smoothness constraint cosine similarity loss and the spectral intensity loss are greater than the weighting coefficients of the photoelectric conversion efficiency loss.

[0010] In one embodiment, the spectral fit is used as the evaluation index for the spectral shape prediction result, and the photoelectric conversion efficiency prediction result is evaluated for accuracy by combining the mean absolute error and root mean square error of the photoelectric conversion efficiency.

[0011] In one embodiment, the multi-dimensional parameters include the dominant wavelength, chip size, voltage, optical power, chip manufacturer, operating current, and encapsulation adhesive.

[0012] In one embodiment, the model is trained using an optimizer and learning rate adjustment strategy, and an early stopping mechanism is set.

[0013] In one embodiment, the training model further includes ablation experiments to determine the optimal model configuration; the ablation experiments include feature selection ablation, network structure ablation, and loss function weight ablation.

[0014] The LED spectrum prediction method based on a one-dimensional convolutional neural network provided in this application solves the problem of feature redundancy by performing feature selection during data preprocessing and improves the model's generalization ability by reducing noise input. By constructing a one-dimensional convolutional neural network model with three independent prediction branches, the problem of existing deep learning models using a single fully connected structure and failing to fully exploit the local correlation information of input features is solved. By using a decoupled loss function to train the model, the problem of existing loss functions having a single design and being unable to simultaneously constrain multiple prediction targets is solved. Attached Figure Description

[0015] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0016] Figure 1 This is a flowchart of the LED spectrum prediction method based on a one-dimensional convolutional neural network in this application; Figure 2 This is a schematic diagram of the structure of the Spectral Decoupled CNN model in this application; Figure 3 This is a trend graph showing the number of ablation experimental features and the spectral fit degree in this application; Figure 4 This is a trend graph showing the prediction error between the number of ablation experimental features and photoelectric conversion efficiency in this application. Figure 5 This is a comparison chart of the predicted spectrum and the actual spectrum in this application; Figure 6 This is a graph showing the average absolute error distribution of the photoelectric conversion efficiency of this application. Detailed Implementation

[0017] To better understand the above-mentioned objectives, features, and advantages of this application, embodiments of this application will be further described below. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0018] Numerous specific details are set forth in the following description to provide a thorough understanding of this application, but this application may also be implemented in other ways than those described herein. Clearly, the embodiments described in the specification are only a portion of, and not all, of the embodiments of this application.

[0019] To address the shortcomings of existing deep learning-based LED spectrum prediction techniques, such as neglecting input feature redundancy, insufficient feature extraction capabilities of deep learning models, and the inability of loss functions to simultaneously constrain multiple prediction targets, this application provides an LED spectrum prediction method based on a one-dimensional convolutional neural network, referring to... Figure 1 As shown, it includes the following steps: Step S1: Obtain the data of the monochrome LED chip, which includes multi-dimensional parameters and corresponding spectral data; Step S2: Preprocess the data of the monochrome LED chip, including spectral wavelength axis alignment, smoothing filtering, feature standardization, feature selection and output data decoupling processing, to obtain the training dataset; Step S3: Construct a one-dimensional convolutional neural network model, which includes a shared convolutional feature extractor based on residual blocks and three independent prediction branches, which output three prediction results: normalized spectral shape, spectral intensity and photoelectric conversion efficiency, respectively. Step S4: Train the model using the decoupling loss function, which includes spectral shape loss, spectral smoothness constraint loss, spectral intensity loss, and photoelectric conversion efficiency loss.

[0020] First, execute step S1 to obtain the data of the monochrome LED chip. The data of the monochrome LED chip includes multi-dimensional parameters and corresponding spectral data.

[0021] In one embodiment, the multi-dimensional parameters are the inherent properties and operating parameters of the LED on the model input side, including but not limited to one or more of the following: dominant wavelength, chip size, voltage, optical power, chip manufacturer, operating current, and encapsulation adhesive. These multi-dimensional parameters cover data from various types, sizes, and packages of monochrome LED chips to ensure the diversity and representativeness of the training data, enabling the model to learn common features across different chip types and improve its generalization ability. In step S1, diverse data acquisition provides a sufficient data foundation for subsequent feature selection and model training.

[0022] Next, step S2 is executed to preprocess the data of the monochrome LED chip, including spectral wavelength axis alignment, smoothing filtering, feature standardization, feature selection, and output data decoupling processing, to obtain the training dataset.

[0023] In one embodiment, spectral wavelength axis alignment includes aligning all spectral data to a uniform band. Smoothing filtering includes outlier handling to eliminate measurement noise. Smoothing filtering employs at least one of the following: Savitzky-Golay filter, moving average filter, Gaussian filter, median filter, or low-pass filter. Specifically, outlier samples can be removed using statistical criteria and replaced with the average value of samples with the same parameters from the same batch. By removing outliers, the interference of measurement errors and process anomalies on model training is eliminated, improving data quality.

[0024] In one embodiment, feature standardization includes feature transformation and standardization processing. Feature transformation includes numerical feature transformation and non-numerical feature transformation, such as converting chip size into area values, and converting non-numerical features into numerical features using sequential encoding. It should be noted that chip area reflects the physical characteristics of the chip better than size numbers, and sequential encoding preserves the order relationship of classification features. The input features are standardized using a standardizer, for example, mapping the transformed feature values ​​to the [0,1] interval.

[0025] In one embodiment, feature selection is used to remove redundant features. Feature selection employs at least one of the following methods: Copula entropy-based feature selection, mutual information method, analysis of variance method, LASSO regression method, random forest feature importance method, or principal component analysis method. By performing feature selection, the problem of existing methods not considering feature redundancy is addressed, and the model's generalization ability is improved by reducing noise input.

[0026] In one embodiment, the Copula entropy-based feature selection method measures the independence and redundancy of features by calculating the Copula entropy between input features, eliminating redundant features, and thus selecting the optimal feature subset. Specifically, a Copula entropy (CE)-based feature selection method is used to filter the multidimensional original input features. The Copula entropy is calculated as follows: (Equation 1) in, Represents the Copula density function, The cumulative distribution function value represents the marginal distribution, and m represents the number of features in the current subset.

[0027] Copula entropy is an information-theoretic metric that measures the independence and redundancy of random variables. It quantifies the information contribution of each feature to the prediction target by calculating the Copula entropy between each feature and the target variable, and identifies redundant features by calculating the Copula entropy between features. For example, the original input features include seven dimensions: dominant wavelength, chip size, voltage, optical power, chip manufacturer, operating current, and encapsulation adhesive. After sorting based on Copula entropy and removing the redundant feature of encapsulation adhesive (pkg_glue), a final optimal subset of six features is selected.

[0028] In one embodiment, the output data decoupling process includes applying at least one of normalization, standardization, or logarithmic transformation to the spectral shape, spectral intensity, and photoelectric conversion efficiency, respectively. Specifically, the output data decoupling process includes spectral shape normalization, logarithmic scaling of spectral intensity, and scaling of photoelectric conversion efficiency. By decoupling the output data, the prediction targets with different dimensions are brought into a unified numerical range, facilitating joint optimization.

[0029] In one embodiment, the preprocessed monochrome LED chip data is divided proportionally into a training set, a validation set, and a test set to obtain a training dataset.

[0030] In one embodiment, a data augmentation step is also included to enhance the noise of low-current (e.g., 0~10mA) samples to improve the model's predictive ability for low-current scenarios.

[0031] Next, step S3 is executed to construct a one-dimensional convolutional neural network model. The model includes a shared convolutional feature extractor based on residual blocks and three independent prediction branches, which output three prediction results: normalized spectral shape, spectral intensity, and photoelectric conversion efficiency, respectively.

[0032] In one embodiment, refer to Figure 2 As shown, the shared convolutional feature extractor includes an ingress convolutional layer and multiple residual blocks. Each residual block contains two layers of one-dimensional convolution, normalization, activation function, and regularization, and includes skip connections to achieve residual learning. When the number of input and output channels of the residual block is inconsistent, the dimension is adjusted through convolution. The one-dimensional convolution can capture the local correlation information of the input feature sequence, exhibiting stronger feature extraction capabilities compared to fully connected networks. The residual blocks alleviate the gradient vanishing problem in deep networks through skip connections, enabling the network to learn deeper feature representations. By enhancing feature extraction capabilities through residual convolutional networks, the problem of insufficient feature extraction in existing models is solved.

[0033] In one embodiment, normalization employs at least one of batch normalization layers, layer normalization layers, or group normalization layers. The activation function employs at least one of SiLU, ReLU, LeakyReLU, PReLU, ELU, or GELU functions. Regularization employs at least one of Dropout regularization or L2 weight decay regularization.

[0034] In one embodiment, refer to Figure 2 As shown, the three independent prediction branches are the spectral shape prediction branch, the spectral intensity prediction branch, and the photoelectric conversion efficiency prediction branch.

[0035] The spectral shape prediction branch comprises multiple fully connected hidden layers, each followed by normalization, activation functions, and regularization, ultimately outputting normalized spectral shape data through a linear mapping. These multiple hidden layers provide ample nonlinear mapping capabilities, enabling the learning of complex mapping relationships from shared features to spectral shapes. The prediction of normalized spectral shapes avoids interference from absolute intensity on waveform learning. Independent prediction of normalized spectral shapes improves the accuracy of spectral waveform prediction.

[0036] The spectral intensity prediction branch outputs spectral intensity values. The photoelectric conversion efficiency prediction branch outputs photoelectric conversion efficiency values. Each of these branches contains an independent fully connected hidden layer.

[0037] By predicting spectral shape, intensity, and photoelectric conversion efficiency separately through three independent prediction branches, gradient interference between prediction targets with different dimensions is avoided. Joint optimization of spectral shape, intensity, and photoelectric conversion efficiency is achieved through decoupling branches, enabling simultaneous constraint of multiple prediction targets.

[0038] Optionally, the outputs of the three prediction branches are concatenated along the feature dimension, simultaneously including the normalized spectral shape, spectral intensity, and photoelectric conversion efficiency predictions, allowing each branch to focus on its respective regression task. It should be noted that concatenating the outputs enables the model to output all predictions simultaneously in a single forward propagation, improving inference efficiency.

[0039] In one embodiment, refer to Figure 2As shown, a Spectral Decoupled CNN model is built based on the PyTorch framework, with the following specific configuration: The shared one-dimensional convolutional feature extractor consists of an ingress convolutional layer and two residual blocks (ResidualBlock1D), with a base hidden dimension (hidden_dim) of 64; the ingress convolution is configured with in_channels=1, out_channels=64, kernel_size=3, padding=1, followed by BatchNorm1d batch normalization and SiLU activation function; each residual block contains two one-dimensional convolutional layers (in_channels=64, out_channels=64, kernel_size=3, padding=1), each followed by BatchNorm1d batch normalization, SiLU activation function, Dropout(0.1) regularization, and includes skip connections (Shortcut). The connection enables residual learning; the output of the convolutional layer is flattened by the Flatten layer to obtain the shared feature vector. The dimensionality transformation process is [B,1,6]→[B,64,6]→[B,64,6]→[B,64,6]→[B,384].

[0040] The spectral shape prediction branch adopts a two-layer fully connected structure of [1024,1024]. Each layer is followed by BatchNorm1d batch normalization, SiLU activation function, and Dropout(0.1) regularization. The output layer outputs 401-dimensional normalized spectral shape data through linear mapping, and the dimension transformation is [B,384]→[B,1024]→[B,1024]→[B,401].

[0041] The spectral intensity prediction branch adopts a single-layer fully connected structure

[128] followed by BatchNorm1d batch normalization and SiLU activation function. The output layer outputs a 1D spectral intensity value through linear mapping, with the dimension transformation being [B,384]→[B,128]→[B,1]. The WPE prediction branch adopts a single-layer fully connected structure

[128] followed by BatchNorm1d batch normalization and SiLU activation function. The output layer outputs a 1D WPE value through linear mapping, with the dimension transformation being [B,384]→[B,128]→[B,1].

[0042] The 401-dimensional output of the spectral shape prediction branch, the 1-dimensional output of the spectral intensity prediction branch, and the 1-dimensional output of the WPE prediction branch are concatenated in the feature dimension (dim=1) to obtain the final output of the model [B,403].

[0043] Next, step S4 is executed, in which the model is trained using a decoupling loss function, which includes spectral shape loss, spectral smoothness constraint loss, spectral intensity loss, and photoelectric conversion efficiency loss.

[0044] In one embodiment, the decoupling loss function specifically includes four terms: spectral shape root mean square error (RMSE) loss, spectral smoothness constraint cosine similarity loss, spectral intensity mean square error (MSE) loss, and photoelectric conversion efficiency mean square error (MSE) loss. These loss terms are weighted and summed using weighted coefficients. Specifically, the spectral shape RMSE loss constrains the prediction accuracy of the normalized spectral shape; the spectral smoothness constraint cosine similarity loss constrains the overall trend consistency of the spectral waveform; the spectral intensity MSE loss constrains the prediction accuracy of the intensity; and the photoelectric conversion efficiency MSE loss constrains the prediction accuracy of the efficiency. The RMSE loss is more sensitive to overall spectral shape deviations than the L1 loss, and can more evenly penalize prediction errors at each wavelength point. The cosine similarity loss ensures the accuracy of the overall waveform trend by measuring the directional consistency between the predicted and true spectra; and the MSE loss has the advantage of gradient stability in regression tasks.

[0045] In the decoupling loss function, the weighting coefficients for spectral smoothness constraint loss and spectral intensity loss are greater than the weighting coefficient for photoelectric conversion efficiency loss. By decomposing the spectral prediction task into multiple independent sub-tasks using the decoupling loss function, and constraining each prediction objective separately through different loss functions and weighting coefficients, dual optimization of spectral waveform consistency and numerical accuracy is achieved. The loss terms are weighted and summed using weighting coefficients determined through ablation experiments.

[0046] The decoupled loss function is used to calculate the difference between the model's predictions and the actual results. The total loss function is: (Equation 2) Among them, L RMSE L represents the RMSE loss of spectral shape. cos λ represents the cosine similarity loss constrained by spectral smoothness. cos L represents the weighting coefficient of the spectral smoothness constraint cosine similarity loss. int λ represents the spectral intensity MSE loss. int L represents the weighting coefficient for the spectral intensity MSE loss. wpe λ represents the photoelectric conversion efficiency (MSE) loss. wpe The weighting coefficient represents the loss of photoelectric conversion efficiency (MSE).

[0047] In one embodiment, the RMSE shape loss weight coefficient is set to 1.0, and the cosine similarity loss weight coefficient λcos =5.0, spectral intensity MSE loss weighting coefficient λ int =5.0, photoelectric conversion efficiency (MSE) loss weighting coefficient λ wpe =2.0.

[0048] In one embodiment, the spectral fit (GFC) can also be used as an evaluation metric for the spectral shape prediction results. The spectral fit (GFC) is obtained based on the true spectral vector and the predicted spectral vector. Specifically, the value of the spectral fit (GFC) is equal to the absolute value of the dot product of the true spectral vector and the predicted spectral vector, divided by the product of the magnitudes of the true and predicted spectral vectors. The dot product of the true and predicted spectral vectors reflects the "isotropy" of the two spectra at each wavelength; the product of the magnitudes of the true and predicted spectral vectors represents the overall intensity level of the spectrum.

[0049] (Equation 3) Among them, S real,i S represents the intensity value of the true spectrum corresponding to the i-th wavelength point. pred,i This represents the intensity value of the predicted spectrum corresponding to the i-th wavelength point, where i represents the i-th wavelength sampling point, and N represents the number of wavelength sampling points.

[0050] In one embodiment, the photoelectric conversion efficiency prediction result can also be evaluated for accuracy by combining the mean absolute error (MAE) and root mean square error (RMSE) of the photoelectric conversion efficiency (WPE).

[0051] When using both spectral fit (GFC) and mean absolute error (MAE) and root mean square error (RMSE) of power photoelectric conversion efficiency (WPE), GFC measures the similarity of spectral waveforms through the inner product of normalized vectors, forming a complementary evaluation with the RMSE loss and cosine similarity loss in the loss function, ensuring that the model achieves high accuracy in both waveform consistency and numerical precision.

[0052] In one embodiment, training the model using a decoupled loss function further includes ablation experiments to determine the optimal model configuration. The ablation experiment design includes feature selection ablation, network structure ablation, and loss function weight ablation.

[0053] In one embodiment, a Copula entropy-based feature selection method is used to filter input features, and an ablation experiment is conducted to determine the optimal model configuration. Copula entropy feature selection is based on information theory principles, objectively quantifying the information contribution of features and avoiding the subjectivity of manual selection. The ablation experiment systematically verifies the effectiveness of each component using a controlled variable method. Through objective feature selection and ablation experiments, the optimality of the model configuration is ensured.

[0054] In one embodiment, ablation experiments were conducted using 7-dimensional, 6-dimensional, 5-dimensional, and 4-dimensional feature combinations to train the model and verify the effectiveness of Copula entropy feature selection. All experiments used the same network structure and training parameters, with a fixed random seed to ensure reproducibility. (Refer to...) Figure 3 and Figure 4 As shown, the 6-dimensional feature combination (Copula entropy selection) achieves optimal performance on GFC, verifying that Copula entropy feature selection can reduce model complexity while maintaining prediction accuracy. Ablation experiments demonstrate that Copula entropy feature selection effectively eliminates redundant features, and the 6-dimensional feature combination achieves optimal performance.

[0055] In one embodiment, an optimizer and learning rate adjustment strategy are used to train the model, and an early stopping mechanism is set to terminate training when the performance on the validation set shows no improvement for several consecutive rounds. The optimizer includes, but is not limited to, at least one of the AdamW optimizer, Adam optimizer, RMSprop optimizer, momentum-driven stochastic gradient descent optimizer, or AdaGrad optimizer. The learning rate adjustment strategy includes, but is not limited to, at least one of the following: performance-based learning rate adjustment strategy, cosine annealing strategy, step-size decay strategy, exponential decay strategy, or linear warm-up strategy.

[0056] The optimizer combines the regularization effects of adaptive learning rate and weight decay. The learning rate decay strategy allows the model to escape local optima during loss plateaus. The early stopping mechanism terminates training promptly when validation performance no longer improves, avoiding overfitting and saving computational resources. A reasonable optimization strategy ensures the stability and convergence of model training. Specifically, the AdamW optimizer is used, with an initial learning rate of 0.001 and a weight decay coefficient of 1e-4. The ReduceLROnPlateau strategy is employed, halving the learning rate when the validation loss shows no improvement for 15 consecutive rounds. The early stopping patience value is set to 50, terminating training when validation set performance shows no improvement for 50 consecutive rounds. The number of training rounds is 500, and the batch size is 32. Figure 5 The comparison between the test set predictions (dashed line) and the true spectra (solid line) is shown. The final model's average GFC on the test set is 0.9994. Figure 6 The distribution of the mean absolute error (MAE) of the photoelectric conversion efficiency (WPE) is shown, with WPE MAE = 0.007.

[0057] In one embodiment, training the model further includes typical sample testing and model iterative optimization steps. Specifically, a yellow LED chip with a size of 27 mil, a main wavelength of 570 nm, and an operating current of 20 mA is selected for testing. The GFC between the predicted spectrum and the true spectrum is 0.9995, and the error between the predicted WPE and the true value is 0.007, indicating excellent fitting performance. The output results of the prediction model and the measured data are obtained, and the bias parameter is calculated. If the GFC of a sample is 0.98, which is lower than the preset threshold of 0.99, the sample is added to the training set to update the dataset. The model is retrained based on the updated dataset to obtain the updated prediction model. The above process is repeated until the bias parameter of all samples is lower than the preset threshold, ensuring continuous optimization and accuracy improvement of the model.

[0058] When applying the above model to chip development, by inputting parameters such as chip size, dominant wavelength, and operating current, the model can quickly output the absolute spectrum in the 380nm~780nm band. This allows for chip spectral characteristic evaluation without the need for physical samples, shortening the new chip development cycle. When applying the model to LED chip mixing, the single-chip spectrum output by the model is combined with packaging parameters to replace traditional mixing and superposition calculations, quickly outputting the spectral characteristics of multi-chip LED chips and supporting rapid optimization of LED chip formulations. When applying the model to lighting fixture development, based on the spectral prediction results of single chips and LED chips, an end-to-end virtual testing model for lighting fixtures can be constructed. This allows for the prediction of lighting fixture spectral and luminous efficacy performance without the need for physical LED strips, significantly reducing lighting fixture development costs and testing cycles.

[0059] The LED spectrum prediction method based on a one-dimensional convolutional neural network provided in this application solves the problem of feature redundancy by performing feature selection during data preprocessing and improves the model's generalization ability by reducing noise input. By constructing a one-dimensional convolutional neural network model with three independent prediction branches, the problem of existing deep learning models using a single fully connected structure and failing to fully exploit the local correlation information of input features is solved. By using a decoupled loss function to train the model, the problem of existing loss functions having a single design and being unable to simultaneously constrain multiple prediction targets is solved.

[0060] 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 thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to the process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0061] The above descriptions are merely embodiments of this application, which enable those skilled in the art to understand and implement this application. Various modifications to the embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of this application. Therefore, this application is not to be limited to the embodiments described herein, but is to be accorded the widest scope consistent with the principles and features disclosed herein.

Claims

1. A method for LED spectrum prediction based on one-dimensional convolutional neural network, characterized in that, include: Acquire data from a monochrome LED chip, the data of which includes multi-dimensional parameters and corresponding spectral data; The data of the monochrome LED chip is preprocessed, including spectral wavelength axis alignment, smoothing filtering, feature standardization, feature selection, and output data decoupling, to obtain a training dataset. A one-dimensional convolutional neural network model is constructed, which includes a shared convolutional feature extractor based on residual blocks and three independent prediction branches, which output three prediction results: normalized spectral shape, spectral intensity and photoelectric conversion efficiency, respectively. The model is trained using a decoupling loss function, which includes spectral shape loss, spectral smoothness constraint loss, spectral intensity loss, and photoelectric conversion efficiency loss.

2. The LED spectral prediction method according to claim 1, characterized in that, The spectral wavelength axis alignment aligns all spectral data to a unified band. The smoothing filter is used to eliminate measurement noise, and the smoothing filter adopts at least one of Savitzky-Golay filter, moving average filter, Gaussian filter, median filter or low-pass filter. The feature selection is used to remove redundant features, and the feature selection adopts at least one of the following methods: Copula entropy-based feature selection method, mutual information method, analysis of variance method, LASSO regression method, random forest feature importance method, or principal component analysis method. The output data decoupling process includes: applying at least one of normalization, standardization, or logarithmic transformation to the spectral shape, spectral intensity, and photoelectric conversion efficiency, respectively.

3. The LED spectrum prediction method according to claim 2, characterized in that, The Copula-based feature selection method measures the independence and redundancy of features by calculating the Copula entropy between input features, quantifies the information contribution of features to the prediction target, and identifies redundant features by calculating the Copula entropy between features in order to eliminate redundant features.

4. The LED spectrum prediction method according to claim 1, characterized in that, The shared convolutional feature extractor includes an ingress convolutional layer and multiple residual blocks. Each residual block contains two layers of one-dimensional convolution, normalization, activation function and regularization, and includes skip connections to achieve residual learning. When the number of input and output channels of the residual block is inconsistent, the dimension is adjusted by convolution.

5. The LED spectrum prediction method according to claim 1, characterized in that, The three prediction branches are as follows: The spectral shape prediction branch contains multiple fully connected hidden layers and outputs normalized spectral shape data through linear mapping. The spectral intensity prediction branch outputs the spectral intensity value. The photoelectric conversion efficiency prediction branch outputs the photoelectric conversion efficiency value.

6. The LED spectral prediction method according to claim 1, characterized in that, The decoupling loss function includes root mean square error loss of spectral shape, cosine similarity loss of spectral smoothness constraint, and root mean square error loss of spectral intensity and photoelectric conversion efficiency. Each loss term is weighted and summed by weighting coefficients. In the decoupling loss function, the weighting coefficients of the spectral smoothness constraint cosine similarity loss and the spectral intensity loss are greater than the weighting coefficient of the photoelectric conversion efficiency loss.

7. The LED spectral prediction method according to claim 1, characterized in that, The spectral fit degree is used as the evaluation index for the spectral shape prediction results, and the accuracy of the photoelectric conversion efficiency prediction results is evaluated by combining the mean absolute error and root mean square error of the photoelectric conversion efficiency.

8. The LED spectrum prediction method according to claim 1, characterized in that, The multi-dimensional parameters include the dominant wavelength, chip size, voltage, optical power, chip manufacturer, and operating current.

9. The LED spectrum prediction method according to claim 1, characterized in that, It also includes training the model using an optimizer and learning rate adjustment strategy, and setting an early stopping mechanism.

10. The LED spectral prediction method according to claim 1, characterized in that, Training the model also includes ablation experiments to determine the optimal model configuration; the ablation experiments include feature selection ablation, network structure ablation, and loss function weight ablation.

Images