LED package aging trend prediction method based on time series modeling

Abstract

This invention discloses a method for predicting the aging trend of LED packaging based on time-series modeling, comprising the following steps: S1, collecting multi-dimensional operational data to construct a raw time-series dataset; S2, cleaning and preprocessing the data, and generating an aging feature time series based on sliding window statistics; S3, inputting the aging feature time series into an improved Crossformer model, sequentially executing cross-feature and time attention mechanisms to extract the aging trend and output predicted values; S4, predicting the aging trend within the future time window based on current data; S5, comparing the predicted values ​​with a preset threshold to determine whether an aging warning is triggered; S6, introducing the AdaBelief optimization algorithm to adaptively fine-tune the model parameters based on the prediction error. This invention possesses advantages such as high prediction accuracy, good optimization stability, and a sound error feedback mechanism, significantly improving practicality, interpretability, and intelligence.

Description

Technical Field

[0001] This invention relates to the field of time series modeling and prediction, and in particular to a method for predicting the aging trend of LED packaging based on time series modeling. Background Technology

[0002] Traditional methods for assessing the aging status of LED packages mostly rely on static rules or empirical thresholds, making it difficult to accurately depict the aging process over time. With the diversification of packaging technologies and material properties, aging exhibits significant nonlinear temporal characteristics, making it difficult to achieve precise modeling and reliable prediction of the aging process based solely on single-point data or linear trend analysis.

[0003] While commonly used prediction models such as LSTM and GRU have certain time series modeling capabilities, they often suffer from insufficient modeling capabilities and low feature extraction efficiency when faced with long-term dependencies, multivariate interactions, and complex aging patterns. This results in large prediction biases and poor robustness, making them unsuitable for life management scenarios of precision electronic devices.

[0004] In addition, some models are based solely on packaging structure parameters or a single aging index, failing to integrate multi-source aging observations such as temperature, current, and luminous flux. They lack a deep understanding and collaborative analysis of key aging factors, making it difficult to fully reflect the true degradation state of LED packaging.

[0005] In practical deployments, most existing methods do not consider the model's ability to adapt to changes in input samples, making it difficult to maintain stable performance under new packaging structures or changing aging environments. Furthermore, they lack dynamic feedback mechanisms for prediction errors, failing to correct biases or self-adjust model parameters in a timely manner, thus affecting prediction accuracy and practicality.

[0006] Therefore, how to provide a method for predicting the aging trend of LED packaging based on time-series modeling is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0007] One objective of this invention is to propose a time-series modeling-based method for predicting the aging trend of LED packages. This invention combines an improved Crossformer model with the AdaBelief optimization algorithm to construct an aging trend prediction process capable of multivariate feature modeling, sliding position-aware embedding, long-term dependency capture, and adaptive error feedback updates. By constructing multidimensional aging observations of the package in actual operation as a time series and introducing sliding position encoding to enhance the model's ability to characterize temporal evolution information, the Crossformer model is used to achieve interactive modeling and trend extraction among multiple variables. Simultaneously, a residual evaluation mechanism is constructed between the model output and actual observations, and the prediction model parameters are dynamically adjusted based on the AdaBelief optimization algorithm. The algorithm combines first-order gradient information and instantaneous gradient variance to achieve stable and rapid adaptive learning rate updates, thereby enhancing the model's adaptability to input perturbations and changes in aging behavior. This invention possesses advantages such as strong modeling capabilities, high prediction accuracy, good adaptability, and error feedback optimization, effectively solving the problems of inaccurate model characterization of the aging process and difficulty in updating and optimizing existing technologies.

[0008] An LED packaging aging trend prediction method based on time-series modeling according to an embodiment of the present invention includes the following steps: S1. Collect multi-dimensional operational data of LED packaging during use or aging test, and collect the data according to timestamps to form a multi-dimensional time series raw dataset. S2. Perform data cleaning and preprocessing on the original multidimensional time series dataset, and generate feature vectors based on sliding window statistics to output the aging feature time series; S3. Input the aging feature time series into the improved Crossformer model, divide and embed it according to time segments, and then execute cross-feature attention and time attention mechanisms in sequence to extract multivariate interaction relationships and aging trends, and fuse and output the aging trend prediction value. S4. Input the current aging characteristic time series of the target LED package into the improved Crossformer model, and output the aging trend prediction value within the future preset time window. S5. Compare the output aging trend prediction value with the preset aging threshold, mark whether the LED package is in an aging warning state and output the aging trend analysis result. S6. Evaluate the error between the newly collected aging observations and the aging trend analysis results during actual operation. If the average prediction residual exceeds the set error tolerance threshold, perform parameter fine-tuning based on the AdaBelief optimization algorithm and adaptively adjust the update step size according to the gradient fluctuation.

[0009] Optionally, S1 specifically includes: S11. After the LED package enters the usage stage or accelerated aging test environment, start the data acquisition sensor and collect four raw data of the package: luminous flux, junction temperature, color temperature and driving current at preset time intervals. S12. Add a corresponding timestamp to each piece of raw data to generate raw data entries with time tags, and arrange them in order of collection time. S13. Align the four raw data items of luminous flux, junction temperature, color temperature and driving current according to the timestamp to construct a multidimensional time series raw dataset.

[0010] Optionally, S2 specifically includes: S21. Perform integrity checks on the four raw data items of luminous flux, junction temperature, color temperature and driving current in the multidimensional time series raw dataset, mark the location of missing values, and use time series interpolation methods to fill in the missing values. S22. Calculate the rate of change of each data point between adjacent time points. When the rate of change exceeds the set abnormal threshold, mark the data point as an abnormal point and replace the abnormal point with the average of the two normal points before and after. S23. Perform min-max normalization on the four raw data of luminous flux, junction temperature, color temperature and driving current respectively. Construct a sliding window of a set length on the normalized data sequence, extract local time segments in turn, and calculate the mean, variance and rate of change in each window to generate feature vectors. S24. Concatenate all the feature vectors extracted from the sliding window in timestamp order to form an aging feature time series.

[0011] Optionally, S3 specifically includes: S31. Divide the aging feature time series into sliding time windows of a preset length according to the time sequence. Each window retains continuous time step data of a preset size. Arrange the four original data corresponding to the time step row by row to form an original feature matrix with a preset number of rows and 4 columns, representing the aging state within the time period. S32. For each row of the original feature matrix in each time window, input a set of preset weight vectors with the same dimension to perform dot product operation on the four feature values ​​and add bias terms. Map each original four-dimensional feature vector to a one-dimensional vector. Concatenate the one-dimensional vectors of all time steps in order to construct the embedding vector. Add position encoding according to the time step index in the embedding vector to preserve the sequential features of the time series. S33. Input the embedding vector into the encoder of the improved Crossformer model, perform cross-feature attention mechanism processing on multiple feature vectors in each time step, use the embedding vector of each feature vector as the query vector, calculate the dot product similarity between it and the embedding vectors of other feature vectors in the same time step, and obtain the attention weight after normalization using the Softmax function. S34. Multiply these attention weights by the embedding vectors of the corresponding feature vectors and sum them to obtain the representation of each feature vector after incorporating the influence of other feature vectors at the current time step, and output the cross-feature representation sequence. S35. Taking the cross-feature representation sequence as input, the multi-head attention mechanism is executed sequentially along the time dimension. The dot product similarity of the embedding vectors of each set of feature vectors at different time steps is calculated to construct the time dependency matrix. The historical state of the current feature vector at different time points is weighted according to the attention weight to generate the trend evolution feature within the entire time window and output the time context representation vector sequence of each feature vector. S36. The cross-feature representation sequence and the temporal context representation vector sequence are concatenated according to the same time step index, and the concatenated vector is compressed in dimension using a linear fusion transformation. At the same time, a residual connection structure is introduced, and the embedded vector before concatenation is added to the fusion output vector. LayerNorm normalization is performed to obtain the fused aging trend feature representation tensor. S37. In the decoder, the fused aging trend feature representation tensor is transformed by a feedforward neural network and passed through two linear layers and one activation function layer in sequence to output the aging trend prediction value containing the luminous flux maintenance rate and color temperature change amount of multiple subsequent time steps.

[0012] Optionally, the improved Crossformer model specifically includes: The aging characteristic time series after cleaning is divided into multiple time segments of equal length according to time sequence. Each time segment contains a set number of time steps. Each time step records four values: luminous flux, junction temperature, color temperature and driving current. In each time segment, a position code is generated for each time step according to its relative position number in the current time segment. The encoding method adopts trigonometric function transformation, that is, the time step number is input into the sine function and the cosine function respectively to form a set of continuously changing encoding vectors, representing the relative position of each time step in the local window. For each time step feature value, the corresponding luminous flux, junction temperature, color temperature and driving current four feature values ​​are standardized and concatenated into a one-dimensional vector. The one-dimensional vector is then concatenated and combined with the sliding position encoding vector corresponding to the time step to form an embedded entry with position-aware information. The embedding entries of all time steps within each time segment are combined into a three-dimensional tensor containing the number of time steps, feature dimension, and batch size, and then fed into the improved Crossformer model.

[0013] Optionally, S4 specifically includes: organizing the current aging characteristic time series of the target LED package in chronological order and inputting it into the improved Crossformer model, extracting the correlation and trend characteristics between multiple variables in sequence, and outputting the predicted values ​​of luminous flux, junction temperature, color temperature and driving current within a future preset time window to form an aging trend prediction value.

[0014] Optionally, S5 specifically includes: S51. Read the aging trend prediction values ​​output by the improved Crossformer model, organize the prediction value sequence in time order, and extract the prediction values ​​of luminous flux, junction temperature, color temperature and drive current for each time step. S52. Read the preset aging threshold standards respectively, compare each predicted value with the corresponding aging threshold, and determine whether the aging threshold conditions are exceeded. If any one of them is exceeded, mark the current time step as the aging warning time point. S53. Statistically analyze all aging warning time points, calculate the continuous warning duration and cumulative exceedance. If the continuous exceedance time exceeds the preset duration threshold, mark the corresponding LED package as being in aging warning state. S54. Based on the changing trends of various predicted values, draw visualization charts of luminous flux decline curve and junction temperature rise trend, and generate aging trend analysis results in chronological order to assess the lifespan stage and health status of LED packages.

[0015] Optionally, S6 specifically includes: S61. The newly collected aging observation values ​​are compared one-to-one with the aging trend prediction values ​​predicted by the improved Crossformer model in chronological order. The residual between the predicted aging trend value and the actual value is calculated. The average value of all residuals is calculated and compared with the preset aging error tolerance threshold. If the aging error tolerance threshold is exceeded, the improved Crossformer model update process is started. S62. Combine the aging feature time series used for prediction in the previous stage with the newly collected observations to form an input-output sample pair, and input it into the current round of fine-tuning training. At the same time, initialize the learning rate and update step size required for the AdaBelief optimization algorithm. S63. For each trainable parameter in the Crossformer model, calculate the gradient direction of the current parameter in the error function. Then, take a weighted average of the gradient value of each parameter with the historical gradients to obtain a smooth trend direction. S64. The AdaBelief optimization algorithm calculates the difference between the current gradient and the trend direction, that is, the degree to which the current gradient deviates from the trend, and uses this deviation difference as a confidence index. If the deviation difference is greater than the preset deviation threshold, it returns to update; if the deviation difference is less than the preset deviation threshold, it continues to update. S65. Read the instantaneous gradient value of the current parameter and the previously accumulated average gradient trend. Take the difference between the two as the deviation of the current gradient. After squaring the deviation, take it as the variance estimate. Preset a very small constant. Divide the preset learning rate by the sum of the variance estimate and the constant to obtain the update step size of the current parameter. S66. Adjust all parameters in the direction of reducing error according to the calculated step size. After completing one round of parameter update, continue to predict the aging trend of the updated and improved Crossformer model, and record the prediction performance of the new round. Repeat the above fine-tuning process for several rounds until the prediction residual falls back to within the tolerance range.

[0016] The beneficial effects of this invention are: This invention constructs an intelligent modeling and dynamic optimization process for predicting the aging trend of LED packaging by integrating the improved Crossformer model and the AdaBelief optimization algorithm, which significantly improves the overall performance of the prediction model in terms of multi-time-series feature modeling, adaptive optimization and trend response capabilities.

[0017] First, the improved Crossformer model introduces a sliding position encoding mechanism in the input layer, which integrates the temporal dimension position with local window information to enhance the model's sensitivity to the temporal evolution patterns and abrupt changes in aging data. At the same time, the Crossformer model structure enables cross-modeling of multivariate aging features, improving the ability to characterize the long-term dependence and periodic changes of aging behavior, effectively overcoming the problem of insufficient capture ability of traditional models.

[0018] Secondly, during the model training and deployment phase, the AdaBelief optimization algorithm is introduced to update the parameters of the Crossformer model. By dynamically constructing the ratio of the first-order gradient bias to the instantaneous variance, the learning rate is adaptively adjusted, enabling the model to maintain a large step size to accelerate convergence in the early stage of training and to achieve stable convergence and reduce overfitting in the later stage of training. The optimization strategy not only improves the robustness of the model under different packaging materials, operating currents and temperatures, but also enhances its adaptability to predicting aging trends under multiple operating conditions.

[0019] Furthermore, in practical prediction applications, this invention supports dynamic parameter fine-tuning based on prediction residuals and set deviation thresholds, realizing error feedback and model self-correction mechanisms, maintaining the stability and continuity of prediction results, and ensuring that trend predictions maintain high accuracy in long-term deployment scenarios.

[0020] In summary, this invention has the advantages of high prediction accuracy, strong time modeling capability, good optimization stability and perfect error feedback mechanism, which can significantly improve the practicality, interpretability and intelligence level of LED packaging aging prediction system in actual industrial environment. Attached Figure Description

[0021] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of an LED packaging aging trend prediction method based on time-series modeling proposed in this invention; Figure 2 This is a flowchart of the improved Crossformer multivariate aging feature modeling and trend prediction based on sliding position encoding proposed in this invention. Figure 3 This is a flowchart of the adaptive parameter update and dynamic model adjustment based on the AdaBelief optimization algorithm proposed in this invention. Detailed Implementation

[0022] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0023] refer to Figures 1-3 A method for predicting the aging trend of LED packaging based on time-series modeling includes the following steps: S1. Collect multi-dimensional operating data of LED packaging during use or aging test. The operating data includes four items: luminous flux, junction temperature, color temperature and driving current. Collect the data according to timestamp to form a multi-dimensional time series raw dataset. S2. Perform data cleaning and preprocessing on the original multidimensional time series dataset, including missing value imputation, outlier removal and numerical normalization, and generate feature vectors based on sliding window statistics, outputting the cleaned aging feature time series. S3. Input the aging feature time series into the improved Crossformer model, divide and embed it according to time segments, and then execute the cross-variable attention and time attention mechanisms in sequence to extract the multivariate interaction relationship and aging trend, and fuse and output the aging trend prediction value. S4. Input the current aging characteristic time series of the target LED package into the improved Crossformer model, and output the aging trend prediction value within the future preset time window. S5. Compare the output aging trend prediction value with the preset aging threshold, mark whether the LED package is in an aging warning state and output the aging trend analysis result. S6. Evaluate the error between the newly collected aging observations and the aging trend analysis results during actual operation. If the average prediction residual exceeds the set aging error tolerance threshold, perform parameter fine-tuning based on the AdaBelief optimization algorithm, and adaptively adjust the learning rate according to the gradient fluctuation to improve the model's fitting ability and generalization performance for non-stationary aging trends, and update the aging trend prediction capability.

[0024] This implementation significantly improves the accuracy, lead time, and adaptability of LED packaging aging trend prediction. By constructing multi-dimensional time-series features such as luminous flux, junction temperature, color temperature, and drive current, and combining sliding window and normalization processing, the aging behavior change trend is effectively extracted. Based on this, an improved Crossformer model is introduced, which fully models the interaction relationships between multiple variables and the temporal evolution during the aging process through cross-variable attention and temporal attention mechanisms, improving the accuracy and stability of trend prediction. To address prediction biases that may be caused by non-stationary aging trends, the AdaBelief optimization algorithm is used to adaptively adjust the learning rate when the error exceeds the threshold, dynamically fine-tuning the model parameters and enhancing the model's generalization ability and robustness under various aging modes. This ensures that the prediction results are more consistent with actual observations, demonstrating good engineering applicability and deployment prospects.

[0025] In this embodiment, S1 specifically includes: S11. After the LED package enters the usage stage or accelerated aging test environment, start the data acquisition sensor and collect four raw data of the package: luminous flux, junction temperature, color temperature and driving current at preset time intervals. S12. Add a corresponding timestamp to each piece of raw data to generate raw data entries with time tags, and arrange them in order of collection time. S13. Align the four raw data items of luminous flux, junction temperature, color temperature and driving current according to the timestamp to construct a multidimensional time series raw dataset.

[0026] In this embodiment, S2 specifically includes: S21. Perform integrity checks on the four raw data items of luminous flux, junction temperature, color temperature and driving current in the multidimensional time series raw dataset, mark the location of missing values, and use time series interpolation methods to fill in the missing values. S22. Calculate the rate of change of each data point between adjacent time points. When the rate of change exceeds the set abnormal threshold, mark the data point as an abnormal point and replace the abnormal point with the average of the two normal points before and after. S23. Perform min-max normalization on the four raw data of luminous flux, junction temperature, color temperature and driving current respectively. Construct a sliding window of a set length on the normalized data sequence, extract local time segments in turn, and calculate the mean, variance and rate of change in each window to generate feature vectors. S24. Concatenate all the feature vectors extracted from the sliding window in timestamp order to form an aging feature time series.

[0027] In this embodiment, S3 specifically includes: S31. Divide the aging feature time series into sliding time windows of a preset length according to the time sequence. Each window retains continuous time step data of a preset size. Arrange the four original data corresponding to the time step row by row to form an original feature matrix with a preset number of rows and 4 columns, representing the aging state within the time period. S32. For each row of the original feature matrix in each time window, input a set of preset weight vectors with the same dimension to perform dot product operation on the four feature values ​​and add bias terms. Map each original four-dimensional feature vector to a one-dimensional vector. Concatenate the one-dimensional vectors of all time steps in order to construct the embedding vector. Add position encoding according to the time step index in the embedding vector to preserve the sequential features of the time series. S33. Input the embedding vector into the encoder of the improved Crossformer model, perform cross-feature attention mechanism processing on multiple feature vectors in each time step, use the embedding vector of each feature vector as the query vector, calculate the dot product similarity between it and the embedding vectors of other feature vectors in the same time step, and obtain the attention weight after normalization using the Softmax function. S34. Multiply these attention weights by the embedding vectors of the corresponding feature vectors and sum them to obtain the representation of each feature vector after incorporating the influence of other feature vectors at the current time step, and output the cross-feature representation sequence. S35. Taking the cross-feature representation sequence as input, the multi-head attention mechanism is executed sequentially along the time dimension. The dot product similarity of the embedding vectors of each set of feature vectors at different time steps is calculated to construct the time dependency matrix. The historical state of the current feature vector at different time points is weighted according to the attention weight to generate the trend evolution feature within the entire time window and output the time context representation vector sequence of each feature vector. S36. The cross-feature representation sequence and the temporal context representation vector sequence are concatenated according to the same time step index, and the concatenated vector is compressed in dimension using a linear fusion transformation. At the same time, a residual connection structure is introduced, and the embedded vector before concatenation is added to the fusion output vector. LayerNorm normalization is performed to obtain the fused aging trend feature representation tensor. S37. In the decoder, the fused aging trend feature representation tensor is transformed by a feedforward neural network and passed through two linear layers and one activation function layer in sequence to output the aging trend prediction value containing the luminous flux maintenance rate and color temperature change amount of multiple subsequent time steps.

[0028] This implementation significantly improves the modeling accuracy and timing prediction capability of LED package aging trends by introducing an improved Crossformer model. During the encoding stage, the model first executes a cross-feature attention mechanism to fully capture the multivariate correlation features between luminous flux, junction temperature, color temperature, and drive current. In the decoding stage, it constructs long-term dependencies through multi-head temporal attention, enhancing the perception and fitting ability of aging trend changes. Simultaneously, strategies such as sliding window segmentation, positional encoding embedding, and residual fusion are introduced, giving the model stronger sequential modeling and contextual understanding capabilities, effectively addressing the nonlinear aging modes exhibited by LED packages under complex operating conditions. The final prediction output accurately reflects the future luminous flux maintenance rate and color temperature change trajectory, providing a reliable basis for preventative maintenance and product life assessment.

[0029] In this embodiment, the improved Crossformer model specifically includes: The aging characteristic time series after cleaning is divided into multiple time segments of equal length according to time sequence. Each time segment contains a set number of time steps. Each time step records four values: luminous flux, junction temperature, color temperature and driving current. In each time segment, a position code is generated for each time step according to its relative position number in the current time segment. The encoding method adopts trigonometric function transformation, that is, the time step number is input into the sine function and the cosine function respectively to form a set of continuously changing encoding vectors, representing the relative position of each time step in the local window. For each time step feature value, the corresponding luminous flux, junction temperature, color temperature and driving current four feature values ​​are standardized and concatenated into a one-dimensional vector. The one-dimensional vector is then concatenated and combined with the sliding position encoding vector corresponding to the time step to form an embedded entry with position-aware information. The embedding entries of all time steps within each time segment are combined into a three-dimensional tensor containing the number of time steps, feature dimension, and batch size, and then fed into the improved Crossformer model.

[0030] In this embodiment, S4 specifically includes: organizing the current aging characteristic time series of the target LED package in chronological order and inputting it into the improved Crossformer model, extracting the correlation and trend characteristics between multiple variables in sequence, and outputting the predicted values ​​of luminous flux, junction temperature, color temperature and driving current within a future preset time window to form an aging trend prediction value.

[0031] In this embodiment, S5 specifically includes: S51. Read the aging trend prediction values ​​output by the improved Crossformer model, organize the prediction value sequence in time order, and extract the prediction values ​​of luminous flux, junction temperature, color temperature and drive current for each time step. S52. Read the preset aging threshold standards respectively, compare each predicted value with the corresponding aging threshold, and determine whether the aging threshold conditions are exceeded. If any one of them is exceeded, mark the current time step as the aging warning time point. S53. Statistically analyze all aging warning time points, calculate the continuous warning duration and cumulative exceedance. If the continuous exceedance time exceeds the preset duration threshold, mark the corresponding LED package as being in aging warning state. S54. Based on the changing trends of various predicted values, draw visualization charts of luminous flux decline curve and junction temperature rise trend, and generate aging trend analysis results in chronological order to assess the lifespan stage and health status of LED packages.

[0032] In this embodiment, S6 specifically includes: S61. The newly collected aging observation values ​​are compared one-to-one with the aging trend prediction values ​​predicted by the improved Crossformer model in chronological order. The residual between the predicted aging trend value and the actual value is calculated. The average value of all residuals is calculated and compared with the preset aging error tolerance threshold. If the aging error tolerance threshold is exceeded, the improved Crossformer model update process is started. S62. Combine the aging feature time series used for prediction in the previous stage with the newly collected observations to form an input-output sample pair, and input it into the current round of fine-tuning training. At the same time, initialize the learning rate and update step size required for the AdaBelief optimization algorithm. S63. For each trainable parameter in the Crossformer model, calculate the gradient direction of the current parameter in the error function. Then, take a weighted average of the gradient value of each parameter with the historical gradients to obtain a smooth trend direction. S64. The AdaBelief optimization algorithm calculates the difference between the current gradient and the trend direction, that is, the degree to which the current gradient deviates from the trend, and uses this deviation difference as a confidence index. If the deviation difference is greater than the preset deviation threshold, it means that the current gradient is not very reliable and returns to update. If the deviation difference is less than the preset deviation threshold, it means that the current gradient direction is consistent with the previous trend and continues to update. S65. Read the instantaneous gradient value of the current parameter and the previously accumulated average gradient trend. Take the difference between the two as the deviation of the current gradient. Squar the deviation and take it as the variance estimate. Preset a very small constant to prevent division by zero error. Divide the preset learning rate by the sum of the variance estimate and the constant to obtain the update step size of the current parameter. S66. Adjust all parameters in the direction of reducing error according to the calculated step size. After completing one round of parameter update, continue to predict the aging trend of the updated and improved Crossformer model, and record the prediction performance of the new round. Repeat the above fine-tuning process for several rounds until the prediction residual falls back to within the tolerance range.

[0033] This implementation introduces the AdaBelief optimization algorithm, achieving sensitive response to the prediction residuals of the Crossformer model and adaptive parameter update adjustment, significantly improving the long-term stability and generalization ability of aging trend prediction. The algorithm dynamically evaluates the residuals between the actual and predicted values ​​after each prediction round. If the error exceeds the tolerance threshold, a fine-tuning process is initiated, calculating the confidence level based on the degree of gradient deviation from the trend and adjusting the learning rate accordingly. Compared to traditional optimization algorithms, AdaBelief focuses more on the confidence deviation of the gradient direction, avoiding over-updates caused by abnormal gradient perturbations, thereby improving the model's prediction robustness during the aging process. Through multiple rounds of iterative optimization, the model can continuously conform to the actual aging evolution trajectory, maintaining high prediction accuracy even when dealing with complex aging modes such as non-stationary, slowly changing, or abrupt changes, effectively supporting the dynamic adjustment of LED package life assessment and early warning response strategies.

[0034] Example 1: To verify the feasibility and accuracy of this invention in predicting the aging trend of LED packaging in a real production environment, the method of this invention was deployed in the aging test production line of an LED packaging manufacturing company. This company has a complete LED chip packaging and testing process, with an average daily shipment of more than 180,000 LED devices. It also has a relatively complete data acquisition system and aging test bench, which can provide multi-dimensional and continuous operational data support for this invention.

[0035] In this scenario, the company employed a continuous 120-hour LED accelerated aging test system, encompassing luminous flux decay testing, junction temperature control, color temperature stability verification, and drive current variation monitoring. Each LED packaged device collected four data metrics every 10 minutes during the test: luminous flux, junction temperature, color temperature, and drive current, resulting in a total of 720 multivariate records per LED unit throughout the entire cycle. The company also configured three test channels, capable of concurrently testing over 3000 devices per day.

[0036] Traditional trend judgment methods rely on manually set rules, such as setting the light flux to decrease by more than 10% after 72 hours as premature aging. This method cannot dynamically adapt to differences in materials from different batches or environmental disturbances, resulting in a high misjudgment rate, especially in large-scale concurrent testing where the accuracy drops significantly.

[0037] In the actual deployment of this invention, the system imports aging test data into the preprocessing module in real time. First, it performs operations such as missing value completion, outlier removal, and sliding window construction to form a standardized feature time series. Then, it inputs this data into an improved Crossformer model to perform multivariate attention modeling and time series trend extraction, outputting an aging trend prediction curve for the next 24 hours.

[0038] After comparing the predicted results with the set thresholds, the model automatically marks the aging warning points and generates trend analysis charts for engineers to view. For some devices with large prediction deviations, the system automatically calls the model parameter fine-tuning mechanism based on the AdaBelief optimization algorithm to improve the fitting ability to nonlinear or abrupt aging characteristics. During the experiment, five batches of LED packages were tested, with approximately 600 LEDs in each batch, totaling more than 210,000 data points. The specific results are shown in the table below, compared with the company's original baseline rule method: Table 1. Performance Comparison of the Invention Method and Traditional Rule-Based Methods in LED Packaging Aging Prediction Tasks

[0039] Based on the comparative data shown in Table 1, it can be seen that the LED packaging aging trend prediction method proposed in this invention, based on the improved Crossformer and AdaBelief optimization algorithms, is significantly superior to traditional rule-based methods in several key performance indicators, such as prediction accuracy, response efficiency, early warning capability, and control of false alarms and missed alarms, demonstrating comprehensive technical advantages.

[0040] In terms of prediction accuracy, the method of this invention maintained an accuracy of over 94.6% across all five batches, reaching a maximum of 96.5%, far exceeding the traditional rule-based methods which generally operate at 68%–72%. This significantly improves the ability to accurately identify aging trends.

[0041] In terms of response timeliness, this invention significantly reduces the time required for the model to complete prediction and judgment. Traditional methods have an average response time of over 82 seconds, while the method of this invention controls it to around 20 seconds. For example, the response time for batch B is 19.7 seconds, only one-quarter of the original method. This rapid response capability is particularly crucial for high-frequency LED product quality monitoring, enabling earlier detection of potential failure risks.

[0042] In terms of prediction lead time and early warning success rate, this invention has early warning capabilities that traditional rule-based methods do not possess at all. It can identify aging failure trends 8.6 to 10.1 hours in advance, and the early warning success rate generally exceeds 89%. For example, the early warning success rate of batch C reaches 92.8%, which can give enterprises sufficient processing time and optimize production scheduling and maintenance plans.

[0043] Regarding the control of false alarm and false negative rates, traditional rule-based methods often have a false alarm rate as high as 13% to 15% and a false negative rate close to 10%. In contrast, the method of this invention controls the average false alarm rate to within 3% and the false negative rate to less than 2.5%. For example, in batch A, the false alarm rate is only 3.1% and the false negative rate is 1.8%, effectively reducing the risk of ineffective intervention and undetected failures, and significantly improving the stability and practicality of the system.

[0044] In summary, the method of this invention not only achieves higher accuracy, shorter response time, and stronger early warning capability in LED packaging aging prediction tasks, but also significantly optimizes the performance of misjudgment control. It is suitable for intelligent quality inspection deployment in large-scale LED production lines and has good prospects for promotion and application.

[0045] The above description is only a preferred embodiment of the present invention, but 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 inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for predicting the aging trend of LED packages based on time-series modeling, characterized in that, Includes the following steps: S1. Collect multi-dimensional operational data of LED packaging during use or aging test, and collect the data according to timestamps to form a multi-dimensional time series raw dataset. S2. Perform data cleaning and preprocessing on the original multidimensional time series dataset, and generate feature vectors based on sliding window statistics to output the aging feature time series; S3. Input the aging feature time series into the improved Crossformer model, divide and embed it according to time segments, and then execute cross-feature attention and time attention mechanisms in sequence to extract multivariate interaction relationships and aging trends, and fuse and output the aging trend prediction value. S4. Input the current aging characteristic time series of the target LED package into the improved Crossformer model, and output the aging trend prediction value within the future preset time window. S5. Compare the output aging trend prediction value with the preset aging threshold, mark whether the LED package is in an aging warning state and output the aging trend analysis result. S6. Evaluate the error between the newly collected aging observations and the aging trend analysis results during actual operation. If the average prediction residual exceeds the aging error tolerance threshold, perform parameter fine-tuning based on the AdaBelief optimization algorithm and adaptively adjust the update step size according to the gradient fluctuation.

2. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S1 specifically includes: S11. After the LED package enters the usage stage or accelerated aging test environment, start the data acquisition sensor and collect four raw data of the package: luminous flux, junction temperature, color temperature and driving current at preset time intervals. S12. Add a corresponding timestamp to each piece of raw data to generate raw data entries with time tags, and arrange them in order of collection time. S13. Align the four raw data items of luminous flux, junction temperature, color temperature and driving current according to the timestamp to construct a multidimensional time series raw dataset.

3. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S2 specifically includes: S21. Perform integrity checks on the four raw data items of luminous flux, junction temperature, color temperature and driving current in the multidimensional time series raw dataset, mark the location of missing values, and use time series interpolation methods to fill in the missing values. S22. Calculate the rate of change of each data point between adjacent time points. When the rate of change exceeds the set abnormal threshold, mark the data point as an abnormal point and replace the abnormal point with the average of the two normal points before and after. S23. Perform min-max normalization on the four raw data of luminous flux, junction temperature, color temperature and driving current respectively. Construct a sliding window of a set length on the normalized data sequence, extract local time segments in turn, and calculate the mean, variance and rate of change in each window to generate feature vectors. S24. Concatenate all the feature vectors extracted from the sliding window in timestamp order to form an aging feature time series.

4. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S3 specifically includes: S31. Divide the aging feature time series into sliding time windows of a preset length according to the time sequence. Each window retains continuous time step data of a preset size. Arrange the four original data corresponding to the time step row by row to form an original feature matrix with a preset number of rows and 4 columns, representing the aging state within the time period. S32. For each row of the original feature matrix in each time window, input a set of preset weight vectors with the same dimension to perform dot product operation on the four feature values ​​and add bias terms. Map each original four-dimensional feature vector to a one-dimensional vector. Concatenate the one-dimensional vectors of all time steps in order to construct the embedding vector. Add position encoding according to the time step index in the embedding vector to preserve the sequential features of the time series. S33. Input the embedding vector into the encoder of the improved Crossformer model, perform cross-feature attention mechanism processing on multiple feature vectors in each time step, use the embedding vector of each feature vector as the query vector, calculate the dot product similarity between it and the embedding vectors of other feature vectors in the same time step, and obtain the attention weight after normalization using the Softmax function. S34. Multiply these attention weights by the embedding vectors of the corresponding feature vectors and sum them to obtain the representation of each feature vector after incorporating the influence of other feature vectors at the current time step, and output the cross-feature representation sequence. S35. Taking the cross-feature representation sequence as input, the multi-head attention mechanism is executed sequentially along the time dimension. The dot product similarity of the embedding vectors of each set of feature vectors at different time steps is calculated to construct the time dependency matrix. The historical state of the current feature vector at different time points is weighted according to the attention weight to generate the trend evolution feature within the entire time window and output the time context representation vector sequence of each feature vector. S36. The cross-feature representation sequence and the temporal context representation vector sequence are concatenated according to the same time step index, and the concatenated vector is compressed in dimension using a linear fusion transformation. At the same time, a residual connection structure is introduced, and the embedded vector before concatenation is added to the fusion output vector. LayerNorm normalization is performed to obtain the fused aging trend feature representation tensor. S37. In the decoder, the fused aging trend feature representation tensor is transformed by a feedforward neural network and passed through two linear layers and one activation function layer in sequence to output the aging trend prediction value containing the luminous flux maintenance rate and color temperature change amount of multiple subsequent time steps.

5. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 4, characterized in that, The improved Crossformer model specifically includes: The aging characteristic time series after cleaning is divided into multiple time segments of equal length according to time sequence. Each time segment contains a set number of time steps. Each time step records four values: luminous flux, junction temperature, color temperature and driving current. In each time segment, a position code is generated for each time step according to its relative position number in the current time segment. The encoding method adopts trigonometric function transformation, that is, the time step number is input into the sine function and the cosine function respectively to form a set of continuously changing encoding vectors, representing the relative position of each time step in the local window. For each time step feature value, the corresponding luminous flux, junction temperature, color temperature and driving current four feature values ​​are standardized and concatenated into a one-dimensional vector. The one-dimensional vector is then concatenated and combined with the sliding position encoding vector corresponding to the time step to form an embedded entry with position-aware information. The embedding entries of all time steps within each time segment are combined into a three-dimensional tensor containing the number of time steps, feature dimension, and batch size, and then fed into the improved Crossformer model.

6. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S4 specifically includes: organizing the current aging characteristic time series of the target LED package in chronological order and inputting it into the improved Crossformer model, extracting the correlation and trend characteristics between multiple variables in sequence, and outputting the predicted values ​​of luminous flux, junction temperature, color temperature and driving current within the future preset time window to form the aging trend prediction value.

7. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S5 specifically includes: S51. Read the aging trend prediction values ​​output by the improved Crossformer model, organize the prediction value sequence in time order, and extract the prediction values ​​of luminous flux, junction temperature, color temperature and drive current for each time step. S52. Read the preset aging threshold standards respectively, compare each predicted value with the corresponding aging threshold, and determine whether the aging threshold conditions are exceeded. If any one of them is exceeded, mark the current time step as the aging warning time point. S53. Statistically analyze all aging warning time points, calculate the continuous warning duration and cumulative exceedance. If the continuous exceedance time exceeds the preset duration threshold, mark the corresponding LED package as being in aging warning state. S54. Based on the changing trends of various predicted values, draw visualization charts of luminous flux decline curve and junction temperature rise trend, and generate aging trend analysis results in chronological order to assess the lifespan stage and health status of LED packages.

8. The method for predicting the aging trend of LED packaging based on time-series modeling according to claim 1, characterized in that, S6 specifically includes: S61. The newly collected aging observation values ​​are compared one-to-one with the aging trend prediction values ​​predicted by the improved Crossformer model in chronological order. The residual between the predicted aging trend value and the actual value is calculated. The average value of all residuals is calculated and compared with the preset aging error tolerance threshold. If the aging error tolerance threshold is exceeded, the improved Crossformer model update process is started. S62. Combine the aging feature time series used for prediction in the previous stage with the newly collected observations to form an input-output sample pair, and input it into the current round of fine-tuning training. At the same time, initialize the learning rate and update step size required for the AdaBelief optimization algorithm. S63. For each trainable parameter in the Crossformer model, calculate the gradient direction of the current parameter in the error function. Then, take a weighted average of the gradient value of each parameter with the historical gradients to obtain a smooth trend direction. S64. The AdaBelief optimization algorithm calculates the difference between the current gradient and the trend direction, that is, the degree to which the current gradient deviates from the trend, and uses this deviation difference as a confidence index. If the deviation difference is greater than the preset deviation threshold, it returns to update; if the deviation difference is less than the preset deviation threshold, it continues to update. S65. Read the instantaneous gradient value of the current parameter and the previously accumulated average gradient trend. Take the difference between the two as the deviation of the current gradient. After squaring the deviation, take it as the variance estimate. Preset a very small constant. Divide the preset learning rate by the sum of the variance estimate and the constant to obtain the update step size of the current parameter. S66. Adjust all parameters in the direction of reducing error according to the calculated step size. After completing one round of parameter update, continue to predict the aging trend of the updated and improved Crossformer model, and record the prediction performance of the new round. Repeat the above fine-tuning process for several rounds until the prediction residual falls back to within the tolerance range.

Images