Method for early failure diagnosis and life prediction of LED based on spectral power distribution

Abstract

The application discloses a kind of LED early fault diagnosis and life prediction method based on spectral power distribution;The present application is based on similarity detection method to carry out fault diagnosis, extracts spectral characteristic value from statistical model, is reduced dimension by principal component analysis, and is clustered using K-means++ method, finally using distance and threshold comparison determines abnormal time.On this basis, long short-term recurrent neural network is established to predict spectral characteristic value, remodel spectrum, and predict remaining useful life.The present application combines spectral power distribution, and carries out fault diagnosis and life prediction to accelerated step aging white light LED based on long short memory recurrent LSTM neural network method, which greatly improves the accuracy of LED early abnormal detection and remaining life prediction.

Description

Technical Field

[0001] This invention belongs to the field of LED testing technology, specifically, it relates to a method for early fault diagnosis and lifetime prediction of LEDs based on spectral power distribution. Background Technology

[0002] As a semiconductor light source, the lifespan of LEDs is a crucial indicator of product quality and reliability. The length of LED lifespan directly impacts product performance stability, lifespan, and maintenance costs. Therefore, accurate early fault detection and lifespan prediction are essential for improving LED product quality and performance. Traditional LED lifespan testing typically requires lengthy experiments and observations, hindering rapid iteration in product development and manufacturing. Consequently, researchers have focused on early fault detection—the ability to accurately detect potential faults and problems early in the LED's lifespan. Early fault detection allows for proactive measures to repair or replace faulty LEDs, thereby extending product lifespan and improving product quality.

[0003] Indicators used for LED fault diagnosis include luminance, chromaticity, radiation, and electrothermal parameters. There are four main methods for fault diagnosis of LED luminaires: measuring and analyzing the electrical parameters of the LED luminaire, such as voltage and current; measuring and analyzing the optical parameters of the LED luminaire, such as luminous flux and color shift; measuring and analyzing the temperature of the LED chip; and measuring and analyzing the fusion results of optical, temperature, and electrical parameter information of the LED luminaire. Luminance and chromaticity parameters are widely used as LED health indicators, with the most critical parameter being lumen maintenance rate data. This data is mostly obtained through integrating spheres, which brings great inconvenience to manufacturers and users in monitoring.

[0004] Spectral power distribution refers to the distribution of radiant power emitted by a light source within the visible wavelength range, encompassing the most fundamental principles and mechanisms of light emission. It is commonly used as an important indicator for studying the optical performance of light-emitting diodes (LEDs), including their luminous and chromatic characteristics. Their luminous and chromatic parameters, such as luminous flux, chromaticity coordinates, correlated color temperature, color rendering index, and chromaticity coordinates, can all be calculated from the spectral distribution function. Therefore, early fault diagnosis can be performed by measuring the luminous flux obtained from the spectral power distribution.

[0005] Compared to traditional light sources, measuring and evaluating the lifespan of long-life LEDs is time-consuming, resulting in significant time costs. Typically, based on the IES LM-80-08 standard and the US DOS's "Energy Star Program Requirements: Luminaire Product Specifications," a testing period of 6000 hours is required. Obtaining sufficient data to support reliability assessment before product launch and estimating remaining useful life (RUL) in relatively short lifespan tests is a major challenge for manufacturers. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the present invention aims to provide a method for early fault diagnosis and lifetime prediction of LEDs based on a spectral power distribution time series model. This invention combines spectral power distribution, similarity detection, and a long short-term memory (LSTM) loop audit network algorithm to achieve fault diagnosis and lifetime prediction during the accelerated temperature aging process of white LEDs. Statistical functions are used to extract feature parameters for early fault diagnosis, health data is combined with LSTM for RUL prediction, and spectral reshaping provides more information about the LED, enabling dynamic monitoring of LED faults and lifetime.

[0007] The technical solution of the present invention is described in detail below.

[0008] This invention provides a method for early fault diagnosis and lifetime prediction of LEDs based on spectral power distribution, comprising the following steps:

[0009] Step 1: Every few hours, use an integrating sphere to collect the photoelectric parameters of the LED sample under accelerated step aging test conditions. The photoelectric parameters include luminous flux and spectral power distribution curves.

[0010] Step 2: Extract features from the acquired spectral power distribution curves using spectral decomposition methods;

[0011] Step 3: Dimensionality reduction of the extracted feature parameters using principal component analysis;

[0012] Step 4: Utilize the similarity between samples to perform anomaly detection to identify potential faulty samples and achieve early fault diagnosis of LEDs;

[0013] Step 5: Determine the minimum amount of data required for anomaly detection by evaluating the accuracy and training time of the similarity model used for anomaly detection in Step 4, thereby establishing a threshold point for predicting the remaining lifespan of the LED, which corresponds to the earliest time when the failure begins and the Euclidean distance at that time.

[0014] Step 6: Using the health data before the threshold point obtained in Step 5, the extracted spectral feature parameters are modeled using an LSTM model to capture the dynamic pattern of LED parameters changing over time. The predicted parameters are then spectrally reshaped to estimate the remaining lifespan of the LED.

[0015] In this invention, in step one, photoelectric parameters are collected every 168 hours under accelerated step aging test conditions; the accelerated step aging test adopts an accelerated step aging test under temperature stress.

[0016] In this invention, in step one, during the accelerated aging test, the temperature of the constant temperature chamber is gradually increased from 55°C to 115°C, with a heating rate of 10°C every 504 hours. Photoelectric parameters are tested on the samples after every 168 hours of aging test, and the aging test is stopped when the lumen maintenance rate drops to about 70%.

[0017] In this invention, during step two, the entire LED SPD is considered to be composed of a superposition of individual bell-shaped spectra, as shown in equation (2.1):

[0018]

[0019] Among them, SPD LED SPD represents the spectrum of an LED. i This represents the i-th SPD spectrum obtained from the decomposition, and n represents the total number of SPDs obtained from the decomposition, where n is 2;

[0020] One of the following methods—double Gaussian function, double Lorentz function, and double asymmetric double S-shaped function—is used as the spectral distribution model to extract spectral feature values. Preferably, the double Gaussian function is used as the spectral distribution model to extract spectral feature values.

[0021] In this invention, during step four, when detecting similarity anomalies, the K-means++ method is used for clustering, and finally, distance and threshold comparisons are used to determine anomalies; the steps are as follows:

[0022] Step 1: Perform K-means++ clustering on the dataset to divide it into different clusters, and calculate the cluster radius and centroid of each cluster;

[0023] Step 2: Calculate the Euclidean distance from each data point to be measured to the corresponding centroid of the cluster;

[0024] Step 3: Define the nearest cluster as the cluster S containing the data point to be tested. j The corresponding cluster centroid μ j The Euclidean distance from the data point to be measured to the corresponding centroid of the cluster is D. j ;

[0025] Step 4: Calculate the threshold T for each cluster j :

[0026]

[0027] Where R(μ) j ) represents cluster S j The corresponding cluster radius, σ(μ) j ) represents cluster S j The corresponding standard deviation describes the degree of bias, and d represents the dimensionality factor;

[0028] Step 5: In chronological order, sequentially add items at distance D. j With the corresponding cluster S j Threshold T j Compare, if D j >T j It is judged as abnormal; if D j ≤T j Proceed to the next data point comparison, and repeat this step until all data points have been compared.

[0029] In this invention, in step five, the similarity model is the K-means++ clustering model.

[0030] In this invention, in step six, the degree of change of the luminous flux of the SPD with the parameters in the spectral power distribution model is examined. Based on parameter sensitivity analysis, the parameters with a large influence are selected as the input parameters of the LSTM-RNN model. The spectral power distribution model is selected from one of the double Gaussian function, double Lorentz function, and double asymmetric double S-shaped function, which is the same type as that in step two.

[0031] In this invention, in step six, the parameters predicted by the LSTM-RNN model are used to backtrack the spectral power distribution model and reconstruct the LED spectrum. The luminous flux variation curve over time is obtained by integrating the spectrum, thereby predicting the remaining lifespan. Preferably, the spectral power distribution model is the same type used in the spectral decomposition in step two.

[0032] The present invention proposes a method based on a spectral power distribution feature value extraction model, combining a K-means++ model for anomaly detection and an LSTM-RNN method for spectral parameter prediction to predict lifetime. This method aims to improve the anomaly detection accuracy and remaining lifetime prediction capability of white LEDs in accelerated step-aging tests. Compared with existing technologies, the advantages of this invention are:

[0033] 1) Similarity-based measurement tests significantly reduce the time required for LED identification and the amount of data required for accelerated degradation testing by using predictive techniques. Time-series-based lifetime prediction methods provide more luminous and chromatic information about LEDs by reshaping the spectrum, offering a feasible solution for LED lifetime prediction.

[0034] 2) The proposed LSTM-based method for predicting and reconstructing spectra based on parameter degradation takes into account the dynamics and uncertainties of the data. The luminous flux prediction error is within 3%, enabling accurate prediction of luminous flux.

[0035] 3) By analyzing the features extracted from the SPD using a Gaussian model, it was found that in this accelerated step aging test, the failure occurred at 40%-50% of the lifetime (1176h-1344h), and phosphor aging was the main failure mechanism of the sample. Attached Figure Description

[0036] Figure 1 For HONGLITRONIC white LED: (a) LED structure diagram; (b) actual picture of a packaged single LED bead; (c) actual picture of the chip.

[0037] Figure 2 To accelerate the step aging test: (a) test system; (b) ambient temperature change.

[0038] Figure 3 This is a flowchart for fault diagnosis and life prediction.

[0039] Figure 4 Fit curves for the three models.

[0040] Figure 5 Dimensionality reduction results for 40 training data points (4 data points for each LED).

[0041] Figure 6 This is the intra-cluster sum of squares for 40 datasets.

[0042] Figure 7 K-means++ clustering results for 40, 60, and 80 training data points.

[0043] Figure 8 The results of monitoring 30% of the training data (LED5).

[0044] Figure 9 The fault diagnosis results are for 40% of the training data.

[0045] Figure 10 Anomaly detection results for training data: (a) 30% of training data; (b) 40% of training data.

[0046] Figure 11 The degeneracy curves of normalized feature parameters are extracted using a double Gaussian model: (a) y0; (b) A1; (c) A2; (d) ω1; (e) ω2; (f) xc1; (g) xc2.

[0047] Figure 12 For optical vision functions (CIE 2006).

[0048] Figure 13 The area ratio extracted by SPD.

[0049] Figure 14The graph shows the luminous flux as a function of parameters: (a) all parameters; (b) blue light chip parameters; (c) yellow phosphor parameters (where y0 is the same parameter in (b) and (c)).

[0050] Figure 15 The prediction results for five parameters based on LSTM-RNN are: (a) normalized y0; (b) normalized A1; (c) normalized A2; (d) normalized ω1; (e) normalized ω2.

[0051] Figure 16 This is the result of the luminous flux prediction. Detailed Implementation

[0052] The technical solution of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0053] 1. Accelerated step aging test and data collection

[0054] 1.1 Test Samples

[0055] This section mainly introduces the samples used in the accelerated step aging experiment and the experimental scheme involved. The 3000K color temperature HONGLITRONIC white LED (HL-A-2835H489W-S1-08L-HR3) used in this experiment consists of six main parts: phosphor, leads, support, chip, electrodes, and substrate. Figure 1 As shown. Phosphor is an important component of surface mount packages. It is located on top of the LED chip and is used to convert blue or ultraviolet light into visible light. The color of the phosphor determines the color of the LED's emission. Leads: Leads are the conductive connections in surface mount packages and are usually made of metal. Frame: The frame is the structural framework of the surface mount package, used to support and protect the LED chip and other components. Chip: The chip is the core part of the surface mount package and is usually made of semiconductor materials (such as gallium arsenide). Electrodes: Electrodes are used to provide current input and output. Substrate: The substrate is the bottom part of the surface mount package and is usually a circuit board.

[0056] The LED chip uses the 2835 standard chip, packaged in a surface-mount (SMD) format. SMD packaging is more compact than traditional through-hole packaging, allowing for smaller physical dimensions. SMD packaging establishes a reliable physical connection by directly soldering or bonding the chip to the circuit board. This connection method has lower resistance and inductance, helping to reduce signal loss and interference during transmission. Furthermore, SMD packaging is better resistant to mechanical stress and vibration, improving device reliability and stability. Therefore, it offers advantages such as small size, high reliability, wide beam angle, and low cost. The white LED uses InGaN as the chip material, emits white light by exciting yellow phosphors with blue light, and uses a yellow diffusing colloid material to improve light distribution and softness.

[0057] 1.1.1 Introduction to the Experiment

[0058] Traditional LED sample aging tests require a long aging process. A more efficient and time-saving method has been proposed: the Step-stress accelerated thermal aging test (SSATAT), which reduces the test time by changing the magnitude of the temperature stress.

[0059] LED sample HL-A-2835H489W-S1-08L-HR3 underwent accelerated aging test with temperature stress in a constant temperature chamber. Figure 2 (a) A total of 15 LED samples were used. Each sample was driven with a constant rated current (I = 120mA) and subjected to accelerated aging tests at 7 different temperature stress levels. During the test, the temperature of the incubator was gradually increased from 55°C to 115°C. A temperature increase of 10°C was performed every 504 hours (three weeks). After every 168 hours (one week) of aging test, the samples were removed from the incubator, and photoelectric parameter tests were performed using an integrating sphere for 1 hour. Ambient temperature changes were as follows: Figure 2 As shown in (b). The aging test was stopped when the lumen maintenance rate dropped to about 70%. A total of 22 cycles (3552h) of aging data were collected. Data from 10 LED samples were collected in this test for subsequent analysis.

[0060] 2. Construction of an early fault diagnosis model based on data-driven algorithms

[0061] We propose a flowchart for fault detection and lifespan prediction, as follows: Figure 3As shown, we first used an integrating sphere to collect photoelectric parameters of LED samples every 168 hours under accelerated step aging test conditions, including parameters such as luminous flux and spectral power distribution curves. Then, we used the collected spectral power distribution curves to extract features through spectral decomposition, performed principal component analysis for dimensionality reduction, and then used a similarity detection model for fault detection. We used training data with different training sample sizes and determined the minimum amount of data required for anomaly detection by evaluating the model's accuracy and training time, thus establishing a threshold point for predicting the remaining lifetime. Finally, we used the obtained threshold to predict the extracted parameters using an LSTM model, reconstructed the spectrum of the predicted parameters, estimated the remaining lifetime, and compared it with the true values ​​to verify its effectiveness.

[0062] Compared to traditional industry standards, the proposed method demonstrates superior performance and application potential in LED fault diagnosis and lifespan prediction. These findings have significant theoretical and practical implications for improving the reliability and lifespan management of LED devices.

[0063] 2.1 Extracting Feature Values ​​Based on Statistical Methods

[0064] In mixed light, the entire SPD of a white pc-LED can be considered as a superposition of a single bell-shaped spectrum, as shown in (2.1):

[0065]

[0066] SPD LED SPD represents the spectrum of an LED. i Let represent the i-th SPD spectrum obtained from the decomposition, and n represent the total number of SPDs obtained from the decomposition. Three widely used statistical functions are used to describe the SPD decomposition model, including the Gaussian function, the Lorentz function, and the asymmetric double sig function. The white LED's bimodal function (where n = 2) obtained by exciting yellow phosphors with blue light is the result of superimposing the above three functions twice, as shown in the model below:

[0067] 1) Bigaussian function:

[0068]

[0069] Where x represents wavelength, xc1 and xc2 represent peak wavelength, ω1 and ω2 represent full width at half maximum (FWHM), and A1 and A2 represent spectral area.

[0070] 2) Bi-Lorentz function:

[0071]

[0072] Where x represents wavelength, xc1 and xc2 represent peak wavelength, ω1 and ω2 represent full width at half maximum (FWHM), and A1 and A2 represent spectral area.

[0073] 3) Asymmetric double-S model:

[0074]

[0075] Where B1 and B2 represent amplitudes, xc1 and xc2 represent peak wavelengths, and ω 11 and ω 12 ω represents the full width at half maximum (FWHM). 31 and ω 32 Indicates low side energy variance, ω 21 and ω 22 This represents the variance of the high-side energy.

[0076] The collected SPD data were fitted using spectral distribution models (2.1)-(2.4), and the fitting curves using the three models are shown as follows. Figure 4 As shown, the goodness of fit was evaluated using the coefficient of determination (2.5). The extracted parameters can reflect the optical and colorimetric characteristics of the LED.

[0077]

[0078] Table 1. Fitting results of the three models

[0079]

[0080] The fitting results are shown in Table 1. The R-values ​​of the three models are... 2 The values ​​are all close to 1, indicating that the three statistical functions have a good fitting effect on the spectral power distribution curve of this type of pc-WLED. Among them, the fitting degree of the double Gaussian function and the asymmetric double S function both reach 0.99. It is worth noting that the number of decompositions in formula (2.1) and the number of parameters in (2.2)-(2.4) determine the accuracy of the function. The superposition of multiple SPD decompositions can provide higher fitting accuracy, but at the same time, it will increase the complexity of the model. From the perspective of SPD construction, too many parameters are detrimental to subsequent parameter analysis and spectrum reshaping, because each parameter affects the shape of the SPD. Under the premise of ensuring the fitting degree, the complexity of the model should be reduced and the number of parameters decomposed from the model should be reduced as much as possible. Therefore, the double Gaussian function is used for subsequent modeling, with a total of seven parameters. The initial parameter values ​​are shown in Table 2.

[0081] Table 2. Feature extraction and initial parameter results

[0082]

[0083] 2.2 Dimensionality reduction based on principal component analysis and K-means++ clustering

[0084] After extracting the feature parameters, we applied Principal Component Analysis (PCA) for dimensionality reduction and K-means++ clustering to process and analyze the proposed feature parameters. PCA is a commonly used dimensionality reduction technique that transforms the original features into a new set of principal components through linear transformation to retain the maximum information in the data. By applying PCA, we can reduce the dimensionality of the feature parameters and extract the most representative features, thereby simplifying the data analysis process and improving the efficiency and interpretability of the model. Simultaneously, we used the K-means++ clustering algorithm to divide the samples into different clusters. This algorithm iteratively calculates the distance between samples and assigns the samples to the nearest cluster. The K-means++ algorithm has good convergence and stability and can effectively handle large-scale datasets. We applied K-means++ clustering to group the data into clusters with similar features to reveal the inherent structure and patterns of the data. Through PCA and K-means++ clustering, we can analyze and process the feature parameters in the process of white LED anomaly detection and RUL prediction, providing a more accurate and reliable foundation for subsequent data modeling and prediction.

[0085] 2.2.1 Principal Component Analysis

[0086] Principal component analysis (PCA) is a widely used data dimensionality reduction algorithm. It is employed in many fields to simplify datasets, reduce feature dimensionality, and extract the most representative features. By mapping the original data from a high-dimensional space to a low-dimensional space, redundant information is removed while preserving the most relevant information from the original data, making the data easier to process and analyze.

[0087] The core of the algorithm is to achieve data dimensionality reduction and feature extraction by decomposing the covariance matrix into eigenvalues, thus obtaining the corresponding eigenvectors. For α variables, X consists of n data points. α×n A vector set. After centering the matrix, the covariance matrix can be calculated, with the eigenvalues ​​of the variable α arranged in descending order as μ1, μ2, μ3…μ α The corresponding feature vectors are A1, A2, A3…A α The original matrix contains α variables. After principal component analysis, the number of variables is reduced. The number of principal components is selected based on their contribution rates. The contribution rate of the j-th principal component is:

[0088]

[0089] Generally, the more principal components retained, the more useful information they contain. The cumulative contribution rate of the first m principal components is:

[0090]

[0091] When the cumulative contribution rate reaches 85%, the first m principal components can be considered to retain the effective information of the original data. For 10 LEDs, 20%, 30%, and 40% of the data from each LED's 22 data points are extracted as training data, i.e., 4, 6, and 8 data points are selected from time 0 for each LED, and the remaining data are used as test data for fault diagnosis. The seven feature parameters extracted from the SPD model are reduced to three dimensions through principal component analysis. Figure 5 The results are shown for 40 training datasets (4 datasets per LED). The x-axis represents the feature value μ. j The y-axis represents the eigenvalue μ. j The corresponding contribution rate p j (The total contribution rate is 10). When accumulating to the third principal component, the cumulative contribution rate exceeding 85% is what we need. The results for 40, 60, and 80 data points are all three-dimensional.

[0092] 2.2.2 K-means++ model

[0093] K-means is an unsupervised clustering method that divides a dataset into multiple clusters by calculating the distance between sample points. Its main goal is to make the sample points within each cluster as close and compact as possible, while keeping the sample points relatively separated between different clusters. The algorithm works by iteratively updating the cluster centroids until convergence is achieved. Traditional K-means algorithms can only obtain the number of clusters K through brute-force search and manual trial and error; therefore, a heuristic method (elbow rule) is used to determine the value of K. The basic idea is to try different values ​​of K and calculate the variance of the clustering results corresponding to each K value. As the value of K increases, the number of sample points in each cluster decreases, thus reducing the overall variance. However, as the value of K continues to increase, the impact of the smaller number of sample points in each cluster gradually weakens, and the improvement in overall variance gradually diminishes. The clustering effect is evaluated by considering the sum of squared errors within the clusters, as shown in Equation (4.3).

[0094]

[0095] Where x i Represents the training value, μ i Represents x i Corresponding cluster S i The center of mass. In this experiment, x i and μ i All data are three-dimensional vectors. The training sets of 40, 60, and 80 data points were divided into five classes. Figure 6 Elbow plot of intra-cluster sum of squares for 40 datasets.

[0096] In the improved K-means++ algorithm, the initial cluster centers are not randomly selected as in the traditional K-means algorithm. The steps for selecting initial cluster centers are as follows:

[0097] Step 1: Randomly select a sample point from the M datasets as the initial cluster center μ1;

[0098] Step 2: Calculate the distance from the remaining M-1 datasets to the initial cluster centers. The furthest one is the second cluster center, μ2;

[0099] Step 3: Calculate the distance of each sample to the cluster center and assign the sample to the cluster with the closest distance. Add the sample with the furthest distance from its own cluster to the new cluster center.

[0100] Step 4: Repeat step 3 until there are 5 cluster centers, then restart the traditional K-means algorithm process.

[0101] Clustering results for 40, 60, and 80 training data points are as follows: Figure 7 As shown in the figure. The x-axis, y-axis, and z-axis represent principal component 1, principal component 2, and principal component 3 after dimensionality reduction by principal component analysis, respectively.

[0102] 2.3 Early Fault Diagnosis Based on Similarity Detection

[0103] For datasets x1, x2…x n Perform K-means++ clustering, resulting in n clusters, S1, S2…S… n The corresponding cluster centroids are μ1, μ2…μ n The remaining data is then processed using a similarity-based fault detection method. Similarity-based fault detection is a commonly used unsupervised fault detection method that utilizes the similarity between samples to identify potential faulty samples. The steps are as follows:

[0104] Step 1: Calculate cluster S i The corresponding cluster radius R(μ) i ) and cluster centroid μ i .

[0105]

[0106]

[0107] Step 2: Calculate the remaining data into clusters S1, S2...S respectively. n Corresponding centroids μ1, μ2…μ n Euclidean distance;

[0108] Step 3: The nearest cluster is the cluster S to which this point belongs. jThe corresponding cluster centroid is cluster centroid μ. j The distance is D j ;

[0109] Step 4: Calculate the threshold T for each cluster. j Where σ(μ) j ) represents cluster S j The corresponding standard deviation describes the degree of bias, and d represents the dimension factor.

[0110]

[0111] Step 5: In chronological order, sequentially add items at distance D. j With the corresponding cluster S j Threshold T j Compare, if D j >T j It is judged as abnormal; if D j ≤T j Proceed to the next data point comparison, and repeat this step until all data points have been compared.

[0112] It is worth noting that the threshold in step 4 is the improved result. According to... Figure 7 We can see that the clustering results are ellipsoidal, with different radii for the major and minor axes. Directly using the radius to replace the threshold will result in significant errors. When the threshold is too large, false negatives will occur; when the threshold is too small, the fault detection rate will increase. To improve accuracy and consider the distribution of detection points, the new threshold is adjusted. At this point, we obtain the approximate time of the fault occurrence, which can be used for monitoring.

[0113] 2.4 Fault Diagnosis Results and Discussion

[0114] 20%, 30%, and 40% of the data were used for training, and the remaining data were used for similarity fault diagnosis. False alarms were defined as fault detection times within 5% of the actual fault time (ATF); false misses were defined as LEDs failing to detect anomalies at the end of their lifespan (luminous flux decay to 70% or color shift to 0.007).

[0115] When training with 30% of the data, the threshold T j The value is 0.615. LED5 detects an anomaly at time 0 and issues a false warning, such as... Figure 8 As shown. 40% training data, fault diagnosis results for ten LEDs are as follows. Figure 9 As shown.

[0116] Table 3 Abnormal Diagnostic Results

[0117]

[0118] As shown in Table 3, the accuracy of anomaly detection improved with the increase of training data. Figure 10 As shown in (a), with 30% of the training data, the fault diagnosis rate is 80%, and the fifth and ninth LEDs malfunction at time 0. Figure 10 (b) When 40% of the training data was available, the fault diagnosis rate reached 100%, with no premature warnings or missed detections. Furthermore, the fault diagnosis time was consistently between 40% and 50% of the actual lifespan, i.e., fault diagnosis time = actual lifespan × (40%–50%). This indicates that we should begin fault detection at 1176 hours (the time corresponding to 40% of the data).

[0119] 3. Time series-based lifetime prediction models

[0120] The time-series-based lifetime prediction model aims to predict the remaining lifetime of white LEDs under accelerated step aging tests by analyzing extracted parameters and using an LSTM-RNN model for time-series modeling. Before modeling and prediction, based on the anomaly detection results from the previous section, we screened the data before the threshold (40% of the training data), normalized it, and then trained the model. Using the LSTM-RNN model, we can capture the dynamic patterns of LED parameters changing over time and predict future lifetime status. The model effectively processes time-series data and considers the influence of past information in the sequence on future predictions. Using the parameters obtained from the time-series prediction, we can backtrack the spectral power distribution model and reconstruct the LED spectrum. By integrating the spectrum, we can obtain the luminous flux variation curve over time, thereby predicting the remaining lifetime and providing insights into the potential causes of LED degradation and the sensitivity of luminous flux to parameters.

[0121] 3.1 Time Series Forecasting Model Based on LSTM-RNN

[0122] LSTM (Long Short-Term Memory) and RNN (Recurrent Neural Network) are both commonly used neural network architectures in sequence data modeling and processing. Traditional RNN models suffer from vanishing and exploding gradients when processing sequence data, making it difficult for them to capture long-term dependencies. To address this issue, LSTM introduces a gating mechanism, using memory units to selectively remember, forget, and output information between different time steps. Therefore, compared to traditional RNN models, LSTM can better handle long-term dependencies, resulting in better performance in sequence modeling tasks. LSTM-RNN can capture long-term dependencies in time series data. By recursively learning past information in the sequence, LSTM-RNN can extract important patterns and trends and predict future data. LSTM-RNN adapts to sequence data of varying lengths by adaptively adjusting the network's state and weights, giving it an advantage in handling variable-length sequences.

[0123] The three control units of LSTM are the forget gate, the input gate, and the output gate. These three gates are used to remember and forget network data, thereby protecting and changing the cell state.

[0124] 1) Forget Gate: Determines which older information should be forgotten from the internal state, controlling the outflow of older information to retain information more relevant to the current time step. Based on the input x at the current time step... t The hidden state h of the previous time step t-1 Perform a nonlinear mapping of the sigmoid function and calculate the value f of the forget gate. t (Between 0 and 1, 0 represents forgetting and 1 represents retention), related to cell state C. t-1 Multiply.

[0125] f t =σ(W f ·[h t-1 x t ]+b f )#(4.7)

[0126] Among them W f It is the weighting factor, σ is the sigmoid activation function, and b f It is the bias term of the forget gate.

[0127] 2) Input gate: mainly updates cell information, including two parts: sigmoid activation and candidate value calculation.

[0128] i t =σ(W i ·[h t-1x t ]+b i )#(4.8)

[0129]

[0130] Where i t For the input gate, W i b i These represent the bias terms of the weight factors of the input gate; For candidate state information, W c b c These represent the weight factor and bias term of the candidate state update layer, respectively.

[0131] The output i of the activation function t Output of the candidate value calculation section Element-wise multiplication is performed to obtain the final updated value, which is added to the internal state to reflect the importance of the current input data and the influence of candidate values. Based on the input of the current time step and the hidden state of the previous time step, a new internal state, C, is calculated. t-1 Update to C t .

[0132]

[0133] 3) Output gate: Determines the output information. Based on the input x at the current time step. t The hidden state h of the previous time step t-1 The result of the nonlinear mapping of Sigmoid is multiplied by the cell state after processing with the tanh activation function to obtain the final result.

[0134] o t =σ(W o [h t-1 x t ]+b o )#(4.11)

[0135] h t =o t *tanh(C t )#(4.12)

[0136] Among them o t It is the output at the current moment, W o b o The bias term, h, represents the weighting factor of the output gate. t It is the currently hidden state.

[0137] This paper utilizes the PyTorch deep learning framework for LSTM-RNN prediction, with a hidden layer size of 64, a recursive layer count of 5 in the LSTM model, and a batch size of 1. RMSE and training time are used as performance evaluation metrics. PyTorch is a dynamic graph framework that allows users to define and modify computation graphs in a natural way. This makes implementing LSTM models in PyTorch highly flexible, facilitating easy model building, modification, and debugging. Components can be freely added, modified, or removed according to specific problems and needs to achieve optimal performance and results. Furthermore, it provides high-level abstraction, making model definition and training more concise and intuitive, and leveraging the powerful computing capabilities of GPUs to significantly improve training and inference speeds.

[0138] 3.2 Parameter degradation prediction and spectral reshaping

[0139] After normalizing the seven parameters y0, A1, A2, ω1, ω2, xc1, and xc2 obtained from the double Gaussian model, as follows: Figure 11 As shown, xc1 and xc2 remain almost unchanged, assuming these two features do not degenerate in this case. Using LSTM-RNN to predict the remaining five parameters, the fault diagnosis results yield healthy data, meaning parameter degradation prediction begins when 40% of the data is available. The sliding window is modified, weights are optimized and searched, and the optimal model corresponding to the minimum RMSE is saved.

[0140] The degradation prediction of five parameters y0, A1, A2, ω1, ω2, xc1, and xc2 can be substituted back into the double Gaussian function model in formula (2.2) to predict the SPD evolution process of white LEDs. Formula (4.13) can be used to predict the luminous flux decay curve of white LEDs, and the accuracy of this method is judged by comparing the predicted values ​​with experimental values. Simultaneously, the remaining lifespan of the LED can be predicted based on the luminous flux decay to 70%.

[0141]

[0142] Where Φ is the luminous flux and λ represents the wavelength, SPD LED V(λ) is the spectral power distribution function of the LED, and V(λ) is the spectral luminous function, which describes the sensitivity of the human eye to different wavelengths of light. The specific function is as follows: Figure 12 As shown:

[0143] 3.3 Lifetime Prediction Results and Discussion

[0144] SPDs (Surface Mount Devices) can be used to characterize the optical performance of LEDs, and changes in their parameters can be used to observe characteristic aging patterns within the LED package. The failure mechanism of the package is believed to be related to the area of ​​the blue chip and yellow phosphor portions in the SPD curve, which depend on the light energy emitted by the LED chip and phosphor, respectively. White LED failure modes are mainly divided into three types: LED chip degradation, phosphor degradation, and package material failure.

[0145] In the long-term operation of LEDs, failure is rarely a single event; it is usually accompanied by other failures, resulting in a coupled failure model of three types: 1) LED chip degradation: The phosphor conversion efficiency depends on the blue light energy emitted by the chip. Chip degradation manifests as degradation in both the emission and conversion spectra; 2) Phosphor degradation: Phosphor degradation leads to a significant reduction in the conversion spectral area without affecting the emission spectrum; 3) Degradation of the encapsulating silicone polymer: LED encapsulation typically uses silicone polymers to protect the LED chip and phosphor. Over prolonged use or in high-temperature environments, the encapsulating silicone polymer may degrade, leading to optical performance degradation. Emission spectrum degradation is more pronounced than phosphor degradation because the polymer materials are primarily materials such as silicone and epoxy resins, which have a greater impact on short wavelengths, thus resulting in more significant emission spectrum degradation.

[0146] exist Figure 13 The image shows the area parameters extracted from the model, specifically the area ratio A1 / A2 of the blue LED chip and the yellow phosphor. The exponential growth observed in the image indicates that phosphor aging is the primary degradation mechanism in this accelerated aging test. The chemical components of the phosphor may be affected by oxidation, decomposition, or other chemical reactions, leading to structural changes and consequently affecting luminescence performance. This may be related to factors such as prolonged use and high-temperature environments promoting the oxidation or decomposition of the phosphor.

[0147] This study examines the luminous flux of a Spotlight Detector (SPD) as a function of seven parameters: y0, A1, A2, ω1, ω2, xc1, and xc2. A spectrum is constructed using the values ​​of any LED at time 0. Then, individual parameters are increased proportionally while others remain constant, and the luminous flux variation is observed and plotted. Figure 14 The variation of luminous flux with seven parameters is shown in Figure 14(a). Figure 14 (b) and Figure 14 (c) The effects of the blue light chip and the yellow phosphor on the luminous flux are shown respectively. We can see that A2, xc1, and xc2 have the greatest impact on the luminous flux. Among them, xc1 and xc2 have an exponential effect on the luminous flux, while y0, A1, and ω1 have a smaller impact on the luminous flux.

[0148] Table 4 shows the specific values ​​of the five parameters predicted based on LSTM-RNN.

[0149]

[0150]

[0151] Table 5. Parameter Mean Square Error

[0152]

[0153] Table 4 shows the specific values ​​of the LSTM-RNN training results after normalizing the feature values ​​extracted from the double Gaussian function. Figure 15 The results are visualized graphically. The average of the five normalized eigenvalues ​​predicted by the LSTM method is relatively close to the actual measured data marked with black dots. Table 5 shows the mean square error of the parameters. Parameter y0 exhibits significant fluctuations and abrupt changes, making accurate prediction of its changes difficult. However, the errors of all parameters can be controlled within 5%. Sensitivity analysis indicates that y0 has a relatively small impact on luminous flux, therefore it will not significantly affect luminous flux prediction.

[0154] Based on the predicted parameters, a Gaussian model is reconstructed to predict the SPD of the LED after 1176 hours, and the percentage error between the predicted value and the actual measured value is calculated.

[0155]

[0156] Where Φ Predicted X represents the predicted luminous flux. Real This indicates the light flux obtained in the experiment.

[0157] Table 6. Specific values ​​for predicted luminous flux

[0158]

[0159]

[0160] Table 6 shows the specific values ​​of the predicted luminous flux. Figure 16 The predicted luminous flux and percentage error are shown. The results indicate that the predicted luminous flux curve matches the experimental measurements very well, with an error within 3%. Furthermore, the predicted value reaches the L70 lifetime (70% luminous flux degradation) earlier than the experimental value, and the predicted L70 lifetime is 3102 hours based on the luminous flux degradation curve. This demonstrates that lifetime prediction based on a time series can be effectively used in accelerated step aging tests, and the spectral power distribution curve can provide more information on LED optoelectronic parameters. Lifetime prediction can improve the performance, quality, and reliability of LED products, meeting user needs and expectations.

Claims

1. A method for early fault diagnosis and lifetime prediction of LEDs based on spectral power distribution, characterized in that, Includes the following steps: Step 1: Every few hours, use an integrating sphere to collect the photoelectric parameters of the LED sample under accelerated step aging test conditions. The photoelectric parameters include luminous flux and spectral power distribution curves. Step 2: Extract features from the acquired spectral power distribution curves using spectral decomposition methods; Step 3: Dimensionality reduction of the extracted feature parameters using principal component analysis; Step 4: Utilize the similarity between samples to perform anomaly detection to identify potential faulty samples and achieve early fault diagnosis of LEDs; Step 5: Determine the minimum amount of data required for anomaly detection by evaluating the accuracy and training time of the similarity model used for anomaly detection in Step 4, thereby establishing a threshold point for predicting the remaining lifespan of the LED, i.e. the earliest time when a failure begins. Step Six: Using the health data before the threshold point obtained in Step Five, extract the spectral feature parameters, remove those with no fluctuations and select those with large fluctuations. Use an LSTM model for time series modeling to capture the dynamic patterns of LED parameters changing over time. Then, reshape the predicted parameters to estimate the remaining lifespan of the LED; where: In step two, during spectral decomposition, the entire LED SPD is considered as a superposition of individual bell-shaped spectra: ； in, Indicates wavelength. Indicates the spectrum of an LED. This represents the i-th SPD spectrum obtained from the decomposition, and n represents the total number of SPDs obtained from the decomposition, where n is 2; One of the following methods—double Gaussian function, double Lorentz function, and double asymmetric double S-shaped function—is used as the spectral power distribution model to extract spectral feature values. In step four, during similarity anomaly detection, the K-means++ method is used for clustering, and finally, distance and threshold comparisons are used to determine anomalies; the steps are as follows: Step 1: Perform K-means++ clustering on the dataset to divide it into different clusters, and calculate the cluster radius and centroid of each cluster; Step 2: Calculate the Euclidean distance from each data point to be measured to the corresponding centroid of the cluster; Step 3: Set the cluster with the closest distance as the cluster where the data point to be tested belongs. Corresponding cluster centroid The Euclidean distance from the data point to be measured to the corresponding centroid of the cluster is ; Step 4: Calculate the threshold for each cluster : ； in, Representative cluster The corresponding cluster radius, Representative cluster The corresponding standard deviation describes the degree of bias, and d represents the dimensionality factor; Step 5: Arrange the distances in chronological order. With the corresponding cluster threshold If a comparison is made, It is judged as abnormal; if Proceed to the next data point comparison, and repeat this step until all data points have been compared; In step six, the parameters predicted by the LSTM-RNN model are used to backtrack the spectral power distribution model and reconstruct the LED spectrum. By integrating the spectrum, the luminous flux variation curve over time is obtained, thereby predicting the remaining lifespan.

2. The method for early LED fault diagnosis and lifetime prediction based on spectral power distribution according to claim 1, characterized in that, In step one, photoelectric parameters are collected every 168 hours under accelerated step aging test conditions; the accelerated step aging test adopts an accelerated step aging test under temperature stress.

3. The method for early LED fault diagnosis and lifetime prediction based on spectral power distribution according to claim 2, characterized in that, In step one, during the accelerated aging test, the temperature of the constant temperature chamber is gradually increased from 55℃ to 115℃, with a temperature gradient of 10℃ every 504 hours. Photoelectric parameters are tested on the samples after every 168 hours of aging test. The aging test is stopped when the lumen retention rate drops to about 70%.

4. The method for early LED fault diagnosis and lifetime prediction based on spectral power distribution according to claim 1, characterized in that, In step two, a double Gaussian function is used as the spectral distribution model to extract spectral feature values.

5. The method for early LED fault diagnosis and lifetime prediction based on spectral power distribution according to claim 1, characterized in that, In step six, the degree of change of the luminous flux of the SPD with the parameters in the spectral power distribution model is examined. Based on parameter sensitivity analysis, the parameters with a large influence are selected as the input parameters of the LSTM-RNN model. The spectral power distribution model is selected from one of the double Gaussian function, double Lorentz function, and double asymmetric double S-shaped function, which is the same type as that in step two.

6. The method for early LED fault diagnosis and lifetime prediction based on spectral power distribution according to claim 1, characterized in that, In step six, the spectral power distribution model is the same type used in step two when decomposing the spectrum.

Images