System and method for lifetime simulation of an LED driving power supply

Abstract

The application discloses a kind of LED drive power life simulation system and method, belong to LED drive power life simulation technical field.The method includes: real-time acquisition LED drive power when running electrical performance data and extract initial electrical characteristic sequence;Build digital twin, extract multi-scale electrical transient disturbance features from it, generate heat-electricity coupling topology graph containing thermal disturbance field and electric stress field using graph neural network, and build multidimensional state evolution model accordingly;The electrical transient disturbance features are input into the evolution model for simulation, and the dynamic load spectrum of multiple stress-sensitive elements is obtained;Based on the dynamic load spectrum, the degradation rate of each element is calculated, and the degradation state data is obtained;The inflection point of degradation is extracted through self-attention mechanism, and the remaining life of each element is determined by combining the material degradation threshold, and the overall simulation life of the LED drive power is obtained by the minimum order statistics.

Description

Technical Field

[0001] This invention relates to the field of LED driver power supply lifespan simulation technology, specifically to a system and method for LED driver power supply lifespan simulation. Background Technology

[0002] LED driver power supply lifetime assessment typically relies on accelerated aging tests and degradation models based on statistical regression. Accelerated aging tests record the degradation of key parameters under constant high temperature or rated electrical stress, and the median lifetime is obtained by extrapolation. Some solutions introduce physical simulation, using finite element analysis to analyze the junction temperature of power devices or capacitor hot spots. When dealing with driver power supplies composed of multiple stress-sensitive components, these methods still treat the thermal and electric fields as quasi-static or decoupled objects, and only use the aging trend of a single component to represent the behavior of the entire system. When the driver power supply operates under conditions of frequent start-stop and transient load changes, transient thermal shocks and electromagnetic stress crosstalk will occur between switching transistors, electrolytic capacitors, and magnetic components due to electrical transient disturbances. Existing decoupled assessment methods cannot reconstruct this coupling process in both time and space dimensions. The resulting degradation data often lacks the evolution details of the thermal-electric coupling topology, and the simulated dynamic load spectrum deviates significantly from the actual stress amplitude and cycle frequency experienced by each component. In the degradation state identification stage, common methods determine failure by setting fixed performance degradation thresholds or using linear degradation trajectories, failing to perform correlation-focused analysis on abrupt changes in the degradation sequence, making it difficult to identify accelerated degradation inflection points from within the time-series data. Furthermore, lifetime summarization for multi-component systems often employs simple averaging of individual component lifetimes or taking thresholds and then taking the union of the results, neglecting the worst-case determination characteristics of weak links in competitive failures, as reflected by order statistics. The resulting overall remaining lifetime is often optimistic and does not match the sudden failure modes encountered in actual operation. To improve the lifetime simulation accuracy of LED driver power supplies in complex power electronic environments, it is necessary to address the problems of accurate dynamic characterization of thermo-electric coupling topology and overall lifetime prediction based on self-focus on degradation inflection points and the inference of the worst-case component order. Summary of the Invention

[0003] The purpose of this invention is to provide a system and method for simulating the lifespan of LED driver power supplies, in order to solve the problems that existing simulation methods cannot accurately characterize the dynamic characteristics of thermal-electric coupling of multi-stress sensitive components under transient conditions, and do not effectively utilize degradation inflection point identification and competitive failure sequence statistics to infer the overall lifespan of the power supply.

[0004] To achieve the above objectives, the present invention provides the following technical solution: The present invention provides a method for simulating the lifetime of an LED driver power supply, comprising: real-time acquisition of electrical performance data of the LED driver power supply during operation, and extraction of an initial electrical feature sequence from the electrical performance data; wherein the LED driver power supply includes multiple stress-sensitive elements; constructing a digital twin of the LED driver power supply, extracting multi-scale electrical transient disturbance features from the initial electrical feature sequence, and generating a thermal-electric coupling topology graph corresponding to the multiple stress-sensitive elements through a graph neural network based on the electrical transient disturbance features, the thermal-electric coupling topology graph containing the thermal disturbance field and electrical stress field of each stress-sensitive element, and constructing multiple [other components] based on the adjacency relationship of each node in the thermal-electric coupling topology graph. A multi-dimensional state evolution model is used to simulate the transient response of each stress-sensitive element. The electrical transient disturbance characteristics are input into the multi-dimensional state evolution model, and the time evolution process of the multi-dimensional state evolution model is simulated to obtain the dynamic load spectrum of the multiple stress-sensitive elements. Based on the dynamic load spectrum, the degradation rate of the multiple stress-sensitive elements is calculated to obtain the degradation state data of the LED driver power supply. According to the degradation state data, the degradation inflection point of the multiple stress-sensitive elements is extracted through a self-attention mechanism, and combined with the material degradation threshold of each stress-sensitive element, the remaining lifetime of each stress-sensitive element is obtained. Then, based on the remaining lifetime, the simulated lifetime of the LED driver power supply is obtained through the minimum order statistic.

[0005] As a technical solution of the present invention, the step of extracting the initial electrical feature sequence from the electrical performance data includes: performing a joint time-domain and frequency-domain transformation on the electrical performance data to obtain a time-spectrum graph; and using a convolutional autoencoder to extract high-dimensional electrical features from the time-spectrum graph to obtain the initial electrical feature sequence. Through joint time-frequency analysis, the transient information of non-stationary electrical signals can be fully preserved. The convolutional autoencoder automatically learns deep representations in an unsupervised manner, making the initial electrical feature sequence highly robust to noise and fully reflecting the differences in electrical behavior of the LED driver power supply under various operating conditions.

[0006] As a technical solution of the present invention, the construction of a digital twin of the LED driver power supply, extracting multi-scale electrical transient disturbance features from the initial electrical feature sequence, includes: establishing a physical parameter library of the LED driver power supply; associating the initial electrical feature sequence with the physical parameter library to generate an electrical transient representation network of the digital twin; performing sliding window decomposition on the electrical transient representation network; and extracting energy fluctuations in different frequency bands through wavelet packet transform to obtain the multi-scale electrical transient disturbance features. This approach combines a physical model with data-driven processing, enabling the digital twin to accurately reflect component-level electrothermal transient behavior. The multi-scale decomposition ensures simultaneous capture of rapid disturbances and slow drifts, improving the dynamic resolution of lifetime simulation.

[0007] As a technical solution of the present invention, the step of generating a thermo-electric coupling topology graph corresponding to the plurality of stress-sensitive elements through a graph neural network based on the electrical transient disturbance characteristics includes: mapping the plurality of stress-sensitive elements as nodes of the graph neural network, using the electrical transient disturbance characteristics as the initial attributes of the nodes; constructing edges between nodes according to the heat conduction paths and electrical connection paths between each node; and dynamically updating the weights of each edge through a graph attention mechanism to obtain the thermo-electric coupling topology graph. By using the graph attention mechanism to adaptively learn the thermal and electrical interaction strength between each element, the topology graph can realistically reproduce the multi-physics coupling relationship and dynamically adjust with changes in operating conditions, thereby accurately characterizing the thermal stress concentration and electrical stress distribution of weak points.

[0008] Preferably, the step of constructing a multidimensional state evolution model based on the adjacency relationships of nodes in the thermo-electric coupling topology includes: aggregating the thermal perturbation fields and electrical stress fields of adjacent nodes of each node to form a local state vector for each node; performing nonlinear mapping on each local state vector; and constructing the state transition function of each node at different time steps using a recurrent neural network to obtain the multidimensional state evolution model. Through the aggregation of local states and the temporal modeling capabilities of the recurrent neural network, the multidimensional state evolution model can simulate the nonlinear transient response of each stress-sensitive element over continuous time, accurately predicting the temporal evolution of the state caused by thermo-electric coupling.

[0009] Furthermore, the step of inputting the electrical transient disturbance characteristics into the multidimensional state evolution model and simulating the time evolution process of the multidimensional state evolution model to obtain the dynamic load spectrum of the multiple stress-sensitive elements includes: inputting the electrical transient disturbance characteristics as initial conditions into the multidimensional state evolution model; iteratively calculating the state values ​​of each node in continuous time steps according to the state transition function to obtain the time series of each node; performing power spectral density analysis on each time series to extract the amplitude and frequency components of each node, thereby generating the dynamic load spectrum. The dynamic load spectrum obtained in this way not only includes the stress amplitude but also covers the energy distribution in the frequency domain, providing more comprehensive load characteristics for degradation analysis.

[0010] As a technical solution of the present invention, the step of calculating the degradation rate of the plurality of stress-sensitive elements based on the dynamic load spectrum to obtain the degradation state data of the LED driver power supply includes: extracting the stress amplitude and cycle number associated with each stress-sensitive element from the dynamic load spectrum to construct a stress-lifetime curve; sampling the parameters of the stress-lifetime curve using a low-difference sequence to generate multiple degradation path samples, and calculating the degradation rate of each stress-sensitive element based on each degradation path sample to obtain the degradation state data. Low-difference sequence sampling improves the uniformity of parameter space coverage, obtains a high-confidence degradation rate estimate with fewer samples, and thus efficiently quantifies the uncertainty of the degradation state.

[0011] As a technical solution of the present invention, the step of extracting the degradation inflection points of the plurality of stress-sensitive elements based on the degradation state data through a self-attention mechanism includes: encoding the degradation state data in the time dimension to generate degradation state embedding vectors; calculating the correlation weights between each embedding vector through a self-attention mechanism, and identifying the time point where the correlation weights change abruptly as the degradation inflection point. The self-attention mechanism can capture dependency mutations in long time series, realize automatic and accurate calibration of degradation acceleration points, and avoid subjective errors caused by manually setting thresholds.

[0012] Furthermore, the remaining lifetime of each stress-sensitive element is obtained by combining its material degradation threshold. Then, based on these remaining lifetimes, the simulated lifetime of the LED driver power supply is obtained using a minimum order statistic. This includes: comparing the degradation inflection point of each stress-sensitive element with its corresponding material degradation threshold; recording the corresponding time point when the degradation state data reaches the material degradation threshold to obtain the remaining lifetime of each stress-sensitive element; sorting the remaining lifetimes and taking the minimum value as the simulated lifetime of the LED driver power supply. This method uses the lifetime of the weakest component as the overall lifespan, conforming to the reliability logic of a series system. The simulation results are more conservative and secure, suitable for the long lifespan and high reliability requirements of LED driver power supplies.

[0013] The present invention also provides a system for simulating the lifespan of an LED driver power supply, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, wherein the processor executes the computer program to implement the steps of the LED driver power supply lifespan simulation method.

[0014] The technical effects and advantages provided by the present invention in the above technical solution are as follows: Multi-scale electrical transient disturbance features are extracted from the initial electrical feature sequence. Based on these features, a graph neural network is used to generate a thermo-electric coupling topology corresponding to multiple stress-sensitive elements. Each stress-sensitive element is mapped as a graph node, with the electrical transient disturbance features serving as the initial attributes of the nodes. Edges between nodes are constructed along the heat conduction path and the electrical connection path. The edge weights are dynamically updated using a graph attention mechanism, enabling the topology to reflect the coupling and transmission of thermal disturbance and electrical stress fields between elements in real time. The thermo-electric states of adjacent nodes are aggregated according to their adjacency relationships to form a local state vector, which is then nonlinearly mapped. A multi-dimensional state evolution model is constructed using a recurrent neural network. The electrical transient disturbance features are used as initial conditions input into this evolution model, and the state values ​​of each node at continuous time steps are iteratively calculated to obtain a dynamic load spectrum containing amplitude and frequency components. This process does not require pre-setting decoupling conditions, thus accurately reproducing the process of electrical transient disturbances propagating through the thermo-electric coupling topology network and forming the true dynamic stress spectrum of each element. This significantly improves the physical fidelity of the load spectrum, allowing subsequent degradation rate calculations to be based on highly reliable stress inputs. After obtaining degradation state data for multiple stress-sensitive components, the degradation state data is encoded along the time dimension to generate degradation state embedding vectors. A self-attention mechanism is used to calculate the correlation weights between each embedding vector, and the time point where the correlation weights change abruptly is identified as the degradation inflection point. This method can automatically focus on long-range dependencies and anomalous jumps in the degradation sequence, overcoming the limitation of the fixed threshold method in capturing the early acceleration of degradation. After obtaining the degradation inflection point of each component, it is compared with the material degradation threshold to accurately define the remaining lifetime of each component. The remaining lifetimes of all stress-sensitive components are then sorted, and the minimum value of the minimum order statistic is taken as the simulated lifetime of the LED driver power supply. This synthesis path of overall system lifetime inference based on competing failures and worst-case scenarios of weak points ensures that the simulated lifetime is consistent with the physical fact that failure is dominated by the most sensitive component in actual operation. This avoids the artificially inflated expectations caused by simple averaging after independent evaluation, improving the robustness and reference value of the lifetime prediction results. Attached Figure Description

[0015] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0016] Figure 1 This is a flowchart of the LED driver power supply lifespan simulation method; Figure 2 This is a flowchart of electrical feature extraction based on time-spectral graphs and convolutional autoencoders; Figure 3 This is a graph showing the amplitude variation of key frequency components in the time spectrum of electrical performance data; Figure 4 It is a graph showing the normalized intensity of the thermal disturbance field and the electric stress field of the stress-sensitive element as a function of time. Figure 5 It is a stress-life curve diagram of multiple stress-sensitive elements; Figure 6 It is a degradation state curve of the degradation path sample. Detailed Implementation

[0017] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] See Figure 1This invention provides a method for simulating the lifetime of an LED driver power supply, specifically for LED driver power supplies containing multiple stress-sensitive elements. The method involves real-time acquisition of electrical performance data of the LED driver power supply during operation, and extraction of an initial electrical feature sequence from this data. A digital twin of the LED driver power supply is constructed, and multi-scale electrical transient disturbance features are extracted from the initial electrical feature sequence. Based on these features, a graph neural network is used to generate a thermo-electric coupling topology corresponding to the multiple stress-sensitive elements. This topology contains the thermal disturbance field and electrical stress field of each stress-sensitive element. A multi-dimensional state evolution model is constructed based on the adjacency relationships of nodes in the topology to simulate the transient response of each stress-sensitive element. The electrical transient disturbance features are input into the multi-dimensional state evolution model, and the time evolution process of the model is simulated to obtain the dynamic load spectrum of the multiple stress-sensitive elements. Based on the dynamic load spectrum, the degradation rate of the multiple stress-sensitive elements is calculated, yielding the degradation state data of the LED driver power supply. Based on the degradation state data, the degradation inflection points of multiple stress-sensitive elements are extracted through a self-attention mechanism. Combined with the material degradation threshold of each stress-sensitive element, the remaining lifetime of each stress-sensitive element is obtained. Then, based on each remaining lifetime, the simulated lifetime of the LED driver power supply is obtained through the minimum order statistic.

[0019] Example 1: In specific implementation, please refer to Figure 2 The electrical performance data is subjected to a joint time-domain and frequency-domain transformation to obtain a time-spectrum graph. The electrical performance data is collected by a sampling device at a fixed sampling frequency, capturing instantaneous voltage and current values ​​during the operation of the LED driver power supply, forming a discrete-time series. The joint time-domain and frequency-domain transformation employs a short-time Fourier transform (SFT). The SFT segments the discrete-time series using a window function, and performs a Fourier transform on each segment. A Hanning window is used, with its length set based on the minimum duration of the electrical transient disturbance feature to be captured. During the SFT, the window function slides along the time axis with an overlap step size set to half the window length, ensuring overlap between adjacent windows and improving the temporal continuity of the time-spectrum graph. The amplitude spectrum of each windowed signal segment is calculated, and the amplitude spectra of all signal segments are arranged chronologically to form a two-dimensional time-spectrum graph. One dimension of the time-spectrum graph is time, and the other is frequency; each element in the time-spectrum graph represents the amplitude value at a specific time and frequency. The time-spectrum is stored in matrix form, with rows corresponding to frequency components and columns corresponding to time frames.

[0020] In some embodiments, a convolutional autoencoder (CAE) is used to extract high-dimensional electrical features from the time-spectrum image to obtain an initial electrical feature sequence. The CAE consists of an encoder and a decoder. The encoder is used during high-dimensional electrical feature extraction, and the decoder assists the encoder in learning feature representations during the training phase. The encoder contains a hierarchical structure with multiple convolutional and pooling layers connected alternately. Each convolutional layer contains several convolutional kernels, the size of which is set to adapt to the time-frequency local correlation pattern of the time-spectrum image. In one embodiment, the first convolutional layer has a 5×5 kernel size, 32 kernels, and a stride of 1. After performing a convolution operation on the time-spectrum image, it outputs a first feature map. The first pooling layer performs max pooling on the first feature map, with a pooling window size of 2×2 and a stride of 2, outputting a first downsampled feature map. The second convolutional layer has a 3×3 kernel size, 64 kernels, and a stride of 1. After performing a convolution operation on the first downsampled feature map, it outputs a second feature map. The second pooling layer performs max pooling on the second feature map with a window size of 2×2 and a stride of 2, outputting the second downsampled feature map. The third convolutional layer has a kernel size of 3×3, 128 kernels, and a stride of 1. After performing convolution on the second downsampled feature map, it outputs the third feature map. The third pooling layer performs max pooling on the third feature map with a window size of 2×2 and a stride of 2, outputting the third downsampled feature map. The third downsampled feature map is unfolded into a one-dimensional vector, which is then mapped through a fully connected layer to a high-dimensional electrical feature vector of a preset length, set to 128.

[0021] The training process of the convolutional autoencoder uses a set of sample time-frequency spectra, which is generated from historically acquired LED driver power supply electrical performance data through a short-time Fourier transform. During training, the encoder extracts features from the input sample time-frequency spectra to obtain a high-dimensional electrical feature vector, and the decoder reconstructs the reconstructed time-frequency spectra from the high-dimensional electrical feature vector. The decoder contains upsampling layers and convolutional layers symmetrical to the encoder. The training loss function uses the mean squared error between the input sample time-frequency spectra and the reconstructed time-frequency spectra. The formula for calculating the mean squared error is:

[0022] in, This represents the mean squared error loss value. This represents the total number of frequency components in the time-frequency spectrum. This represents the total number of time frames in the time spectrum. The index represents the frequency component, with a value ranging from 1 to... Each integer in Indicates the index of a time frame, with a value ranging from 1 to... Each integer in In the spectrum of the input sample, the first... Line number The element values ​​of the column, Indicates the first spectral element in the reconstruction spectrum. Line number The column contains element values. The Adam optimizer is used for training, with a learning rate of 0.001, a batch size of 16, and the number of training iterations determined based on the convergence of the loss value. After training, the parameters of the encoder are saved and used for feature extraction from the time-spectrum graph generated from the real-time acquired electrical performance data.

[0023] In practice, the real-time generated time-spectrum graph is input into the trained encoder. The encoder sequentially performs convolution, pooling, and fully connected mapping operations, outputting the corresponding high-dimensional electrical feature vectors. Following the order of time frames in the time-spectrum graph, the time-spectrum region corresponding to each time frame is input into the encoder, or the entire time-spectrum graph is divided into several sub-graphs along the time dimension, and each sub-graph is input into the encoder, sequentially obtaining the high-dimensional electrical feature vectors corresponding to each time segment. All high-dimensional electrical feature vectors are arranged in chronological order to form the initial electrical feature sequence. Each high-dimensional electrical feature vector in the initial electrical feature sequence constitutes a sequence element for subsequent processing.

[0024] See Figure 3 The figure shows the amplitude trends of the fundamental frequency component (50Hz), switching frequency component (20kHz), and transient disturbance frequency component (100kHz) of the LED driver power supply as a function of time during operation. The horizontal axis represents time in seconds, ranging from 0 to 0.1 seconds; the vertical axis represents amplitude in volts, ranging from approximately 0.1V to 0.85V.

[0025] The fundamental frequency component is represented by a solid blue line. Its amplitude remains relatively stable throughout the entire time interval, ranging from approximately 0.79V to 0.83V, exhibiting a stable state with low amplitude fluctuations. This reflects the stability of the 50Hz fundamental frequency characteristic in the electrical performance data of the LED driver power supply.

[0026] The switching frequency component is represented by a red dashed line, with its amplitude varying between 0.45V and 0.55V, exhibiting a slow, fluctuating trend. The curve changes relatively smoothly along the time axis, showing the continuous presence of the 20kHz switching frequency component and its slight amplitude fluctuations, reflecting the stable contribution of the switching frequency to the electrical characteristics.

[0027] The frequency components of the transient disturbance are represented by green dashed lines. The baseline amplitude is approximately 0.13V, but significant amplitude peaks appear at multiple time points, with the highest peak amplitude reaching approximately 0.75V. The peaks are concentrated at approximately 0.008 seconds, 0.062 seconds, 0.086 seconds, and 0.094 seconds, indicating the presence of multiple transient disturbance events. The 100kHz transient disturbance produced significant electrical disturbance responses at these time points.

[0028] Example 2: In practical implementation, a physical parameter library for the LED driver power supply is established. This library is stored in the form of a structured data table, which includes fields for stress-sensitive element identification, parameter name, parameter value, and parameter unit. The stress-sensitive element identification field records the unique identifier for each stress-sensitive element in the LED driver power supply. These elements include electrolytic capacitors, power MOSFETs, transformer windings, rectifier diodes, and output filter inductors. The parameter name field records the name of the physical parameter corresponding to each stress-sensitive element, the parameter value field records the numerical value of the physical parameter, and the parameter unit field records the SI unit of the physical parameter. For electrolytic capacitors, the physical parameters include nominal capacitance, equivalent series resistance, thermal resistance, thermal capacitance, rated operating temperature, and rated voltage. For power MOSFETs, the physical parameters include on-resistance, gate charge, junction-to-case thermal resistance, junction-to-ambient thermal resistance, thermal capacitance, and maximum junction temperature. For transformer windings, the physical parameters include winding DC resistance, winding inductance, core thermal resistance, and inter-winding parasitic capacitance. For rectifier diodes, the physical parameters include forward voltage drop, reverse recovery charge, junction-to-case thermal resistance, and maximum junction temperature. For the output filter inductor, the physical parameters include DC resistance, winding inductance, core thermal resistance, and saturation current. The physical parameter values ​​for each stress-sensitive element in the physical parameter library are obtained from component datasheets, measured data, or simulation models provided by the manufacturer.

[0029] An initial electrical feature sequence is associated with a physical parameter library to generate an electrical transient representation network for the digital twin. The initial electrical feature sequence contains high-dimensional electrical feature vectors arranged chronologically, with each vector corresponding to a time segment. The association process binds each high-dimensional electrical feature vector in the initial sequence to a stress-sensitive element node in the physical parameter library. During binding, a mapping relationship between the feature vector dimension and the physical parameter is established in the digital twin's data model. The electrical transient representation network is represented using a graph structure containing a set of nodes and a set of edges. Each node in the node set corresponds to a stress-sensitive element, and its attributes consist of two parts: the first part is the high-dimensional electrical feature vector corresponding to the time segment, and the second part is a physical parameter vector composed of the physical parameter values ​​of the corresponding stress-sensitive element extracted from the physical parameter library. The order of elements in the physical parameter vector matches the order of the parameter name fields in the physical parameter library. Each edge in the edge set corresponds to a physical connection between two stress-sensitive elements, and these edges are pre-established based on the circuit schematic and PCB layout during the initialization of the electrical transient representation network. The electrical transient representation network is organized in units of time frames, with each time frame corresponding to a graph instance. The node attributes of the graph instance use the high-dimensional electrical feature vector of the corresponding time segment.

[0030] A sliding window decomposition is performed on the electrical transient representation network. In the specific implementation, the window length and sliding step size are set. The window length is determined based on the minimum duration of the transient event, which is 10 microseconds, corresponding to the single cycle width of a 100kHz transient disturbance signal. The window length is set to 10 time frames to ensure that the window covers the complete waveform of the transient event. The sliding step size is set to 5 time frames, so that adjacent windows overlap by half the window length. The electrical transient representation network is arranged into a time series according to the time frame sequence. A fixed-length window is used to truncate the time series from the beginning. Each truncation includes all the graph instances contained within the window, which is called a window graph sequence. After completing one truncation, the window is shifted forward along the time axis by the sliding step size to truncate the next window graph sequence, until the window exceeds the end of the time series.

[0031] Energy fluctuations in different frequency bands are extracted using wavelet packet transform to obtain multi-scale electrical transient disturbance characteristics. For each window sequence obtained from sliding window decomposition, a single-dimensional value of the high-dimensional electrical feature vector in the node attributes of each stress-sensitive element is extracted, forming a one-dimensional discrete signal at each time frame within the window. Wavelet packet transform is then performed on this one-dimensional discrete signal. The wavelet packet transform uses the Daubechies4 wavelet basis function, with a decomposition level of 3, dividing the frequency band of the one-dimensional discrete signal into 8 uniform sub-bands. After the 3rd level of wavelet packet decomposition, 8 sets of wavelet packet decomposition coefficients are obtained, each set corresponding to a sub-band. The frequency range of the sub-bands is uniformly distributed from the DC component to the Nyquist frequency. For the ... Wavelet packet decomposition coefficient sequence within each frequency band Calculate the energy of sub-bands The calculation formula is:

[0032] in, Indicates the first Energy value of each sub-band This is the index of the sub-band number. The value of is any integer from 1 to 8. Indicates the first The total number of wavelet packet decomposition coefficients within each sub-band. Indicates the first Index of wavelet packet decomposition coefficients within each sub-band The value range is 1 to Each integer in Indicates the first Within each sub-band Each wavelet packet decomposition coefficient is used. For each stress-sensitive element within each window, an 8-dimensional energy feature vector is formed by the energy values ​​of 8 sub-bands. The 8-dimensional energy feature vectors of all windows of all stress-sensitive elements are collected to form a multi-scale electrical transient disturbance feature.

[0033] Example 3: In practical implementation, multiple stress-sensitive elements are mapped to nodes in a graph neural network. These stress-sensitive elements include electrolytic capacitors, power MOSFETs, transformer windings, rectifier diodes, and output filter inductors. Each stress-sensitive element corresponds to an independent node in the graph neural network, and the node identifier is consistent with the stress-sensitive element identifier in the physical parameter library. Electrical transient disturbance features are used as the initial attributes of the nodes. These features are multi-scale energy feature vectors obtained from sliding window decomposition and wavelet packet transform. The initial attribute vector of each node consists of the 8-dimensional energy feature vector of the corresponding stress-sensitive element within a time window.

[0034] Based on the thermal conduction paths and electrical connection paths between nodes, edges are constructed between nodes. Thermal conduction paths are established based on the component packaging of stress-sensitive elements and the copper foil layout of the printed circuit board. The thermal resistance network connections between stress-sensitive elements are extracted from the thermal design documents and PCB design documents of the LED driver power supply. When two stress-sensitive elements are directly connected by thermal resistance, an edge is established between the corresponding two nodes in the graph neural network. This edge has an edge type attribute, and its type is thermal conduction edge. Electrical connection paths are determined by the circuit network topology. The electrical interconnection relationships between stress-sensitive elements are extracted from the circuit schematic netlist of the LED driver power supply. When two stress-sensitive elements are directly connected by a wire or connected through a passive component in the circuit network, an edge is established between the corresponding two nodes in the graph neural network, and its type is electrical connection edge. Both thermal conduction edges and electrical connection edges can exist between two nodes simultaneously, and these are merged into a multi-relationship edge. The multi-relationship edge carries feature vectors representing the characteristics of both paths. The path characteristic vector is composed of the thermal resistance value, the inductance value of the connecting line, and parasitic parameters of the two nodes.

[0035] The weights of each edge are dynamically updated using a graph attention mechanism. This mechanism employs a multi-head graph attention network structure, containing multiple parallel attention heads, each independently calculating the attention coefficients between nodes. For each node... and adjacent nodes In the Among the attention heads, the attention coefficient The calculation method is as follows: [The node is missing from the original text] Node feature vectors and nodes Node feature vectors Each through a shared linear transformation matrix Perform a linear mapping to obtain the mapped vector. and ;Map the vector With the mapped vector Perform splicing, then combine with nodes and nodes Path characteristic vector of the edges between The vectors are concatenated to obtain a combined vector; a learnable attention weight vector is then used. The inner product operation is performed on the combined vectors, and the original attention values ​​are obtained by applying the LeakyReLU activation function with leakage correction. For the nodes... After calculating the original attention value for all neighboring nodes, the original attention value is normalized using the softmax function to obtain the attention coefficient. The calculation formula is:

[0036] in, Indicates the first Nodes in each attention head For nodes Attention coefficient; This represents the activation function of the linear unit with leakage correction, and the negative slope coefficient is set to 0.2; Indicates the first The learnable attention weight vector for each attention head. The dimension is set to , Node feature vectors Dimensions The linear transformation matrix represents the dimension of the concatenated vector. The output dimension is also set to Therefore, the dimension of the mapped vector is By concatenating the mapped vectors, we obtain... Then concatenate the path characteristic vector Path characteristic vector The dimension is set to The total dimension of the combined vector is ; Indicates the first The linear transformation matrix of each attention head, the linear transformation matrix The dimension is ; Represents a node The node feature vector, with dimension . ; Represents a node The node feature vector, with dimension . ; Represents a node With nodes The path characteristic vector is composed of the thermal resistance of the heat conduction path and the connection resistance and inductance of the electrical connection path. The dimension of the path characteristic vector is... Dimensions that are insufficient are filled with zeros; symbol This represents a vector concatenation operation; Represents a node The set of adjacent nodes; Represents a node A neighboring node in the set of neighboring nodes; Indicates adjacent nodes Node feature vectors; Represents a node With neighboring nodes Path characteristic vectors between them. Number of attention heads. Set to 4, each attention head calculates its attention coefficient independently. The final attention coefficient is obtained by averaging the calculation results of each attention head. The attention coefficients are ultimately used for edge weight updates. The parameters of the graph attention mechanism include the linear transformation matrix. and attention weight vector It is determined through an end-to-end training process.

[0037] The training process of the graph attention mechanism is as follows: A training sample set is constructed, consisting of electrical transient perturbation features from multiple time windows and the corresponding true values ​​of physically simulated stress distributions. The true values ​​of physically simulated stress distributions are obtained through finite element thermal simulation and circuit simulation tools, labeled with the thermal perturbation field value and electrical stress field value for each node. The electrical transient perturbation features are input into the graph neural network, and forward propagation yields the updated thermal-electric coupling features for each node. The mean absolute error loss is calculated by combining the thermal-electric coupling features with the true values ​​of physically simulated stress distributions. The loss function is the sum of the prediction errors of the thermal perturbation field and the electrical stress field for each node. The training optimizer uses an adaptive moment estimation optimizer, with a learning rate of 0.0005, a batch size of 32 time windows, and 200 training iterations. After training, the parameters of the graph attention mechanism are fixed and used for real-time inference to generate the thermal-electric coupling topology graph.

[0038] After dynamically updating the edge weights, for each node, the feature vectors of neighboring nodes and path characteristic vectors are aggregated, and the node feature vector is updated by weighted summation to form a higher-order representation of the node. The updated node feature vector contains two components: a thermal perturbation field and an electric stress field. The thermal perturbation field component takes the first half of the node feature vector, and the electric stress field component takes the second half, resulting in a thermal-electric coupling topology graph. Each node in the thermal-electric coupling topology graph contains its own features as well as the thermal perturbation field and electric stress field aggregated from its neighboring nodes.

[0039] Based on the adjacency relationships of nodes in the thermo-electric coupling topology, the thermal perturbation fields and electric stress fields of adjacent nodes are aggregated to form the local state vector of each node. For any node in the thermo-electric coupling topology... Local state vector By node Its own thermal perturbation field vector, nodes It is composed of four parts: its own electric stress field vector, the average value of the thermal disturbance field vectors of all adjacent nodes, and the average value of the electric stress field vectors of all adjacent nodes.

[0040] A nonlinear mapping is applied to each local state vector, and a multi-layer fully connected network is used to extract high-order coupling features. The multi-layer fully connected network consists of one input layer, three hidden layers, and one output layer. The number of nodes in the input layer is consistent with the dimension of the local state vector, which is [dimension 1]. ,in Let be the dimension of the thermal disturbance field vector or the electric stress field vector. The number of neurons in the first hidden layer is set to 64. The first hidden layer contains 128 neurons, using the modified linear unit activation function (MRU). The second hidden layer contains 256 neurons, also using the MRU. The third hidden layer contains 128 neurons, again using the MRU. The output layer outputs a high-order coupled feature vector with a dimension of 128. The parameters of the multilayer fully connected network are obtained through training. The training data uses historical data from the thermal-electric coupling topology. The training objective is to minimize the reconstruction error between the high-order coupled feature vector and the actual transient response label of the stress-sensitive element.

[0041] A recurrent neural network (RNN) is used to construct the state transition function for each node at different time steps. The RNN employs a gated recurrent unit (GRU) structure, where each GRU contains update and reset gates. The state transition function for each node at each time step is then constructed. The higher-order coupled eigenvectors are used as inputs to the gated recurrent unit. The gated loop unit reads the input. and the hidden state of the previous time step Output the hidden state at the current time step. Hidden state The mixing ratio of old and new information is determined by the update gate. The gating parameters of the gated recurrent unit are obtained through training. During training, the historical electrical transient disturbance feature sequence is input into the multidimensional state evolution model. Using the hidden state of the previous time step and the input of the current time step as conditions, the hidden state of the next time step is predicted. Mean squared error loss is used for training, and the training optimizer is an adaptive moment estimation optimizer with a learning rate of 0.001 and 150 training iterations. The trained gated recurrent units constitute the state transition functions of each node at different time steps. The state transition functions of all nodes together constitute the multidimensional state evolution model, which is used to simulate the transient response of each stress-sensitive element.

[0042] See Figure 4The figure shows the normalized intensity variations of the thermal perturbation and electrical stress fields of two stress-sensitive components, the electrolytic capacitor and the power MOSFET, over a 50-millisecond time period. The horizontal axis represents time in milliseconds (ms), and the vertical axis represents the normalized intensity, reflecting the transient response amplitude of each stress field. In the curves, the solid blue line represents the thermal perturbation field of the electrolytic capacitor, the dashed red line represents the electrical stress field of the electrolytic capacitor, the dotted green line represents the thermal perturbation field of the power MOSFET, and the dotted orange line represents the electrical stress field of the power MOSFET.

[0043] As can be seen from the curve trends, the electrical stress field intensity of the electrolytic capacitor is generally higher than its thermal perturbation field intensity, and it rapidly rises to a normalized intensity of approximately 0.3 within the first 20 milliseconds before stabilizing, reflecting the rapid excitation and stabilization process of the electrolytic capacitor's electrical stress field. The thermal perturbation field intensity of the electrolytic capacitor starts at a low value, begins to rise significantly around 13 milliseconds, and then stabilizes after 20 milliseconds, with a maximum normalized intensity of approximately 0.2. The thermal perturbation field and electrical stress field intensity of the power MOSFET both show a similar upward trend, but their peak intensities are lower than those of the corresponding curves for the electrolytic capacitor. The maximum normalized intensity of the thermal perturbation field remains around 0.15, and the maximum normalized intensity of the electrical stress field is approximately 0.28. All four curves experience a rapid rise phase within the first 10 milliseconds, followed by a slow, fluctuating stabilization phase, indicating that the system's thermal-electric coupling response gradually reaches dynamic equilibrium after initial excitation.

[0044] Example 4: In practical implementation, electrical transient perturbation features are used as initial conditions input into the multidimensional state evolution model. The multidimensional state evolution model consists of state transition functions for each node at different time steps. These state transition functions employ a gated recurrent unit (GRU) structure, and the parameters of the GRUs have been trained and fixed. The electrical transient perturbation features are multi-scale features, containing an 8-dimensional energy feature vector corresponding to each stress-sensitive element after sliding window decomposition. For each node in the multidimensional state evolution model, the electrical transient perturbation features of the corresponding stress-sensitive element in the initial time window are used as the initial hidden state value of the node in the first time step, and the electrical transient perturbation features in subsequent time windows are used as inputs for each time step. The state values ​​of each node in consecutive time steps are iteratively calculated according to the state transition function. In time step State value The gated loop unit determines the characteristics of the input electrical transient disturbance. and the state value of the previous time step Calculations show that ,in, This represents the forward propagation function of the gated loop unit. From time step... Begin by iterating sequentially to the last time step of the electrical transient perturbation feature sequence. , obtain node Sequence of state values ​​at all time steps Each state value vector in the sequence has a dimension of 128, and the state value sequence serves as a node. The time series of multiple stress-sensitive elements is obtained by performing the same iterative process on all nodes.

[0045] In practice, power spectral density analysis is performed on the time series data of each node, and the power spectrum of each node is calculated using the periodogram method. For each node... The time series was used to select the first dimension representing the thermal perturbation field intensity and the second dimension representing the electric stress field intensity from the 128-dimensional state value vector, which respectively constituted the thermal perturbation intensity scalar sequence. and scalar sequence of electrical stress intensity The power spectral density of the thermal perturbation intensity scalar sequence is estimated using the periodogram method. The thermal perturbation intensity scalar sequence is divided into equal parts of length [missing information]. of Non-overlapping data segments The value of is set to 256, if the total length of the sequence is... If the value is not divisible by 256, the remaining portion is truncated. For each data segment, multiply by the Hanning window function, calculate the windowed Discrete Fourier Transform of each segment, obtain the complex values ​​of each frequency component, calculate the square of the modulus of the complex value, and divide by the data segment length. The product of the window function energy coefficient and the timescale coefficient yields the periodogram estimate for each data segment. The window function energy coefficient is the average of the sum of squares of the elements in the Hanning window sequence. The arithmetic mean of the periodogram estimates for each data segment is used to obtain the power spectral density estimate vector of the thermal perturbation intensity sequence. The length of this vector is equal to... The frequencies are uniformly discretely distributed from 0 to the Nyquist frequency. The power spectral density estimation vector is calculated for the electrical stress intensity scalar sequence in the same manner. Peak points with amplitudes exceeding 10% of the maximum amplitude are extracted from the power spectral density estimation vector. The frequency and amplitude corresponding to each peak point are recorded, with the amplitude unit being the square unit of the power spectral density per Hertz. These frequency and amplitude values ​​are used to construct the thermal perturbation load spectrum component and the electrical stress load spectrum component, and then merged to generate nodes. The dynamic load spectrum is obtained by sequentially performing the above process on all nodes to obtain the dynamic load spectrum of multiple stress-sensitive elements.

[0046] In practical implementation, the stress amplitude and cycle number associated with each stress-sensitive element are extracted from the dynamic load spectrum, and the load history is statistically analyzed using the cycle counting method. For nodes... The dynamic load spectrum, including the thermal disturbance load spectrum and the electrical stress load spectrum, each contains a set of frequency-amplitude pairs. Each frequency-amplitude pair is converted into a time-domain equivalent stress cycle, and the stress cycles are statistically analyzed using the rainflow counting method. The execution process of the rainflow counting method is as follows: the frequency-amplitude pairs of the dynamic load spectrum are synthesized into a time-domain load waveform through inverse Fourier transform. The sampling frequency of the synthesized time-domain load waveform is set to four times the Nyquist frequency, and the synthesis duration is determined according to the common period of the frequency components. The time-domain load waveform data points are arranged sequentially, and the peaks and valleys in the load waveform are identified. The peak-valley sequence is input into the rainflow counting algorithm, which counts the load cycles according to the stress-strain hysteresis loop rule, outputting multiple full cycles and half cycles. Each full cycle contains stress amplitude, average stress, and cycle count 1, and each half cycle contains stress amplitude and average stress. The stress amplitude of the full cycle is half the difference between the maximum and minimum stress values ​​in the cycle, and the average stress is half the sum of the maximum and minimum stress values ​​in the cycle. Using the stress amplitudes of full-cycle and half-cycle stresses as the total stress amplitude, count each full-cycle stress as 1 and each half-cycle stress as 0.5. Summarize all full-cycle and half-cycle stresses to obtain the stress amplitude distribution and corresponding cycle counts, and construct nodes. The corresponding stress-life curve data points of the stress-sensitive element, the data point pairs include stress amplitude. and the corresponding allowed number of loops The stress-life curve data points were obtained by fitting the stress amplitude and cycle number extracted from the dynamic load spectrum with the stress-life relationship obtained from fatigue test data of the stress-sensitive element material. The least squares method was used for fitting, and the fitting function was in the form of: ,in, Indicates the stress amplitude. Indicates stress amplitude The fatigue life cycle count is below. The fatigue strength coefficient, This refers to the fatigue strength index. Fatigue strength coefficient. and fatigue strength index The value is obtained from the material fatigue property database based on the material type of the stress-sensitive element. When the stress-sensitive element is an electrolytic capacitor, the fatigue strength coefficient is... Values Fatigue strength index The value is 3.5; when the stress-sensitive element is a power MOSFET, the fatigue strength coefficient is... Values Fatigue strength index The value is 2.7; when the stress-sensitive element is a transformer winding, the fatigue strength coefficient is... Values Fatigue strength index The value is 3.2; when the stress-sensitive element is a rectifier diode, the fatigue strength coefficient is... Values Fatigue strength index The value is 2.9; when the stress-sensitive element is the output filter inductor, the fatigue strength coefficient is... Values Fatigue strength index The value is set to 3.0. The above parameter values ​​are based on the fitting results of standard fatigue test data for the corresponding materials.

[0047] Low-difference sequences were used to sample the parameters of the stress-life curve, generating multiple degradation path samples. The low-difference sequences were Sobol sequences, which have a generation dimension of 2, corresponding to the fatigue strength coefficients in the stress-life curve. and fatigue strength index Two parameters. The Sobol sequence is generated as follows: based on a preset initial table of highest polynomial degree and direction number, it is generated recursively using bit XOR. The Sobol sequence points are uniformly distributed across intervals. The number of Sobol sequence points is set to 200. For each Sobol sequence point... ,in, and All real numbers within, will Mapped to fatigue strength coefficient The sampled values ​​are mapped using the following formula: ,in, The nominal fatigue strength coefficient is obtained from a materials fatigue property database. The range of values ​​is .Will Mapped to fatigue strength index The sampled values ​​are mapped using the following formula: ,in, The nominal fatigue strength index is obtained from a materials fatigue property database. The range of values ​​is Through 200 sets of sampled values 200 different stress-life curves were generated as 200 degradation path samples.

[0048] Based on the samples from each degradation path, calculate the cumulative damage per unit time. For nodes... For each degradation path sample, the stress amplitude extracted from the dynamic load spectrum is used. and the corresponding number of loops The damage per unit time is calculated using Miner's linear cumulative damage rule. The calculation formula is:

[0049] in, This represents the total cumulative damage per unit time. This represents the total number of different stress amplitudes extracted from the dynamic load spectrum. The index represents the ordinal number of the stress amplitude category. The value range is 1 to Each integer in Indicates the first The number of cycles corresponding to each stress amplitude category Indicates the first Stress amplitude category The number of fatigue life cycles on the stress-life curve corresponding to the degradation path sample. When the total accumulated damage A value of 1 indicates that the stress-sensitive element has reached fatigue failure. The total cumulative damage per unit time is calculated for each degradation path sample. This represents the degradation rate under this degradation path sample. For nodes... Degradation rates are calculated for all 200 degradation path samples, resulting in 200 degradation rate values. The set of these degradation rate values ​​constitutes a node. The degradation state data is obtained by performing the above calculation process on all nodes. The degradation state data includes multiple degradation rate sample values ​​for each stress-sensitive element.

[0050] See Figure 5 The vertical axis of the figure represents the stress amplitude S, in megapascals (MPa), expressed on a logarithmic scale, ranging from approximately 10^1 to 10^3. The horizontal axis represents the number of cycles N, also expressed on a logarithmic scale, ranging from approximately 10^1 to 10^9. The figure shows the stress-life curves for five stress-sensitive components: electrolytic capacitors, power MOSFETs, transformer windings, rectifier diodes, and output filter inductors. All curves exhibit a typical negative power function relationship, reflecting the fatigue life characteristics between stress amplitude and the allowable number of cycles.

[0051] The curves descend from the upper left to the lower right, indicating that the stress amplitude decreases with increasing cycle count, consistent with the theoretical law of fatigue life curves. The electrolytic capacitor curve starts at a lower stress amplitude, and its overall position is in the lower left region of the graph, indicating that it has a lower cycle life for higher stress amplitudes. The power MOSFET curve is located in the middle, showing that its fatigue strength and exponent are between those of other components. The transformer winding curve is close to that of the electrolytic capacitor, but slightly higher overall, indicating that its fatigue performance is slightly better. The rectifier diode curve is the lowest, indicating that this component has a higher cycle life under the same stress amplitude, indicating that its material fatigue performance is better. The output filter inductor curve is located at the highest position, reflecting its strongest fatigue performance, capable of withstanding a larger stress amplitude and cycle count.

[0052] Example 5: In practical implementation, the degradation state data is encoded along the time dimension to generate a degradation state embedding vector. The degradation state data contains multiple degradation rate sample values ​​for each stress-sensitive element. For any stress-sensitive element, all degradation rate sample values ​​in the corresponding degradation state data are arranged according to the generation order of the degradation path samples, forming a degradation rate sequence. The time dimension of the degradation rate sequence is represented by the index number of the degradation path sample, and the degradation rate value of each degradation path sample is considered as a state value at a time step. The time dimension encoding consists of two parts: position encoding and linear transformation. The position encoding adopts a sinusoidal position encoding method. For the th... Degradation rate value at each location, location encoding vector The dimension is set to Location encoding vector The Middle peacekeeping The element values ​​of each dimension are calculated using the following formulas:

[0053] in, Indicates the position index of the degradation rate value in the degradation rate sequence. The value ranges from 1 to the total number of degenerate path samples. Each integer in The value is 200; Indicates the index of the dimension pair in the positional encoding vector. The value of is any integer from 0 to 31; This represents the total dimension of the position encoding vector. The value is 64; Indicates the position index is The position encoding vector of the first The element value of a dimension; Indicates the position index is The position encoding vector of the first The element value of the dimension. Index of each position. Corresponding positional encoding vector and degradation rate value The fusion is performed by combining the degradation rate values. After a linear transformation matrix, the vector is mapped to a 64-dimensional feature vector. Then, the 64-dimensional feature vector is combined with the position encoding vector. By adding elements one by one, we obtain the degenerate state embedding vector. The above encoding process is performed sequentially on all 200 degenerate path sample locations to obtain 200 degenerate state embedding vectors, forming a sequence of degenerate state embedding vectors.

[0054] In practical implementation, a self-attention mechanism is used to calculate the correlation weights between the embedding vectors of each degenerate state. The self-attention mechanism employs a scaled dot product attention structure, comprising a query vector generation module, a key vector generation module, and a value vector generation module. The query vector generation module consists of a linear transformation matrix. Implement, linear transformation matrix The dimension is Embed the degenerate state into a vector Through linear transformation matrix Perform a linear mapping to obtain the query vector. The key vector generation module consists of a linear transformation matrix. Implement, linear transformation matrix The dimension is Embed the degenerate state into a vector Through linear transformation matrix Perform a linear mapping to obtain the key vector. The value vector generation module consists of a linear transformation matrix. Implement, linear transformation matrix The dimension is Embed the degenerate state into a vector Through linear transformation matrix Perform a linear mapping to obtain a value vector. Linear transformation matrix Linear transformation matrix and linear transformation matrix The parameters are determined through training.

[0055] The training process of the self-attention mechanism is as follows: A degradation state history dataset is constructed, consisting of degradation rate sequences of multiple LED driver power supply samples over their complete lifespan. Each degradation rate sequence is labeled with a known degradation inflection point. The inflection point is manually determined based on the shape of the degradation curve. The labeling rule is that when the rate of change of three consecutive sampling points relative to the previous sampling point exceeds 20%, the first sampling point among the three consecutive sampling points is marked as the degradation inflection point. The degradation rate sequences in the degradation state history dataset are encoded using the above method to generate a degradation state embedding vector sequence, which is then used as the training input. During forward propagation, for any position in the sequence... and location Calculate the query vector With key vector The scaled dot product, where the scaling factor is the dimension of the position encoding vector. The square root of the value is 8, which gives the original attention score. This original attention score is then normalized using the softmax function to obtain the position. Position The relevance weights are applied to the value vector. We obtain the position by weighted summation. The context vector is then input into a binary classifier consisting of a fully connected layer and a sigmoid activation function. The fully connected layer has an input dimension of 64 and an output dimension of 1. The sigmoid activation function outputs a real number between 0 and 1, representing the location. This represents the probability value of the degradation inflection point. The loss function used is binary cross-entropy loss, the training optimizer is an adaptive moment estimator, the learning rate is set to 0.0001, the batch size is set to 8 degradation rate sequences, and the number of training iterations is set to 300. After training, the linear transformation matrix... Linear transformation matrix Linear transformation matrix The parameters of the binary classifier are fixed and stored.

[0056] In practical applications, the degenerate state to be analyzed is embedded into a vector sequence and input into a trained self-attention mechanism to calculate the value at any position. With position The correlation weights between these elements form a correlation weight matrix. The time points where abrupt changes in correlation weights are identified are used as inflection points of degradation. The method for determining abrupt changes in correlation weights is as follows: for position... Extracting the position in the correlation weight matrix The corresponding row vector, which represents the position. Relevance weight distribution for all positions in the sequence; calculate position row vectors and positions The cosine distance between row vectors is defined as 1 minus the cosine similarity of the two row vectors. When the position... When the corresponding cosine distance exceeds the preset abrupt change threshold, the position is determined. These are the weighted mutation points. The preset mutation threshold is set to 0.35. This value is based on the cosine distance distribution of all known degradation inflection point positions statistically analyzed from the historical degradation state dataset, with the lower quartile of the distribution (0.35) used as the threshold. The time positions corresponding to all weighted mutation points are considered degradation inflection points. There is a definite correspondence between the time values ​​corresponding to the degradation inflection points and the actual physical time of the degradation state data. The physical time is obtained through index mapping of degradation path samples.

[0057] In practice, the degradation inflection point of each stress-sensitive element is compared with the corresponding material degradation threshold. The material degradation threshold is stored in a material degradation characteristic database. This database assigns a material degradation threshold to each failure mode of each stress-sensitive element, and the unit of the material degradation threshold is a dimensionless value consistent with the degradation state data. For electrolytic capacitors, the material degradation threshold is set to the degradation state value corresponding to a 20% reduction in their initial capacitance. When the degradation state data reaches the material degradation threshold, it indicates that the electrolytic capacitor has degraded and failed. For power MOSFETs, the material degradation threshold is set to the degradation state value corresponding to a 50% increase in their on-resistance. For transformer windings, the material degradation threshold is set to the degradation state value corresponding to a 15% increase in their winding DC resistance. For rectifier diodes, the material degradation threshold is set to the degradation state value corresponding to a 200% increase in their reverse leakage current. For output filter inductors, the material degradation threshold is set to the degradation state value corresponding to a 15% decrease in their inductance. The degradation status data of the stress-sensitive element are arranged chronologically. Starting from the degradation inflection point, the degradation status data value is compared with the material degradation threshold point by point. When the degradation status data value first exceeds or equals the material degradation threshold, the corresponding time point is recorded. The difference between the corresponding time point and the initial service time of the stress-sensitive element is the remaining life of the stress-sensitive element, recorded in hours. The above comparison process is performed on all stress-sensitive elements to obtain the remaining life of each stress-sensitive element.

[0058] In practice, the remaining lifetime of all stress-sensitive components is sorted in ascending order of numerical values. The minimum value in this ascending order is then taken as the simulated lifetime of the LED driver. The simulated lifetime of the LED driver represents the end time of the driver's life determined by the component that first degrades and fails under current operating conditions. The stress-sensitive component identifier corresponding to the minimum value, along with the simulated lifetime value, is output as the lifetime simulation result.

[0059] In a specific implementation, a system for simulating the lifespan of an LED driver power supply is also provided, comprising a memory, a processor, and a computer program stored in the memory and running on the processor. When the processor executes the computer program, it performs the following steps: real-time acquisition of electrical performance data during the operation of the LED driver power supply; extraction of an initial electrical feature sequence from the electrical performance data; joint transformation of the electrical performance data in the time and frequency domains to obtain a time-frequency spectrum; extraction of high-dimensional electrical features from the time-frequency spectrum using a convolutional autoencoder to obtain the initial electrical feature sequence; establishment of a physical parameter library and association with the initial electrical feature sequence to generate an electrical transient characterization network; sliding window decomposition of the electrical transient characterization network and extraction of multi-scale electrical transient disturbance features using wavelet packet transform; mapping stress-sensitive elements to graph neural network nodes; dynamically updating edge weights through a graph attention mechanism to obtain a thermal-electric coupling topology graph; and aggregating the thermal disturbance field and electrical stress of adjacent nodes. A force field forms a local state vector, and a multidimensional state evolution model is constructed using a recurrent neural network. Electrical transient disturbance characteristics are input into the multidimensional state evolution model, and the state values ​​of each node are iteratively calculated to obtain a time series. Power spectral density analysis is then performed to generate a dynamic load spectrum. Stress amplitude and cycle count are extracted from the dynamic load spectrum to construct a stress-life curve. Low-difference sequence sampling is used to generate degradation path samples, and the degradation rate is calculated to obtain degradation state data. The degradation state data is encoded in the time dimension to generate a degradation state embedding vector. A self-attention mechanism is used to calculate correlation weights and identify weight mutation points as degradation inflection points. The degradation inflection points are compared with the material degradation threshold to obtain the remaining lifetime of each stress-sensitive element. The remaining lifetimes are sorted, and the minimum value is taken to obtain the simulated lifetime.

[0060] See Figure 6 The horizontal axis of the graph represents the operating time of the LED driver power supply in hours, ranging from 0 to 50,000 hours; the vertical axis represents the degradation state value of the stress-sensitive element in percentage, ranging from 0% to 20%. The graph shows sample curves for five degradation paths and the material degradation threshold line. The five path sample curves are marked with different colors and line types: the blue solid line is degradation path sample 1, the red dashed line is degradation path sample 2, the green dotted line is degradation path sample 3, the orange dotted line is degradation path sample 4, the purple solid line is degradation path sample 5, and the black dashed line represents the material degradation threshold, which is constant at 20%.

[0061] As shown in the figure, the degradation state of each degradation path sample exhibits varying degrees of fluctuation and an upward trend with increasing running time. The degradation state values ​​of degradation path samples 1 and 4 remain consistently close to zero, showing no obvious degradation trend. These are marked as points around 35,000 hours and 37,000 hours respectively, indicating that no substantial increase occurred at the degradation inflection point at these times, and the degradation state is close to stable. The degradation state value of degradation path sample 2 gradually increases from the initial value, reaches a peak, and then shows a significant downward trend after approximately 28,000 hours, eventually dropping to near zero. This is marked as a point around 36,000 hours, indicating that this path sample experienced a certain degree of degradation followed by recovery or measurement fluctuations. The degradation state values ​​of degradation path samples 3 and 5 showed a continuous increasing trend. In particular, degradation path sample 3 rapidly approached 10% after about 45,000 hours. The degradation state value of degradation path sample 5 was generally high and fluctuated significantly in the range of 30,000 to 40,000 hours. Corresponding markers appeared in both cases, reflecting the identification of degradation inflection points.

[0062] The material degradation threshold is fixed at 20%, significantly higher than the degradation state values ​​of each path sample, indicating that the degradation path samples in the figure have not yet reached the material's set failure threshold. According to the technical solution of Embodiment 5, degradation state data is analyzed and identified using time-dimensional encoding and a self-attention mechanism to pinpoint degradation inflection points. The marked points of each path sample in the figure represent the locations of these degradation inflection points confirmed by this mechanism. The identification of degradation inflection points provides a time-based basis for remaining lifetime assessment. Combined with a comparison of the corresponding material degradation thresholds, the remaining lifetime of each stress-sensitive component can be accurately determined.

[0063] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for simulating the lifespan of an LED driver power supply, characterized in that, include: The electrical performance data of the LED driver power supply during operation is collected in real time, and an initial electrical characteristic sequence is extracted from the electrical performance data; wherein, the LED driver power supply includes multiple stress-sensitive elements; A digital twin of the LED driver power supply is constructed, and multi-scale electrical transient disturbance features are extracted from the initial electrical feature sequence. Based on the electrical transient disturbance features, a thermal-electric coupling topology graph corresponding to the multiple stress-sensitive elements is generated through a graph neural network. The thermal-electric coupling topology graph contains the thermal disturbance field and electrical stress field of each stress-sensitive element. Based on the adjacency relationship of each node in the thermal-electric coupling topology graph, a multi-dimensional state evolution model is constructed. The multi-dimensional state evolution model is used to simulate the transient response of each stress-sensitive element. The electrical transient disturbance characteristics are input into the multidimensional state evolution model, and the time evolution process of the multidimensional state evolution model is simulated to obtain the dynamic load spectrum of the multiple stress-sensitive elements. Based on the dynamic load spectrum, the degradation rate of the plurality of stress-sensitive elements is calculated to obtain the degradation state data of the LED driver power supply; Based on the degradation state data, the degradation inflection points of the multiple stress-sensitive elements are extracted through a self-attention mechanism, and combined with the material degradation threshold of each stress-sensitive element, the remaining lifetime of each stress-sensitive element is obtained. Then, based on the remaining lifetime, the simulated lifetime of the LED driver power supply is obtained through the minimum order statistic.

2. The method for simulating the lifespan of an LED driver power supply according to claim 1, characterized in that, The step of extracting an initial electrical feature sequence from the electrical performance data includes: The electrical performance data is subjected to a joint time-domain and frequency-domain transformation to obtain a time-spectrum diagram; A high-dimensional electrical feature sequence is obtained by extracting high-dimensional electrical features from the time-spectrum map using a convolutional autoencoder.

3. The method for simulating the lifespan of an LED driver power supply according to claim 1, characterized in that, The construction of the digital twin of the LED driver power supply extracts multi-scale electrical transient disturbance features from the initial electrical feature sequence, including: A physical parameter library for the LED driver power supply is established, and the initial electrical characteristic sequence is associated with the physical parameter library to generate the electrical transient characterization network of the digital twin. The electric transient characterization network is decomposed by a sliding window, and energy fluctuations in different frequency bands are extracted by wavelet packet transform to obtain the multi-scale electric transient perturbation features.

4. The method for simulating the lifespan of an LED driver power supply according to claim 1, characterized in that, The step of generating a thermo-electric coupling topology map corresponding to the plurality of stress-sensitive elements using a graph neural network based on the electrical transient disturbance characteristics includes: The plurality of stress-sensitive elements are mapped as nodes of a graph neural network, and the electrical transient disturbance features are used as the initial attributes of the nodes. Based on the heat conduction path and electrical connection path between each node, edges between nodes are constructed, and the weights of each edge are dynamically updated through a graph attention mechanism to obtain the heat-electric coupling topology graph.

5. The method for simulating the lifespan of an LED driver power supply according to claim 4, characterized in that, The construction of a multidimensional state evolution model based on the adjacency relationships of nodes in the thermo-electric coupling topology includes: The thermal perturbation field and electric stress field of the adjacent nodes of each node are aggregated to form the local state vector of each node; The local state vectors are nonlinearly mapped, and the state transition functions of each node at different time steps are constructed through a recurrent neural network to obtain the multidimensional state evolution model.

6. The method for simulating the lifespan of an LED driver power supply according to claim 5, characterized in that, The step of inputting the electrical transient disturbance characteristics into the multidimensional state evolution model, simulating the time evolution process of the multidimensional state evolution model, and obtaining the dynamic load spectrum of the multiple stress-sensitive elements includes: The electrical transient disturbance characteristics are used as initial conditions and input into the multidimensional state evolution model. The state values ​​of each node at continuous time steps are iteratively calculated according to the state transition function to obtain the time series of each node. Power spectral density analysis is performed on each of the time series to extract the amplitude and frequency components of each node, thereby generating the dynamic load spectrum.

7. The method for simulating the lifespan of an LED driver power supply according to claim 1, characterized in that, The process of calculating the degradation rate of the plurality of stress-sensitive elements based on the dynamic load spectrum to obtain degradation state data of the LED driver power supply includes: The stress amplitude and cycle number associated with each stress-sensitive element are extracted from the dynamic load spectrum to construct a stress-life curve; The parameters of the stress-life curve are sampled using a low-difference sequence to generate multiple degradation path samples. The degradation rate of each stress-sensitive element is calculated based on each degradation path sample to obtain the degradation state data.

8. The method for simulating the lifespan of an LED driver power supply according to claim 1, characterized in that, The step of extracting the degradation inflection points of the plurality of stress-sensitive elements through a self-attention mechanism based on the degradation state data includes: The degradation state data is encoded along the time dimension to generate a degradation state embedding vector; The correlation weights between the embedding vectors are calculated using a self-attention mechanism, and the time points at which the correlation weights change abruptly are identified as the degradation inflection points.

9. The method for simulating the lifespan of an LED driver power supply according to claim 8, characterized in that, The remaining lifetime of each stress-sensitive element is obtained by combining the material degradation threshold of each stress-sensitive element. Then, based on each remaining lifetime, the simulated lifetime of the LED driver power supply is obtained through a minimum order statistic, including: The degradation inflection point of each stress-sensitive element is compared with the corresponding material degradation threshold. When the degradation state data reaches the material degradation threshold, the corresponding time point is recorded to obtain the remaining lifetime of each stress-sensitive element. The remaining lifetimes of each item are sorted, and the minimum value is taken as the simulated lifetime of the LED driver power supply.

10. A system for simulating the lifespan of an LED driver power supply, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the method for simulating the lifespan of an LED driver power supply as described in any one of claims 1 to 9.

Images