Online prediction method for residual life of electronic device under operation condition

Abstract

The invention discloses an electronic device residual life online prediction method and system under an operation condition, a storage medium and electronic equipment, and the method comprises the steps: building a Gamma state space model, and carrying out the processing through an expectation maximization algorithm, and obtaining an initial parameter estimation value of the Gamma state space model; and based on the initial parameter estimation value and pre-acquired real-time monitoring data of the field operation device, dynamically updating the real health state of the electronic device through a kernel smoothing particle filter algorithm, and estimating new parameter estimation of the Gamma state space model by using an expectation maximization algorithm to obtain a final Gamma state space model. According to the method, the degradation information is extracted and the initial degradation model is established based on the accelerated failure experiment data, and then the degradation information and the model parameters are dynamically updated by adopting the kernel smoothing particle filter algorithm in combination with the field operation data, so that the real-time life prediction under the field operation condition is realized.

Description

Technical Field

[0001] This application relates to the field of device reliability assessment, and particularly to an online prediction method, system, storage medium, and electronic device for the remaining life of an electronic device under operating conditions, with the invention name Background Art

[0002] Data-driven methods for predicting the remaining life of electronic devices rely on accelerated degradation data. As an effective experimental method, accelerated degradation testing is widely used to obtain degradation datasets of electronic devices with high reliability and long life. However, since accelerated degradation experiments place electronic devices under extremely harsh conditions, which are far from the actual operating conditions of electronic devices, there are often significant differences between the inferred results in the laboratory and the results of on-site failure analysis. Therefore, it is of great significance to study the method for predicting the remaining life of electronic devices under on-site operating conditions and conduct on-site reliability analysis.

[0003] Through a search of the current literature, it is found that most existing techniques for predicting the remaining life of electronic devices do not combine degradation test data with real-time monitoring data. One of the existing methods for predicting the remaining life mainly uses the state monitoring data of a test device before the prediction time point to identify model parameters, and then predicts the remaining life based on the estimated model parameters. Such methods do not make full use of laboratory test data. For example, Chen et al. developed a remaining life prediction method that combines a fractional grey model and an unscented particle filter. This method uses an adaptive mutation particle swarm optimizer to update model parameters based on data before different prediction starting points. Ahwiadi et al. proposed an enhanced particle filtering technique to improve the accuracy of battery state of health monitoring and remaining life prediction. During training, this method uses the limited state monitoring data from a single battery to identify the particle filter model, and during prediction, the identified model is used for remaining life prediction. The drawback of such methods is that they only utilize the limited state monitoring data of the test device while ignoring the large amount of information contained in the accelerated degradation test dataset. Another existing method for predicting the remaining life completely relies on historical test degradation datasets, including methods based on traditional support vector regression and methods based on neural networks. These methods learn network or model parameters from offline test datasets and then directly apply them to on-site operating datasets. However, since the training dataset is usually obtained through accelerated degradation testing and the operating environment of the test dataset is relatively mild, it is difficult to ensure that the operating conditions of the training dataset and the test dataset are the same, which limits the application of the above methods.

[0004] The common point of the above two types of methods is that they do not combine degradation test data with real-time monitoring data and only utilize one data source, which will cause the prediction method to be inapplicable to on-site operating devices or result in inaccurate remaining life prediction when applied to on-site operating devices. Summary of the Invention

[0005] The main object of the present application is to provide an online prediction method, system, storage medium and electronic device for the remaining life of electronic devices under operating conditions, aiming to propose a method that can not only make full use of historical data sets, but also use the status monitoring data of on-site operating devices to adjust the model to make it applicable to the device.

[0006] To achieve the above object, the first aspect of the present application provides an online prediction method for the remaining life of electronic devices under on-site operating conditions, including: in the offline stage, using the Gamma stochastic process to model the degradation process of the electronic device, and embedding the physical prior model as the shape parameter into the Gamma stochastic process to obtain the Gamma state space model, and using the expectation maximization algorithm to process the pre-obtained accelerated degradation experiment data set to obtain the initial model after parameter estimation; in the online stage, inputting the real-time monitoring data of the on-site operating device into the state and parameter joint estimation algorithm of the kernel smoothing particle filter to obtain the accelerated degradation information of the dynamically updated electronic device and the parameters of the initial model of the electronic device; according to the final updated initial model, the randomized failure threshold distribution and the accelerated degradation information, calculate the remaining life of each electronic device.

[0007] Optionally, in the offline stage, using the Gamma stochastic process to model the degradation process of the electronic device, and embedding the physical prior model as the shape parameter into the Gamma stochastic process to obtain the Gamma state space model, includes: obtaining the probability density function of the Gamma stochastic process, where the probability density function includes a shape parameter, a scale parameter and a measurement noise parameter, and the shape parameter has a degradation state increment; representing the degradation state increment as a random variable related to the device individual to obtain the probability density function of the device individual, and determining the true degradation state of the electronic device based on the probability density function of the device individual; based on the true degradation state of each electronic device after adding Gaussian noise, correspondingly determining the Gamma state space model of each electronic device.

[0008] Optionally, using the expectation maximization algorithm to process the pre-obtained accelerated degradation experiment data set to obtain the initial model after parameter estimation, includes: determining the unknown parameters to be estimated based on the Gamma state space model of each electronic device, where the unknown parameters include state transition model parameters and noise parameters; performing a particle filter and smoothing algorithm on the pre-obtained historical degradation measurement data set of each electronic device to obtain a smoothed particle set; determining the log-likelihood function of the unknown parameters based on the smoothed particle set, and determining the expectation of the log-likelihood based on the log-likelihood function; estimating the unknown parameters based on the expectation of the log-likelihood to obtain the initial model after parameter estimation.

[0009] Optionally, input the real-time monitoring data of the on-site operating device in the state and parameter joint estimation algorithm of kernel smoothed particle filter to obtain the accelerated degradation information of the electronic device and the parameters of the initial model of the electronic device after dynamic update, including: taking the parameter particle vector at any moment as the parameter particle; using the particle filter algorithm based on kernel smoothing to shrink the parameter particle towards the mean to obtain the first particle parameter; adding random noise to the first particle parameter to obtain the second particle parameter; obtaining the updated particle parameter according to the first particle parameter and the second particle parameter; generating the real health state data of each device based on the real-time monitoring data of the on-site operating device and the updated particle parameter; using the expectation maximization algorithm to process the real health state data to obtain the updated parameters of the initial model, and using the updated initial model to process the real health state data to obtain the accelerated degradation information of the electronic device.

[0010] Optionally, obtaining the updated particle parameter according to the mean particle parameter and the diversity particle parameter includes: determining the expression of the parameter state transition equation according to the function expression of the first particle parameter and the function expression of the second particle parameter; calculating the updated particle parameter according to the expression of the parameter state transition equation.

[0011] Optionally, calculating the remaining life of each electronic device according to the finally updated initial model, the randomized failure threshold distribution and the accelerated degradation information includes: determining the marginal survival function of each electronic device according to the updated initial model; and obtaining the marginal survival function based on the survival function, approximating the marginal survival function by using the Gauss-Hermite quadrature method to obtain the life probability density function of the electronic device in the current degradation state; calculating the remaining life of each electronic device based on the life probability density function of the electronic device, the current accelerated degradation information and the randomized failure threshold.

[0012] Optionally, determining the marginal survival function of each electronic device according to the updated initial model includes: determining the probability density function of the shape parameter degradation increment under the condition of the updated particle parameter based on the updated initial model, where the probability density function of the shape parameter degradation increment is a function of the failure threshold; determining the survival function of each electronic device driven by the Gamma process based on the probability density function of the shape parameter degradation increment; integrating the survival function with respect to the failure threshold to obtain the marginal survival function of each electronic device.

[0013] The second aspect of the present application provides an online prediction system for the remaining life of electronic devices under on-site operating conditions, which is characterized by including the following steps: an offline parameter estimation module, which is used in the offline stage to model the degradation process of electronic devices by using a Gamma stochastic process, and embed a physical prior model as a shape parameter into the Gamma stochastic process to obtain a Gamma state space model, and use the expectation maximization algorithm to process the pre-acquired accelerated degradation experiment data set to obtain an initial model after parameter estimation; an online parameter update module, which is used in the online stage to input the real-time monitoring data of on-site operating devices into the state and parameter joint estimation algorithm of kernel smoothing particle filtering to obtain the accelerated degradation information of the dynamically updated electronic devices and the parameters of the initial model of the electronic devices; a life estimation module, which is used to calculate the remaining life of each of the electronic devices according to the final updated initial model, the randomized failure threshold distribution, and the accelerated degradation information.

[0014] The third aspect of the present application provides a computer-readable storage medium, which includes instructions that, when running on a computer, cause the computer to execute the online prediction method for the remaining life of electronic devices under on-site operating conditions provided in the first aspect.

[0015] The fourth aspect of the present application provides an electronic device, which includes: at least one processor, a memory, and an input / output unit; wherein, the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the online prediction method for the remaining life of electronic devices under on-site operating conditions according to any one of claims 1 to 7.

[0016] An online prediction method, system, storage medium and electronic device for the remaining life of an electronic device under operating conditions proposed in an embodiment of the present application model the degradation process of the electronic device by using a Gamma stochastic process to obtain a Gamma state space model, wherein the Gamma state space model has a shape parameter, the shape parameter is a physical prior model, and the expectation maximization algorithm is used to process a pre-acquired accelerated degradation experiment data set to obtain an initial parameter estimate value of the Gamma state space model; based on the initial parameter estimate value of the Gamma state space model and real-time monitoring data of on-site operating devices pre-acquired, the true health state of the electronic device is dynamically updated by a kernel smoothing particle filter algorithm, and based on the true health state, the expectation maximization algorithm is used to determine a new parameter estimate of the Gamma state space model to obtain a final Gamma state space model; according to the final Gamma state space model, a randomized failure threshold distribution and the true health state, the remaining life of each electronic device is calculated. The present application accurately estimates the initial model parameters by using the EM algorithm based on accelerated failure experiment data, extracts degradation information and establishes an initial degradation model, and then combines on-site operating data and uses a kernel smoothing particle filter algorithm to dynamically update the degradation information and model parameters, so as to realize real-time life prediction under on-site operating conditions. This method fully combines the advantages of accelerated failure experiment data and on-site operating data, not only improves the accuracy and reliability of life prediction, but also enhances the applicability and engineering practical value of the method. Description of the Drawings

[0017] Figure 1 is the flowchart of the method provided by an embodiment of the present application;

[0018] Figure 2 is the implementation framework diagram provided by an embodiment of the present application;

[0019] Figure 3 is the degradation trajectory diagram of the IGBT device provided by an embodiment of the present application;

[0020] Figure 4 is the parameter convergence process in the offline parameter estimation stage provided by an embodiment of the present application;

[0021] Figure 5 is the prediction trajectory diagram of sample 1 at different time points provided by an embodiment of the present application;

[0022] Figure 6 is the prediction trajectory diagram of sample 3 at different time points provided by an embodiment of the present application;

[0023] Figure 7 is the RUL prediction result of eight devices at 80% of their life cycle provided by an embodiment of the present application;

[0024] Figure 8 The prediction errors and 95% confidence intervals of different devices from the start of operation to the failure time point provided by an embodiment of the present application;

[0025] Figure 9 It is a system framework diagram provided by an embodiment of the present application.

[0026] The realization, functional features and advantages of the objectives of the present application will be further described in conjunction with the embodiments with reference to the accompanying drawings. Specific Embodiments

[0027] It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0028] Referring to Figure 1 and Figure 2 An online prediction method for the remaining life of electronic devices under on-site operating conditions provided by the first embodiment of the present application. This method can be executed by a processor of a terminal or a server. The online prediction method for the remaining life of electronic devices under on-site operating conditions includes:

[0029] S10. In the offline stage, use the Gamma random process to model the degradation process of the electronic device, and embed the physical prior model as the shape parameter into the Gamma random process to obtain the Gamma state space model. Use the expectation maximization algorithm to process the pre-acquired accelerated degradation experiment data set to obtain the initial model after parameter estimation;

[0030] S20. In the online stage, input the real-time monitoring data of the on-site operating device into the state and parameter joint estimation algorithm of kernel smoothing particle filter to obtain the accelerated degradation information of the dynamically updated electronic device and the parameters of the initial model of the electronic device;

[0031] S30. Calculate the remaining life of each electronic device according to the final updated initial model, the randomized failure threshold distribution, and the accelerated degradation information.

[0032] Among them, the present application includes an offline process and an online process. In the offline process, the parameters of the degradation model are estimated using the accelerated failure experiment data set, and the obtained model describes the degradation characteristics of the device population in the data set; in the online process, the model parameters are continuously updated according to the status monitoring data of the on-site operating device, and the model is customized according to the test device to accurately describe the degradation process of the test device.

[0033] In this embodiment, to address the adaptability issue of the model from the accelerated failure test environment to the on-site operating environment, the present application models the degradation process of the device through the Gamma process and incorporates the physical prior model as the shape parameter into the Gamma process model. Based on the accelerated failure test data, the Expectation Maximization (EM) algorithm is used to estimate the initial model parameters, extract the degradation information from the accelerated failure dataset, and establish an initial degradation model. Based on the on-site operating data, the kernel smoothed particle filter algorithm is used to extract the degradation information of the on-site operating devices and update the model parameters. The present application makes full use of the accelerated failure test data and on-site operating data to achieve online prediction of the remaining life of electronic devices under on-site operating conditions, effectively solving the problem of insufficient adaptability of the model from the accelerated failure test environment to the on-site operating environment. The present application enhances the physical interpretability and applicability of the model in theory. In terms of method implementation, real-time life prediction under on-site operating conditions is achieved. This method fully combines the advantages of the accelerated failure test data and on-site operating data, not only improving the accuracy and reliability of life prediction, but also enhancing the applicability and engineering practical value of the method.

[0034] In the embodiment of the present application, step S10 includes the following specific implementation processes:

[0035] S101, obtain the probability density function of the Gamma random process, where the probability density function includes a shape parameter, a scale parameter, and a measurement noise parameter, and the shape parameter has a degradation state increment;

[0036] S102, represent the degradation state increment as a random variable related to the device individual, obtain the probability density function of the device individual, and determine the true degradation state of the electronic device based on the probability density function of the device individual;

[0037] S103, based on the true degradation state of each electronic device after adding Gaussian noise, correspondingly determine the Gamma state space model of each electronic device.

[0038] Among them, the processor realizes the modeling of the offline model by executing step S101 - step S103.

[0039] This application first constructs an offline model for the degradation process. Specifically, during the degradation process of an electronic device, its health state deteriorates continuously over time. Therefore, the change in the device's health state is considered monotonic, and this monotonicity reflects the irreversible performance loss or functional degradation of the electronic device during long-term use. To fully consider this inherent monotonicity, a Gamma random process is used to model the device's degradation process. According to the definition of the Gamma random process, the probability density function (PDF) of the health state x(t) can be expressed as:

[0040]

[0041] where x(t) represents the health state, and x(t), v(t), and u are the shape and scale parameters of the Gamma distribution. is the Gamma function when a > 0. Theoretically, v(t) determines the degradation trajectory and non-linearity, and u determines the scale size of the distribution, that is, stretching or compressing the interval of the probability density function. Let v(t) be a non-decreasing, right-continuous, real-valued function for t > 0. The Gamma process with v(t) > 0 and u > 0 has the following properties: 1) x(t + Δt) - x(t) ~ Ga(v(t + Δt) - v(t), u), where Δt > 0. According to the characteristics of the Gamma distribution, the mean of x(t + Dt) - x(t) is [v(t + Dt) - v(t)]u, and the variance is [v(t + Dt) - v(t)]u 2 , and the distribution of x(t + Dt) - x(t) describes the transfer function of the degradation state over time. 2) x(t) has independent increments, that is, for any state monitoring times 0 < t1 < t2 <... < ∞, the corresponding degradation state increments Dx(0, t1), Dx(t1, t2),... are all independent random variables.

[0042] Due to factors such as the material properties and manufacturing tolerances of the device, even if the device types are the same, each device has its unique degradation mode, and this difference between devices is called inter-device heterogeneity. Theoretically, v(t) determines the non-linearity of the degradation trajectory. v(t) can be replaced by an empirical model or physical model of device degradation. By establishing the parameters in v(t) as device-specific random variables to characterize the variance between devices, the parameters in v(t) are written as θ, and it is assumed that θ ~ N(μ θ , Σ θ ). Any single device can be regarded as a sample of the overall distribution. Assume that there are m devices in the historical device population. Without loss of generality, based on the independent increment property, the PDF of the degradation state increment Δx(t) of a specific device numbered i can be developed as:

[0043]

[0044] Among them, Dx i (t) = x i (t + Dt) - x i (t) and Dv i (t) = v i (t + Dt) - v i (t) are the degradation state increment and shape parameter of device i at time t, respectively. v i (t) = v(t; θ i ) and θ i is the specific individual parameter of device i.

[0045] Measurement noise is considered by introducing a noise term into the measurement function of the state space model. For the i-th device, the measurement equation formula is:

[0046] y i (t) = x i (t) + ε i (t) (3)

[0047] Among them, y i (t) is the measurement value, x i (t) is the true degradation state hidden under the measurement value, and ε i (t) represents the measurement noise, which follows a normal distribution with a mean of 0 and a variance of σ 2 , and it is assumed that the noises at different times are independent of each other. Equation (2) and Equation (3) constitute the state space model as the state transition equation and the measurement equation, respectively.

[0048] In the embodiment of the present application, the use of the expectation-maximization algorithm to process the accelerated degradation experiment data set to obtain the initial parameter estimation value of the Gamma state space model includes:

[0049] S104, determining the state transition model parameters and noise parameters to be estimated based on the Gamma state space model of each of the electronic devices;

[0050] S105, performing a particle filtering and smoothing algorithm on the pre-acquired historical degradation measurement data sets of each of the electronic devices to obtain a smoothed particle set;

[0051] S106, determining the log-likelihood function of the unknown parameters based on the smoothed particle set;

[0052] And determining the expectation of the log-likelihood based on the log-likelihood function of the unknown parameters;

[0053] S107, estimating the unknown parameters based on the expectation of the log-likelihood to obtain an initial model after parameter estimation.

[0054] Among them, the processor realizes the parameter estimation of the offline model by executing step S104 - step S107.

[0055] Specifically, assume that there are m devices in the historical degradation dataset, and these m devices run from the time of being put into use until failure. During the entire operation period, their degradation signals are recorded intermittently and non-periodically, and these degradation signals are used for the estimation of the state space model parameters. The two core problems to be solved in the model parameter estimation stage are: 1) The parameters θ = {θ 1:m , λ, u, σ 2} of the established state space model are unknown, and appropriate statistical estimation methods need to be introduced to estimate them based on the historical degradation dataset. 2) The state transition equation of the Gamma state space model is non-linear, which poses challenges to the estimation of the degradation state. Traditional linear filtering methods (such as Kalman filtering) are difficult to handle complex non-linear systems, and non-linear filtering techniques based on Monte Carlo need to be applied to accurately estimate the degradation state.

[0056] Due to the existence of measurement noise, the true degradation state of the device is unobservable, and the state transition equation has a complex form. Therefore, the Particle Filtering (PF) method is used to extract the true degradation state of the device. The PF method is a numerical implementation of the recursive Bayesian equation through Monte Carlo simulation. The PF method models the state transition equation as a first-order Markov process and performs state estimation based on the state transition model and the observation model. The PF method has been widely applied in practice, especially for non-linear systems or cases where an analytical model cannot be established. Without loss of generality, the state transition equation and the measurement equation are defined in the following form:

[0057]

[0058] where x k is the hidden state to be estimated, y k is the measurement value at the k-th moment, w k and v k are the process noise and the measurement noise respectively.

[0059] The Markov process has an initial distribution p(x0) and a transition equation p(x t |x t-1 ). The posterior probability density function p(x k |y 1:k ) can be represented as a set of N particles The corresponding weights are calculated from the conditional likelihood of each particle given the current observation result. The posterior PDF at the k-th time point can be approximated as:

[0060]

[0061] where δ(·) is the Dirac function, and y 1:k is the noise measurement sequence, and y 1:k = {y1, y2,..., y k}. The weights can be recursively updated using methods such as density importance sampling. The particle set is extracted from the prior PDF, and the prior PDF is usually used as the proposal distribution:

[0062]

[0063] The importance weights of the particles can be recursively updated according to the following formula:

[0064]

[0065] where is used to generate the importance density of the particles, and the weights of the particles can be normalized to:

[0066]

[0067] A potential problem in the above algorithm is the particle degeneracy problem. The weights of the particles are updated recursively, and there may be a phenomenon where the importance of large-weight particles increases continuously while the importance of small-weight particles decreases continuously. At this time, only a few particles in the particle set that affect the distribution are left, and this phenomenon is called particle degeneracy. The particle degeneracy phenomenon will cause waste of computing resources and make the calculation fall into a local optimal solution. To solve this problem, a resampling algorithm is introduced. The idea of the resampling algorithm is that after the normalized calculation of the particle weights is completed, resampling is performed according to the weight sizes of the particles. The probability of a particle being sampled is proportional to the size of its normalized weight. In this way, some particles with smaller weights can be randomly eliminated, and particles with larger weights are replicated multiple times, and the number of particles after sampling remains N. In this application, to reduce the computational amount, the basic importance sampling particle filter is used to estimate the degradation state.

[0068] Suppose the degradation processes of m devices are monitored collectively, where the monitoring time series of the i-th device is The degradation measurement sequence is The particle set obtained by the PF algorithm at time j is denoted as The corresponding weights are denoted as When approximating the posterior distribution, the particle filter only considers the observations up to the current moment. In the historical degradation dataset, the degradation data of the device from the time of commissioning to failure is known. To make full use of the information in the observations, this application applies backward particle smoothing after completing the forward particle filtering step to further improve the estimation accuracy of the health state. The particle set at time j obtained by forward particle filtering is At initialization, the smoothed particle set is set to be the same as the filtered particle set, and the number of smoothed particles is the same as the number of filtered particles. Denote the smoothed particle set as Then the weights of the backward smoothed particles can be calculated in the following way:

[0069]

[0070] where is the weight of the k-th particle of the i-th device at time j after the PF algorithm, is the weight corrected by using the observation information at time j + 1 through the particle smoothing technique.

[0071] It should be noted that the above degradation state estimation is carried out under the assumption of known model parameters Θ. Generally, in the case of known observations, the model parameters can be obtained by maximizing the likelihood function. However, this application takes into account the influence of observation noise. The noisy observations do not represent the true degradation state and the maximum likelihood estimation algorithm cannot be directly applied. Therefore, the EM algorithm is used to estimate the parameters of the Gamma-based state space model. Given the degradation measurements of m historical devices and the true degradation state The log-likelihood function of the unknown parameter Θ can be expressed as:

[0072]

[0073] where represents the PDF of the degradation observation given the true degradation state, which is only related to the measurement noise, while represents the PDF of the true degradation state, which is only related to the parameters in the state transition equation. The expectation of the log-likelihood can be divided into:

[0074]

[0075] where Θ1 = σ 2 is the noise parameter, and Θ2 = {θ 1:m , λ, u} are the state transition model parameters. The above formula can be further derived as:

[0076]

[0077]

[0078] where Dx i,j = x i (t i,j+1 ) - x i (t i,j ), Dv i,j = v i (t i,j+1 ) - v i (t i,j ). It is expected that the relevant terms can be calculated using the particles and their weights obtained from particle smoothing. Specifically, they can be calculated by the following formula:

[0079]

[0080] In this way, the expectation of the log-likelihood function can be calculated, and then the noise parameters and state transition model parameters are obtained by maximizing Equation (14) using an optimization algorithm. The EM algorithm is an iterative process, and the parameters estimated in each round are used for particle filtering and smoothing to obtain the true health state, and the obtained true health state is used for maximizing the expectation function.

[0081] In the embodiment of the present application, step S20 includes the following specific execution process:

[0082] S201. Use the initial parameter estimate as the initial value of the particle set to generate parameter particles;

[0083] S202. Use the particle filter algorithm based on kernel smoothing to shrink the parameter particles towards the mean to obtain the first particle parameters;

[0084] S203. Add random noise to the first particle parameters to obtain the second particle parameters;

[0085] S204. Obtain the updated particle parameters according to the first particle parameters and the second particle parameters;

[0086] S205. Based on the real-time monitoring data of the on-site operating devices and the updated particle parameters, generate the true health state data of each device;

[0087] S206. Use the expectation maximization algorithm to process the true health state data to obtain the parameters of the updated initial model, and use the updated initial model to process the true health state data to obtain the accelerated degradation information of the electronic device.

[0088] Among them, the processor realizes the online update of the model parameters by executing steps S201 - S206.

[0089] In the embodiment of the present application, step S204 includes the following specific execution process;

[0090] S2041. Determine the expression of the parameter state transition equation according to the function expressions of the first particle parameter and the second particle parameter;

[0091] S2042. Calculate the updated particle parameters according to the expression of the parameter state transition equation.

[0092] The obtained parameter offline estimation values represent the degradation characteristics of the accelerated life test data set. However, due to the differences between the operating environment of the in-service devices and the accelerated life test environment, the degradation paths of the devices will vary with the changes in operating conditions. Meanwhile, the heterogeneity among devices further exacerbates the differences in degradation paths, which makes the model established based on the accelerated life test data set unable to be directly used for predicting the RUL (Remaining Useful Life, RUL) of in-service devices. To achieve the adaptive update of the model from the accelerated failure environment to the in-service environment, this application proposes a method for joint state and parameter estimation based on kernel smoothing particle filtering. This method can not only accurately estimate the degradation state of in-service devices but also realize the online update of model parameters to ensure that the model can dynamically adapt to the in-service operating conditions. It should be noted that when conducting degradation modeling, the individual parameter θ in the model is established as a random variable, and θ ∼ N(μ θ , Σ θ ). The sample mean and sample covariance of the offline estimated θ 1:m are regarded as the estimated values of μ θ and Σ θ . For the in-service device p, its model individual parameter θ p can be regarded as a sample drawn from the distribution N(μ θ , Σ θ ). As the monitoring data of the degradation state of device p gradually increases, θ p is continuously updated through the kernel smoothing particle filtering method to achieve the adaptive adjustment of individual parameters, thereby effectively reflecting the actual degradation state of in-service devices.

[0093] The essence of Kernel Smoothing Particle Filtering (KS-PF) is an improvement of the enhanced particle filter. The idea of the enhanced particle filter is to regard the model parameters as elements of the state vector estimated by PF and introduce artificial evolution to solve the problem of particle impoverishment. The state transition equation of the parameter θ can be expressed as:

[0094]

[0095] where, is artificial noise, is device p at time t p,jThe k-th parameter particle vector. This method is less used in practice because it is difficult to determine artificial noise. Using a smaller The parameter particles converge to the true value very slowly and cannot solve the problem of particle depletion. Instead, using a larger will prevent the parameter particles from converging to the actual value. To address the above challenges, this application considers a PF method combined with kernel smoothing. The evolution of the parameter particles is achieved through two steps. The first step is contraction, and the first particle parameter is expressed as:

[0096]

[0097] where is the mean of the particle set , and s is the kernel parameter. The contraction step makes the parameter particles contract towards their mean, and the degree of contraction is controlled by s. The larger s is, the greater the degree of contraction towards the mean, and the smaller s is, the smaller the degree of contraction. The second step is perturbation, and the second particle parameter is expressed as:

[0098]

[0099] where V p,j is the variance of the particle set . The perturbation step involves adding random noise to the parameter particles, thereby enhancing particle diversity. In addition, the special setting of the random noise variance (the random noise variance is set to s 2 V p.j ) keeps the variance of the parameter particles unchanged after contraction and perturbation. This PF variant is called KS-PF. Based on equations (16) and (17), the state transition equation of the parameter can be derived as:

[0100]

[0101] In the embodiment of this application, step S30 includes the following specific execution process:

[0102] S301, model the failure threshold as a random variable obeying a normal distribution, where the mean and variance of the random variable are calculated through the historical failure data set;

[0103] S301, according to the updated initial model, determine the marginal survival function of each of the electronic devices;

[0104] S302, obtain the marginal survival function based on the survival function, approximate the marginal survival function using the Gauss-Hermite quadrature method, and obtain the life probability density function of the electronic device in the current degradation state;

[0105] S303. Calculate the remaining useful life of each of the electronic devices based on the life probability density function of the electronic devices, the current accelerated degradation information, and the randomized failure threshold.

[0106] In an embodiment of the present application, step S301 includes the following specific implementation processes:

[0107] S3011. Determine the probability density function of the shape parameter degradation increment under the condition of updated particle parameters based on the updated initial model, where the probability density function of the shape parameter degradation increment is a function of the failure threshold;

[0108] S30112. Determine the survival function of each of the electronic devices driven by the Gamma process based on the probability density function of the shape parameter degradation increment;

[0109] S30113. Integrate the survival function with respect to the failure threshold to obtain the marginal survival function of each of the electronic devices.

[0110] Among them, the processor realizes the online estimation of the URL of the electronic device by executing step S301-step S303.

[0111] Considering that when different devices fail, the thresholds reached by the degradation signals are different, the present application incorporates the uncertainty of the failure threshold into the prediction model. When the degradation state x(t) exceeds the threshold x F at that time, it is considered that the device fails, and the threshold x F is established as a random variable subject to a normal distribution, that is μ F and can be calculated using the actual failure time of the devices in the offline historical dataset. Denote the number of devices in the training set as m, then the failure threshold dataset is expressed as By using the sample mean and sample variance as estimates, the mean and variance of the estimated failure threshold can be obtained as:

[0112]

[0113] The duration from the current monitoring time to the failure moment is regarded as the RUL of the device at the time point . Since for the Gamma degradation process, the degradation process at each moment is a random variable, and the failure threshold is also regarded as a random variable, the life T F of the device is also a random variable, and its value is affected by multiple uncertainty factors. Once the parameter update of the in-situ operating device is completed based on the KS-PF algorithm, the RUL of the device can be predicted. For the sake of simplicity of expression, hereinafter is written as f(Dx(t)), where Under the condition of the updated parameters is the probability density function of the degradation increment Dx(t). Then the survival function of the device driven by the Gamma process can be written as:

[0114]

[0115] where z is the integration variable and has no physical meaning, is the lower incomplete Gamma function when x≥0 and a>0. Among them is the remaining useful life (RUL) prediction time of device p, is the health state of device p at the prediction time, and Dv(t) is the shape parameter increment.

[0116] Based on Equation (20), the marginal survival function obtained by integrating x F is:

[0117]

[0118] where p(x F ) is the probability density function (PDF) of the failure threshold x F . Equation (21) has no closed-form solution. Using the Gauss-Hermite quadrature method for approximation, when v(t) is differentiable, the conditional probability density function of the lifetime under the current degradation state of can be derived as:

[0119]

[0120] where Dv'(t) is the derivative of Dv(t), and the function is the derivative of the logarithm of the Gamma function.

[0121] At the prediction time , the degradation state estimated by the KS-PF algorithm is represented by a set of degradation particles. Given the estimated degradation state particles, denoted as and the model parameters obtained by the KS-PF algorithm, based on Equation (22), the PDF of the lifetime of the p-th device state under the condition that the observed degradation state is can be derived as:

[0122]

[0123] The effect of this application is verified through the accelerated failure dataset of insulated-gate bipolar transistors (IGBTs).

[0124] 1. Dataset Introduction

[0125] In this application, power cycle tests were conducted on 8 IGBT test samples, and these data fully reflect the degradation behavior of IGBTs under accelerated failure conditions. A total of 4 test conditions were designed in the experiment, and their detailed parameters are shown in Table 1, where T jmax is the maximum junction temperature during the experiment, and DT j is the change in junction temperature during heating and cooling, and t on and t off are the heating and cooling durations respectively. For each test condition, 2 devices were tested, and the thermal stress generated by the cyclic turn-on and turn-off of the power cycle current was used to accelerate the aging process. During the test, the on-state saturation voltage V CE of each cycle was recorded to characterize the device degradation state. The increase in V CE is considered to be one of the important characteristics of IGBT degradation and can reflect the performance degradation trend of the device under long-term stress.

[0126] Table 1 IGBT Test Conditions and Numbers

[0127]

[0128] For simplicity, the following uses serial numbers to represent the 4 test conditions. The test conditions, actual failure cycle numbers, and percentage increases in V CE of 8 IGBT test devices are shown in Table 2. The figure shows the increase path of V CE for each sample, and the endpoints of the trajectory represent the actual failure cycle numbers of the devices. This chapter develops an adaptive remaining life prediction algorithm based on this dataset to provide a solution for IGBT reliability management in actual application scenarios.

[0129] Table 2 True Failure Cycle Numbers of IGBT Samples and Degradation Signal Values at the Failure Moment

[0130]

[0131] The degradation process of IGBTs can be modeled by a physical degradation model, and the parameters in the model have specific physical meanings. The change in the on-state saturation voltage V CE during the IGBT degradation process can be expressed as:

[0132]

[0133] where i c is the collector current, ρ is the resistivity, c is the equivalent radius of the bond wire contact area, c0 is the initial value of c, h is the thickness of the aluminum metallization layer, b is its equivalent radius, and w is the factor that increases the resistivity caused by the reconstruction of the reactive metallization layer.

[0134] V CEAll the unknown parameters to be estimated in the varying physical model are \(c_0\), \(k_1\sim k_5\), and \(b\), which can be divided into two groups according to whether they are related to the test conditions. The first group includes the individual parameters \(\theta=\{k_1\sim k_5\}\), which vary with the test conditions. The second group includes the population parameters \(\lambda = \{c_0, b\}\), which are independent of the test conditions. As shown in the figure, although the degradation of all samples follows the physical model determined by Equation (24), there are significant differences in the degradation trajectories of different samples. In particular, due to the relatively mild test conditions, the failure times of Samples 3 and 4 are much longer than those of other samples. In degradation modeling, this difference is reflected by the variation of the individual parameters in the model.

[0135] 2. Degradation Modeling and Model Parameter Estimation

[0136] During the operation of the IGBT device, it degrades continuously, and its degradation behavior is modeled by the Gamma state - space model. Considering that the mean of the Gamma process at time \(t\) is \(u\cdot v(t)\), where \(u\) is the scale parameter and is independent of time. The above physical model is integrated into \(v(t)\) as prior knowledge to improve the performance of the degradation model. The unknown parameters in the state - space model are \(k_1\sim k_5\) and \(c_0\), \(b\), \(u\), \(\sigma\) 2 . Assume that \(\theta=\{k_1\sim k_5\}\) is a multivariate random variable with mean \(\mu\) θ and covariance \(\Sigma\) θ to cover the variations in operating conditions and device - to - device heterogeneity.

[0137] The power cycle dataset contains four test conditions, and two samples are tested under each condition, for a total of eight devices. To simulate the difference between the accelerated failure experiment environment and the actual operating conditions, when performing offline estimation, two devices under a specific condition are removed from the degradation dataset, and then these two devices are used as field - operating devices to simulate the actual working situation, continuously collect their degradation signals, and then perform RUL prediction on these two devices. The purpose is to construct a prediction model that better meets the actual application requirements by excluding the information related to the field - operating conditions in the historical degradation dataset, thus providing more powerful verification for the effectiveness of subsequent prediction model updates.

[0138] Execute the EM algorithm to obtain the individual parameters and the overall parameters \(c_0\), \(b\), \(u\), \(\sigma\) 2 . Then calculate the sample mean \(\mu\) according to and the sample variance \(\Sigma\) θ θ ​As the mean and covariance matrix of θ. During the experiment, sample 5 was randomly selected as the on-site operating device, and the convergence process of the estimated population parameters by offline estimation is shown in the figure. It can be seen from the figure that all parameters converge within 49 rounds of iteration. Among them, the parameter u converges the slowest, indicating that the estimation of parameter u is the most complex. After analysis, this is because the parameter, as the scale parameter of the Gamma process, scales the shape parameter and affects the variance of the Gamma process at the same time. The contribution of this parameter to the degradation trajectory is coupled with the size of the shape parameter, and the contribution to the randomness of the degradation state at each moment is coupled with the measurement noise. It is a complex parameter, so the convergence process is slower.

[0139] 3. Model Parameter Update

[0140] Given a test device, the individual parameter θ in the model parameters will be updated according to its status monitoring data, so that the prediction model is specific to this device. This application simulates the CM data collection process by assuming that the degradation signal of the test device is only known before the prediction time point. The parameters obtained by the EM algorithm in the offline parameter estimation stage are used as the initial parameters, and the degradation status of the on-site operating device is estimated according to the KS-PF algorithm introduced above, and at the same time, the individual parameters in the offline estimation model are updated. Taking devices 1 and 3 as examples, the degradation paths predicted at different points during the update process of the proposed model are illustrated.

[0141] Using the models updated by KS-PF at different time points, the predicted degradation paths of device 1 and device 3 at different prediction time points are plotted in the figures. The prediction time points are selected as 40%, 60% and 80% of the device life cycle, which represent the early, middle and late stages of the device life cycle respectively. The 95% confidence interval in the figure comes from time uncertainty. For the Gamma process, the increment at each time point is random. By randomly generating 1000 Gamma increments at each time point, the estimated value of the degradation state at the next time point is obtained, and the 97.5% and 2.5% quantile values of these values are taken to obtain the 95% confidence interval of the degradation trajectory prediction. When the prediction time is 40% and 60% of the life, there are obvious deviations between the predicted path and the real path. This is because the estimated parameters obtained in the offline process represent the overall characteristics of the entire historical data set, and there are deviations between the degradation paths of the devices in the training data set and device 1 or device 3. At these two time points, the status monitoring data of device 1 or device 3 is very little, and the adjustment of the offline estimation model is very limited, so the degradation trajectory of device 1 or device 3 cannot be accurately predicted. As the prediction time increases, more and more degradation data of the on-site operating device is collected, and the prediction model is adjusted to the on-site operating environment. It can also be seen from the figure that when the prediction time is 80% of the life, the predicted degradation trajectory is more consistent with the actual path, which indicates that the proposed algorithm effectively uses the on-site operating data to update the model trained offline.

[0142] 4. Analysis of RUL Prediction Results

[0143] Generally, the failure criterion of IGBT is determined according to the growth percentage of the on-state saturation voltage. Since the physical model adopted in this application describes the increase in the on-state saturation voltage, multiplying the calculated failure threshold random variable by the initial voltage of the test sample can obtain the failure threshold distribution of the on-state saturation voltage increase. Assuming the initial voltage of a certain IGBT device is V0, then μ F and need to be reassigned to μ F ·V0 and Extract 100 sample points from the distribution as the failure thresholds for calculating RUL, and then perform RUL prediction based on the method introduced above. Rotate the eight devices under four working conditions as the test devices in turn, and perform RUL prediction at 80% of the life of each device. It should be noted that the state monitoring data of all devices are discrete, so it is impossible to obtain RUL prediction at the exact 80% of their life cycles. For example, for device 1, the proportion of the life cycle corresponding to the actual prediction time is 81.6%.

[0144] The figure shows the results of RUL prediction for different test devices at 80% of their life. Each scatter point in the figure represents the predicted RUL value corresponding to a certain failure threshold sample, the violin plot describes the overall distribution of RUL, and the red asterisks represent the actual RUL values of each device. It can be seen from the figure that the actual RUL always falls within the distribution range of the predicted RUL, indicating that the proposed method can perform accurate RUL prediction in the later stage of the device life while considering the uncertainty of the failure threshold.

[0145] This application realizes the adaptive adjustment of the prediction model from the accelerated failure test environment to the field operation environment through the update of individual parameters. The update process is carried out with the continuous collection of the state monitoring data of the devices in the field operation. To verify the effectiveness of the update process, this application analyzes how the prediction error changes with the prediction start time. Normalize the prediction start time and the prediction error according to the life of each device, and the error changes of different devices during the prediction process are shown in the figure. It can be seen from the figure that the initial prediction results show relatively large prediction errors, and as the prediction progresses, the normalized prediction error decreases significantly.

[0146] To verify the necessity and effectiveness of the model adaptive update step, Table 3 shows the comparison of the prediction accuracy of the model before and after adopting the adaptive mechanism. It can be seen from the table that the prediction accuracy before adopting the adaptive mechanism does not improve with the passage of the prediction time, and the overall prediction accuracy is relatively low. After adopting the adaptive mechanism, with the passage of the prediction time, the prediction accuracy becomes higher and higher. When the normalized prediction time is 0.8, the prediction accuracy reaches 0.965.

[0147] Table 3 Comparison of Prediction Accuracy of the Model before and after Adopting the Adaptive Mechanism

[0148]

[0149] Through case verification, the following conclusions can be obtained:

[0150] Through the adaptive adjustment of the prediction model from the accelerated failure experiment environment to the field operation environment in this application, high-precision RUL prediction of electronic devices under field operation conditions is achieved. Specifically, the prediction error decreases with the passage of the prediction time. When predicting at 80% of the device life, the prediction accuracy reaches 0.965. This method overcomes the challenge that the prediction model trained in the accelerated failure experiment environment cannot be used in the field operation environment through the adaptive update of the model, and has high practical significance and engineering application value.

[0151] Reference Figure 9 , on the basis of the above embodiments, this application also proposes an online remaining life prediction system for electronic devices under field operation conditions. The online remaining life prediction system 1000 for electronic devices includes:

[0152] The offline parameter estimation module 1001 is used in the offline stage to model the degradation process of the electronic device by using the Gamma stochastic process, and embed the physical prior model as the shape parameter into the Gamma stochastic process to obtain the Gamma state space model. The expectation maximization algorithm is used to process the pre-obtained accelerated degradation experiment data set to obtain the initial model after parameter estimation; the online parameter update module 1002 is used in the online stage to input the real-time monitoring data of the field operation device into the state and parameter joint estimation algorithm of the kernel smoothed particle filter to obtain the accelerated degradation information of the dynamically updated electronic device and the parameters of the initial model of the electronic device; the life estimation module 1003 is used to calculate the remaining life of each electronic device according to the final updated initial model, the randomized failure threshold distribution, and the accelerated degradation information.

[0153] On the basis of the above embodiments, this application also proposes a computer-readable storage medium, which includes instructions that, when running on a computer, cause the computer to execute the online remaining life prediction method for electronic devices under field operation conditions provided in any one of the foregoing embodiments.

[0154] Based on the above embodiments, the present application further provides an electronic device, which includes: at least one processor, a memory, and an input / output unit; wherein, the memory is used to store a computer program, and the processor is used to call the computer program stored in the memory to execute the on-line prediction method for the remaining life of an electronic device under on-site operating conditions provided in any of the foregoing embodiments.

[0155] The above are only the preferred embodiments of the present application, and do not limit the patent scope of the present application. Any equivalent structure or equivalent process transformation made by using the content of the specification and drawings of the present application, or directly or indirectly applied in other related technical fields, shall be equally included in the patent protection scope of the present application.

Claims

1. An online prediction method for the remaining life of electronic devices under on-site operating conditions, characterized in that, Including: In the offline stage, a Gamma random process is used to model the degradation process of electronic devices, and a physical prior model is embedded as a shape parameter into the Gamma random process to obtain a Gamma state space model. The expectation-maximization algorithm is used to process the pre-acquired accelerated degradation experiment data set to obtain an initial model after parameter estimation; In the online stage, real-time monitoring data of in-service devices is input into the state and parameter joint estimation algorithm of kernel smoothed particle filter to obtain the accelerated degradation information of the dynamically updated electronic devices and the parameters of the initial model of the electronic devices; According to the final updated initial model, the randomized failure threshold distribution, and the accelerated degradation information, the remaining life of each electronic device is calculated.

2. In the embodiment of the present application, the on-line prediction method for the remaining life of an electronic device under the on-site operating conditions as described in claim 1, wherein In the offline stage, a Gamma random process is used to model the degradation process of electronic devices, and a physical prior model is embedded as a shape parameter into the Gamma random process to obtain a Gamma state space model, including: Obtain the probability density function of the Gamma random process, where the probability density function includes a shape parameter, a scale parameter, and a measurement noise parameter, and the shape parameter has a degradation state increment; Express the degradation state increment as a random variable related to the device individual to obtain the probability density function of the device individual, and determine the true degradation state of the electronic device based on the probability density function of the device individual; Based on the true degradation state of each electronic device after adding Gaussian noise, the Gamma state space model of each electronic device is correspondingly determined.

3. The online prediction method for the remaining life of an electronic device under the on-site operating conditions as described in claim 1 in the embodiments of the present application, wherein, The use of the expectation-maximization algorithm to process the pre-acquired accelerated degradation experiment data set to obtain an initial model after parameter estimation includes: Determine the unknown parameters to be estimated based on the Gamma state space model of each electronic device, where the unknown parameters include state transition model parameters and noise parameters; Perform a particle filter and smoothing algorithm on the historical degradation measurement data set of each pre-acquired electronic device to obtain a smoothed particle set; Determine the log-likelihood function of the unknown parameters based on the smoothed particle set, and determine the expectation of the log-likelihood based on the log-likelihood function; Estimate the unknown parameters based on the expectation of the log-likelihood to obtain an initial model after parameter estimation.

4. The online prediction method for the remaining life of an electronic device under on-site operating conditions according to claim 1, characterized in that, Inputting the real-time monitoring data of in-service devices into the state and parameter joint estimation algorithm of kernel smoothed particle filter to obtain the accelerated degradation information of the dynamically updated electronic devices and the parameters of the initial model of the electronic devices includes: Taking the parameter particle vector at any moment as a parameter particle; Using the particle filter algorithm based on kernel smoothing to shrink the parameter particle towards the mean to obtain a first particle parameter; Adding random noise to the first particle parameter to obtain a second particle parameter; Obtaining an updated particle parameter according to the first particle parameter and the second particle parameter; Generating the true health state data of each device based on the real-time monitoring data of the in-service device and the updated particle parameter; Process the real health state data by using the expectation maximization algorithm to obtain the parameters of the updated initial model, and process the real health state data by using the updated initial model to obtain the accelerated degradation information of the electronic device.

5. The online prediction method for the remaining life of an electronic device under on-site operating conditions according to claim 4, characterized in that, The obtaining of the updated particle parameters according to the first particle parameter and the second particle parameter includes: Determine the expression of the parameter state transition equation according to the function expressions of the first particle parameter and the second particle parameter; Calculate the updated particle parameters according to the expression of the parameter state transition equation.

6. The on-line prediction method for the remaining life of an electronic device under on-site operating conditions according to claim 1, characterized in that The calculating of the remaining life of each electronic device according to the finally updated initial model, the randomized failure threshold distribution and the accelerated degradation information includes: Determine the marginal survival function of each electronic device according to the updated initial model; Obtain the marginal survival function based on the survival function, approximate the marginal survival function by using the Gauss-Hermite quadrature method to obtain the life probability density function of the electronic device under the current degradation state; Calculate the remaining life of each electronic device based on the life probability density function of the electronic device, the current accelerated degradation information and the randomized failure threshold.

7. The on-line prediction method for the remaining life of an electronic device under on-site operating conditions as claimed in claim 6, characterized in that, The determining of the marginal survival function of each electronic device according to the updated initial model includes: Based on the updated initial model, determine the probability density function of the shape parameter degradation increment under the condition of the updated particle parameters, where the probability density function of the shape parameter degradation increment is a function of the failure threshold; Determine the survival function of each electronic device driven by the Gamma process based on the probability density function of the shape parameter degradation increment; Integrate the survival function with respect to the failure threshold to obtain the marginal survival function of each electronic device.

8. An online prediction system for the remaining life of electronic devices under on-site operating conditions, characterized in that, It includes the following steps: An offline parameter estimation module, which is used in the offline stage to model the degradation process of the electronic device by using a Gamma stochastic process, embed the physical prior model as the shape parameter into the Gamma stochastic process to obtain a Gamma state space model, and process the pre-obtained accelerated degradation experiment data set by using the expectation maximization algorithm to obtain the initial model after parameter estimation; An online parameter update stage, which is used in the online stage to input the real-time monitoring data of the on-site operating device into the state and parameter joint estimation algorithm of the kernel smoothing particle filter to obtain the accelerated degradation information of the dynamically updated electronic device and the parameters of the initial model of the electronic device; A life estimation module, which is used to calculate the remaining life of each electronic device according to the finally updated initial model, the randomized failure threshold distribution and the accelerated degradation information.

9. A computer-readable storage medium, characterized in that, It includes instructions that, when running on a computer, cause the computer to execute the method for online prediction of the remaining life of an electronic device under the on-site operating conditions described in any one of claims 1 to 7.

10. An electronic device, characterized in that, The electronic device includes: At least one processor, a memory and an input / output unit; Wherein, the memory is used for storing a computer program, and the processor is used for calling the computer program stored in the memory to execute the online prediction method for the remaining life of the electronic device under the on-site operating conditions according to any one of claims 1 to 7.

Images