Method for predicting residual service life of MOSFET device by combining neural network and random process

Abstract

The invention belongs to the technical field of residual service life prediction of electronic components, and relates to an MOSFET device residual service life prediction method based on combination of a neural network and a random process. According to the invention, the sample interactive convolutional network is adopted to effectively learn the degradation trend of historical data, and the long-term degradation trend of the target equipment is predicted. And updating the degradation state by adopting a nonlinear fractal Brownian motion random degradation model with a measurement error and a long-term dependence characteristic. Combining a degradation measured value with a predicted value, and constructing a state space model based on nonlinear fractal Brownian motion; a state parameter joint estimation method for synchronously updating parameters based on a Kalman filtering algorithm, an expectation maximization algorithm and a maximum likelihood estimation algorithm is adopted. The method can also be popularized to RUL prediction of different devices such as lithium batteries, and theoretical basis and technical support are provided for intelligent operation and maintenance and maintenance decision making of electronic products. The technology provided by the invention has important value for reducing energy consumption and operation cost.

Description

Technical Field

[0001] The present invention belongs to the technical field of remaining useful life (RUL) prediction of electronic components, and relates to a remaining useful life prediction method for MOSFET devices combining a neural network with a random process, and is a prediction method combining a sample interactive convolution network, a nonlinear fractal Brownian motion and a state-parameter joint estimation method. Background Art

[0002] MOSFET devices are one of the basic components of electronic products. MOSFET has the advantages of simple circuit, fast switching speed, good thermal stability, and low conduction loss. It is widely used in transportation, photovoltaic power generation, industrial control and other fields, which can significantly reduce power loss and improve system efficiency. However, during use, as the conduction time of MOSFET increases, the power and performance will gradually decrease, seriously affecting the safe operation of the entire system. Therefore, the RUL prediction of MOSFET is crucial to ensure the safety and stability of the equipment, and has important value in reducing energy consumption and operating costs.

[0003] In recent years, most of the life research on power MOSFET devices is based on two categories: finite element simulation analysis and data-driven. The former uses finite element simulation modeling and analysis to study the failure of MOSFET devices in multi-physical fields and analyze their fatigue life. However, this type of method requires the establishment of an accurate geometric dimension model, and secondly, it is impossible to realize the online acquisition of parameters, so it is impossible to understand the changes in the electrical parameters of the device. The data-driven approach has the advantage that only the historical operation data of the device needs to be collected to predict the future changes of the device, so it has become a trend in the study of MOSFET device life prediction. Patent CN 114186492 A discloses a SiC MOSFET device life prediction method based on a gated recurrent unit neural network. The method first establishes a gated recurrent unit neural network model; secondly, establishes a model evaluation mechanism and trains the neural network model; and finally predicts the remaining service life of the SiC MOSFET device. However, under the combined action of internal and external stresses, the degradation of MOSFET devices will be uncertain. The method proposed in the patent can only obtain a point estimate of the RUL prediction, rather than a probability density function (PDF) to describe this uncertainty. Therefore, it is necessary to propose a new prediction method for the RUL of MOSFET devices. Summary of the invention

[0004] The present invention proposes a hybrid prediction method for the remaining useful life of a MOSFET that combines a neural network with a nonlinear fractal Brownian motion. A sample interactive convolutional network (SCINet) is used to effectively learn the degradation trend of historical data and predict the long-term degradation trend of the target device. A nonlinear fractal Brownian motion (FBM) random degradation model with measurement error and long-term dependence characteristics is used to update the degradation state. The measured degradation value is combined with the predicted value, a state space model is constructed based on the FBM, and a state parameter joint estimation method based on the Kalman filter (KF) algorithm, the expectation maximization (EM) algorithm and the maximum likelihood estimation (MLE) algorithm is used to synchronously update the parameters. Compared with the existing technology, the present invention has strong innovation and high prediction accuracy.

[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is:

[0006] A method for predicting the remaining useful life of a MOSFET device by combining a neural network with a random process comprises the following steps:

[0007] Step 1: Based on the power cycle test aging test, the on-state voltage Vds(on) is selected as the aging parameter of the MOSFET device, and the on-state voltage change of the MOSFET device is collected as the degradation data to characterize the device degradation. The moving average filtering method is used to eliminate the degradation data noise, smooth the signal, and set the failure threshold.

[0008] Step 2: Based on the degradation data collected in step 1, the training set and the test set are divided based on the leave-one-out method. Based on the training set, the sample interaction convolutional network SCINet is trained to output the future prediction value of the degradation data, thereby learning the historical trend of MOSFET device degradation.

[0009] Step 3: Based on the trained SCINet network, iteratively predict the test set data until the prediction exceeds the failure threshold and obtains the predicted observation results of the long-term degradation trend.

[0010] Step 4: Based on the nonlinear fractal Brownian motion FBM, a nonlinear random degradation model is constructed; based on the measurement error, a measurement equation is constructed; based on the measurement equation and the nonlinear degradation model, a state space model considering the measurement error is constructed.

[0011] Step 5: Based on the degradation data of the test set, the Hurst exponent is estimated in combination with the rescaled range (R / S) analysis method. Then, the state-parameter joint estimation strategy is implemented, in which the Kalman filter (KF) is used to estimate the hidden degradation state. In this process, the predicted value predicted by the SCINet network is used as a new observation value and brought into the state space model. Then, the expectation maximization (EM) algorithm is used to estimate the relevant parameters. Finally, the maximum likelihood estimation (MLE) is used to correct the parameter estimation results.

[0012] Step 6: Based on the concept of first arrival time, define the remaining useful life (RUL) at any time and define the corresponding reliability function. Based on the parameter estimation results of step 5 and the statistical characteristics of the FBM diffusion term increment in step 4, a recursive strategy is used to approximately perform the extrapolation process. The final RUL and the corresponding probability density distribution (PDF) are obtained based on the Monte Carlo (MC) method.

[0013] Beneficial effects of the present invention:

[0014] Aiming at the RUL prediction problem of MOSFET devices, the present invention proposes a hybrid prediction method based on SCINet and nonlinear FBM. Among them, the SCINet network effectively learns the degradation trend of historical data and predicts the long-term degradation trend of the target device. At the same time, a random degradation model is constructed using nonlinear fractional Brownian motion (FBM) with measurement error and LRD characteristics to update the degradation state. Combining the existing measured values ​​and predicted values, based on the state space model, combined with KF, EM and MLE algorithms, a state-parameter joint estimation method is used for synchronous update. The MC method is used to realize RUL and PDF acquisition. The present invention overcomes the shortcomings of random process methods and artificial intelligence methods in RUL prediction, can quantify the uncertainty problem of neural network RUL prediction and the randomness of the degradation process, and achieve better prediction accuracy.

[0015] The hybrid prediction method for the remaining useful life of MOSFET devices combined with a neural network and a random process provided by the present invention can not only be applied to the online and offline prediction of the remaining useful life of MOSFETs devices, but can also be extended to the RUL prediction of different devices such as lithium batteries, providing a theoretical basis and technical support for the intelligent operation and maintenance of electronic products. The technology of the present invention should reduce energy consumption and operating costs, and has important value. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is the overall technical path of the remaining useful life hybrid prediction method that combines the neural network and the random process adopted in the present invention.

[0017] Figure 2 It is the degradation data of the on-voltage change of the MOSFET device collected in the embodiment of the invention.

[0018] Figure 3 1 is the actual degradation path of device 1 in the embodiment of the present invention and the multi-step advance prediction results of SCINet at different starting points.

[0019] Figure 4The figures are the prediction results of RUL and PDF of MOSFET devices 1-3 at different cycle periods using the proposed method and the comparative method in the embodiments of the present invention.

[0020] Figure 5 The RUL prediction results of MOSFET devices 1-3 at different cycle times using the proposed method and the comparative method in the embodiments of the present invention are shown. DETAILED DESCRIPTION

[0021] The specific implementation of the present invention is further described below in conjunction with the accompanying drawings and technical solutions.

[0022] like Figure 1 As shown, a method for predicting the remaining useful life of a MOSFET device by combining a neural network with a random process comprises the following steps:

[0023] Step 1, based on the power cycle test aging test, select the on-voltage Vds(on) as the aging parameter of the MOSFET device, collect the on-voltage change of the MOSFET device as the degradation data characterizing the device degradation, use the moving average filtering method to eliminate the noise of the degradation data, smooth the signal, and set the failure threshold. In order to retain the characteristics of the data, the filter queue length U is set to 50 in this embodiment. In addition, when filtering all data, use consistent filtering parameters.

[0024] Step 2: Based on the degradation data collected in step 1, the training set and the test set are divided based on the leave-one-out method. Based on the training set, the sample interaction convolution network SCINet is constructed and trained. The input is the historical degradation data, and the output is the future prediction value of the degradation data. Through the SCINet network, the historical trend of MOSFET device degradation is learned. In the implementation process, a multi-step prediction strategy is adopted. And a grid search is performed on all adjustable hyperparameters in the SCINet network to determine the optimal parameters.

[0025] Among them, the SCINet network can learn effective representations with enhanced predictability by iteratively extracting and exchanging information at multiple time resolutions. The SCINet network is stacked with multiple SCI-Net sub-blocks, so that it can be used to predict more complex time series. Each SCI-Net sub-block is added with residual connections, so the gradient disappearance caused by the network being too deep can be avoided, thereby ensuring the network training effect. In addition, a relay supervision mechanism is used at the output of each SCI-Net sub-block to simplify the learning of intermediate time series features. The SCI-Net sub-block consists of SCI-Blocks with a binary tree structure. In each SCI-Block, the time series is divided into two parts. As the depth of the binary tree increases, the "time resolution" of the time series continues to decrease, and more detailed time series information will also be extracted. In this way, both short-term and long-term dependencies in the degraded sequence will be extracted. After passing through the L-layer SCI-Block, the time series sequence is reassembled based on Concat&Realign to generate a new sequence, and it is added to the original sequence using residual connections. Finally, the prediction result or feature is obtained through the fully connected layer. In SCI-Block, the time series is divided into S odd and S even Two subsequences (separate the elements in the original sequence according to odd and even positions to form two new subsequences. odd contains the elements with odd positions in the original sequence, and S even The interactive learning between the two subsequences is completed through the following strategy:

[0026] Using two different one-dimensional convolution modules c1 and x2, S odd and S even They are mapped to two hidden states respectively, and then converted into exponential form for element-wise multiplication.

[0027]

[0028] in, Represent odd and even subsequences; denotes the updated odd and even subsequences after the interactive learning step. ⊙ denotes the Hadamard product, i.e., element-wise multiplication. exp denotes the exponential function.

[0029] Using two different one-dimensional convolution modules c3 and c4, continue Mapped to two hidden states respectively, and then perform addition or subtraction operations to obtain the final sub-feature S′ odd and S′ even .

[0030]

[0031] Among them, S′ odd and S′ even It indicates the updated odd and even subsequences output after the interactive learning process. ± indicates that addition or subtraction operations can be performed.

[0032] In view of the effectiveness of the SCINet network in timing prediction tasks, the present invention uses the SCINet network to extract the long-term degradation trend exhibited by MOSFET devices. Specifically, the degradation trend predicted by the SCINet network based on historical observations is regarded as a future observation and is input into the subsequent stochastic process model.

[0033] Step 3: Based on the trained SCINet network, iteratively predict the test data until the prediction exceeds the failure threshold, and obtain the predicted observation results of the long-term degradation trend. Specifically, the input sequence length of the SCINet network is defined as p, and the degradation data z0, z1, …, z up to time p-1 are input. p-1 Enter the SCINet network and observe the value z p is predicted, i.e. Then form a new input Next, at time p+1, the observed value of the next time step is predicted This iterative prediction continues until the predicted value of the observation exceeds the predefined fault threshold within the prediction range, and the predicted observation results of the long-term degradation trend can be obtained.

[0034] Step 4: Based on fractal Brownian motion (FBM), a nonlinear random degradation model is constructed; based on the measurement error, a measurement equation is constructed; based on the measurement equation and the nonlinear degradation model, a state space model considering the measurement error is constructed.

[0035] 4.1) The nonlinear random degradation model can be expressed as follows:

[0036]

[0037] Wherein, Y(0) represents the initial degradation state of the MOSFET device. For the sake of generality, it is assumed that Y(0) = 0. Y(t) represents the degradation state at time t. represents a nonlinear drift function that depends on the integration time s and the unknown parameter vector ρ, and σ represents the diffusion coefficient, which reflects the degree of fluctuation of the degradation process. H (t) represents the standard FBM, which is used to characterize the time-varying uncertainty of degraded devices. H is the Hurst exponent.

[0038] 4.2) Based on formula (3), the measurement equation (Z(t) = Y(t) + ε) shown in formula (4) is introduced to solve the unknown measurement noise. Therefore, at any time t, the MOSFET device degradation Z(t) directly monitored by the power cycle test aging test can be expressed as the following state space degradation model:

[0039]

[0040] where ε represents the measurement error, and is the variance of the measurement error. The nonlinear drift function is defined as the power law function λt b , where λ and b are the parameters to be estimated in the power law function.

[0041] Step 5: Based on the degradation data of the test set, the Hurst exponent Hurst is estimated by combining the rescaling range (R / S) analysis method; then the state-parameter joint estimation strategy is implemented. In the implementation of the state-parameter joint estimation strategy, the Kalman filter (KF) is used to estimate the hidden degradation state y and the unknown parameter set Ψ = {b, λ, σ, σ z}, the expectation maximization (EM) algorithm is used to estimate the relevant parameters, and the maximum likelihood estimation (MLE) is used to correct the parameter estimation results; in the implementation process, after the actual observation value is obtained at the latest moment, the predicted degradation value of the SCINet network is used as the new observation value and substituted into the state space degradation model; based on the state-parameter joint estimation method, multiple parameters and degradation states are continuously updated until the fault threshold is reached; the specific steps are as follows:

[0042] 5.1) First, use the KF algorithm to recursively estimate the hidden degradation state. At the nth monitoring time, let Ψ n = {b n ,λ n ,σ n ,σ z,n} represents the parameter set to be estimated. n , the time update equation and the measurement update equation are expressed as formula (5) and (6) respectively:

[0043]

[0044] Among them, y n∣n and n+1∣n Represents the predicted value of the hidden state based on the information of step n at step n and step n+1 respectively. n∣n and C n+1∣n They represent the covariance prediction values ​​of the hidden state estimation based on the n-step information at the n-step and n+1-step, respectively. n ) and h(t n+1 ) are respectivelyn and t n+1 The covariance function value of the process noise in time.

[0045]

[0046] Among them, y n+1∣n+1 Represents the updated estimated value (filtered state) of the hidden state based on the n+1 step measurement information at step n+1. n+1∣n+1 represents the updated covariance (filtered covariance) of the hidden state estimate based on the n+1 step measurement information at step n+1. n+1 Represents the actual measured value at step n+1. represents the measurement noise variance at step n.

[0047] 5.2) In order to cooperate with KF, the EM algorithm is used to n Finally, the estimated value of the parameter is updated in the d+1th iteration They can be expressed as:

[0048]

[0049] in, represents the hidden state of the jth time step estimated based on the observed data in the dth iteration; The square of z j represents the actual measurement value at the jth time step.

[0050]

[0051] Among them, h(t j-1 ) and h(t j ) respectively represent the j-1 and t j The covariance function value of the process noise at time t. They represent the hidden states of the j-1th and jth time steps estimated based on the observed data in the dth iteration, respectively. represents the estimated value of b at the nth step in the d+1th iteration; middle, is the power index.

[0052]

[0053] In addition, although There is no analytical solution, but an approximate solution can be achieved through numerical methods based on MATLAB.

[0054] 5.3) The discretization of the state space degradation model ignores some characteristics of FBM, and the direct coupling of model parameters may lead to inaccurate estimation results. In order to correct the deviation in parameter estimation, maximum likelihood estimation is applied. Unlike the EM algorithm, MLE directly uses the state estimate obtained by KF as the observation value. At this time, the degradation state, diffusion term, and drift term can be expressed as: Y = [y 1∣1 … y m∣M ]′, B H =[B H (t1) … B H (t M )]′、 The superscript “′” indicates the transposition operation of the matrix.

[0055] Then Y~N(λg,σ 2 D), where D represents B H The covariance matrix of . The log-likelihood function of the relevant parameters F(b,λ,σ|Y) can be expressed as:

[0056]

[0057] Where F is the total time step. In addition, the estimate of b is independent of the FBM and no longer requires an update process. Therefore, only λ and σ in the log-likelihood function need to be updated. The estimated values ​​of λ and σ can be further updated as:

[0058]

[0059]

[0060] Among them, g T is the transpose of g, D -1 is the inverse matrix of D.

[0061] Step 6: Based on the concept of first arrival time, define the RUL at any time and define the corresponding reliability function. Based on the parameter estimation results of step 5 and the statistical characteristics of the FBM diffusion term increment, a recursive strategy is used to approximately perform the extrapolation process. The final RUL and the corresponding PDF are obtained based on the Monte Carlo (MC) method.

[0062] 6.1) According to the concept of first arrival time, when the failure threshold is ω, the degraded MOSFET device will n The RUL at a monitoring moment is defined as:

[0063] L n = inf {r n :y(t n +r n )≥ω|y(t n )≤ω} (13)

[0064] Among them, L n Remaining useful life, r n Indicates that from the current time t n The time difference to a certain point in the future. The symbol inf represents the operation of taking the maximum lower bound. n ) and y(t n +r n ) represent t n With t n +r n The status of the system at the moment.

[0065] The corresponding reliability function can be defined as:

[0066]

[0067] Among them, R n (r n ) is the reliability function. Where P represents probability, represents the estimated values ​​of the model parameters, Represents an estimate of the system state.

[0068] 6.2) Monte Carlo (MC) method is used to simulate the PDF of RUL. The core idea is to generate a large number of sample paths to approximate the reliability function given by formula (16).

[0069] In order to solve the difficulty of accurate extension of FBM, the present invention is based on the parameter estimation result of step 5 and the diffusion term increment (Y(t) = Y(0) + λt b +σB H (t)), a recursive scheme is used to approximate the extrapolation process. Specifically, the final simulation samples can be generated in the following way:

[0070]

[0071] Where i is the index of the time step, τ is the sampling interval of the extrapolation, and v n is a Gaussian random variable, where Represents the final parameter estimate. By calculating formula (14), the reliability function R of each extrapolation time is directly obtained n (iτ). Afterwards, the PDF of RUL at the nth monitoring time can be obtained by Get, where f n (r n ) is the PDF of RUL at the nth monitoring moment, is a positive integer. In actual calculations, the numerical differentiation method is used to calculate the discrete gradient operator To solve.

[0072] Implementation Cases:

[0073] The power cycle test device constructed in this embodiment is mainly composed of a power supply, a power switch, a temperature sensor, a cooling device, a host control system and a data collection system. The on-voltage Vds(on) is used as the aging parameter of the device, and the on-voltage signal is collected in real time during the power cycle test. According to the highest temperature cycle point set in the specification of the MOSFET device used for the test, the maximum test temperature is set to 200°C and the minimum temperature is set to 40°C. Based on the power cycle test platform, a power cycle aging test is performed on the MOSFET device under two temperature conditions. Six MOSFET devices are tested at 160°C.

[0074] During the implementation process, in order to demonstrate the advancedness of the method proposed in the present invention, multiple mainstream RUL prediction methods are used to compare the prediction results of MOSFETs. The comparison prediction methods are introduced as follows:

[0075] Comparison method 1: Nonlinear adaptive WP model, which can be expressed as Where B(u) represents the standard Brownian motion.

[0076] Comparison method 2: Nonlinear adaptive WP model considering measurement error can be expressed as Here, ε represents the measurement error.

[0077] Comparison method 3: FBM model without SCINet can be expressed as Among them, B H (u) represents fractal Brownian motion.

[0078] In the above method, Z(u) represents the degradation measurement value, represents the diffusion term, is a nonlinear function of the input u and parameter vector θ, both of which use power law functions. σ represents the diffusion coefficient.

[0079] The degradation data of the on-state voltage variation of the MOSFET device collected in this embodiment is as follows: Figure 2 shown.

[0080] The actual degradation path of device 1 in this embodiment and the multi-step advance prediction results of SCINet at different starting points are as follows: Figure 3As shown in the figure. It can be seen that the predicted values ​​of the trained SCINet model are able to characterize the long-term degradation trend, so that its future state can be estimated. Specifically, in the early degradation stage, the predicted trajectory cannot accurately capture the actual degradation path. Therefore, the predicted failure threshold time is earlier or later than the true end-of-life time. However, as time goes by, more and more observation data are available, and the predicted trajectory better describes the true degradation trend. Therefore, the predicted degradation value is closer to the actual value, which means that the RUL prediction gradually converges to the actual RUL.

[0081] In this embodiment, the prediction results of RUL and PDF of MOSFET devices 1-3 at different cycle times using the proposed method are as follows: Figure 4 As shown in the figure, the early prediction of RUL shows a wide distribution and flat PDF shape due to the large parameter estimation error in the initial degradation stage. As the observation data accumulates, the parameter estimation becomes more accurate, resulting in a clearer and narrower PDF curve. This means that the uncertainty of RUL prediction decreases.

[0082] The RUL prediction results of MOSFET devices 1-3 under different cycle times using the proposed method and the comparative method in the embodiment of the present invention are as follows: Figure 5 As shown. It can be seen from the figure that with the emergence of more observations, the RUL value of each model gradually approaches the actual RUL. However, compared with other methods, the predicted value of the method proposed in the present invention is closer to the actual RUL. In addition, through the 30% performance area analysis, it can be seen that only a small number of RUL prediction results of the comparison methods fall within this range. Most of the prediction results of the method proposed in the present invention are within the performance range of 30%, which also proves that it has a high prediction accuracy. Compared with other degradation models, the method proposed in the present invention has significant advantages such as non-Markov, LRD characteristics and full use of historical degradation trends. There is less fluctuation during the prediction process.

[0083] The present invention can be applied to the fault prediction and health management of MOSFET devices, and realize the online and offline prediction of the remaining service life of the equipment. It provides a theoretical basis and technical support for the intelligent operation and maintenance of electronic products. The technical application of the invention can improve the safety, reliability and availability of electronic products and reduce the maintenance cost of equipment.

Claims

1. A method for predicting the remaining useful life of a MOSFET device by combining a neural network with a random process, characterized in that: The following steps are involved: Step 1: Based on the power cycle test aging test, the on-voltage is selected as the aging parameter of the MOSFET device, the on-voltage change of the MOSFET device is collected as the degradation data characterizing the device degradation, and the moving average filtering method is used to eliminate the degradation data noise, smooth the signal, and set the failure threshold; Step 2: Based on the degradation data collected in step 1, the training set and the test set are divided according to the leave-one-out method; Based on the training set, the sample interactive convolutional network SCINet is trained to output the future prediction value of the degradation data, and then learn the historical trend of MOSFET device degradation; Step 3: Based on the trained SCINet network, iteratively predict the test set data until the prediction exceeds the failure threshold, and obtain the predicted observation results of the long-term degradation trend; Specifically, the input sequence length of the SCINet network is defined as p, and the degraded data z0, z1, ..., z up to time p-1 are input. p-1 Enter the SCINet network and observe the value z p is predicted, i.e. Then form a new input Next, at time p+1, the observed value of the next time step is predicted Step 4: Based on the nonlinear fractal Brownian motion FBM, a nonlinear random degradation model is constructed; based on the measurement error, a measurement equation is constructed; based on the measurement equation and the nonlinear degradation model, a state space model considering the measurement error is constructed; 4.1) The nonlinear random degradation model is expressed as follows: Y(t)=Y(0)+∫0 t g(s;ρ)ds+σB H (t) (3) Where, Y(0) represents the initial degradation state of the MOSFET device. In order to maintain generality, it is assumed that Y(0) = 0; Y(t) represents the degradation state at time t; ∫0 t g(s;ρ)ds represents a nonlinear drift function that depends on the integration time s and the unknown parameter vector ρ, σ represents the diffusion coefficient, which reflects the degree of fluctuation of the degradation process; B H (t) represents the standard FBM, which is used to characterize the time-varying uncertainty of degraded devices; H is the Hurst exponent; 4.2) Based on formula (3), the measurement equation shown in formula (4) is introduced to solve the unknown measurement noise; therefore, at any time t, the MOSFET device degradation amount Z(t) directly monitored by the power cycle test aging test is expressed as the following state space degradation model: where ε represents the measurement error, and is the variance of the measurement error; the nonlinear drift function is defined as the power law function λt b , where λ and b are the parameters to be estimated in the power law function; Step 5: Based on the degradation data of the test set, the Hurst exponent is estimated by combining the rescaling range (R / S) analysis method; and then the state-parameter joint estimation strategy is implemented; In the implementation of the state-parameter joint estimation strategy, the Kalman filter KF is used to estimate the hidden degradation state y and the unknown parameter set Ψ = {b, λ, σ, σ z }, the expectation maximization EM algorithm is used to estimate the relevant parameters, and the maximum likelihood estimation MLE is used to correct the parameter estimation results; in the implementation process, after the actual observation value is obtained at the latest moment, the predicted degradation value of the SCINet network is used as the new observation value and substituted into the state space degradation model; the specific steps are as follows: 5.1) First, the KF algorithm is used to recursively estimate the hidden degradation state; at the nth monitoring moment, let Ψ n = {b n ,λ n , σ n , σ z，n } represents the parameter set to be estimated; based on Ψ n , the time update equation and the measurement update equation are expressed as formula (5) and (6) respectively: Among them, y n|n and n+1|n Respectively represent the predicted value of the hidden state based on the information of step n at step n and step n+1; C n|n and C n+1|n Respectively represent the covariance prediction value of the hidden state estimation based on the n-step information at the n-step and n+1-step; h(t n ) and h(t n+1 ) are respectively n and t n+1 The covariance function value of the process noise under time; Among them, y n+1|n+1 represents the updated estimated value of the hidden state based on the n+1 step measurement information at step n+1; C n+1|n+1 represents the updated covariance of the hidden state estimate based on the n+1 step measurement information at step n+1; z n+1 represents the actual measurement value at step n+1; represents the measurement noise variance at step n; 5.2) In order to cooperate with KF, the EM algorithm is used to n The parameters in are estimated; finally, the estimated values ​​of the parameters are updated in the d+1th iteration Respectively expressed as: in, represents the hidden state of the jth time step estimated based on the observed data in the dth iteration; yes The square of z j represents the actual measurement value at the jth time step; Among them, h(t j-1 ) and h(t j ) respectively represent the j-1 and t j The covariance function value of the moment process noise; and They represent the hidden states of the j-1th and jth time steps estimated based on the observed data in the dth iteration respectively; represents the estimated value of b at the nth step in the d+1th iteration; and middle, is the power index; 5.3) MLE directly uses the state estimate obtained by KF as the observation value; at this time, the degradation state, diffusion term, and drift term are expressed as: Y = [y 1|1 …y M|M ]′, B H =[B H (t1)…B H (t M )]′、 Wherein, the superscript "′" indicates the transpose operation of the matrix; Then Y~N(λg,σ 2 D), where D represents B H The covariance matrix of ; the log-likelihood function of the relevant parameters F(b, λ, σ|Y) is expressed as: Where F is the total time step; the estimate of b is independent of the FBM and no longer requires an update process; therefore, only λ and σ in the log-likelihood function need to be updated; the estimated values ​​of λ and σ are further updated as: Among them, g T is the transpose of g, D -1 is the inverse matrix of D; Step 6: Based on the concept of first arrival time, define the remaining useful life RUL at any time and define the corresponding reliability function; based on the parameter estimation results of step 5 and the statistical characteristics of the FBM diffusion term increment of step 4, use a recursive strategy to approximately perform the extrapolation process; obtain the final RUL and the corresponding probability density distribution PDF based on the Monte Carlo method; 6.1) According to the concept of first arrival time, when the failure threshold is ω, the degraded MOSFET device will n The RUL at a monitoring moment is defined as: L n =inf{r n :y(t n +r n )≥ω|y(t n )≤ω} (13) Among them, L n Remaining useful life, r n Indicates that from the current time t n The time difference to a certain point in the future; the symbol inf represents the operation of taking the maximum lower bound; y(t n ) and y(t n +r n ) represent t n With t n +r n The status of the system at the moment; The corresponding reliability function is defined as: Among them, R n (r n ) is the reliability function; where P represents probability, represents the estimated values ​​of the model parameters, represents an estimate of the system state; 6.2) Monte Carlo method is used to simulate the PDF of RUL; Based on the parameter estimation results of step 5 and the statistical characteristics of the diffusion term increment Y(t) in formula (4), a recursive scheme is used to approximately perform the extrapolation process; specifically, the final simulation sample is generated in the following way: Where i is the index of the time step, τ is the sampling interval of the extrapolation, and v n is a Gaussian random variable, where and Represents the final parameter estimate; by calculating formula (14), the reliability function R of each extrapolation time is directly obtained n (iτ); after that, the PDF of RUL at the nth monitoring time is Get, where f n (r n ) is the PDF of RUL at the nth monitoring moment, Is a positive integer.

Images