IGBT service life prediction method of adaptive unscented particle filtering based on time sequence form perception optimization

Abstract

The invention discloses an IGBT life prediction method based on adaptive unscented particle filtering of time sequence form perception optimization, and belongs to the technical field of health prediction and reliability engineering. The method aims at the problems that a traditional filtering algorithm is prone to falling into local optimum in a high-dimensional parameter space and prediction trajectory forms are discontinuous, and comprises the steps that firstly, an IGBT nonlinear degradation model is established through a state-space equation; performing Sigma point sampling and nonlinear propagation at a particle level by adopting an unscented transformation and particle filtering cooperation mechanism, and fusing process noise to realize dynamic state prediction; constructing a Gaussian likelihood function by using the deviation between a predicted value and an actual observed value to generate a system residual sequence; an FCD fusion index is innovatively introduced, the index fuses a Pearson correlation coefficient and a Frechet distance to synchronously quantify trend consistency and form similarity, and self-adaptive genetic optimization is driven accordingly; and repeating the behaviors before a preset termination condition is met, and continuously updating the parameters of the state-space equation.

Description

Technical Field

[0001] This invention belongs to the field of health prediction and reliability engineering technology, and specifically relates to an IGBT lifetime prediction method based on time series morphology-aware optimization and adaptive unscented particle filtering. Background Technology

[0002] The Insulated Gate Bipolar Transistor (IGBT), a fully controllable voltage-driven power semiconductor device, is one of the most critical components in modern power electronics technology. It boasts advantages such as high input impedance, high switching frequency, large conduction current, and low conduction loss, playing a vital role in numerous fields including industry, transportation, and energy. Failures of power conversion devices account for the vast majority of all power electronic system failures, with IGBT failures accounting for 42% of these. Therefore, improving the stability of electrical equipment centered around IGBTs is of practical significance for reducing maintenance costs and minimizing the probability of system failure.

[0003] Currently, in the research on IGBT lifetime prediction based on filtering algorithms, most schemes have two inherent limitations: First, the state exploration mechanism of existing methods is difficult to effectively cope with the global optimization requirements of high-dimensional parameter spaces, causing the algorithm to converge to a local optimum region too early during the iteration process, and failing to fully capture the overall evolution trend of complex degradation processes; Second, the traditional architecture completely ignores the geometric continuity between the historical degradation trajectory and the prediction path, and only updates the particle state based on the current instantaneous observation value, which leads to discontinuous and abrupt evolution of the system's state estimation, weakening the ability to predict future states. Summary of the Invention

[0004] The purpose of this invention is to overcome the problem that existing methods cannot perform global optimization operations, resulting in an inability to fully capture the overall evolution trend of complex degradation processes and thus insufficient prediction capabilities. This invention provides an adaptive unscented particle filter-based IGBT lifetime prediction method based on time series morphology-aware optimization.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: This invention provides an IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering, comprising the following steps: A state-space model is constructed based on the nonlinear characteristics of IGBT health degradation. Based on the constructed state-space model, dynamic state prediction is performed through a collaborative mechanism of unscented transformation and particle filtering to generate Sigma points. Nonlinear propagation is then performed on the Sigma points, and process noise is fused to reconstruct a local Gaussian distribution to obtain new particles. A Gaussian likelihood function is constructed based on the deviation between the predicted state value and the actual observed value of the new particle. The particle weight is corrected by the Gaussian likelihood function, and the particle weight is normalized to establish a complete distribution representation of the posterior probability of the system. The FCD index is calculated based on the observed sequence and the actual observed sequence of the particles after weight correction. The FCD index is used to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability. The state space model is dynamically updated based on an adaptive control mechanism. When the preset termination condition is reached, the estimated value of the state parameters is output and the predicted value of the remaining service life is calculated.

[0006] In the step of constructing a state-space model based on the nonlinear characteristics of IGBT health degradation, the constructed state-space model includes state transition equations and observation equations, specifically expressed as follows:

[0007] Where k is the number of IGBT operating cycles, Let k be the state vector of the parameters of the degenerate model with period k. Let h be the observed value, and h be the observation function. and These are the noise matrices for the state transition equation and the measurement equation, respectively.

[0008] The specific method for obtaining new particles in the steps of dynamically predicting the state based on the constructed state-space model, generating Sigma points through a collaborative mechanism of unscented transformation and particle filtering, performing nonlinear propagation on the Sigma points, fusing process noise, and reconstructing a local Gaussian distribution is as follows: Generate Sigma points and perform deterministic sampling for each particle; Propagation based on the state transition equation is performed on each Sigma point to obtain the state prediction estimate of the i-th particle at time step k; Based on the weighting coefficients of each Sigma point, the noise of the fusion process is fused to reconstruct a local Gaussian distribution, and the state prediction mean and covariance of each particle are calculated. New particles are obtained by importance sampling from the reconstructed local Gaussian distribution.

[0009] The process involves reconstructing a local Gaussian distribution based on the noise from the fusion process using the weighting coefficients of each Sigma point, and calculating the mean state prediction value for each particle. With covariance The specific formula for the steps is as follows:

[0010] The weighting coefficients are:

[0011] in, This is the predicted state estimate of the j-th Sigma point at time step k; The mean weight of the j-th Sigma point; The covariance weight of the j-th Sigma point; This is process noise; The dimension of the state vector; This is the scaling factor.

[0012] In the steps of constructing a Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particle, correcting the particle weights through the Gaussian likelihood function, and normalizing the particle weights to establish a complete distribution representation of the system's posterior probability, the specific formula for constructing the Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particle, and correcting the particle weights through the Gaussian likelihood function, is as follows:

[0013]

[0014] in, This represents the constructed Gaussian likelihood function; This refers to the deviation between the predicted value and the actual observed value. The normalized particle weights; R represents the original weight of the i-th particle at time step k, and R is the covariance of the observation noise.

[0015] Normalizing the particle weights establishes a complete distribution representation of the system's posterior probability, as expressed by the following formula:

[0016] in, Given the observation sequence from the initial time to time k, The total number of particles; For Dirac functions.

[0017] In the step of calculating the FCD index based on the observed sequence and the actual observed sequence of particles after weight correction, and using the FCD index to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability, the adaptive control mechanism is as follows: The FCD index is calculated using the predicted sequence of each particle and the actual historical observation sequence to measure historical consistency. The FCD index is used to drive adaptive genetic optimization, and a weighted FCD roulette wheel selection mechanism is used to determine the parent particles. Based on the characteristics of the parent particles, directional crossover and intelligent mutation operations with enhanced FCD probability are performed on the same generation particles to obtain an optimized particle population. When the adaptive genetic optimization process stalls, a perturbation operation is performed on the current particle population to increase population diversity by replacing non-elite particles in the particle population and applying strong mutations. By integrating the evolutionary trajectories of historical elite particles, new particles with potential optimization capabilities can be created.

[0018] The formula for calculating the FCD indicator is as follows:

[0019] in, The Pearson correlation coefficient measures the consistency of trends. The Fréchet distance is used to measure morphological consistency. This is the scaling factor.

[0020] The specific formula for determining parent particles using the FCD index-driven adaptive genetic optimization and the weighted FCD roulette wheel selection mechanism is expressed as follows:

[0021] in, Let be the weight value of the i-th particle. For the weighted fitness including FCD, The error between the predicted value and the actual observed value for each particle. Let be the FCD value of the i-th particle.

[0022] Using FCD index-driven adaptive genetic optimization, based on the characteristics of parent particles, FCD-enhanced probability-directed crossover and intelligent mutation operations are performed on the same generation particles to obtain the optimized particle population. The specific formula for performing FCD is as follows.

[0023]

[0024] in, For adaptive crossover probability, and This is the crossover probability adjustment parameter. For adaptive mutation probability, and This is a parameter for adjusting the mutation probability.

[0025] The specific methods for increasing population diversity by replacing non-elite particles in the particle swarm and applying enhanced mutations are as follows:

[0026] in, For strongly mutated particles, when the initial particles When it is an elite particle, the scaling factor Take 0.3; when the initial particle When it is an ordinary particle, Take 1; Taken from the standard normal distribution, The standard deviation of the process noise; By performing interpolation operations and fusing the evolutionary trajectories of historical elite particles, new particles with potential optimization capabilities are created. The specific methods are as follows:

[0027] in, For historical elite particles, For the current elite particles, This is a random disturbance term.

[0028] Compared with the prior art, the present invention has the following beneficial effects: This invention provides an IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering, comprising the following steps: constructing a state-space model based on the nonlinear characteristics of IGBT health degradation; based on the constructed state-space model, performing dynamic state prediction through a collaborative mechanism of unscented transformation and particle filtering to generate Sigma points, performing nonlinear propagation on the Sigma points, fusing process noise to reconstruct a local Gaussian distribution, and obtaining new particles; constructing a Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particles, correcting the particle weights through the Gaussian likelihood function, and normalizing the particle weights to establish a complete distribution representation of the system's posterior probability; calculating the FCD index based on the observed sequence and the actual observed sequence of the weighted particles, using the FCD index to fuse sequence morphological similarity and trend correlation, and constructing an adaptive control mechanism for crossover and mutation probabilities; dynamically updating the state-space model based on the adaptive control mechanism, and when a preset termination condition is reached, outputting estimated state parameters and calculating the predicted remaining lifetime value. This invention, based on unscented transformation and adaptive genetic optimization, guides the particle swarm to explore better regions, effectively overcoming the local optimum trap and significantly improving the algorithm's global search capability in complex high-dimensional spaces. Through the synergistic mechanism of unscented transformation and particle filtering, combined with time-series morphology-aware optimization, it significantly improves the accuracy and adaptability of IGBT lifetime prediction. This method effectively captures the nonlinear characteristics of the IGBT degradation process using dynamic Sigma point generation and nonlinear propagation. Through an adaptive control mechanism constructed by correcting particle weights using the Gaussian likelihood function and the FCD index, it achieves accurate characterization of the system's posterior probability distribution and dynamic optimization of model parameters, overcoming the shortcomings of traditional methods in tracking abrupt changes.

[0029] Furthermore, this invention constructs the FCD index based on Fréchet distance and Pearson correlation coefficient, taking into account the sequence morphological similarity and trend correlation between actual and filtered measurements, thereby improving the reliability and continuity of long-term prediction.

[0030] Furthermore, this invention improves the genetic algorithm based on perturbation and interpolation operators, which can better solve the problem of effective local optimization stagnation and maintain long-term trends, thereby improving the global convergence efficiency and long-term prediction stability of the algorithm. Attached Figure Description

[0031] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is the IGBT health status graph of Embodiment 2 of the present invention; Figure 3 This is a trend tracking chart of IGBT health status in Embodiment 2 of the present invention; Figure 4This is a graph showing the predicted remaining lifetime of the IGBT in Embodiment 2 of the present invention; Figure 5 This is a graph showing the prediction error of the remaining life of the IGBT in Embodiment 2 of the present invention. Detailed Implementation

[0032] To further understand the content of this invention, the invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments are merely illustrative and not limiting of the invention.

[0033] Example 1 An adaptive unscented particle filter-based IGBT lifetime prediction method based on time-series morphology-aware optimization includes the following steps: S1: Constructing a state-space model based on the nonlinear characteristics of IGBT health degradation; S2: Based on the constructed state-space model, dynamic state prediction is performed through the collaborative mechanism of unscented transformation and particle filtering to generate Sigma points. Nonlinear propagation is then performed on the Sigma points, and process noise is fused to reconstruct the local Gaussian distribution to obtain new particles. S3: Construct a Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particle, correct the particle weight through the Gaussian likelihood function, normalize the particle weight, and establish a complete distribution representation of the system's posterior probability. S4: Calculate the FCD index based on the observed sequence and the actual observed sequence of the particles after weight correction. Use the FCD index to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability. S5: The state space model is dynamically updated based on the adaptive control mechanism. When the preset termination condition is reached, the estimated value of the state parameters is output and the predicted value of the remaining service life is calculated.

[0034] Specifically, in S1, based on the degradation mechanism of IGBTs, a nonlinear degradation state-space model is constructed using appropriate degradation characteristic parameters to describe the evolution of its health state. This model consists of two closely related parts: the state transition equation and the observation equation. The state transition equation describes the dynamic evolution of the "hidden" state inside the system, while the observation equation maps the internal state to actual measurable information.

[0035] The constructed state-space model is as follows:

[0036] Where k is the number of IGBT operating cycles, Let k be the state vector of the parameters of the degenerate model with period k. Let h be the observed value, and h be the observation function. and These are the noise matrices for the state transition equation and the measurement equation, respectively.

[0037] Specifically, in S2, based on the constructed state-space equation, dynamic state prediction is performed through a collaborative mechanism of unscented transformation and particle filtering to generate Sigma points. Nonlinear propagation is then performed on the Sigma points, and process noise is fused to reconstruct a local Gaussian distribution, thereby obtaining new particles.

[0038] Generate Sigma points for each particle. Perform deterministic sampling:

[0039] in, For the first Each particle time step The state vector is used as the center point of sampling, and the subsequently generated sampling points spread symmetrically around it. The scaling factor for diffusion. , , Controlling the range of Sigma point distribution, For second-order scaling parameters, These are the parameters for parameter tuning. The covariance matrix is ​​decomposed using Cholesky decomposition, i.e. ; For the first i Each particle at time step The covariance matrix, Let j be the j-th column vector of the square root of the covariance matrix, used to traverse the state space. j Disturbances are generated along each principal axis direction to ensure that the diffusion direction is aligned with the direction of uncertainty. The dimension of the state vector; Centered on, representing undisturbed particles; This represents the sampling point along the positive direction of the j-th principal axis; Let be the sampling point along the negative direction of the j-th principal axis.

[0040] Propagation based on the state transition equation is performed at each Sigma point to obtain the state prediction estimate of the i-th particle at time step k. : Based on the weighting coefficients of each Sigma point, a local Gaussian distribution is reconstructed from the fusion process noise, and the mean state prediction value of each particle is calculated. With covariance :

[0041] The weighting coefficients are:

[0042] in, This is the predicted state estimate of the j-th Sigma point at time step k; The mean weight of the j-th Sigma point; The covariance weight of the j-th Sigma point; For process noise, The dimension of the state vector; This is the scaling factor.

[0043] Importance sampling is performed to extract new particles from the reconstructed local Gaussian distribution.

[0044] Specifically, in S3, a Gaussian likelihood function is constructed based on the deviation between the predicted state value and the actual observed value of the new particle. The particle weight is corrected by the Gaussian likelihood function, and the particle weight is normalized to establish a complete distribution representation of the system's posterior probability.

[0045] Utilizing the deviation between predicted and actual observed values Constructing the Gaussian likelihood function And perform weight normalization. ;

[0046]

[0047] in, R represents the original weight of the i-th particle at time step k, and R is the covariance of the observation noise.

[0048] By using the Gaussian likelihood function to perform short-term local corrections on particle weights, the problem of real-time matching between observed data and predicted values ​​can be solved.

[0049] A complete posterior probability distribution representation is constructed based on the normalized weight values.

[0050]

[0051] in, Given the observation sequence from the initial time to time k, The total number of particles; For Dirac functions.

[0052] Specifically, in S4, the FCD index is calculated based on the observed sequence of particles after weight correction and the real historical observation sequence. The FCD index is used to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability, guiding the particle swarm to evolve towards historical consistency.

[0053] The adaptive control mechanism is as follows: S41: Calculate the FCD index using the predicted sequence of each particle and the actual historical observation sequence to measure historical consistency;

[0054] in, The Pearson correlation coefficient measures the consistency of trends. The Fréchet distance is used to measure morphological consistency. This is the scaling factor.

[0055] S42: Uses FCD index to drive adaptive genetic optimization, and executes a weighted FCD roulette wheel selection mechanism to determine the parent particles;

[0056] in, Let be the weight value of the i-th particle. For the weighted fitness including FCD, The error between the predicted value and the actual observed value for each particle. Let be the FCD value of the i-th particle.

[0057] S43: Using FCD index-driven adaptive genetic optimization, based on the characteristics of parent particles, FCD-enhanced probability-directed crossover and intelligent mutation operations are performed on the same generation particles to obtain an optimized particle population.

[0058]

[0059] in, For adaptive crossover probability, and This is the crossover probability adjustment parameter. For adaptive mutation probability, and This is a parameter for adjusting the mutation probability.

[0060] When the algorithm stagnates (i.e. there is no significant improvement for several consecutive generations), a perturbation operation is performed on the current particle population: by replacing non-elite particles in the particle population and applying strong mutations, the population diversity is increased, thereby escaping the local optimum;

[0061] in, For strongly mutated particles, when the initial particles When it is an elite particle, the scaling factor Take 0.3; when the initial particle When it is an ordinary particle, Take 1. Taken from the standard normal distribution, This represents the standard deviation of the process noise.

[0062] Furthermore, in order to maintain the degradation trend in long-term predictions, an interpolation operation is performed: the evolutionary trajectories of historical elite particles are merged to create new particles with potential optimization capabilities. ;

[0063] in, For historical elite particles, For the current elite particles, This is a random disturbance term.

[0064] Specifically, in S5, S2~S4 are iteratively executed, and the state-space model is dynamically adjusted using newly arrived observation data to achieve online tracking and adaptive updating of the actual degradation trajectory of the IGBT. When the loop process meets the preset termination condition, the state observation value of the model parameters at the prediction starting point is output. Based on the state observation value at this point, the future failure endpoint with a health level below 0.1 is predicted. The remaining lifetime is the difference between the failure endpoint and the prediction starting point.

[0065] Example 2 This embodiment verifies the effectiveness of the invention by combining IGBT accelerated life test data. The IGBT under test was subjected to accelerated degradation experiments on an accelerated degradation test platform. During the IGBT degradation process, key characteristic signals were collected using voltage and current sensors at a sampling frequency of 50Hz. The health curve obtained after fusing various degradation characteristic parameters is shown below. Figure 2 As shown.

[0066] To verify the effectiveness of this invention, a comparison was made between the traditional particle filtering method and the unscented particle filtering method. The specific results are as follows: IGBT health degradation trend tracking based on different filtering algorithms, such as Figure 3 As shown, compared to the other two methods, the time-series morphology-aware optimized adaptive unscented particle filter method proposed in this study exhibits significant advantages in IGBT degradation tracking. Figure 4 The comparison of the bearing remaining life prediction results shows that the prediction curve of this method accurately matches the actual degradation trajectory, and the prediction variance is small, indicating stronger robustness. Figure 5 The comparison of prediction errors further confirms that the absolute average error of this method is only 1.593 (the absolute average error of the traditional particle filter method is 7.831, and the absolute average error of the unscented particle filter method is 6.169).

[0067] In summary, the adaptive unscented particle filtering method for time-series morphology-aware optimization provided by this invention, compared with other methods, overcomes the challenge of modeling IGBT's strong nonlinear degradation by constructing a state-space equation including a double exponential decay model and integrating a collaborative prediction mechanism of unscented transformation and particle filtering. It innovatively introduces an FCD dynamic weight fusion architecture, simultaneously considering Fréchet distance and correlation coefficient in genetic optimization, achieving two-dimensional optimization of global parameter search and historical trajectory morphology constraints. This improves the accuracy of remaining lifetime prediction and provides a standardized prediction solution for IGBT and other strongly nonlinear degradation systems.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A method for predicting IGBT lifetime based on adaptive unscented particle filtering with time-series morphology-aware optimization, characterized in that, Includes the following steps: A state-space model is constructed based on the nonlinear characteristics of IGBT health degradation. Based on the constructed state-space model, dynamic state prediction is performed through a collaborative mechanism of unscented transformation and particle filtering to generate Sigma points. Nonlinear propagation is then performed on the Sigma points, and process noise is fused to reconstruct a local Gaussian distribution to obtain new particles. A Gaussian likelihood function is constructed based on the deviation between the predicted state value and the actual observed value of the new particle. The particle weight is corrected by the Gaussian likelihood function, and the particle weight is normalized to establish a complete distribution representation of the posterior probability of the system. The FCD index is calculated based on the observed sequence and the actual observed sequence of the particles after weight correction. The FCD index is used to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability. The state space model is dynamically updated based on an adaptive control mechanism. When the preset termination condition is reached, the estimated value of the state parameters is output and the predicted value of the remaining service life is calculated.

2. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 1, characterized in that, In the step of constructing a state-space model based on the nonlinear characteristics of IGBT health degradation, the constructed state-space model includes state transition equations and observation equations, specifically expressed as follows: Where k is the number of IGBT operating cycles, Let k be the state vector of the parameters of the degenerate model with period k. Let h be the observed value, and h be the observation function. and These are the noise matrices for the state transition equation and the measurement equation, respectively.

3. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 2, characterized in that, The specific method for obtaining new particles in the steps of dynamically predicting the state based on the constructed state-space model, generating Sigma points through a collaborative mechanism of unscented transformation and particle filtering, performing nonlinear propagation on the Sigma points, fusing process noise, and reconstructing a local Gaussian distribution is as follows: Generate Sigma points and perform deterministic sampling for each particle; Propagation based on the state transition equation is performed on each Sigma point to obtain the state prediction estimate of the i-th particle at time step k; Based on the weighting coefficients of each Sigma point, the noise of the fusion process is fused to reconstruct a local Gaussian distribution, and the state prediction mean and covariance of each particle are calculated. New particles are obtained by importance sampling from the reconstructed local Gaussian distribution.

4. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 3, characterized in that, The process involves reconstructing a local Gaussian distribution based on the noise from the fusion process using the weighting coefficients of each Sigma point, and calculating the mean state prediction value for each particle. With covariance The specific formula for the steps is as follows: The weighting coefficients are: in, This is the predicted state estimate of the j-th Sigma point at time step k; The mean weight of the j-th Sigma point; The covariance weight of the j-th Sigma point; This is process noise; The dimension of the state vector; This is the scaling factor.

5. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 3, characterized in that, In the steps of constructing a Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particle, correcting the particle weights through the Gaussian likelihood function, and normalizing the particle weights to establish a complete distribution representation of the system's posterior probability, the specific formula for constructing the Gaussian likelihood function based on the deviation between the predicted state value and the actual observed value of the new particle, and correcting the particle weights through the Gaussian likelihood function, is as follows: in, This represents the constructed Gaussian likelihood function; The deviation between the predicted value and the actual observed value; The normalized particle weights; R represents the original weight of the i-th particle at time step k, and R is the covariance of the observation noise. Normalizing the particle weights establishes a complete distribution representation of the system's posterior probability, as expressed by the following formula: in, Given the observation sequence from the initial time to time k, The total number of particles; For Dirac functions.

6. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 5, characterized in that, In the step of calculating the FCD index based on the observed sequence and the actual observed sequence of particles after weight correction, and using the FCD index to fuse sequence morphological similarity and trend correlation to construct an adaptive control mechanism for crossover mutation probability, the adaptive control mechanism is as follows: The FCD index is calculated using the predicted sequence and the actual observed sequence for each particle to measure historical consistency. The FCD index is used to drive adaptive genetic optimization, and a weighted FCD roulette wheel selection mechanism is used to determine the parent particles. Based on the characteristics of the parent particles, directional crossover and intelligent mutation operations with enhanced FCD probability are performed on the same generation particles to obtain an optimized particle population. When the adaptive genetic optimization process stalls, a perturbation operation is performed on the optimized particle population to increase population diversity by replacing non-elite particles in the particle population and applying strong mutations. By integrating the evolutionary trajectories of historical elite particles, new particles with potential optimization capabilities can be created.

7. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 6, characterized in that, The formula for calculating the FCD indicator is as follows: in, The Pearson correlation coefficient measures the consistency of trends. The Fréchet distance is used to measure morphological consistency. This is the scaling factor.

8. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 7, characterized in that, The specific formula for determining parent particles using the FCD index-driven adaptive genetic optimization and the weighted FCD roulette wheel selection mechanism is expressed as follows: in, Let be the weight value of the i-th particle. For the weighted fitness including FCD, The error between the predicted value and the actual observed value for each particle. Let be the FCD value of the i-th particle.

9. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 8, characterized in that, Based on the characteristics of the parent particles, the specific formula for performing FCD-enhanced directional crossover and intelligent mutation operations on the same generation particles to obtain an optimized particle population is as follows. in, For adaptive crossover probability, and This is the crossover probability adjustment parameter. For adaptive mutation probability, and This is a parameter for adjusting the mutation probability.

10. The IGBT lifetime prediction method based on time-series morphology-aware optimization and adaptive unscented particle filtering according to claim 9, characterized in that, The specific method for increasing population diversity by replacing non-elite particles in the optimized particle swarm and applying enhanced mutations is as follows: in, For strongly mutated particles, when the initial particles When it is an elite particle, the scaling factor Take 0.3; when the initial particle When it is an ordinary particle, Take 1; Taken from the standard normal distribution, The standard deviation of the process noise; By performing interpolation operations and fusing the evolutionary trajectories of historical elite particles, new particles with potential optimization capabilities are created. The specific methods are as follows: in, For historical elite particles, For the current elite particles, This is a random disturbance term.

Images