IGBT service life prediction method based on improved trust region reflection nonlinear least square method and distributed parameter degradation model

Abstract

The invention discloses an IGBT service life prediction method based on an improved trust region reflection nonlinear least square method and a distributed parameter degradation model, and belongs to the technical field of IGBT device service life prediction methods. Parameters of the distributed parameter life model are estimated in combination with an improved TRRLS algorithm, accurate life prediction is realized while parameter dynamic drift characteristics under IGBT'electricity-heat-force 'multi-field coupling are considered, the influence of noise interference and invalid parameters on a prediction result is reduced, the depiction capability of the distributed parameter model on local degradation can be brought into play, and the reliability of life prediction is improved. And the accuracy and the reliability of IGBT service life prediction are improved.

Description

Technical Field

[0001] This invention relates to the technical field of IGBT device lifetime prediction methods, specifically an IGBT lifetime prediction method based on an improved trust region reflection nonlinear least squares method and a distributed parameter degradation model. Background Technology

[0002] As the core power device of power electronic systems, IGBTs (Insulated Gate Bipolar Transistors) are widely used in key fields such as new energy power generation, rail transit, and industrial frequency conversion. Their service life directly determines the reliability and operation and maintenance costs of the system. Therefore, IGBT life prediction technology has become a core research direction for ensuring stable system operation and realizing predictive maintenance.

[0003] Current mainstream IGBT lifetime prediction methods not only face the problem of conduction voltage drop data being easily interfered with by noise, but also suffer from the inherent limitations of "fixed-parameter modeling." Existing methods generally treat degradation model parameters as fixed values, using lumped-parameter models to construct the degradation relationship of conduction voltage drop over operating time. This modeling approach ignores the individual differences in IGBT devices during manufacturing and operation, resulting in the model's inability to characterize the lifetime dispersion characteristics of the same model of IGBT devices. In actual engineering, the failure lifetime of the same batch of IGBTs under the same operating conditions often exhibits significant range fluctuations, while "fixed-parameter models" can only output a single lifetime prediction value, failing to reflect the probability distribution of lifetime and making it difficult to support risk-level-based operation and maintenance decisions in engineering, further exacerbating the deviation between lifetime prediction results and actual operating conditions.

[0004] To address the issue of noise in degradation feature data, some studies have attempted to use data filtering techniques to optimize data quality. However, traditional filtering methods struggle to dynamically track the nonlinear changes in the IGBT degradation process and cannot effectively separate degradation trends from noise interference.

[0005] Furthermore, in the lifetime model parameter estimation stage, while the commonly used Levenberg-Marquardt (LM) algorithm has certain advantages in solving nonlinear least squares problems, it suffers from inherent defects such as being prone to getting trapped in local optima and being sensitive to initial parameter values. It also cannot impose physical constraints on model parameters, potentially leading to invalid parameters that do not conform to the actual degradation patterns of IGBTs, severely impacting the physical rationality of the model. More importantly, this algorithm is designed for "fixed-parameter models" and cannot adapt to the statistical characteristics of distributed parameters. If forcibly applied to distributed parameter estimation, it is difficult to fit the probability distribution of parameters and quantify the impact of parameter fluctuations on lifetime dispersion. Therefore, we propose an IGBT lifetime prediction method based on an improved trust-region reflection nonlinear least squares method and a distributed parameter degradation model to alleviate or solve the aforementioned problems.

[0006] The information disclosed above in this background section is only for enhancing the understanding of the background section of this invention, and therefore may include prior art that is not known to those skilled in the art. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides an IGBT lifetime prediction method based on an improved trust region reflection nonlinear least squares method and a distributed parameter degradation model. This method aims to solve one of the problems in existing technologies, such as the difficulty in effectively suppressing noise interference from conduction voltage drop data, the inability of fixed parameter modeling to characterize the dispersion of IGBT lifetime, and the tendency of LM algorithm parameter estimation to be inaccurate and unsuitable for distributed parameters.

[0008] To achieve the above objectives, this invention provides an IGBT lifetime prediction method based on an improved trust region reflection nonlinear least squares method and a distributed parameter degradation model, specifically including the following steps:

[0009] Step 1: Based on the characteristic parameters of the statistical law of device dispersion, establish the degradation model of IGBT device and treat the model parameters as normal random variables that depend on the operating condition characteristics;

[0010] Step 2: Standardize the experimental data and perform Kalman filtering data preprocessing to suppress noise interference.

[0011] Step 3: Use the improved trust region reflection nonlinear least squares method to estimate the parameters of the degraded model after the preprocessing in Step 2, and ensure the rationality of the parameters through physical constraints and iterative optimization;

[0012] Step 4: Predict the lifetime distribution of IGBT devices using the Monte Carlo method and output the lifetime probability distribution law.

[0013] Furthermore, step 1 specifically includes:

[0014] (a) Establishing a degradation model for IGBT devices: The degradation process of IGBT aging damage is described using a monotonic function related to operating time, expressed as: ,in, This is the initial value of the on-state voltage drop. and To characterize the device degradation process, model parameters, Runtime;

[0015] (b) Parameters and Treated as a normally distributed random variable related to operating conditions: based on Bayesian statistical theory, and It follows a normal distribution, and its expression is: Where Θ represents the current operating condition parameter of the IGBT. , They are respectively The mean and variance of the normal distribution. , They are respectively The mean and variance of the normal distribution.

[0016] Furthermore, in step (b), the calculation and The expressions for the mean and variance are: ,in, Corresponding parameters and , This represents the total amount of data obtained in the experiment. Indicates the current number Group data.

[0017] Furthermore, the standardization process in step 2 involves standardizing the acquired IGBT on-state voltage drop to eliminate initial differences, resulting in a standardized on-state voltage drop. The expression is: ,in, The measured value of the on-state voltage drop at any given time. This is the initial value of the on-state voltage drop.

[0018] Furthermore, the Kalman filter data preprocessing in step 2 specifically includes:

[0019] (a) Initialize Kalman filter parameters: including process noise covariance Measurement of noise covariance State transition matrix Observation matrix and initial covariance ;

[0020] (b) Data filtering: Save the data of the first time the IGBT degradation state crosses the failure threshold, and do not perform Kalman filtering on this part of the data;

[0021] (c) Define system state: Select the standardized on-state voltage drop As a state variable, System state at time 1 for ,in for Time-standardized on-state voltage drop;

[0022] (d) Establish the state and observation equations for the linear system: ,in, For process noise, This is the measured value of the on-state voltage drop. For measuring noise;

[0023] (e) Perform Kalman filter iterative calculations until all data points have been processed.

[0024] Furthermore, the iterative calculation in step (e) specifically includes:

[0025] (e1), take the initial state variables. , As the initial optimal estimate The on-state voltage drop is the normalized value at the initial moment.

[0026] (e2), State prediction: ,in, This is the predicted state value at the current moment. This is the optimal estimate from the previous moment;

[0027] (e3), Calculation of prior error covariance: ,in, for The posterior error covariance matrix at time t. The prior error covariance matrix at the current time;

[0028] (e4), Kalman gain calculation: ;

[0029] (e5), Status Update: To obtain the optimal estimate at the current time. ;

[0030] (e6), posterior error covariance update: ,in, It is an identity matrix.

[0031] Furthermore, step 3 specifically includes:

[0032] (a) Initialize and improve TRRLS parameters: Set degradation model parameters according to the physical laws of IGBT aging. and upper limit Lower limit ;

[0033] (b) Define the basic optimization problem: ,in, It is a nonlinear function. The coefficients to be determined are: For input data, The output data is N, where N is the number of measurements.

[0034] (c) Transformation Objective Function: Combining the standardized on-state voltage drop degradation model, the objective function is: ,in, Let j be the number of iterations at the j-th measurement time. The normalized on-state voltage drop after filtering at time j;

[0035] (d) Solve the trust region subproblem to obtain the optimal search step size. The trust region subproblem is: ,in, For the first The search step size for the next iteration. For the objective function exist The Hessian matrix at that location, It is a diagonal scale matrix. For the first Trust region scale of the next iteration The gradient of the objective function at that point. This indicates the number of iterations. The parameters are then updated iteratively until the convergence condition is met, at which point the optimal parameters are output. and .

[0036] Furthermore, in step (d), the iterative parameter update includes: calculating candidate iteration points based on the optimal search step size. , The step size is used as the step size coefficient. When a candidate point exceeds the parameter range, the trust region scale is adjusted based on the gradient direction. If the gradient points into the feasible region, the scale is increased; if it points out of the region, the scale is decreased. Then, a one-sided reflection is performed on the out-of-bounds dimension to adjust it back to the trust region and the constraint range. The objective function value of the reflection point is calculated. If it meets the objective, the original iteration point is updated to the reflection point. When the candidate point does not exceed the parameter range, the objective function value of the candidate point is directly calculated. If it meets the objective, the original iteration point is updated to the candidate point. During the iteration process, the number of parameter out-of-bounds occurrences is recorded, and the out-of-bounds frequency is calculated. When the out-of-bounds frequency is less than 0.05, the boundary is shrunk; when it is greater than or equal to 0.05, the boundary is expanded. When both the change in the objective function residual and the parameter update step size are less than the convergence threshold, the algorithm terminates.

[0037] Furthermore, step 4 specifically includes:

[0038] (a) Define IGBT lifetime: lifetime L is the time it takes for the first time the device in a degraded state to cross the failure threshold, expressed as: ,in, =1.05 is the failure threshold, and inf{·} is the function to find the minimum lower bound of the set;

[0039] (b) Deriving the lifetime formula: Substituting the standardized degradation model into the failure threshold condition, we obtain... ;

[0040] (c) Monte Carlo numerical solution: based on parameters and By taking advantage of the normal distribution characteristics, a parameter sample set is generated through random sampling, which is then substituted into the lifetime formula to calculate the lifetime value. Statistical analysis of the lifetime value is then performed to obtain the lifetime distribution.

[0041] Furthermore, step (c) specifically includes:

[0042] (c1) Establishing a parametric probability model: Constructing lifetime model parameters , The corresponding mean and variance The functional relationship between the device's current operating condition parameter Θ and the device's operating condition parameter Θ.

[0043] (c2), Random sampling: Determine the number of samples to be sampled, and perform random sampling based on the probability distribution model of random variables to generate a set of parameter samples that conforms to the distribution law of the population;

[0044] (c3), Lifetime calculation and statistics: Substitute each group of samples into the lifetime formula to calculate the lifetime, and perform statistical analysis on all lifetimes to obtain the IGBT lifetime distribution.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] This invention employs Kalman filtering to preprocess IGBT on-state voltage drop data, enabling the construction of a linear system state equation based on the dynamic characteristics of the IGBT degradation process. This allows for real-time tracking of the nonlinear changes in the degradation trend, accurately separating noise such as temperature fluctuations and sensor errors from the actual degradation signal. Simultaneously, by recursively iteratively updating the state estimate, the problem of over-smoothing the degradation trend or residual noise interference is avoided.

[0047] The improved trust-region reflective nonlinear least squares method introduced in this invention constructs a trust-region constrained optimization space, limiting the parameter update step size within a reliable range, thus avoiding the algorithm from getting trapped in local optima and significantly reducing its sensitivity to initial parameter values. Furthermore, it sets upper and lower limits for parameters based on the physical laws of IGBT aging, and through intelligent reflective adjustment and elastic boundary update mechanisms, it prevents the generation of invalid parameters with positive degradation rates, ensuring that each set of estimated parameters conforms to the actual physical laws of "electro-thermal" coupled aging. Moreover, it adapts to the statistical fitting requirements of distributed parameters, stably solving the normal distribution characteristics of model parameters, providing a reliable parameter basis for subsequent lifetime dispersion modeling.

[0048] This invention treats the degradation model parameters as normal random variables dependent on operating conditions based on Bayesian statistical theory, capturing individual differences in IGBTs caused by manufacturing and operation; then, it combines the Monte Carlo method to numerically solve the lifetime distribution, and finally outputs a complete lifetime probability distribution including mean lifetime, characteristic lifetime, and reliable lifetime.

[0049] The above overview is for illustrative purposes only and is not intended to be limiting in any way. In addition to the illustrative aspects, embodiments, and features described above, further aspects, embodiments, and features of the invention will become readily apparent from the accompanying drawings and the following detailed description. Attached Figure Description

[0050] Figure 1 This is a graph showing the fitting results of the improved TRRLS algorithm before Kalman filtering in this invention;

[0051] Figure 2 This is a graph showing the fitting results of the improved TRRLS algorithm after Kalman filtering in this invention;

[0052] Figure 3 The image shows the estimation results of the traditional LM algorithm in this invention.

[0053] Figure 4 This is one of the parameter distribution diagrams of the degradation model of 36 IGBTs under 6 different working conditions in this invention;

[0054] Figure 5 This is the second of the parameter distribution diagrams for the degradation model of 36 IGBTs under 6 different working conditions according to the present invention.

[0055] Figure 6 The graph shows the IGBT lifetime prediction results based on the method of this invention under a 10s operating condition.

[0056] Figure 7 The figure shows the IGBT lifetime prediction results based on the method of this invention under a 5s operating condition;

[0057] Figure 8 The figure shows the IGBT lifetime prediction results based on the method of this invention under the 2s operating condition. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. It should be noted that the drawings are schematic and not illustrated to scale. For clarity and convenience, the relative sizes and proportions of the parts shown in the drawings have been exaggerated or reduced in size. Any size is only illustrative and not limiting.

[0059] The applicable IGBT application scenarios for this method include new energy storage converters (PCS) and rail transit traction converters. The corresponding prediction system needs to include an "electric-thermal" coupled aging test platform, a data acquisition module, a Kalman filter module, a TRRLS parameter estimation module, and a lifetime calculation module. The IGBT degradation data comes from the accelerated aging test of the test platform, which does not require hardware modification of the actual operating equipment and has non-invasiveness and engineering feasibility.

[0060] The IGBT device lifetime prediction optimization method proposed in this invention uses the IGBT collector-emitter on-state voltage drop (... This study focuses on the core degradation characteristics and overcomes the limitations of traditional methods, such as large data noise interference and parameter estimation that easily deviates from physical laws. On the one hand, it uses Kalman filtering to analyze the acquired data. Data preprocessing effectively suppresses noise caused by temperature fluctuations and sensor errors. On the other hand, an improved trust region reflection nonlinear least squares method is introduced to estimate the degradation model parameters. Combined with Bayesian theory, the parameters are set as normal random variables that depend on the operating conditions, ensuring that the parameters conform to the physical laws of IGBT “electric-thermal” coupled aging, and finally achieving high-precision life prediction.

[0061] Example 1: An IGBT lifetime prediction method based on an improved trust region reflection nonlinear least squares method and a distributed parameter degradation model, comprising the following steps:

[0062] Step 1: Model characteristic parameters based on the statistical regularity of device dispersion.

[0063] Based on the model of the IGBT device to be predicted, an "electro-thermal" coupled aging test platform was built. The core of the platform includes: a power cycling module, a temperature control module, and a high-precision data acquisition module.

[0064] The experiment involved six different temperature cycling cycles on 36 IGBT modules of the same model. Aging test, The time intervals were 10s, 5s, 2s, 1s, 0.8s, and 0.59s, respectively. In all experiments, the junction temperature fluctuation was maintained. and average junction temperature The basic values ​​are the same, namely 80℃ and 102℃ respectively.

[0065] The experimental platform was run continuously, with each group accumulating 10,000 cycles; one set of samples was collected every 100 cycles. A total of 100 sets of experimental data were obtained.

[0066] During the service life of IGBT devices, aging damage gradually intensifies with operating time. Therefore, the degradation process of the device can be described by a time-dependent monotonic function, expressed as:

[0067] (1)

[0068] in, This is the initial value of the on-state voltage drop. and These are degradation model parameters used to characterize the degradation process of the device.

[0069] Next, firstly, six different temperature cycles will be used. Consider it as 6 different working conditions Θ.

[0070] Secondly, based on relevant theories of Bayesian statistics, the parameters are... and Considering it as a set of normally distributed random variables related to the same working condition, the expression is:

[0071] (13)

[0072] in, and It follows a normal distribution, where Θ represents the current operating condition parameter of the device. and These are the mean and variance of the normal distribution they follow, respectively.

[0073] For different working conditions and Perform normal distribution fitting and calculate the normal distribution. and The corresponding mean and variance are expressed as follows:

[0074] (14)

[0075] in, They represent , Two model parameters, This represents the total amount of data obtained in the experiment. Indicates the current number Group data.

[0076] Step 2: Perform standardization and Kalman filtering preprocessing on the experimental data in sequence.

[0077] To eliminate the initial The impact of differences on model fitting, on the collected data To standardize, the expression is:

[0078] (2)

[0079] Then standardize Perform Kalman filtering data preprocessing. Specifically:

[0080] First, the Kalman filter parameter process noise covariance Measure noise covariance State transition matrix Observation matrix covariance initialization.

[0081] definition The system state at any given time is the standardized on-state voltage drop. Secondly, the state and observation equations of the linear system are established, and their expressions are as follows:

[0082] (3)

[0083] in, For process noise, For the measured values ​​of the variables, For measurement noise; Q is the process noise covariance, R is the measurement noise covariance, A is the state transition matrix, and C is the observation matrix;

[0084] Secondly, iterative calculations of the Kalman filter are performed. Specifically:

[0085] Take the state variables at the initial time. , as the optimal estimate at the initial moment, is used to predict the state variables at the next moment, where The normalized value of the on-state voltage drop acquired at the initial moment;

[0086] The expression for predicting the current state using the optimal estimate from the previous time step is:

[0087] (4)

[0088] in, These are the predicted values ​​of the state variables. It is the optimal estimate from the previous moment;

[0089] Based on the previous moment, that is posterior error covariance matrix at time 1 Calculate the prior error covariance matrix at the current time. The expression is:

[0090] (5)

[0091] in, yes The posterior error covariance matrix at time t. It is the prior error covariance matrix at the current moment;

[0092] The gain is determined based on the prior error covariance matrix and the measurement noise covariance matrix. The expression is:

[0093] (6)

[0094] By combining measured values ​​with the corrected predicted state, the optimal estimated value for the current time can be obtained. The expression is:

[0095] (7)

[0096] To provide parameters for the next iteration, the expression is:

[0097] (8)

[0098] in, It is the identity matrix;

[0099] Iterate through equations (4)-(8) until all 100 sets of data have been processed, and obtain the filtered result. The data provides high-quality data for subsequent parameter estimation.

[0100] The standardized on-state voltage drop data before and after filtering were fitted using an improved TRRLS algorithm. The relevant parameters of the data fitted by the improved TRRLS algorithm before filtering were as follows: =-0.014951、 =-0.116267, RMSE=0.0058, R=0.98544, as shown Figure 1 As shown, some noise still exists in the data, resulting in limited fit between the fitted curve and the measured values. After Kalman filtering preprocessing, the relevant parameters of the improved TRRLS algorithm are optimized as follows: =0.014950、 =-0.006267, RMSE=0.0058, R=0.98546, as Figure 2 As shown, the filtered data noise is significantly reduced, and the fitted curve is consistent with the filtered data. The fit is significantly improved.

[0101] Step 3: Use the improved TRRLS algorithm to estimate the parameters of the degenerate model after the preprocessing in Step 2.

[0102] First, initialize the improved TRRLS algorithm parameters, and set the lower limit of the parameters based on the actual physical meaning of IGBT aging. upper limit ;

[0103] Secondly, the core of improving the TRRLS algorithm is solving for the coefficients of the nonlinear function. The basic optimization problem is defined as follows:

[0104] (10)

[0105] in, It is a nonlinear function. Let be the coefficients to be solved. Given the input data; The observed output data is N, where N is the number of data measurements.

[0106] Then, combining the IGBT standardized on-state voltage drop degradation model (2), the basic optimization problem is transformed into an objective function for this model. The expression is:

[0107] (11)

[0108] in, Let j be the number of iterations at the j-th measurement time. This is the filtered on-state voltage drop measurement value corresponding to the j-th measurement time.

[0109] Then solve the trust region subproblem to obtain the optimal search step size. Specifically:

[0110] Initial iteration point And satisfy the upper and lower limits of the parameters.

[0111] in, It is a parameter initial value, It is a parameter The initial value;

[0112] Constructing the trust region subproblem: at the current iteration point The improved TRRLS algorithm finds the optimal search step size by solving the trust region subproblem. The expression is:

[0113] (12)

[0114] in, For the first The search step size for each iteration; For the objective function exist The Hessian matrix at that location; It is a diagonal scale matrix; For the first The trust region scale of the next iteration; The gradient of the objective function at that point. Indicates the number of iterations;

[0115] Solving for the optimal search step size: Solving the trust region subproblem shown in equation (12), in " Within the constraints of ", find the function Minimum search step size This ensures that the parameter update step size is within the "reliable" range of the algorithm, avoiding iteration divergence due to excessively large step sizes.

[0116] Then, candidate iteration points are calculated, and intelligent reflection adjustment and parameter iterative updates are performed. Specifically:

[0117] First, calculate the candidate iteration points. ( (This is the step size coefficient, subject to parameter range constraints).

[0118] Then, intelligent reflection adjustment is performed; more specifically: first, candidate iteration points are detected. Whether it exceeds the parameter range (initially the preset boundary, and during iteration the elastically updated boundary);

[0119] If the candidate iteration point exceeds the boundary: first adjust the trust region scale based on the gradient direction. If the gradient points into the feasible region, increase the trust region scale; if the gradient points out of the feasible region, decrease the trust region scale.

[0120] Then by checking the out-of-bounds dimension r ( This indicates that the parameter is currently being tested. ; This indicates that the parameter is currently being tested. Performing a one-sided reflection "reflects" the parameters back into the trust region and constraints. If ,but ;like ,but Calculate the objective function value of the candidate points after reflection. ,like If the reflection point is correct, then accept the reflection point; otherwise, keep the original iteration point. Keep the trust region size unchanged and reduce it. ;

[0121] If the candidate iteration point does not exceed the boundary: If If so, accept the candidate point and increase the trust region size. Otherwise, keep the original iteration point. Keep the trust region size unchanged and reduce it. ;

[0122] Then, the parameters are constrained and the elastic boundaries are updated. Specifically:

[0123] First, perform parameter boundary detection: record the number of times parameters exceed the limit, which will be used for subsequent elastic boundary updates;

[0124] Secondly, perform elastic boundary updates: calculate the out-of-bounds frequency of each parameter (number of out-of-bounds / total number of iterations). If the out-of-bounds frequency is less than 0.05, slightly shrink the boundary; if the out-of-bounds frequency is greater than or equal to 0.05, slightly expand the boundary.

[0125] Finally, if both the change in the objective function residual and the parameter update step size are less than the convergence threshold, the algorithm terminates and outputs the optimal parameters. , Otherwise, return to the step of finding the optimal search step size and continue iterating.

[0126] Finally, 36 sets of optimal parameters were obtained. , .

[0127] To verify the parameter estimation advantages of the improved TRRLS algorithm, the traditional LM algorithm was also used to estimate the parameters of the same batch of data. The parameters estimated by the LM algorithm were: =0.014501、 =-0.006267, RMSE=0.0062, R=0.98535, as Figure 3 As shown, its fitting residuals are higher than those of the improved TRRLS algorithm, and some data points deviate from the fitting curve; while the estimation results of the improved TRRLS algorithm not only have lower fitting residuals, but also have a higher overall fit between the fitting curve and the measured values, and better parameter estimation accuracy.

[0128] Finally, 36 sets of optimal parameters were obtained. , The degradation model parameters of 36 IGBTs under 6 different operating conditions were statistically analyzed. The distribution of parameters is as follows Figure 4 As shown, with the change of temperature cycle period, different operating conditions affect the model parameters. The impact is not significant. Regardless of operating conditions; The distribution of parameters is as follows Figure 5 As shown, its overall fluctuation range is concentrated in -7×10 -3 to -3.5×10 -3 The range, and also showing a pattern of increasing absolute value with increasing severity of the working conditions, different working conditions affect the model parameters. It has a significant impact. Related to the working conditions, therefore, it is possible to further... , This is expressed by equation (15).

[0129] Step 4: Predict the lifetime distribution of IGBT devices using the Monte Carlo method.

[0130] First, based on the statistical results of multiple sets of model parameters, a probabilistic model of the random distribution of lifetime model parameters is established, i.e., the lifetime model parameters. , The mean and standard deviation of the same working condition Perform function fitting, the expression is:

[0131] (15)

[0132] Next, the lifetime of an IGBT device is defined as the time it takes for its degradation state to first cross the failure threshold, and its lifetime L is expressed as:

[0133] (16)

[0134] in, The specified device failure threshold is set to 1.05 according to the failure standard, L is the device lifetime, and inf{·} is a function used to find the minimum lower bound of the set.

[0135] Secondly, by substituting the standardized post-degradation model (2) into the failure threshold condition, the lifetime L and random parameters are obtained. , The relationship is expressed as:

[0136] (17)

[0137] The Monte Carlo method (MCM) is used to numerically calculate the lifetime distribution: First, based on , 10,000 sets of parameter samples are generated; then, each set of samples is substituted into formula (17) to calculate the lifetime value; finally, the 10,000 sets of lifetime values ​​are statistically analyzed to obtain six sets of IGBT lifetime prediction charts under different working conditions.

[0138] Among them, the IGBT life prediction results under the 10s operating condition are as follows: Figure 6 As shown, the life prediction results under the 5s operating condition are as follows: Figure 7 As shown, the life prediction results under the 2s operating condition are as follows: Figure 8 As shown in the figure, IGBT devices exhibit good consistency in the initial aging stage. However, as the aging process deepens, the differences in their degradation states gradually increase, ultimately leading to a significant dispersion in the device lifetime distribution. Fitting the predicted lifetime distribution reveals that it roughly follows a normal distribution, and the distribution characteristics are influenced by operating conditions. Furthermore, with… As the pressure decreases, the lifetime of IGBT devices gradually increases, and the lifetime distribution range gradually narrows. The proposed method can accurately predict the lifetime distribution range of IGBT devices under a 95% confidence interval (CI), with an accuracy rate exceeding 95%, demonstrating the effectiveness of the method.

[0139] Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for predicting the lifetime of IGBTs based on an improved trust region reflection nonlinear least squares method and a distributed parameter degradation model, characterized in that, Specifically, the following steps are included: Step 1: Based on the characteristic parameters of the statistical law of device dispersion, establish the degradation model of IGBT device and treat the model parameters as normal random variables that depend on the operating condition characteristics; Step 2: Standardize the experimental data and perform Kalman filtering data preprocessing to suppress noise interference. Step 3: Use the improved trust region reflection nonlinear least squares method to estimate the parameters of the degraded model after the preprocessing in Step 2, and ensure the rationality of the parameters through physical constraints and iterative optimization; Step 4: Predict the lifetime distribution of IGBT devices using the Monte Carlo method and output the lifetime probability distribution law.

2. The IGBT device lifetime prediction method according to claim 1, characterized in that, Step 1 specifically includes: (a) Establishing a degradation model for IGBT devices: The degradation process of IGBT aging damage is described using a monotonic function related to operating time, expressed as: ,in, This is the initial value of the on-state voltage drop. and To characterize the device degradation process, model parameters, Runtime; (b) Parameters and Treated as a normally distributed random variable related to operating conditions: based on Bayesian statistical theory, and It follows a normal distribution, and its expression is: Where Θ represents the current operating condition parameter of the IGBT. , They are respectively The mean and variance of the normal distribution. , They are respectively The mean and variance of the normal distribution.

3. The IGBT device lifetime prediction method according to claim 2, characterized in that, In step (b), the calculation , The expressions for the mean and variance are: ,in, Corresponding parameters and , This represents the total amount of data obtained in the experiment. Indicates the current number Group data.

4. The IGBT device lifetime prediction method according to claim 1, characterized in that, In step 2, the standardization process involves standardizing the acquired IGBT on-state voltage drop to eliminate initial differences and obtain the standardized on-state voltage drop. The expression is: ,in, The measured value of the on-state voltage drop at any given time. This is the initial value of the on-state voltage drop.

5. The IGBT device lifetime prediction method according to claim 4, characterized in that, Step 2, specifically the Kalman filter data preprocessing, includes: (a) Initialize Kalman filter parameters: including process noise covariance Measurement of noise covariance State transition matrix Observation matrix and initial covariance ; (b) Data filtering: Save the data of the first time the IGBT degradation state crosses the failure threshold, and do not perform Kalman filtering on this part of the data; (c) Define system state: Select the standardized on-state voltage drop As a state variable, System state at time 1 for ,in, for Time-standardized on-state voltage drop; (d) Establish the state and observation equations for the linear system: ,in, For process noise, This is the measured value of the on-state voltage drop. For measuring noise; (e) Perform Kalman filter iterative calculations until all data points have been processed.

6. The IGBT device lifetime prediction method according to claim 5, characterized in that, The iterative calculation in step (e) specifically includes: (e1), take the initial state variable , As the initial optimal estimate The on-state voltage drop is the normalized value at the initial moment. (e2), State prediction: ,in, This is the predicted state value at the current moment. This is the optimal estimate from the previous moment; (e3), Calculation of prior error covariance: ,in, for The posterior error covariance matrix at time t. The prior error covariance matrix at the current time; (e4), Kalman gain calculation: ; (e5), Status Update: To obtain the optimal estimate at the current time. ; (e6), posterior error covariance update: ,in, It is an identity matrix.

7. The IGBT device lifetime prediction method according to claim 1, characterized in that, Step 3 specifically includes: (a) Initialize and improve TRRLS parameters: Set degradation model parameters according to the physical laws of IGBT aging. and upper limit Lower limit ; (b) Define the basic optimization problem: ,in, It is a nonlinear function. The coefficients to be determined are: For input data, The output data is N, where N is the number of measurements. (c) Transformation Objective Function: Combining the standardized on-state voltage drop degradation model, the objective function is: ,in, Let j be the number of iterations at the j-th measurement time. The normalized on-state voltage drop after filtering at time j; (d) Solve the trust region subproblem to obtain the optimal search step size. The trust region subproblem is: ,in, For the first The search step size for the next iteration. For the objective function exist The Hessian matrix at that location, It is a diagonal scale matrix. For the first Trust region scale of the next iteration The gradient of the objective function at that point. This indicates the number of iterations; the parameters are then updated iteratively until the convergence condition is met, at which point the optimal parameters are output. and .

8. The IGBT device lifetime prediction method according to claim 7, characterized in that, In step (d), the iterative parameter update includes: calculating candidate iteration points based on the optimal search step size. , The step size is used as the step size coefficient. When a candidate point exceeds the parameter range, the trust region scale is adjusted based on the gradient direction. If the gradient points into the feasible region, the scale is increased; if it points out of the region, the scale is decreased. Then, a one-sided reflection is performed on the out-of-bounds dimension to adjust it back to the trust region and the constraint range. The objective function value of the reflection point is calculated. If it meets the objective, the original iteration point is updated to the reflection point. When the candidate point does not exceed the parameter range, the objective function value of the candidate point is directly calculated. If it meets the objective, the original iteration point is updated to the candidate point. During the iteration process, the number of parameter out-of-bounds occurrences is recorded, and the out-of-bounds frequency is calculated. When the out-of-bounds frequency is less than 0.05, the boundary is shrunk; when it is greater than or equal to 0.05, the boundary is expanded. When both the change in the objective function residual and the parameter update step size are less than the convergence threshold, the algorithm terminates.

9. The IGBT device lifetime prediction method according to claim 1, characterized in that, Step 4 specifically includes: (a) Define IGBT lifetime: lifetime L is the time it takes for the first time the device in a degraded state to cross the failure threshold, expressed as: ,in, =1.05 is the failure threshold, and inf{·} is the function to find the minimum lower bound of the set; (b) Deriving the lifetime formula: Substituting the standardized degradation model into the failure threshold condition, we obtain... ; (c) Monte Carlo numerical solution, and obtain lifetime distribution.

10. The IGBT device lifetime prediction method according to claim 9, characterized in that, Step (c) specifically includes: (c1) Establishing a parametric probability model: Constructing lifetime model parameters , The corresponding mean and variance The functional relationship between the device's current operating condition parameter Θ and the device's operating condition parameter Θ. (c2), Random sampling: Determine the number of samples to be sampled, and perform random sampling based on the probability distribution model of random variables to generate a set of parameter samples that conforms to the distribution law of the population; (c3), Lifetime calculation and statistics: Substitute each group of samples into the lifetime formula to calculate the lifetime, and perform statistical analysis on all lifetimes to obtain the IGBT lifetime distribution pattern.

Images