IGBT Junction Temperature Prediction Method Based on ISFO-SVM Model

Abstract

The present invention proposes a method for predicting the IGBT junction temperature. The steps are as follows: Simulate the aging process of IGBT through the IGBT aging acceleration test to obtain data such as IGBT junction temperature, saturation voltage drop, collector current, number of aging times, etc., and normalize the data; Set the parameters of the improved sailfish algorithm and the support vector machine model; Run the improved sailfish algorithm to obtain the optimal penalty factor in the support vector machine model and the optimal parameters of the kernel function; Substitute the optimized optimal parameters into the support vector machine model, and train the support vector machine (ISFO-SVM) model optimized by the improved sailfish algorithm; Input the prediction data into the ISFO-SVM model to obtain the prediction result, and denormalize the prediction result. The results show that under the RMSE, MAPE, and R2 performance indicators, the prediction performance of the ISFO-SVM model is better, and the fitting degree between the predicted junction temperature value and the actual junction temperature is higher, making up for the deficiency of the low prediction accuracy of the existing IGBT junction temperature prediction method and realizing the effective prediction of the IGBT junction temperature.

Description

Technical Field

[0001] The technical solution of the present invention belongs to the technical field of IGBT reliability, and specifically relates to a method for predicting the IGBT junction temperature based on the ISFO-SVM model. Background Art

[0002] IGBT is a key component for power conversion in new energy systems. The reliability of IGBT greatly restricts the safe and reliable operation of the working system. Due to the different thermal expansion coefficients of the materials of each layer of the IGBT module, the temperature cycle fluctuations cause different thermal stresses at its internal joints, resulting in the failure of the IGBT module. Especially in a high-frequency and high-power working environment, IGBT will generate large switching losses, causing large junction temperature fluctuations in the IGBT module. The higher the working environment temperature, the greater the failure probability of the IGBT module. When the power generated by the IGBT chip cannot be dissipated in time, the junction temperature will continue to rise until the device fails. Since the aging degree of the IGBT module has a certain impact on the conversion efficiency of the power conversion device and the safety of its system, accurately predicting the IGBT junction temperature is of great significance for evaluating its reliability and accurately judging the safety of the system.

[0003] There are two methods for obtaining the IGBT junction temperature: the relevant experimental equipment measurement method and the calculation model method; the laboratory measurement method mainly obtains the IGBT junction temperature through the thermal sensor method, the optical fiber detection method, and the temperature-sensitive parameter method. However, this will damage the packaging structure of the IGBT and is not suitable for predicting the IGBT junction temperature under working conditions; the calculation model method calculates the IGBT junction temperature by establishing a junction temperature calculation model based on electrothermal coupling or a junction temperature prediction model based on an intelligent algorithm, and predicts the IGBT junction temperature by establishing a machine learning model, so as to evaluate the reliability of the IGBT.

[0004] In the prior art, it is difficult for a simple machine learning model to achieve high-precision prediction. Therefore, an intelligent evolutionary algorithm is often used to optimize the machine learning model to construct a combined model for prediction. However, these algorithms have problems such as poor optimization ability and premature convergence, and the accuracy and accuracy of the final junction temperature prediction results still need to be improved. Therefore, seeking an IGBT junction temperature prediction method with high prediction accuracy and small prediction error is of great significance for the replacement and maintenance of IGBTs and the safety and reliability of their systems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a method for predicting the IGBT junction temperature. This method is based on an improved sailfish algorithm to optimize the support vector machine model (ISFO-SVM) for predicting the IGBT junction temperature. An adaptive non-linear iterative factor is introduced to improve the optimization ability of sailfish individuals, and the Levy flight strategy is used to improve the diversity of the search space. In the search process, in order to avoid the algorithm falling into local convergence, the DE / current to best / 1 strategy in the differential mutation strategy is introduced to increase the diversity of the population.

[0006] The technical solution adopted by the present invention to solve the above technical problem is to provide a method for predicting the IGBT junction temperature. Specifically, it is a method for predicting the IGBT junction temperature based on an improved sailfish algorithm to optimize the support vector machine model (ISFO-SVM). The steps are as follows:

[0007] Step 1: Obtain data such as the IGBT junction temperature, saturation voltage drop, collector current, and aging times through the IGBT power cycle aging acceleration test, and normalize the data.

[0008] Step 2: Set the parameters of the improved sailfish algorithm and the support vector machine model.

[0009] Step 3: Run the improved sailfish algorithm to obtain the optimal penalty factor in the support vector machine model and the optimal parameters of the kernel function.

[0010] Step 4: Substitute the optimized optimal penalty factor and the optimal parameters of the kernel function into the support vector machine model, and train the support vector machine (ISFO-SVM) model optimized by the improved sailfish algorithm.

[0011] Step 5: Input the prediction data into the ISFO-SVM model to obtain the prediction result, and denormalize the prediction result.

[0012] Step 6: Display and output the IGBT junction temperature prediction result.

[0013] Further, the specific implementation method of step 1 includes the following steps:

[0014] Step 1.1: In order to obtain the degradation data of the IGBT module in its full life cycle, the present invention designs an IGBT power cycle aging test and a single-pulse test to obtain a data set S including the saturation voltage drop, collector current, junction temperature, and aging cycle times of the IGBT.

[0015] Step 1.2: Randomly shuffle and normalize the obtained data set S, and divide it into training data and test data according to a ratio. Use the power cycle times N, saturation voltage drop Vce, and collector current Ic as the input part of the model; and the junction temperature Tj as the output part of the model.

[0016] Step 1.3, normalize the above data;

[0017]

[0018] Where A is the value of the variable to be normalized, such as the number of power cycles N, saturation voltage drop Vce, collector current Ic, and junction temperature T j , A min is the minimum value of the variable, A max is the maximum value of the variable, A N is the normalized value of the variable;

[0019] Furthermore, the parameters set in step 2 include: the population size, the maximum number of iterations, the population dimension in the improved sailfish algorithm, the search range of the penalty factor C in the support vector machine model, and the search range of the kernel function parameter g;

[0020] Furthermore, the specific implementation method of step 3 includes the following steps:

[0021] Step 3.1, initialize the individual positions of the sailfish population and the sardine population of the improved sailfish algorithm, and calculate the objective function value of each individual;

[0022] Step 3.2: Sort the position and objective function value of each individual, and record the current optimal individual position and optimal objective function value;

[0023] Step 3.3: Introduce an adaptive nonlinear iteration factor to update the position of the sailfish, and introduce a Levy flight strategy to update the position of the sardine.

[0024] In step 3.4, a differential mutation strategy is introduced to continuously search and update the positions of individuals in the swordfish and sardine populations to determine whether optimal convergence has been achieved. If so, the optimal parameters of the model are obtained, and the optimal swordfish individual position X(x1, x2) is output (corresponding to the optimal penalty factor C of the SVM model and the optimal parameter g of the kernel function, respectively). If not, the algorithm returns to step 3.2.

[0025] Furthermore, in step 3, the root mean square error is selected as the objective function, and the internal parameters of the support vector machine model are optimized using the training data;

[0026]

[0027] The improved sailfish algorithm determines whether the updated position is better than the original position based on the objective function value of the individual position at this time, and decides whether to use the updated position in the subsequent search process. The objective function is described as follows:

[0028]

[0029] In the formula, is a random number in [0, 1];

[0030] The specific implementation method of step 3.3 is as follows:

[0031] Step 3.3.1, introduce an adaptive non-linear iterative factor to update the position of sailfish individuals;

[0032] Since the sailfish individuals are randomly distributed in the initial stage of iteration, in order to improve the ability of sailfish individuals in optimization, the present invention introduces an adaptive non-linear iterative factor into the position update formula of sailfish to accelerate the optimization ability of sailfish individuals; among them, the sailfish population is represented by X SF denotes;

[0033] The update formula of the adaptive non-linear iterative factor of the i-th sailfish individual at the t-th iteration is described as follows:

[0034]

[0035] The position update formula of sailfish is:

[0036]

[0037] In the formula, represents the best individual position in the sailfish population at the t-th iteration; represents the best individual position in the sardine population at the t-th iteration; represents the position of the sailfish individual to be updated at the t-th iteration; λ i The definition of the coefficient is shown in formula (6):

[0038] λ i = 2 × rand(0, 1) × PD - PD (6)

[0039] In the formula, PD represents the density of the prey group, which is described in detail by formula (7):

[0040]

[0041] In the formula, N SF represents the number of sailfish, N S represents the number of sardines;

[0042] Step 3.3.2, introduce the Levy flight strategy to update the position of sardine individuals; among them, the sardine population is represented by X F denotes;

[0043] The position update formula of sardines in the original sailfish algorithm is shown in formula (8):

[0044]

[0045] Wherein, represents the adaptive non - linear iterative factor of the i - th sailfish individual at the t - th iteration, represents the position of the best individual in the sailfish population at the t - th iteration, represents the position of the sardine individual to be updated at the t - th iteration; AP represents the attack strength of the sailfish, and its detailed description is shown in formula (9):

[0046] AP = A×(1 - 2×Itr×e) (9)

[0047] Wherein, A and e represent the control coefficients of the sailfish attack strength, which linearly transform the sailfish attack strength from A to 0;

[0048] When AP > 0.5, that is, when the attack strength of the sailfish is strong, update the positions of all sardines using formula (8); when AP < 0.5, at this time the attack strength of the sailfish is low, and only the positions of some sardines need to be updated;

[0049] The range of the positions of some sardines is defined as follows:

[0050] α = N S ×AP (10)

[0051] β = d i ×AP (11)

[0052] Wherein, α represents the number of sardines to be updated, β represents the number of dimensions for sardine update, and d i is the number of variables at the i - th iteration;

[0053] In order to improve the randomness of the sardine population and the diversity of the search space, the present invention introduces the Levy flight strategy, and the position update formula of the sardines at this time is:

[0054]

[0055] Wherein, t is the current iteration number, d is the dimension of the position vector, represents the position of the sardine individual to be updated at the t - th iteration;

[0056] The formula of Levy flight can be described as:

[0057]

[0058] Wherein, r1 and r2 are two random numbers, and the value range is [0, 1], β = 1.5, and σ can be calculated as:

[0059]

[0060] Wherein, Γ(x) = (x - 1)!

[0061] Therefore, first calculate the attack power AP of the sailfish according to formula (9). When AP > 0.5, that is, when the attack power of the sailfish is strong, update the positions of all sardines using the improved sardine position update formula (12); when AP < 0.5, at this time the attack power of the sailfish is low, then calculate the number and dimensions of the sardines that need position update according to formulas (10) and (11), and then update them using the improved sardine position update formula (12).

[0062] The specific implementation method of step 3.4 is as follows:

[0063] To avoid falling into local convergence during the search process, the present invention introduces the DE / currentto best / 1 strategy in the differential mutation strategy to mutate the vectors of the population individuals, and adds the differential mutation strategy in the later stage of each round of search to increase the diversity of the population; the formula is expressed as follows:

[0064]

[0065] Wherein, p1 ≠ p2 ≠ p3, is the differential vector, F ∈ [0.1, 0.9] is the scaling factor, h i，t is the mutation vector at the i-th position in the t-th search. After obtaining the mutation vector, the crossover operation is as follows:

[0066]

[0067] Wherein, v i，t is the crossover variable at the i-th search position, j0 is a random value in the dimension, and each crossover operation only involves one dimension of the individual. pCR ∈ [0, 1] is the crossover probability;

[0068] Perform the selection operation, and retain the vector with the better objective function value as the next-generation individual. The selection operation is expressed as:

[0069]

[0070] According to the above formula, continuously search and update the positions of the population individuals, and judge whether the optimal convergence is achieved; if satisfied, obtain the best parameters of the model and output the optimal sailfish individual position X(x1, x2) (corresponding to the optimal penalty factor C of the SVM model and the optimal parameters g of the kernel function respectively); if not satisfied, return and continue to execute step 3.2;

[0071] Further, the specific implementation method of step 4 is as follows: Input the optimal penalty factor C and the optimal parameter g of the kernel function obtained in step 3 into the support vector machine model to form an improved sailfish algorithm optimized support vector machine (ISFO-SVM) model, and use the optimized optimal penalty factor C and the optimal parameter g of the kernel function to train the support vector machine model.

[0072] Further, the inverse normalization process in step 5 is as shown in formula (18):

[0073] T′ i =T′ soale，i ×(T max -T min )+T min (18)

[0074] In the formula, T′ i is the predicted value of the junction temperature after inverse normalization, T′ soale，i is the predicted value of the normalized junction temperature obtained in step 4, and T max , T min are the maximum and minimum values of the junction temperature variable in step 1.3.

[0075] Further, the specific implementation method of step 6 is as follows: Output the prediction graph of the IGBT junction temperature obtained in step 5 by the ISFO-SVM model on the display screen of the computer, and display the error curve graph and error histogram of the IGBT junction temperature prediction using different models.

[0076] The method of inputting data into the computer in the above steps is a well-known method; the computers, monitors, and MATLAB computer software used are all obtained through commercial purchases.

[0077] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0078] (1) An improved sailfish algorithm including an adaptive non-linear iterative factor, Levy flight strategy, and differential mutation strategy is proposed; the proposed improved sailfish algorithm not only maintains the solution stability and robustness of the original algorithm, but also the obtained optimal solution has good convergence and optimization accuracy;

[0079] (2) In order to make the change amount of the junction temperature meet the requirements of the test setting and accelerate the aging of the IGBT module, the present invention designs an IGBT aging acceleration experiment and a single-pulse test test platform with reference to the IEC60068-2-14 JEDEC standard set by the International Electrotechnical Commission (IEC) for the power cycle test, and uses a 5% increase in the saturation voltage drop as a reference quantity for IGBT failure; data of the saturation voltage drop, collector current, junction temperature, and aging times of the IGBT are obtained through experiments, and an IGBT test data set is constructed;

[0080] (3) An improved sine firefly optimization (ISFO) - optimized support vector machine (SVM) model is proposed, and an IGBT junction temperature prediction model based on ISFO - SVM is constructed. Compared with the SFO - SVM and SVM models, the IGBT junction temperature prediction model based on ISFO - SVM has better performance in predicting the junction temperature and higher fitting degree. In summary, the ISFO - SVM model constructed in the present invention has better prediction results for the IGBT junction temperature. Description of the Drawings

[0081] The present invention will be further described below with reference to the drawings and embodiments.

[0082] Figure 1 It is a schematic diagram of the junction temperature prediction steps of the ISFO - SVM model provided by the present invention;

[0083] Figure 2 It is the circuit diagram of the IGBT power cycle aging test provided by the embodiment of the present invention;

[0084] Figure 3 It is the circuit diagram of the single - pulse test provided by the embodiment of the present invention;

[0085] Figure 4 It is the prediction diagram of the IGBT junction temperature by the ISFO - SVM model provided by the embodiment of the present invention;

[0086] Figure 5 It is the error curve diagram of the IGBT junction temperature prediction by different models provided by the embodiment of the present invention;

[0087] Figure 6 It is the error histogram of the IGBT junction temperature prediction by different models provided by the embodiment of the present invention; Detailed Embodiments

[0088] Figure 1It shows that the general steps of the IGBT junction temperature prediction method provided by the present invention are as follows: start → IGBT power cycle aging test and single pulse test → obtain the saturation voltage drop, junction temperature, and collector current data of the IGBT at the corresponding power cycle aging times → set the IGBT saturation voltage drop, collector current, and aging time as the input data set, and the junction temperature as the output data set → perform preprocessing such as normalization on the obtained data set → initialize the sailfish population, sardine population, and algorithm parameters → select the root mean square error as the objective function, calculate the fitness values of the sailfish and sardine, sort the values of the individual objective function, and record the optimal fitness value and position → introduce an adaptive non-linear iterative factor to update the position of the sailfish individual → introduce the Levy flight strategy to update the sardine position → introduce the DE / current to best / 1 strategy in the differential mutation strategy to increase the diversity of the population → determine whether the optimal convergence is achieved. If satisfied, obtain the best parameters of the model and output the optimal sailfish individual position X(x1, x2) (corresponding to the optimal penalty factor C of the SVM model and the optimal parameter g of the kernel function) → construct an IGBT junction temperature prediction model based on ISFO-SVM → predict the test data and display the output IGBT junction temperature prediction result → end.

[0089] Figure 2 It shows the circuit principle of the IGBT power cycle aging test proposed by the present invention. The test object of this study is 2 IGBT modules; when the switch is closed, the lower transistor is connected to the reverse voltage and continuously turned off; an aging acceleration test is performed on the upper transistor. The gate signal drives the upper transistor IGBT to conduct. At this time, the programmable DC power supply outputs a current of 75A, causing the IGBT junction temperature and case temperature to rise rapidly, achieving the purpose of accelerating IGBT aging. The basic test steps of the IGBT power cycle aging test are as follows:

[0090] (1) First, close the switch S, set the programmable constant current source to output a current of 50A, and the gate to output a driving voltage of 15V to make the upper transistor of the IGBT power module conduct. The test sample module generates power loss, causing the junction temperature and case temperature to rise. Set the initial case temperature value to 40°C;

[0091] (2) Monitor the temperature change through the temperature sensor built into the bottom of the module. When the highest case temperature is 90°C, open the switch S and turn on the air-cooled radiator to work until the IGBT power module cools down rapidly to a case temperature of 40°C. One power cycle aging of the case temperature fluctuation is completed;

[0092] (3) Repeat the above steps (1) and (2) until the module approaches the failure standard and stop the test. This test has completed a total of 6000 power cycle tests, and pause once every 1000 power cycle tests. Remove the module and put it into a constant temperature oven for a short-time single pulse test; record the collector current and saturation voltage drop values of the IGBT module at different junction temperatures.

[0093] Figure 3 It shows the principle of the single-pulse test circuit proposed by the present invention. Through the DSP development board, the amplifier circuit and the driver, a single-pulse driving signal is given to the IGBT power module to inject a single-pulse trigger current into the IGBT. In this test, it is set that the temperature adjustment range of the constant temperature box is [0 °C, 100 °C], the temperature adjustment interval is 10 °C, the set range of the collector current is [25 A, 70 A], and its adjustment interval is 5 A. The steps of the single-pulse test are as follows:

[0094] (1) Place the IGBT power module to be tested in the constant temperature box, adjust the temperature of the constant temperature box, and when the temperature of the constant temperature box is stable, it is considered that the module has reached thermal equilibrium at this time;

[0095] (2) After the thermal equilibrium condition is satisfied, adjust the set values of the temperature of the constant temperature box and the collector current, and record the data of the saturation voltage drop, junction temperature, and collector current at the corresponding number of power cycle aging times in sequence.

[0096] In order to illustrate the technical solution described in the present invention, it will be described below through specific implementation cases;

[0097] Step 1, obtain data such as the IGBT junction temperature, saturation voltage drop, collector current, and aging times through the IGBT power cycle aging acceleration test, and perform normalization processing on the data;

[0098] In order to facilitate the construction of the prediction model and without loss of generality, in the data of the above IGBT aging test, 386 groups of data are randomly selected to establish a data set. The first 70% of this data set is used as the training sample. After the data is randomly scrambled and normalized, it is used for the training of the prediction model. The remaining 30% is used as the test sample to test and verify the effectiveness of the IGBT junction temperature prediction model. The present invention sets the power cycle number N, saturation voltage drop Vce, and collector current Ic as the input part of the model, and the junction temperature T j as the output part of the model; perform normalization processing on the above data according to formula (1);

[0099] Step 2, set the parameters of the improved sailfish algorithm and the support vector machine model;

[0100] Set the population size, maximum number of iterations, population dimension in the improved sailfish algorithm, the search range of the penalty factor C in the support vector machine model, and the search range of the kernel function parameter g;

[0101] In this embodiment, the population size in the improved sailfish algorithm is denoted as 30, the maximum number of iterations is 500, the population dimension is 2, the search range of the penalty factor C in the support vector machine model is [0.1, 1200], the range of the kernel function parameter g of the support vector machine model is [0.01, 100], and the rest of the parameters are default values;

[0102] Step 3: Run the improved sailfish algorithm to obtain the optimal penalty factor in the support vector machine model and the optimal parameters of the kernel function.

[0103] Step 3.1: Initialize the individual positions of the sailfish population and the sardine population of the improved sailfish algorithm, and calculate the objective function value of each individual.

[0104] Select the root mean square error as the objective function according to formula (2), and use the training data to optimize the internal parameters of the support vector machine model.

[0105] Step 3.2: Sort the positions and objective function values of each individual, and record the current optimal individual position and the optimal objective function value.

[0106] Step 3.3: Introduce an adaptive non-linear iteration factor to update the positions of sailfish individuals, and introduce the Levy flight strategy to update the positions of sardine individuals.

[0107] Since the sailfish individuals are randomly distributed in the initial stage of iteration, in order to improve the optimization ability of sailfish individuals, an adaptive non-linear iteration factor is introduced into the position update formula of sailfish to accelerate the optimization ability of sailfish individuals; among them, the position update formula of sailfish is shown in formula (5).

[0108] In order to improve the randomness of the sardine population and the diversity of the search space, the present invention introduces the Levy flight strategy. At this time, the position update formula of sardine is shown in formula (12); first calculate the attack power AP of the sailfish according to formula (9). When AP > 0.5, that is, when the attack power of the sailfish is strong, update the positions of all sardines with the improved sardine position update formula (12); when AP < 0.5, at this time the attack power of the sailfish is low, then calculate the number and dimension of sardines that need to update their positions according to formulas (10) and (11), and then use the improved sardine position update formula (12) to update them.

[0109] Step 3.4: To avoid falling into local convergence during the search process, the present invention introduces the DE / current to best / 1 strategy in the differential mutation strategy to mutate the vectors of the population individuals, and adds the differential mutation strategy in the later stage of each round of search to increase the diversity of the population; the specific description is shown in formula (15); after obtaining the mutated vector, perform the crossover operation according to formula (16); then perform the selection operation according to formula (17), and retain the vector with a better objective function value as the next generation individual, continuously search and update the positions of the population individuals, and judge whether the optimal convergence is reached; if satisfied, obtain the optimal penalty factor C in the support vector machine model and the optimal parameter g of the kernel function; if not satisfied, return and continue to execute Step 3.2.

[0110] Step 4: Input the optimal penalty factor C and the optimal parameter g of the kernel function obtained in Step 3 into the support vector machine model to form an improved sailfish algorithm optimized support vector machine (ISFO - SVM) model, and use the optimized optimal penalty factor C and the optimal parameter g of the kernel function to train the support vector machine model.

[0111] Step 5: Input the prediction data into the ISFO - SVM model to obtain the prediction result, and perform inverse normalization on the prediction result;

[0112] Step 6: Display and output the IGBT junction temperature prediction result;

[0113] On the display screen of the computer, output the prediction graph of the IGBT junction temperature obtained by the ISFO - SVM model in Step 5, and display the error curve graph and error histogram of the IGBT junction temperature prediction using different models.

[0114] In addition, in the embodiment of the present invention, in order to better demonstrate the good performance of the proposed prediction model, the present invention selects the SVM model, the SFO - SVM model and the ISFO - SVM model for comparison, and selects RMSE, MAPE, and R2 as the evaluation indexes of the model; the evaluation index values of the IGBT junction temperature prediction under different models are shown in Table 1;

[0115]

[0116]

[0117]

[0118] where N is the number of test samples; T′ i is the predicted value; T i is the actual value.

[0119] Table 1 Analysis table of junction temperature prediction evaluation indexes

[0120]

[0121] Combined with Table 1 and Figure 5It is found that the RMSE, MAPE, and R2 of the ISFO-SVM model are all better than those of the other two models. The maximum, minimum, and average values of the RMSE and MAPE indicators of the SVM model differ relatively little, indicating that the SVM model has good robustness; the RMSE, MAPE, and R2 of the ISFO-SVM model are all better than those of the other two models, demonstrating good prediction performance and relatively high prediction accuracy for predicting the IGBT junction temperature; compared with the SFO-SVM model, the average value of RMSE of the ISFO-SVM model is reduced by 67.16%, the average value of MAPE is reduced by 56.52%, and R2 is increased by 0.71%, proving that the ISFO-SVM model has better performance in predicting the junction temperature and higher fitting degree.

[0122] According to Figure 6 It can be seen that the prediction errors of the sample individuals in the SVM model are distributed in the range of [-3°C, 3°C] accounting for 64.65%, and the absolute value of the error is in the range of [9°C, 12°C] accounting for 0.86%. The prediction result of the SFO-SVM model for the junction temperature is slightly improved compared with the SVM model and is more fitted to the true value of the junction temperature; by improving the SFO algorithm, the constructed ISFO-SVM model for predicting the junction temperature has relatively small and concentrated errors. The sample individuals mainly distributed in the interval [-3°C, 3°C) account for 91.38% of the total number of samples, and only 8.82% of the sample individuals have an absolute value of the error in the range of [3°C, 6°C); in summary, the IGBT junction temperature prediction method based on the ISFO-SVM model proposed in the present invention effectively improves the prediction accuracy and robustness of the IGBT junction temperature prediction model.

[0123] In the above IGBT junction temperature prediction method based on the ISFO-SVM model, the method of inputting data into the computer is a well-known method; the computer, monitor, and MATLAB computer software used are all obtained through commercial purchase.

[0124] The above is only the specific implementation manner of the method introduced in the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and all should be covered by the protection scope of the present invention.

Claims

1. A method for predicting the IGBT junction temperature based on the ISFO-SVM model, characterized in that, This method predicts the IGBT junction temperature by constructing an improved sailfish algorithm optimized support vector machine model (ISFO - SVM). The specific steps are as follows: Step 1: Obtain the IGBT junction temperature, saturation voltage drop, collector current, and aging times data through the IGBT power cycling aging acceleration test, and normalize the data. Step 2: Set the parameters of the improved sailfish algorithm and the support vector machine model. Step 3: Run the improved sailfish algorithm to obtain the optimal penalty factor in the support vector machine model and the optimal parameters of the kernel function. Step 4: Substitute the optimized optimal penalty factor and the optimal parameters of the kernel function into the support vector machine model, and train the support vector machine (ISFO - SVM) model optimized by the improved sailfish algorithm. Step 5: Input the prediction data into the ISFO - SVM model to obtain the prediction result, and denormalize the prediction result. Step 6: Display and output the IGBT junction temperature prediction result. Furthermore, the specific implementation method of Step 1 includes the following steps: Step 1.1: In order to obtain the degradation data of the IGBT module under the full life cycle, an IGBT power cycling aging test and a single - pulse test are designed to obtain a data set S containing the saturation voltage drop, collector current, junction temperature, and aging cycle times of the IGBT. Step 1.2, perform random shuffling and normalization on the obtained dataset S and divide it into training data and test data according to a ratio, taking the number of power cycles N, the saturation voltage drop Vce, and the collector current Ic as the input part of the model; the junction temperature T i as the output part of the model; Step 1.3: Normalize the above - mentioned data. Where A is the value of the variable to be normalized, A min is the minimum value of the variable, A max is the maximum value of the variable, A N is the value of the variable after normalization; Furthermore, the parameters set in Step 2 include: the population size in the improved sailfish algorithm, the maximum number of iterations of the improved sailfish algorithm, the search range of the penalty factor C in the support vector machine model, the range of the kernel function parameter g in the support vector machine model, and the population dimension in the improved sailfish algorithm. Furthermore, the specific implementation method of Step 3 includes the following steps: Step 3.1: Initialize the individual positions of the sailfish population and the sardine population of the improved sailfish algorithm, and calculate the objective function value of each individual. Step 3.2: Sort the positions and objective function values of each individual, and record the current optimal individual position and the optimal objective function value. Step 3.3: Introduce an adaptive non - linear iterative factor to update the positions of sailfish individuals, and introduce the Levy flight strategy to update the positions of sardine individuals. Step 3.4: Introduce the differential mutation strategy, continuously search for and update the positions of individuals in the sailfish population and the sardine population, and determine whether the optimal convergence is reached. If satisfied, obtain the best parameters of the model, output the optimal sailfish individual position X(x1, x2), which respectively correspond to the optimal penalty factor C of the SVM model and the optimal parameters of the kernel function g. If not satisfied, return and continue to execute Step 3.

2. Furthermore, in Step 3, the root - mean - square error is selected as the objective function to optimize the internal parameters of the support vector machine model using the training data. The improved sailfish algorithm determines whether the updated position is better than the original position according to the objective function value of the individual position at this time, and decides whether to use the updated position in the subsequent search process. The objective function is described as follows: wherein, is a random number in [0, 1]; The specific implementation method of Step 3.3 is as follows: Step 3.3.1: Introduce an adaptive non - linear iterative factor to update the positions of sailfish individuals. Since the sailfish individuals are randomly distributed at the initial stage of iteration, in order to improve the optimization ability of sailfish individuals, an adaptive non-linear iteration factor is introduced into the position update formula of sailfish to accelerate the optimization ability of sailfish individuals. Among them, the sailfish population is represented by X sF denotes; The update formula for the adaptive non-linear iteration factor of the $i$-th sailfish individual at the $t$-th iteration is described as follows: The position update formula of the sailfish is: In the formula, represents the position of the best individual in the sailfish population at the t-th iteration; represents the position of the best individual in the sardine population at the t-th iteration; represents the position of the sailfish individual to be updated at the t-th iteration; λ γ The definition of the coefficient is shown in formula (6): λ i = 2 × rand(0,1) × PD - PD (In Equation (6), PD represents the density of the prey population, which is described in detail by Equation (7): Where N SF represents the number of swordfish, and N S represents the number of sardines; Step 3.3.2, introducing the Levy flight strategy to update the positions of sardine individuals; among them, the sardine population is represented by X F denoted as The position update formula of the sardines in the original sailfish algorithm is shown in formula (8): In the formula, represents the adaptive non-linear iteration factor of the i-th sailfish individual at the t-th iteration, represents the position of the best individual in the sailfish population at the t-th iteration, represents the position of the sardine individual to be updated at the t-th iteration; AP represents the attack strength of the sailfish, and its detailed description is shown in formula (9): AP = A×(1 - 2×Itr×e) (9) In the formula, A and e represent the control coefficients of the sailfish attack strength, which linearly transform the sailfish attack strength from A to 0; When AP > 0.5, that is, when the attack strength of the sailfish is strong, update the positions of all sardines using formula (8); when AP < 0.5, at this time the attack strength of the sailfish is low, and only the positions of some sardines need to be updated; The range of the positions of some sardines is defined as follows: α = N S × AP(10) β = d i × AP(11) Where ɑ represents the number of updated sardines, β represents the number of dimensions for sardine update, and d i The number of variables at the i-th iteration; To improve the randomness of the sardine population and the diversity of the search space, the Levy flight strategy is introduced. At this time, the position update formula of the sardines is: where \(t\) is the current iteration number and \(d\) is the dimension of the position vector, represents the position of the sardine individual to be updated at the \(t\)-th iteration; The formula of Levy flight can be described as: In the formula, r1 and r2 are two random numbers, and the value range is [0, 1], β = 1.5, and σ can be calculated as: In the formula, Γ(x) = (x - 1)! Therefore, first calculate the attack strength AP of the sailfish according to formula (9). When AP > 0.5, that is, when the attack strength of the sailfish is strong, update the positions of all sardines using the improved sardine position update formula (12); when AP < 0.5, at this time the attack strength of the sailfish is low, then calculate the number and dimension of the sardines whose positions need to be updated according to formulas (10) and (11), and then update them using the improved sardine position update formula (12); The specific implementation method of step 3.4 is: To avoid falling into local convergence during the search process, the DE / current to best / 1 strategy in the differential mutation strategy is introduced to mutate the vectors of the population individuals, and the differential mutation strategy is added in the later stage of each round of search to increase the diversity of the population; the formula is expressed as follows: where p1≠p2≠p3, is the difference vector, F∈[0.1, 0.9] is the scaling factor, and h i,t is the mutation vector at the i-th position in the t-th search. After obtaining the mutation vector, the crossover operation is as follows: where u i,t is the crossover variable at the i-th search position, j0 is a random value in the dimension, and each crossover operation only involves one dimension of the individual. pCR ∈ [0, 1] is the crossover probability; Perform the selection operation, and retain the vector with the better objective function value as the next-generation individual. The selection operation is expressed as: According to the above formula, continuously search and update the positions of the population individuals, and judge whether the optimal convergence is reached. If satisfied, the best parameters of the model are obtained, and the optimal sailfish individual position X(x1, x2) is output, which respectively correspond to the optimal penalty factor C of the SVM model and the optimal parameters g of the kernel function; if not satisfied, return and continue to execute step 3.2; Furthermore, the specific implementation method of step 4 is: input the optimal penalty factor C and the optimal parameters g of the kernel function obtained in step 3 into the support vector machine model to form an improved sailfish algorithm optimized support vector machine (ISFO-SVM) model, and use the optimized optimal penalty factor C and the optimal parameters g of the kernel function to train the support vector machine model; Furthermore, the anti-normalization process in step 5 is as shown in formula (18): where T′ i is the predicted value of the junction temperature after anti-normalization, is the normalized predicted value of the junction temperature obtained in Step 4, T max and T min are the maximum and minimum values of the junction temperature variable in Step 1.3; Furthermore, the specific implementation method of step 6 is: output the prediction graph of the IGBT junction temperature by the ISFO-SVM model obtained in step 5 on the computer display screen, and display the error curve graph and error histogram of the IGBT junction temperature prediction using different models.

2. The IGBT junction temperature prediction method based on the ISFO-SVM model according to claim 1, characterized in that: The IGBT aging test dataset was designed by designing an accelerated IGBT power cycle aging test and a single-pulse test to obtain data on the IGBT's saturation voltage drop, collector current, junction temperature, and number of aging cycles. In this test case, the IGBT model is MMG75S-120B, with a rated value of 1200V / 75A. The junction temperature fluctuation ΔTj was set to 100°C, and a 5% increase in the saturation voltage drop Vce was used as the IGBT failure criterion.

3. The IGBT junction temperature prediction method based on the ISFO-SVM model according to claim 2, characterized in that: The IGBT aging test data set is designed by IGBT power cycle aging test and single pulse test; Referring to the IEC60068-2-14 JEDEC standard for power cycling tests set by the International Electrotechnical Commission (IEC), an IGBT power cycling aging test circuit was built. The basic test steps of this test are as follows: (1) First, close the switch S, set the programmable constant current source output current to 50A, and the gate output drive voltage to 15V, so that the upper tube of the IGBT power module is turned on. The test module generates power loss, causing the junction temperature and case temperature to rise. The initial case temperature is set to 40℃. (2) The temperature change is monitored by the temperature sensor built into the bottom of the module. When the highest shell temperature reaches 90°C, the switch S is disconnected and the air cooling radiator is turned on to quickly cool the IGBT power module to a shell temperature of 40°C. The shell temperature fluctuation power cycle aging is completed once. (3) Repeat the above steps (1) and (2) until the module is close to the failure standard and stop the test. This test completes a total of 6000 power cycle tests, and pauses every 1000 power cycle tests. Remove the module and place it in a constant temperature box for a short-time single pulse test; record the collector current and saturation voltage drop value of the IGBT module at different junction temperatures; Considering that after completing one stage of accelerated aging every 1000 times, it is necessary to obtain the saturation voltage drop, junction temperature, and collector current of the IGBT power module in the current aging stage, a single-pulse test platform was built. The temperature adjustment range of the constant temperature box was set to [0°C, 100°C], with a temperature adjustment interval of 10°C; the collector current setting range was set to [25A, 70A], with an adjustment interval of 5A. The specific steps are as follows: (1) Place the IGBT power module to be tested in a constant temperature box and adjust the temperature of the constant temperature box. When the temperature of the constant temperature box stabilizes, it is considered that the module has reached thermal equilibrium. (2) When the thermal equilibrium condition is met, adjust the set values of the constant temperature box temperature and collector current, and record the saturation voltage drop, junction temperature, and collector current data under the corresponding power cycle aging times.

Images