A method for estimating junction temperature of hybrid devices

Abstract

This invention provides a method for estimating the junction temperature of a hybrid device, comprising: acquiring a historical voltage and current signal dataset of the hybrid device; acquiring a historical junction temperature time series of the hybrid device; based on the historical junction temperature time series and the historical voltage and current signal dataset of the hybrid device, processing them accordingly to obtain a junction temperature estimation model for each type of device included in the hybrid device; determining a junction temperature estimation model for the hybrid device based on the junction temperature estimation model corresponding to each type of device; collecting the current voltage and current data of the power device under test in the hybrid device, inputting it into the junction temperature estimation model of the hybrid device, and outputting the estimated junction temperature value of the power device under test. This method successfully solves the problem of inaccurate junction temperature estimation caused by differences in thermal characteristics and thermal dynamic response of power devices made of different materials in hybrid devices, and significantly improves the generalization ability of the estimation model.

Description

Technical Field

[0001] The present invention relates to the field of semiconductors, and in particular to the field of junction temperature estimation of semiconductor hybrid devices. Background Art

[0002] Hybrid devices combine different types of semiconductor devices, usually including devices made of materials such as silicon (Si) and silicon carbide (SiC), to take advantage of their respective advantages to meet the requirements of specific applications. These devices have significant advantages in power electronics and other high-performance applications, and reduce costs. However, obtaining the junction temperature of different monomer devices in a hybrid device is a complex technical issue that requires comprehensive consideration of multiple factors, as it directly affects the reliability and performance of the power electronics system.

[0003] In hybrid power devices, Si-based and SiC-based devices have different thermal characteristics. For example, the thermal resistance and thermal capacitance of silicon insulated gate bipolar transistors (Si IGBTs) may be significantly different from those of silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs). This requires modeling for each type of device separately, which makes the junction temperature estimation more complicated and requires considering thermal coupling and thermal conduction between different devices. In addition, the dynamic thermal response of Si IGBTs and SiC MOSFETs may be quite different. IGBTs usually have higher thermal capacitance and slower thermal response, while MOSFETs respond faster. Therefore, under dynamic operating conditions, the junction temperature estimation must simultaneously consider the different thermal inertia of the two devices, which makes the thermal model more complicated. Some existing junction temperature estimation methods can only estimate the junction temperature for a single type of semiconductor device. Even for the same type of semiconductor, the estimation model has low junction temperature estimation accuracy due to device process errors, and the estimation model does not have generalization. However, there are many types of semiconductor devices. Compared with a single device, the operation process of estimating the junction temperature of a hybrid device will significantly increase many tedious links and complex work. Summary of the invention

[0004] The present invention provides a junction temperature estimation method for a hybrid power device, which is used to solve the problem that the junction temperature estimation model of the junction temperature estimation method in the prior art can only perform a single estimation and cannot be generalized.

[0005] To achieve the above object, the present invention provides a junction temperature estimation method for a hybrid power device under test, comprising:

[0006] Obtain historical voltage and current signal datasets of hybrid devices;

[0007] Obtaining a historical junction temperature time series of the hybrid device;

[0008] Based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device is constructed, and the nonlinear autoregressive neural network model corresponding to each type of device is trained to obtain target training model parameters corresponding to each type of device; wherein the target training model parameters are parameters obtained after the nonlinear autoregressive neural network model is trained on the historical junction temperature time series and the historical voltage and current signal data sets of each type of device included in the hybrid device;

[0009] Using a preset optimization algorithm, the target training model parameters corresponding to each type of device are optimized to obtain a junction temperature estimation model corresponding to each type of device;

[0010] Determining a junction temperature estimation model of the hybrid device based on the junction temperature estimation model corresponding to each type of device;

[0011] The current voltage and current data of the power device to be tested in the hybrid device are collected and input into the junction temperature estimation model of the hybrid device, and the junction temperature estimation value of the power device to be tested is output.

[0012] Preferably, based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device is constructed, including:

[0013] Determining model parameters of each type of device included in the hybrid device based on a historical junction temperature time series of the hybrid device and a historical voltage and current signal data set of the hybrid device;

[0014] Based on the model parameters of each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device included in the hybrid device is constructed.

[0015] Preferably, based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device, for each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device is constructed, including:

[0016] Using the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, construct a formula of the nonlinear autoregressive neural network model corresponding to each type of device;

[0017] The process of constructing the nonlinear autoregressive neural network model formula is as follows:

[0018] In the formula, represents the historical junction temperature time series, represents external input, where , , is the autoregressive order, which indicates the number of past observations considered by the model; is the order of the external input, which indicates the number of external input histories considered by the model; For neural network.

[0019] Among them, the is a neural network, the neural network It includes the input layer, hidden layer, and output layer connected in sequence, specifically:

[0020] Input layer: The input layer is used to receive Dimensional input;

[0021] Hidden layer: The hidden layer Node input It can be expressed as:

[0022] In the formula, The input layer nodes to the hidden layer The weight of each node; The input layer The input of each node; is the hidden layer The bias term of the node. Then the hidden layer Output of the node Through the activation function The calculation results are:

[0023] Output layer: The output layer is used to output the predicted next time step The output layer has only one node, and its input It can be expressed as:

[0024] In the formula, is the hidden layer The weight from each node to the output layer node; is the hidden layer The output of each node; is the bias term of the output layer; is the number of neurons in the hidden layer.

[0025] Preferably, the preset optimization algorithm is an annealing Langevin Dynamics algorithm;

[0026] The preset optimization algorithm is used to optimize the target training model parameters to obtain the junction temperature estimation model parameters of each type of device, specifically:

[0027] The target training model parameters are optimized by using an annealing Langevin Dynamics algorithm to obtain junction temperature estimation model parameters corresponding to each type of device.

[0028] Preferably, the annealing Langevin Dynamics algorithm is used to optimize the target training model parameters to obtain the junction temperature estimation model parameters corresponding to each type of device, including:

[0029] Step S501: Determine the function construction process of annealing Langevin Dynamics optimization, specifically:

[0030] Given a nonlinear function ;

[0031] Determine the function update formula as In the formula, Weight and bias , It is The parameter vector of the step; It is The parameter vector of the step; yes right The Jacobian matrix of is the Jacobian matrix The transpose of is the identity matrix; It is the adjustment parameter in the LM algorithm, which is used to control the step size and the step length; is the observed target value; is the predicted value;

[0032] Step S502: Based on the , introduce a Gaussian distributed random disturbance, namely:

[0033] in, is a random vector drawn from a Gaussian distribution, is the variance of the control disturbance size, according to the temperature Adjustment, the formula is as follows:

[0034] In the formula, is the initial disturbance variance; is the initial temperature; It is Step temperature;

[0035] Step S503: Using the temperature drop coefficient Gradually reduce the temperature , the specific process is as follows:

[0036] In the formula, It is Step temperature;

[0037] Step S504: Determine the temperature Whether the preset convergence condition is met, if so, the iteration is stopped, and the model corresponding to the target training model parameters after convergence is used as the junction temperature estimation model corresponding to each type of device;

[0038] If not satisfied, repeat steps S501-S504.

[0039] Preferably, the junction temperature estimation model is obtained based on the junction temperature estimation model parameters corresponding to each type of device.

[0040] Preferably, the collecting current voltage and current data of the power device to be tested in the hybrid device and inputting them into the junction temperature estimation model, and outputting the junction temperature estimation value of the power device to be tested comprises:

[0041] Collecting current voltage and current data of the power device to be tested in the hybrid device, and inputting the data into a junction temperature estimation model of the hybrid device;

[0042] Determining a junction temperature estimation model corresponding to the power device to be tested based on current voltage and current data of the power device to be tested;

[0043] Based on the junction temperature estimation model corresponding to the power device to be tested, the junction temperature estimation value of the power device to be tested is output.

[0044] The present invention has the following beneficial effects:

[0045] The junction temperature estimation method of the hybrid device provided in this application collects the voltage and current signals in the hybrid device that can reflect the working state of each independent device as the input of the nonlinear autoregressive neural network model (NARX neural network) estimation model, so as to avoid the inaccurate estimation problem caused by the difference in device thermal characteristics and thermal dynamic response caused by the difference in Si and SiC materials. The training process is optimized using a preset optimization algorithm to solve the problem that the loss function is complex and multi-peaked and prone to local optimal solutions, improve the generalization of the model, and enable the model to maintain a high estimation accuracy for the test device.

[0046] In addition to the above-described purposes, features and advantages, the present invention has other purposes, features and advantages. The present invention will be further described in detail below with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0048] Figure 1 It is a schematic flow chart of the junction temperature estimation method of the hybrid device provided by the present invention;

[0049] FIG. 2 (a) is a diagram of a preferred embodiment of the present invention. Schematic diagram of electrical signal data;

[0050] FIG2( b ) is a diagram of the acquisition IGBT of the preferred embodiment of the present invention. Schematic diagram of electrical signal data;

[0051] Figure 3 The neural network of the preferred embodiment of the present invention Schematic diagram of the structure;

[0052] Figure 4 This is an experimental diagram of the training result obtained by using the electrical signal data of the device IGBT5 in the experiment of the present invention;

[0053] FIG5(a) is an experimental diagram showing the result of the junction temperature estimation model of the present application for estimating the junction temperature of IGBT2;

[0054] FIG5( b ) is an experimental diagram of the estimation error results of the junction temperature estimation model of the present application for the IGBT2 junction temperature estimation;

[0055] FIG6( a ) is an experimental diagram showing the result of the junction temperature estimation model of the present application for estimating the junction temperature of IGBT3;

[0056] FIG6( b ) is an experimental diagram of the estimation error results of the junction temperature estimation model of the present application for the IGBT3 junction temperature estimation;

[0057] FIG. 7( a ) is an experimental diagram showing the result of the junction temperature estimation model of the present application for estimating the junction temperature of IGBT4;

[0058] FIG. 7( b ) is an experimental diagram of the estimation error results of the junction temperature estimation model of the present application for IGBT4 junction temperature estimation. DETAILED DESCRIPTION

[0059] The embodiments of the present invention are described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0060] The semiconductor junction temperature refers to the temperature of the PN junction in the semiconductor device, which is crucial to the device performance and reliability. Therefore, in the process of semiconductor operation, the importance of estimating its junction temperature is self-evident.

[0061] In hybrid power devices, Si-based and SiC-based devices have different thermal characteristics. For example, the thermal resistance and thermal capacitance of silicon insulated gate bipolar transistors (Si IGBTs) may be significantly different from those of silicon carbide metal-oxide-semiconductor field-effect transistors (SiC MOSFETs). This requires modeling for each type of device separately, which makes the junction temperature estimation more complicated and requires considering thermal coupling and thermal conduction between different devices. In addition, the dynamic thermal response of Si IGBTs and SiC MOSFETs may be quite different. IGBTs usually have higher thermal capacitance and slower thermal response, while MOSFETs respond faster. Therefore, under dynamic operating conditions, the junction temperature estimation must simultaneously consider the different thermal inertia of the two devices, which makes the thermal model more complicated. Some existing junction temperature estimation methods can only estimate the junction temperature for a single type of semiconductor device. Even for the same type of semiconductor, the estimation model has low junction temperature estimation accuracy due to device process errors, and the estimation model does not have generalization. However, there are many types of semiconductor devices. Compared with a single device, the operation process of estimating the junction temperature of a hybrid device will significantly increase many tedious links and complex work.

[0062] In order to solve the above problems, an embodiment of the present invention provides a junction temperature estimation method of a hybrid device, and the implementation steps are as follows: Figure 1 As shown:

[0063] Step S101: Acquire a historical voltage and current signal data set of a hybrid device;

[0064] Hybrid devices are electronic devices that combine multiple different types of semiconductor elements and other related components. For example, in the field of power electronics, a common hybrid device may contain power transistors, diodes, resistors, capacitors, etc., which work together to achieve a specific function.

[0065] In the present application, a sensor system is used to collect historical voltage and current signal data sets of hybrid devices. It should be noted that when collecting historical voltage and current signal data sets of hybrid devices, the historical voltage and current signal parameters that need to be collected are determined based on the type of hybrid device. Specifically: the sensor system needs to first determine the type of hybrid device, and then determine the historical voltage and current signal parameters that need to be collected based on the type of hybrid device. Because different hybrid devices have different device characteristics, the voltage and current generated are also different. In other words, each hybrid device has unique voltage and current signal parameters that closely match its own characteristics. For example, if the hybrid device collected is an IGBT (Insulated Gate Bipolar Transistor), the historical collector-emitter voltage of the IGBT should be collected at this time. , Historical gate-emitter voltage and the historical collector current If the hybrid device being collected is a silicon carbide MOSFET (metal-oxide-semiconductor field-effect transistor), the historical drain-source voltage is collected. , historical gate-source voltage and the historical drain current .

[0066] This application collects historical voltage and current signal data sets of hybrid devices because in hybrid power devices, Si and SiC devices have different junction temperatures of independent devices due to differences in thermal characteristics and thermal dynamic responses when working. By recording voltage and current data at different time points during the actual operation of the device, dynamic information of its working state can be obtained. These historical data provide a solid foundation for junction temperature estimation. The advantages are mainly reflected in the following aspects: First, it can reflect the electrical characteristics of the device under real working conditions. Junction temperature estimation based on these data can more accurately simulate actual working conditions, and the results are more reliable than theoretical calculations or measurements at a single moment. Secondly, the historical data set covers information under different working modes and environmental conditions, which helps to establish a more comprehensive junction temperature prediction model, improve the adaptability and generalization ability of the model, so as to more accurately predict the junction temperature of hybrid devices in different scenarios, and provide strong support for device reliability assessment, thermal management design, and life prediction, etc., to ensure the stable operation and performance of hybrid devices in complex working environments.

[0067] Step S102: Acquire the historical junction temperature time series of the hybrid device;

[0068] The historical junction temperature time series refers to the data sequence obtained by continuously measuring and recording the junction temperature of semiconductor devices (such as transistors, diodes, IGBTs, etc.) over a period of time. Junction temperature refers to the temperature of the PN junction in a semiconductor device, which is a key factor affecting the performance, reliability and life of the device. For example, during the operation of a hybrid device, the junction temperature is recorded at a certain time interval (such as every second, every minute, etc.). After a period of time (such as several hours, days or even longer), a historical data sequence reflecting the change of junction temperature over time is obtained.

[0069] Generally speaking, in order to improve the accuracy of the model's later estimation, the time interval of the historical data sequence generally collected in this application should not be too long. Because in hybrid devices, many physical quantities such as voltage, current, temperature, etc. may change dynamically. If the collection time interval is too long, the rapid changes of these physical quantities will be missed, resulting in the collected data not accurately reflecting the actual working state of the device, which will in turn affect the measurement accuracy. As shown in Figure 2, this application applies a square wave signal at the IGBT gate. Figure (a) collects the data of four IGBT power devices. Electrical signal data, Figure (b) collects the electrical signal data of four IGBT power devices. Electrical signal data.

[0070] Step S103: Based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, construct the nonlinear autoregressive neural network model corresponding to each type of hybrid device, and train the nonlinear autoregressive neural network model to obtain target training model parameters corresponding to each type of device; wherein the target training model parameters are parameters obtained after the nonlinear autoregressive neural network model is trained on the historical junction temperature time series and the historical voltage and current signal data sets of each type of device included in the hybrid device;

[0071] The construction of a nonlinear autoregressive neural network model (NARX network model) mentioned in this application, in order to reduce the difficulty of reading, the NARX network model is used in the following description to refer to the nonlinear autoregressive neural network model. The NARX network model has significant advantages in building models in the semiconductor field. For junction temperature, it can effectively track its dynamic characteristics that change over time and various factors. It can be accurately modeled based on a variety of related inputs, such as the relationship between current, voltage, and heat dissipation conditions and junction temperature, thereby helping to optimize the design and working conditions of semiconductor devices, ensure stable and reliable operation of devices, improve performance and extend service life.

[0072] The NARX network model uses past observations (the time series’ own delay) as input, which in this application is the junction temperature value, and predicts future values ​​by learning nonlinear mapping. At the same time, in addition to the time series’ own delayed input, the NARX network model also accepts external inputs, such as the electrical signal of the IGBT mentioned above. , gate-emitter voltage and collector current External input enables the network to more fully consider the system's environmental conditions, thereby improving prediction accuracy and generalization capabilities.

[0073] In the present application, since each type of device included in the hybrid device is different, it is necessary to determine the model parameters of the type of device included in the hybrid device after obtaining the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device, and then, based on the model parameters of each type of device included in the hybrid device, for each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device is constructed. That is to say, in the present application, in order to facilitate the construction of the model and the processing of the data by the model, it is necessary to classify, clean, or standardize and normalize the data obtained in the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device, so as to obtain the model parameters that can be processed by each type of device, and then construct a nonlinear autoregressive neural network model corresponding to each type of device according to the model parameters of each type of device. The specific construction process is as follows:

[0074] In the present application, based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device, for each type of device contained in the hybrid device, the nonlinear autoregressive neural network model corresponding to each type of device is constructed, and the specific process is as follows;

[0075] Using the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device to construct a formula for the nonlinear autoregressive neural network model;

[0076] The process of constructing the nonlinear autoregressive neural network model is as follows:

[0077] (1)

[0078] Before explaining formula (1), it should be noted that the premise for formula (1) is: Assume that the junction temperature time series It can be represented by autoregression of past observations and is affected by external inputs impact.

[0079] In the formula, represents the historical junction temperature time series (expressed by autoregression of past observations), represents external input, where , , is the autoregressive order, which indicates the number of past observations considered by the model; is the order of the external input, which indicates the number of external input histories considered by the model; For neural network.

[0080] Among them, the For a neural network, such as Figure 3 As shown: the structure of the neural network 300 includes an input layer 310, a hidden layer 320, and an output layer 330 connected in sequence, specifically:

[0081] Input layer 310: The input layer 310 is used to receive Dimensional input;

[0082] Hidden layer 320: Hidden layer 320 Node input It can be expressed as:

[0083] (2)

[0084] In the formula, The input layer nodes to the hidden layer The weight of each node; The input layer The input of each node (i.e., past observations or external input); is the hidden layer The bias term of the node. Then the hidden layer Output of the node Through the activation function The calculation results are:

[0085] (3)

[0086] Output layer 330: The output layer 330 is used to output the predicted next time step. The output layer has only one node, and its input It can be expressed as:

[0087] (4)

[0088] In the formula, is the hidden layer The weight from each node to the output layer node; is the hidden layer The output of each node; is the bias term of the output layer; is the number of neurons in the hidden layer.

[0089] After the NARX network model corresponding to each type of device is constructed according to the above formula, the NARX network model of each type of device constructed needs to be trained. The historical junction temperature time series of a single type of device in the hybrid device and the historical voltage and current signal data set of a single type of device in the hybrid device are used to perform corresponding training on the constructed NARX model, so as to obtain the target training model parameters of the single type of device. That is to say, the target training model parameters of each type of device in the hybrid device of the present application are the parameters obtained by training the historical junction temperature time series and the historical voltage and current signal data set of each type of device contained in the hybrid device by the nonlinear autoregressive neural network model. For each type of device in the hybrid device, its historical junction temperature time series is used as a part of the information source of the target output, and the historical voltage and current signal data set is used as an exogenous input variable. By inputting these data into the model for training, the model will continuously adjust its own parameters according to the specific training algorithm and objective function to better fit these data and capture the patterns and laws in the data.

[0090] Step S104: using a preset optimization algorithm to optimize the target training model parameters corresponding to each type of device. After obtaining the training model parameters of each type of device, a combined junction temperature estimation model is obtained;

[0091] Since the target training model parameters in this application may have the following problems: there are certain errors, the training data may be overfitted, the performance of new data is poor, and the consumption of computing resources is increased. Therefore, a preset optimization algorithm is used in this application to optimize the target training model parameters to obtain a junction temperature estimation model. The selection criteria of the preset optimization algorithm include problem characteristics, convergence speed, hyperparameters, and resource requirements. In other words, in this application, the preset optimization algorithm can be selected according to the above criteria based on the hybrid power device that needs to be estimated, and no specific restrictions are made here.

[0092] It should be noted that, for the sake of ease of understanding, the preset optimization algorithm used in this application is the annealing Langevin Dynamics algorithm. That is, the annealing Langevin Dynamics algorithm is used in this application to optimize the target training model parameters to obtain the junction temperature estimation model.

[0093] The training of NARX neural network usually adopts supervised learning method, which adjusts the network parameters including weights by minimizing the error between the predicted value and the actual observed value. and bias Therefore, when optimizing the target training model, the weights and biases are mainly adjusted. The specific process is as follows:

[0094] Step S501: Determine the function construction process of annealing Langevin Dynamics optimization, specifically:

[0095] Given a nonlinear function ;

[0096] Determine the function update formula as (5) In the formula, Weight and bias , It is The parameter vector of the step; It is The parameter vector of the step; yes right The Jacobian matrix of is the Jacobian matrix The transpose of is the identity matrix; It is the adjustment parameter in the LM algorithm, which is used to control the step size and the step length; is the observed target value; is the predicted value;

[0097] It should be noted that formula (5) is the update formula used by the Levenberg-Marquardt (LM) algorithm. The Levenberg-Marquardt (LM) algorithm is an optimization algorithm specifically used for nonlinear least squares problems and can be used to fit the nonlinear relationship between the neural network output and the target.

[0098] In this step, given a nonlinear function , the goal is to adjust the parameters To minimize the residual sum of squares, Weight and bias The goal of the LM algorithm is to iteratively adjust the parameters The value of is combined with the gradient information and the approximate Hessian matrix to achieve optimization.

[0099] Since the loss function of neural networks is often complex and multi-peaked, the LM algorithm is not a global optimization algorithm and may fall into a local optimal solution. The simulated annealing algorithm is a global optimization algorithm that gradually reduces the temperature during the search process to find the global optimal solution or a solution close to the optimal solution in the solution space. It can help jump out of the local minimum point and increase the diversity of the search. Langevin Dynamics is an optimization method based on random processes that simulates the movement of physical particles in the gradient field by introducing randomness. During the optimization process, it uses a random process similar to Brownian motion, which helps to explore uniformly in the solution space, making it more likely to find the global optimal solution or a good local solution. By combining the two, it helps to allow the system to explore a larger configuration space in the early stage, and focus on the potential optimal solution in the later stage. Therefore, the annealing Langevin Dynamics algorithm is introduced for optimization.

[0100] Step S502: Based on the , introduce Gaussian distributed random perturbations to simulate the random motion in Langevin Dynamics, that is:

[0101] (6)

[0102] In the formula, is a random vector drawn from a Gaussian distribution, is the variance of the control disturbance size, according to the temperature Adjustment, the formula is as follows:

[0103] (7)

[0104] In the formula, is the initial disturbance variance; is the initial temperature; It is Step temperature;

[0105] Step S503: Using the temperature drop coefficient Gradually reduce the temperature , the specific process is as follows:

[0106] (8)

[0107] In the formula, It is Step temperature;

[0108] It should be noted that the temperature drop coefficient The value of is usually determined according to the scale and characteristics of the problem, ranging from The value of the temperature drop coefficient is inversely proportional to the scale of the problem. In other words, the smaller the scale of the problem, the closer the value of the temperature drop coefficient is to 1.

[0109] Step S504: Determine the temperature Whether the preset convergence condition is met, if so, the iteration is stopped, and the model corresponding to the target training model parameters after convergence is used as the junction temperature estimation model corresponding to each type of device;

[0110] If not satisfied, repeat the above steps S501-S504.

[0111] It should be noted that each time the temperature is obtained After that, you need to Whether the preset convergence conditions are met is judged to determine the subsequent operations. If the convergence conditions are met, it means that the probability of the annealing algorithm accepting a worse solution will also decrease. This is similar to the gradual stabilization of a substance during the cooling process, from liquid to solid. Just like in the solution space of the optimization function, when the temperature drops to a certain level, the annealing algorithm will perform a more detailed search near the currently found better solution, and gradually converge to the global optimal solution or the better local optimal solution. At this point, it means that the annealing algorithm has completed the optimization processing of the target training model, and the model corresponding to the converged target training model parameters at this time is used as the junction temperature estimation model corresponding to each type of device.

[0112] It should be noted that the preset condition here is that the residual reaches a sufficiently small threshold or reaches the maximum number of iterations.

[0113] If not, repeat the above steps 501-504. If the preset convergence condition is not met, steps S501-504 are repeated until the temperature reaches The preset convergence conditions are met.

[0114] It should be noted that when steps 501-504 are processed once, the parameters of the target training model of a single type of device are optimized.

[0115] Step S105: determining a junction temperature estimation model of the hybrid device based on the junction temperature estimation model corresponding to each type of device;

[0116] Based on the junction temperature estimation model parameters corresponding to each type of device, these models can be further combined to obtain the junction temperature estimation model of the hybrid device. First, analyze the role and mutual influence of different types of devices in the hybrid device, and then combine their respective junction temperature estimation results, using weighted average or more complex fusion algorithms to comprehensively consider various factors. At the same time, based on the junction temperature estimation model parameters corresponding to each type of device, these parameters are reasonably integrated and adjusted so that they can better adapt to the overall characteristics of the hybrid device, thereby accurately determining the junction temperature estimation model of the hybrid device and providing a reliable basis for the thermal management and performance evaluation of the hybrid device.

[0117] Step S106: collecting current voltage and current data of the power device to be tested in the hybrid device, inputting them into the junction temperature estimation model of the hybrid device, and outputting the junction temperature estimation value of the power device to be tested.

[0118] In the process of in-depth analysis of hybrid devices, we need to accurately evaluate the junction temperature of the power device to be tested. First, the current voltage and current data of the power device to be tested in the hybrid device are collected and input into the junction temperature estimation model of the hybrid device. The junction temperature estimation model in this application is carefully constructed and optimized, and can perform preliminary analysis based on the input data. Then, based on the current voltage and current data of the power device to be tested, a more accurate junction temperature estimation model corresponding to the power device to be tested is further determined. Finally, the junction temperature estimation model corresponding to the power device to be tested is used to determine the current power device to be tested, and then the junction temperature estimation value of the power device to be tested is output. In this way, we can understand the junction temperature state of the power device to be tested in real time, provide guarantee for its safe operation, and also help to optimize and adjust the performance of the device to ensure the overall stability and efficient operation of the hybrid device.

[0119] In order to verify the effectiveness of the junction temperature estimation model training of the hybrid device of the present application, the electrical signal data of IGBT5 in the hybrid device in Figure 2 is used for training. The training results are shown in Figure 2. Figure 4 As shown, the error of the training estimated junction temperature is within 0~3°C, which is enough to show that the junction temperature estimation model in this application has high accuracy.

[0120] In addition, due to factors such as process errors, the same power devices will get different signal feedback under the same working conditions. Therefore, when verifying the model experiment of this application, the estimated model is used to estimate the junction temperature of IGBT2, IGBT3 and IGBT4 in the hybrid device, respectively, to verify the generalization of the model parameters of a single type of device in the junction temperature estimation model. The experimental results are shown in Figures 5 to 7. As can be seen from the figure, there are differences in the observed junction temperature rise process, steady-state maximum junction temperature and gradient fluctuation amplitude of IGBT2, IGBT3 and IGBT4 in the hybrid device, but the estimation model can obtain accurate estimation results. The estimated junction temperature error range of IGBT2 is -3.59~2.08℃; the estimated junction temperature error range of IGBT3 is -5.53~2.29℃; the estimated junction temperature error range of IGBT4 is -3.91~1.74℃, and IGBT has only one large error value of -7.68. Therefore, the generalization of the model is proved and the method can be effectively applied to junction temperature monitoring of hybrid devices.

[0121] In summary, in this application, the voltage and current signals in the hybrid device that can directly reflect the working state of each independent device are collected as the input of the NARX neural network estimation model, which avoids the problem of inaccurate junction temperature estimation of independent devices due to differences in thermal characteristics and thermal dynamic responses when different devices in the hybrid device are working. At the same time, the annealing LangevinDynamics algorithm is used to optimize the training process, allowing the system to explore a larger configuration space in the early stage of finding the optimal parameters, and focusing on the potential optimal solution in the later stage, solving the problem that the loss function is complex and multi-peaked and easily falls into the steady state of the local optimal solution, improving the generalization of the model, and still having a high estimation accuracy for the test device.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for estimating junction temperature of a hybrid device, characterized in that: include: Obtain historical voltage and current signal datasets of hybrid devices; Obtaining a historical junction temperature time series of the hybrid device; Based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device is constructed, and the nonlinear autoregressive neural network model corresponding to each type of device is trained to obtain target training model parameters corresponding to each type of device; wherein the target training model parameters are parameters obtained after the nonlinear autoregressive neural network model is trained on the historical junction temperature time series and the historical voltage and current signal data sets of each type of device included in the hybrid device; Using a preset optimization algorithm, the target training model parameters corresponding to each type of device are optimized to obtain a junction temperature estimation model corresponding to each type of device; Determining a junction temperature estimation model of the hybrid device based on the junction temperature estimation model corresponding to each type of device; Collecting current voltage and current data of a power device to be tested in a hybrid device, inputting the data into a junction temperature estimation model of the hybrid device, and outputting an estimated junction temperature value of the power device to be tested; Wherein, determining the junction temperature estimation model based on the junction temperature estimation model parameters corresponding to each type of device includes: The junction temperature estimation model is obtained based on the junction temperature estimation model parameters corresponding to each type of device.

2. The junction temperature estimation method according to claim 1, characterized in that: The method comprises: constructing a nonlinear autoregressive neural network model corresponding to each type of device contained in the hybrid device based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device, including: Determining model parameters of each type of device included in the hybrid device based on a historical junction temperature time series of the hybrid device and a historical voltage and current signal data set of the hybrid device; Based on the model parameters of each type of device included in the hybrid device, a nonlinear autoregressive neural network model corresponding to each type of device included in the hybrid device is constructed.

3. The method according to claim 1, characterized in that: The method of constructing a nonlinear autoregressive neural network model corresponding to each type of device contained in the hybrid device based on the historical junction temperature time series of the hybrid device and the historical voltage and current signal data set of the hybrid device comprises: Using the historical junction temperature time series of the hybrid device and the historical voltage and current signal data sets of the hybrid device, for each type of device included in the hybrid device, construct a formula of the nonlinear autoregressive neural network model corresponding to each type of device; The process of constructing the nonlinear autoregressive neural network model formula is as follows: In the formula, represents the historical junction temperature time series, represents external input, where , , is the autoregressive order, which indicates the number of past observations considered by the model; is the order of the external input, which indicates the number of external input histories considered by the model; for neural networks; Among them, the is a neural network, the neural network It includes the input layer, hidden layer, and output layer connected in sequence, specifically: Input layer: The input layer is used to receive Dimensional input; Hidden layer: The hidden layer Node input It can be expressed as: In the formula, The input layer nodes to the hidden layer The weight of each node; The input layer The input of each node; is the hidden layer The bias term of the node; then the hidden layer Output of the node Through the activation function The calculation results are: ， Output layer: The output layer is used to output the predicted next time step , the output layer has only one node, and its input It can be expressed as: In the formula, is the hidden layer The weight from each node to the output layer node; is the hidden layer The output of each node; is the bias term of the output layer; is the number of neurons in the hidden layer.

4. The method according to claim 1, characterized in that: The preset optimization algorithm is the annealing LangevinDynamics algorithm; The preset optimization algorithm is used to optimize the target training model parameters to obtain the junction temperature estimation model parameters corresponding to each type of device, specifically: The target training model parameters are optimized by using an annealing Langevin Dynamics algorithm to obtain junction temperature estimation model parameters corresponding to each type of device.

5. The method according to claim 4, characterized in that The annealing Langevin Dynamics algorithm is used to optimize the parameters of the target training model to obtain the junction temperature estimation model corresponding to each type of device, including: Step S501: Determine the function construction process of annealing Langevin Dynamics optimization, specifically: Given a nonlinear function ; The function update formula is determined as: In the formula, Weight and bias , It is The parameter vector of the step; It is The parameter vector of the step; yes right The Jacobian matrix of is the Jacobian matrix The transpose of is the identity matrix; It is the adjustment parameter in the LM algorithm, which is used to control the step size and the step length; is the observed target value; is the predicted value; Step S502: Based on the , introduce a Gaussian distributed random disturbance, namely: in, is a random vector drawn from a Gaussian distribution, is the variance of the control disturbance size, according to the temperature Adjustment, the formula is as follows: In the formula, is the initial disturbance variance; is the initial temperature; It is Step temperature; Step S503: Using the temperature drop coefficient Gradually reduce the temperature , the specific process is as follows: In the formula, It is Step temperature; Step S504: Determine the temperature Whether the preset convergence condition is met, if so, the iteration is stopped, and the model corresponding to the target training model parameters after convergence is used as the junction temperature estimation model corresponding to each type of device; If not satisfied, repeat steps S501-S504.

6. The method according to claim 1, characterized in that The collecting current voltage and current data of the power device to be tested in the hybrid device, inputting the data into the junction temperature estimation model, and outputting the junction temperature estimation value of the power device to be tested comprises: Collecting current voltage and current data of the power device to be tested in the hybrid device, and inputting the data into a junction temperature estimation model of the hybrid device; Determining a junction temperature estimation model corresponding to the power device to be tested based on current voltage and current data of the power device to be tested; Based on the junction temperature estimation model corresponding to the power device to be tested, the junction temperature estimation value of the power device to be tested is output.

Images