IGBT module rapid online junction temperature estimation method based on neural network

Abstract

The invention relates to an IGBT module rapid online junction temperature estimation method based on a neural network, and the method comprises the following steps: 1) obtaining on-off loss data samples of a large number of IGBT modules through an automatic double-pulse testing machine platform, and building a loss prediction model based on an improved particle swarm optimization BP neural network; 2) based on the transient thermal resistance curve of the Foster network model, obtaining the transient thermal resistance curve of a certain number of samples of the same IGBT module product through a thermal resistance test machine, and realizing the prediction of the thermal resistance of the IGBT module; and 3) combining the loss data obtained by the loss prediction model, the thermal resistance data obtained by the thermal resistance prediction model and junction temperature calculation formulas corresponding to different modulation modes to construct an IGBT junction temperature prediction off-line system, and performing junction temperature prediction of a corresponding IGBT product. According to the method, the problem of inaccurate junction temperature prediction caused by the unicity of the sample size of the IGBT module can be well solved.

Description

Technical Field

[0001] The invention belongs to the technical field of IGBT application, and relates to a method for estimating the fast online junction temperature of an IGBT module based on a neural network. Background Art

[0002] At present, there are four main methods for predicting IGBT junction temperature: contact sensor method, infrared thermal imaging method, thermal impedance model method and thermal sensitive electrical parameter method. The contact sensor method requires the use of a temperature sensor to directly contact and measure the chip monitoring points inside the IGBT module; the infrared imaging method is a non-contact optical measurement method that uses infrared equipment to measure the internal monitoring points of the specially treated IGBT module; the thermal impedance model method converts the thermodynamic model into a circuit model to estimate the junction temperature based on the equivalent heat dissipation form of the multi-layer material structure of the IGBT module; the thermal sensitive electrical parameter method reversely infers the junction temperature of the device based on the corresponding relationship between the device junction temperature and the external electrical parameters. But sometimes, as an IGBT module manufacturer, it is more desirable to calculate the junction temperature of IGBT-related products through an offline system.

[0003] Therefore, a neural network-based IGBT module fast online junction temperature estimation method is designed to overcome the above problems. Summary of the invention

[0004] The purpose of the present invention is to overcome the shortcomings of the prior art and to provide a method for estimating the junction temperature of an IGBT module online quickly based on a neural network, which is suitable for IGBT module manufacturers to predict and evaluate the IGBT junction temperature offline or other users who need to predict the IGBT junction temperature.

[0005] The present invention is implemented by the following technical solution: a method for estimating the fast online junction temperature of an IGBT module based on a neural network, the method comprising the following steps:

[0006] Step 1) obtaining a large number of IGBT module turn-on and turn-off loss data samples of a certain IGBT through an automated double-pulse test machine platform, establishing a loss prediction model based on an improved particle swarm algorithm to optimize the BP neural network, and using the improved particle swarm algorithm to optimize the initial weights and thresholds of the model, and then inputting the IGBT module turn-on and turn-off loss data samples into the model for training and testing, and taking the main influencing factors of the IGBT switching loss as the input vector of the model, taking the turn-on loss and turn-off loss as the output vector, and using the learning rule of the conjugate gradient method to accelerate convergence, so as to achieve effective prediction of the turn-on and turn-off loss of the IGBT module;

[0007] Step 2) Based on the transient thermal resistance curve of the Foster network model, the transient thermal resistance curves of a certain number of samples of the same IGBT module product are obtained through a thermal resistance test machine, which are equivalent to a one-dimensional time series. A one-dimensional convolutional layer is used to construct a mapping model of the thermal resistance of the IGBT module. In order to effectively extract the thermal impedance feature information, a CNN neural network thermal resistance prediction model with two convolutional layers and two fully connected layers is established to realize the prediction of the thermal resistance of the IGBT module;

[0008] Step 3) Combining the loss data obtained by the loss prediction model and the thermal resistance data obtained by the thermal resistance prediction model, corresponding to the junction temperature calculation formula under different modulation modes, an IGBT junction temperature prediction offline system is constructed to perform junction temperature prediction of the corresponding IGBT product.

[0009] As a preferred embodiment: the improved particle swarm algorithm optimization in step 1) is specifically as follows:

[0010] The weights and thresholds of the IGBT module switching loss BP neural network model are initialized into a group of random particles and transmitted to the network for learning and training. After the training is completed, the network is tested using test samples, and the fitness value of the example is calculated based on the test error. Then, the individual extreme value and the global extreme value are determined based on the fitness value, thereby updating the speed and position of the particles. If the termination condition is not met, the above process is repeated until the optimal initial weights and thresholds are found.

[0011] As a preferred embodiment: the loss prediction model based on the improved particle swarm algorithm to optimize the BP neural network in step 1) is specifically as follows: the neural network must include an input layer and an output layer, and the number of hidden layers needs to be determined based on the complexity analysis of the switching loss of the IGBT module. There are five main factors affecting the switching loss of the IGBT module, so the input layer has five nodes; the output layer corresponds to the turn-on and turn-off losses of the output module, so the output layer has two nodes; the optimal number of nodes in the hidden layer needs to consider the effect after network training, the length of training time, and the generalization energy of the model; the activation transfer function of the hidden layer selects the logarithmic S-type function; the activation transfer function of the output layer selects the linear function. The learning rate determines the amount of change in the weights in each cycle of training. The size of the learning rate needs to ensure that the network error does not jump out of the error surface trough and can tend to the minimum error.

[0012] As a preference: the loss prediction model of the BP neural network is as follows: Figure 2 , where the emitter voltage, collector current, drive voltage, drive resistance, and junction temperature are input signals, ω ik is the weight of neuron k, b j is the threshold, the turn-on loss and turn-off loss are the outputs, and the initial weight ω ik and threshold b j Optimized by particle swarm algorithm.

[0013] Preferably, the CNN neural network thermal resistance prediction model in step 2) is composed of a convolutional layer, a pooling layer and a fully connected layer. The local features are gradually extracted through the convolutional layer, the global high-level features are obtained through the pooling layer, and the fully connected layer concatenates the features and outputs the data. Considering that the output feature information of the thermal resistance curve is small, a CNN neural network thermal resistance prediction model with two convolutional layers and two fully connected layers is used.

[0014] Preferably, the CNN neural network thermal resistance prediction model realizes feature extraction by performing convolution operation through a convolution kernel. For the abstraction of thermal impedance curve information, the output of the convolution node can be expressed as:

[0015]

[0016] In the above formula, and Represent the feature maps of the nth and previous convolutional layers, k i，j represents the convolution kernel used between the i-th feature map and the j-th feature map, represents the bias of the neuron, and f(·) represents the activation function. After convolution, CNN abstracts more advanced features through the pooling layer and reduces the vector dimension. The output of the pooling layer is represented as:

[0017]

[0018] In the formula, down(·) represents the pooling method of the pooling layer, which generally uses maximum pooling. represents the weight of the product;

[0019] Finally, the feature structure extracted by multiple convolutional layers and pooling layers is input into the fully connected layer. The fully connected layer neural network learns and summarizes the features and finally outputs the predicted value. The fully connected layer can be expressed as:

[0020] y=f(W T x+b)

[0021] Where x and y are the input and output of the neurons in the fully connected layer, respectively, and W and b are the weights and bias of the neurons in the fully connected layer, respectively.

[0022] Preferably, the different modulation modes mentioned in step 3) are mainly SPWM and SPVWM modulation modes.

[0023] The beneficial effects of the present invention are as follows:

[0024] The neural network-based IGBT module fast online junction temperature estimation method designed in the present invention aims to establish an offline system for IGBT junction temperature prediction. It is suitable for IGBT module manufacturers to offline predict and evaluate IGBT junction temperature or other users who need to predict IGBT junction temperature. It can well solve the problem of inaccurate junction temperature prediction caused by the singleness of the IGBT module sample size. BRIEF DESCRIPTION OF THE DRAWINGS

[0025] Figure 1 It is a flow chart of the particle swarm algorithm optimization process of the present invention.

[0026] Figure 2 This is a schematic diagram of the loss prediction model of the present invention based on the improved particle swarm algorithm to optimize the BP neural network.

[0027] Figure 3 It is a schematic diagram of the thermal resistance prediction model based on the CNN neural network of the present invention. DETAILED DESCRIPTION

[0028] In order to enable those skilled in the art to more clearly understand the objectives, technical solutions and advantages of the present invention, the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0029] In the description of the present invention, it is necessary to understand that the orientation or position relationship indicated by the terms such as "up", "down", "left", "right", "inside", "outside", "horizontal" and "vertical" is based on the orientation or position relationship shown in the accompanying drawings, and is only for the convenience of describing the present invention, rather than indicating or implying that the device or original referred to must have a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0030] The present invention will be described in detail below with reference to the accompanying drawings: Figure 1 As shown, a method for estimating the fast online junction temperature of an IGBT module based on a neural network comprises the following steps:

[0031] Step 1) obtaining a large number of IGBT module turn-on and turn-off loss data samples of a certain IGBT through an automated double-pulse test machine platform, and establishing a loss prediction model based on an improved particle swarm algorithm to optimize the BP neural network. Since the initial weights and thresholds of the model have a great influence on the neural network training, it is necessary to optimize it using the improved particle swarm algorithm. The initial weights and thresholds of the model are optimized using the improved particle swarm algorithm, and then the IGBT module turn-on and turn-off loss data samples are input into the model for training and testing, and the main influencing factors of the IGBT switch loss are used as the input vector of the model, and the turn-on loss and turn-off loss are used as the output vector. The learning rule of the conjugate gradient method is used to accelerate convergence, so as to achieve effective prediction of the turn-on and turn-off loss of the IGBT module;

[0032] In step 1), by analyzing the physical characteristics of the IGBT module, the main factors affecting the switching loss of the IGBT module are: collector voltage, collector current, gate drive voltage, gate drive resistance, and junction temperature. The factors affecting the switching loss have a more complex linear distribution for solving the loss value, and it is difficult to solve using mathematical fitting methods such as power exponential fitting and linear interpolation. The BP neural network can approximate any nonlinear mapping through learning and training while considering more influencing factors. It is composed of a large number of neuronal connections, and learning rules such as the steepest descent method and the conjugate gradient method can be used to continuously adjust the network weights and thresholds through the back propagation of errors to minimize the network error. It needs to be considered that the initial connection weights and thresholds of the BP neural network have a great influence on network training. The BP neural network itself selects the initial weights and thresholds in a random manner, which cannot meet the accurate solution of the switching loss, and the initial weights and thresholds of the network must be optimized.

[0033] The particle swarm algorithm can just use the random solution to evaluate the quality of the initial weights and thresholds of the BP neural network for IGBT module switching loss through fitness, and find the optimal solution for the initial weights and thresholds after multiple iterations.

[0034] Step 2) Based on the transient thermal resistance curve of the Foster network model, the transient thermal resistance curves of a certain number of samples of the same IGBT module product are obtained through a thermal resistance test machine, which are equivalent to a one-dimensional time series. A one-dimensional convolutional layer is used to construct a mapping model of the thermal resistance of the IGBT module. In order to effectively extract the thermal impedance feature information, a CNN neural network thermal resistance prediction model with two convolutional layers and two fully connected layers is established to realize the prediction of the thermal resistance of the IGBT module;

[0035] Step 3) Combine the loss data obtained by the loss prediction model, the thermal resistance data obtained by the thermal resistance prediction model, and the junction temperature calculation formulas under different modulation modes to build an IGBT junction temperature prediction offline system to predict the junction temperature of the corresponding IGBT product. As a preferred embodiment: the improved particle swarm algorithm optimization in step 1) is specifically as follows:

[0036] The weights and thresholds of the IGBT module switching loss BP neural network model are initialized into a group of random particles and transmitted to the network for learning and training. After the training is completed, the network is tested using test samples, and the fitness value of the example is calculated based on the test error. Then, the individual extreme value and the global extreme value are determined based on the fitness value, thereby updating the speed and position of the particles. If the termination condition is not met, the above process is repeated until the optimal initial weights and thresholds are found.

[0037] The loss prediction model based on the improved particle swarm algorithm to optimize the BP neural network in step 1) is specifically as follows: the neural network must include an input layer and an output layer, and the number of hidden layers needs to be determined based on the complexity analysis of the IGBT module switching loss. There are 5 main factors affecting the IGBT module switching loss, so the input layer has 5 nodes; the output layer corresponds to the turn-on and turn-off losses of the output module, so the output layer has 2 nodes; the optimal number of nodes in the hidden layer needs to consider the effect after network training, the length of training time, and the generalization energy of the model; the activation transfer function of the hidden layer selects the logarithmic S-type function; the activation transfer function of the output layer selects the linear function. The learning rate determines the amount of weight change in each cycle of training. The size of the learning rate needs to ensure that the network error does not jump out of the error surface trough and can tend to the minimum error.

[0038] The loss prediction model of the BP neural network is as follows: Figure 2 , where the emitter voltage, collector current, drive voltage, drive resistance, and junction temperature are input signals, ω ik is the weight of neuron k, b j is the threshold, the turn-on loss and turn-off loss are the outputs, and the initial weight ω ik and threshold b j Optimized by particle swarm algorithm.

[0039] For IGBT module product manufacturers or specifications, what is generally provided is the transient thermal resistance curve of the Foster thermal network model, which can be equivalent to a one-dimensional time series. A one-dimensional convolutional layer is used to construct a thermal resistance mapping model. The CNN neural network thermal resistance prediction model in step 2) consists of a convolutional layer, a pooling layer, and a fully connected layer. The convolutional layer gradually extracts local features, the pooling layer obtains global high-level features, and the fully connected layer splices features and outputs data. Considering that the output feature information of the thermal resistance curve is small, a CNN neural network thermal resistance prediction model with two layers of convolution and two layers of fully connected layers is used.

[0040] The CNN neural network thermal resistance prediction model realizes feature extraction by performing convolution operation through convolution kernel. For the abstraction of thermal impedance curve information, the output of the convolution node can be expressed as:

[0041]

[0042] In the above formula, and Represent the feature maps of the nth and previous convolutional layers, k i，j represents the convolution kernel used between the i-th feature map and the j-th feature map, represents the bias of the neuron, and f(·) represents the activation function. After convolution, CNN abstracts more advanced features through the pooling layer and reduces the vector dimension. The output of the pooling layer is represented as:

[0043]

[0044] In the formula, down(·) represents the pooling method of the pooling layer, which generally uses maximum pooling. represents the weight of the product;

[0045] Finally, the feature structure extracted by multiple convolutional layers and pooling layers is input into the fully connected layer. The fully connected layer neural network learns and summarizes the features and finally outputs the predicted value. The fully connected layer can be expressed as:

[0046] y=f(W T x+b)

[0047] Where x and y are the input and output of the neurons in the fully connected layer, respectively, and W and b are the weights and bias of the neurons in the fully connected layer, respectively.

[0048] The different modulation modes mentioned in step 3) are mainly modulation modes such as SPWM and SPVWM. Since the modulation modes are existing technologies, the calculation formulas for junction temperature under the corresponding modulation modes are not introduced.

[0049] The neural network-based IGBT module fast online junction temperature estimation method designed in the present invention aims to establish an offline system for IGBT junction temperature prediction. It is suitable for IGBT module manufacturers to offline predict and evaluate IGBT junction temperature or other users who need to predict IGBT junction temperature. It can well solve the problem of inaccurate junction temperature prediction caused by the singleness of the IGBT module sample size.

[0050] The specific embodiments described herein are merely illustrative of the principles and effects of the present invention and are not intended to limit the present invention. Anyone familiar with the art may modify or alter the above embodiments without departing from the spirit and scope of the present invention. Therefore, all equivalent modifications or alterations made by a person of ordinary skill in the art without departing from the spirit and technical ideas disclosed by the present invention shall still be covered by the claims of the present invention.

Claims

1. A method for estimating the fast online junction temperature of an IGBT module based on a neural network, characterized in that: The method comprises the following steps: Step 1) obtaining a large number of IGBT module turn-on and turn-off loss data samples of a certain IGBT through an automated double-pulse test machine platform, establishing a loss prediction model based on an improved particle swarm algorithm to optimize the BP neural network, and using the improved particle swarm algorithm to optimize the initial weights and thresholds of the model, and then inputting the IGBT module turn-on and turn-off loss data samples into the model for training and testing, and taking the main influencing factors of the IGBT switching loss as the input vector of the model, taking the turn-on loss and turn-off loss as the output vector, and using the learning rule of the conjugate gradient method to accelerate convergence, so as to achieve effective prediction of the turn-on and turn-off loss of the IGBT module; Step 2) Based on the transient thermal resistance curve of the Foster network model, the transient thermal resistance curves of a certain number of samples of the same IGBT module product are obtained through a thermal resistance test machine, which are equivalent to a one-dimensional time series. A one-dimensional convolutional layer is used to construct a mapping model of the thermal resistance of the IGBT module. In order to effectively extract the thermal impedance feature information, a CNN neural network thermal resistance prediction model with two convolutional layers and two fully connected layers is established to realize the prediction of the thermal resistance of the IGBT module; Step 3) Combining the loss data obtained by the loss prediction model, the thermal resistance data obtained by the thermal resistance prediction model, and the junction temperature calculation formula corresponding to different modulation modes, an IGBT junction temperature prediction offline system is constructed to perform junction temperature prediction of the corresponding IGBT product.

2. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1, characterized in that: The improved particle swarm algorithm optimization in step 1) is specifically as follows: The weights and thresholds of the IGBT module switching loss BP neural network model are initialized into a group of random particles and transmitted to the network for learning and training. After the training is completed, the network is tested using test samples, and the fitness value of the example is calculated based on the test error. Then, the individual extreme value and the global extreme value are determined based on the fitness value, thereby updating the speed and position of the particles. If the termination condition is not met, the above process is repeated until the optimal initial weights and thresholds are found.

3. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1, characterized in that: The loss prediction model based on the improved particle swarm algorithm to optimize the BP neural network in step 1) is specifically as follows: the neural network must include an input layer and an output layer, and the number of hidden layers needs to be determined based on the complexity analysis of the IGBT module switching loss. There are 5 main factors affecting the IGBT module switching loss, so the input layer has 5 nodes; the output layer corresponds to the turn-on and turn-off losses of the output module, so the output layer has 2 nodes; the optimal number of nodes in the hidden layer needs to consider the effect after network training, the length of training time, and the generalization energy of the model; the activation transfer function of the hidden layer selects the logarithmic S-type function; the activation transfer function of the output layer selects the linear function. The learning rate determines the amount of weight change in each cycle of training. The size of the learning rate needs to ensure that the network error does not jump out of the error surface trough and can tend to the minimum error.

4. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1 or 3, characterized in that: The loss prediction model of the BP neural network is shown in FIG2 , where the emitter voltage, collector current, drive voltage, drive resistance, and junction temperature are input signals, ω ik is the weight of neuron k, b j is the threshold, the turn-on loss and turn-off loss are the outputs, and the initial weight ω ik and threshold b j Optimized by particle swarm algorithm.

5. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1, characterized in that: The CNN neural network thermal resistance prediction model in the step 2) is composed of a convolutional layer, a pooling layer and a fully connected layer. The local features are gradually extracted through the convolutional layer, the global high-level features are obtained through the pooling layer, and the fully connected layer splices the features and outputs the data. Considering that the output feature information of the thermal resistance curve is small, a CNN neural network thermal resistance prediction model with two convolutional layers and two fully connected layers is used.

6. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1 or 5, characterized in that: The CNN neural network thermal resistance prediction model realizes feature extraction by performing convolution operation through convolution kernel. For the abstraction of thermal impedance curve information, the output of the convolution node can be expressed as: In the above formula, and Represent the feature maps of the nth and previous convolutional layers, k i，j represents the convolution kernel used between the i-th feature map and the j-th feature map, represents the bias of the neuron, and f(·) represents the activation function. After convolution, CNN abstracts more advanced features through the pooling layer and reduces the vector dimension. The output of the pooling layer is represented as: In the formula, down(·) represents the pooling method of the pooling layer, which generally uses maximum pooling. represents the weight of the product; Finally, the feature structure extracted by multiple convolutional layers and pooling layers is input into the fully connected layer. The fully connected layer neural network learns and summarizes the features and finally outputs the predicted value. The fully connected layer can be expressed as: y=f(W T x+b) Where x and y are the input and output of the neurons in the fully connected layer, respectively, and W and b are the weights and bias of the neurons in the fully connected layer, respectively.

7. The method for estimating the fast online junction temperature of an IGBT module based on a neural network according to claim 1, characterized in that: The different modulation modes mentioned in step 3) are mainly SPWM and SPVWM modulation modes.

Images