An IGBT bonding wire aging monitoring method based on BPNN

Abstract

The application discloses an IGBT bonding wire aging monitoring method based on BPNN, and specifically is: through a double pulse experiment, feature information of an IGBT bonding wire aging state under different experimental conditions, namely, on-voltage drop OVD, is acquired; then, the feature information under different working conditions is split into a training set and a test set according to a certain proportion and is used as input of a bonding wire aging monitoring model; on this basis, a suitable learning rate is selected and an L-M method is used to realize minimization of error between reference and output of the BPNN, so that the influence of temperature and load current is excluded, and finally, an optimal model of bonding wire state monitoring is acquired, so as to realize accurate monitoring of the IGBT bonding wire aging state. The application shows satisfactory performance in the bonding wire monitoring aspect and has certain advantages in avoiding the influence of junction temperature fluctuation and load current dependence, and can realize monitoring of the bonding wire aging state in actual industrial application.

Description

Technical Field

[0001] This invention belongs to the field of power device health status monitoring technology, and particularly relates to an IGBT bond wire aging monitoring method based on BPNN. Background Technology

[0002] The increasing demand for new energy conversion paradigms necessitates the reliable operation of efficient, high-power-density converters. Power modules within power converters, such as Insulated Gate Bipolar Transistor (IGBT) modules, are particularly crucial, as they perform both energy conversion and power control. However, IGBT modules are considered a major cause of declining power converter reliability due to their high failure rate. Among the many types of IGBT module failures, bond wire degradation is especially prominent. Therefore, precise bond wire aging monitoring is essential to ensure the reliability of power converters.

[0003] Under the influence of bond wire aging, the electrical parameters of IGBT modules will undergo significant changes. Due to this unique characteristic, indirect monitoring methods accomplish the task of bond wire aging monitoring by monitoring aging-related electrical parameters. These parameters are suitable for online applications and have the advantages of ease of implementation and high flexibility. For indirect bond wire aging monitoring methods, the selection of electrical parameters is crucial, and considerable effort has been made to explore suitable parameters. Among them, on-state voltage drop (OVD) is one of the most commonly used monitoring indicators. However, this bond wire aging monitoring method only works well at near-constant junction temperatures and performs poorly under junction temperature fluctuations and performance deteriorates under different load currents. These issues limit the application of OVD as an indicator in bond wire aging monitoring. Summary of the Invention

[0004] While many methods utilize OVD (Optical Variation Discharge) for aging monitoring of IGBT (Insulated Ga1te Bipolar Transistor) bond wires, most of these methods are plagued by challenging issues such as junction temperature fluctuations and load current dependence. Therefore, further innovative solutions are needed. In view of this, this invention provides an IGBT bond wire aging monitoring method based on Back Propagation Neural Networks (BPNN).

[0005] This invention discloses a BPNN-based IGBT bond wire aging monitoring method. It utilizes the on-state voltage drop (OVD) as a monitoring parameter and combines it with a backpropagation neural network (BPNN) to eliminate the influence of junction temperature fluctuations and load current, providing accurate information on the aging degree of the bond wires for IGBT modules under different operating conditions. Specifically, it includes the following steps:

[0006] Step 1: Data collection.

[0007] Based on the dual-pulse experimental platform, different experimental conditions were set and experiments were conducted to obtain OVD data under different conditions. The experimental conditions included: load current, number of bond wires stripped, and IGBT chip temperature. The number of bond wires stripped was arranged from least to most to simulate the aging degree of the bond wires from small to large.

[0008] Step 2: Preprocessing.

[0009] The OVD data collected from the double-pulse experiment were filtered to reduce the influence of noise, and the features of OVD were extracted from these filtered data. At the same time, the extracted OVD features were normalized to eliminate the influence of dimensionality and eliminate outliers.

[0010] The max-min normalization method is used to normalize the extracted features within the range [0,1], and it is described as follows:

[0011]

[0012] in, x i x min and x max These are the normalized features of the training samples, the i-dimensional features of the training samples, and the minimum and maximum values ​​of the training samples, respectively.

[0013] Step 3: Sample splitting.

[0014] The processing features of OVD are further divided into training and test sets in an appropriate ratio of training set to test set. Here, 80% of the OVD samples are used for training set, while the test set includes 20% of the OVD samples.

[0015] Step 4: Model training.

[0016] First, the architecture and initialization of the BPNN are executed; the number of neurons in the input layer is the dimension of the sample features, while the number of neurons in the hidden layer is determined based on the test results of the BPNN; the activation function shows high importance in the BPNN, and different types of activation functions will be used in the BPNN, as shown in (2)-(4); that is, the sigmod function and the tansig function are used in the input layer and the hidden layer, respectively; while the purelin function is used as the activation function in the output layer.

[0017]

[0018]

[0019] purelin(x) = x(4)

[0020] The initialization of the BPNN mainly focuses on setting the number of training iterations and the training target to 300 and 0.01, respectively; the learning rate is set to 0.02.

[0021] In the training set, the Levenberg-Marquardt method is used to minimize the error between the reference and output of the BPNN. The Gauss-Newton algorithm is used to iteratively optimize the parameters of the BPNN. At the same time, a decay coefficient is introduced into the BPNN to achieve fast convergence.

[0022] Step 5: Model Validation.

[0023] The model test set is used to verify the model's effectiveness. The model's output results are compared with the test set results. If the error in the comparison results is within an acceptable range, the model is considered to have a good monitoring effect.

[0024] The beneficial technical effects of this invention are as follows:

[0025] 1. This invention uses IGBT on-state voltage drop as an indicator for monitoring the aging status of bond wires. In practical applications, only one measurement is required, reducing the demand for measuring equipment, simplifying the measurement process, and facilitating the application of the method under actual working conditions. Simultaneously, the on-state voltage drop exhibits good sensitivity to bond wire aging, ensuring the high accuracy of the proposed bond wire aging monitoring method.

[0026] 2. This invention uses a backpropagation neural network to reduce the influence of objective factors other than aging on the identification results, and achieves accurate estimation of the aging state of the bond wire under load current and junction temperature change conditions.

[0027] 3. This invention has good applicability and can be applied to converter systems composed of different topologies and different power devices. Attached Figure Description

[0028] Figure 1 This is a flowchart of the bonding wire aging monitoring method based on backpropagation neural network of the present invention.

[0029] Figure 2 This is a photograph of the bipulse experimental platform used to collect training data in this invention.

[0030] Figure 3 This invention describes how the error changes with the number of iterations during the BPNN training process.

[0031] Figure 4 This is a verification result of the BPNN trained by this invention. Detailed Implementation

[0032] The present invention will be further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0033] This invention discloses a BPNN-based IGBT bond wire aging monitoring method. It utilizes the on-state voltage drop (OVD) as a monitoring parameter and combines it with a backpropagation neural network (BPNN) to eliminate the influence of junction temperature fluctuations and load current, providing accurate information on the aging degree of the bond wires for IGBT modules under different operating conditions. The process of this invention is as follows: Figure 1 As shown, the specific steps include:

[0034] Step 1: Data collection.

[0035] Using a double-pulse experimental platform, such as Figure 2 As shown, experiments were conducted under different temperatures, load currents, and different numbers of broken bond wires to obtain the transient and steady-state voltage waveforms of the IGBT under different conditions.

[0036] Step 2: Preprocessing.

[0037] The global feature information data collected from the double-pulse experiment is processed by labeling, normalization, and filtering. In this invention, the number of bond wire breaks is used as a marker for the BPNN. OVD features are then extracted from this filtered data. Simultaneously, the extracted OVD features are normalized to eliminate dimensionality effects and outlier samples.

[0038] The max-min normalization method is used to normalize the extracted features within the range [0,1], and it is described as follows:

[0039]

[0040] in, x i x min and x maxThese are the normalized features of the training samples, the i-dimensional features of the training samples, and the minimum and maximum values ​​of the training samples, respectively.

[0041] Step 3: Sample splitting.

[0042] The processing features of OVD are divided into training and test sets according to a certain ratio. The training set is used for model training, and the test set is used for model validation. To improve convergence speed, the training samples need to be further randomly shuffled. Random shuffling avoids overfitting of the model to the training set and improves the ability of the trained BPNN to handle unknown data.

[0043] In the embodiments, 80% of the samples from OVD are used as the training set, while the test set comprises 20% of the samples from OVD (the ratio of training and test set samples is not explicitly required in the method of this invention). A further process of randomly shuffling the training samples is implemented to improve the convergence speed. By doing so, the effects of overfitting are avoided, and the generalization ability of BPNN is improved. Furthermore, as a supervised learning method, the training samples in BPNN should be labeled, and in the proposed method, the number of bond lines stripped is used as the label for the BPNN.

[0044] Step 4: Model training.

[0045] First, the BPNN architecture is initialized. It's worth noting that the number of neurons in the input layer is determined by the dimension of the sample features, while the number of neurons in the hidden layer is determined based on the BPNN test results. It should be pointed out that the selection of activation functions is crucial in BPNNs, and different types of activation functions need to be used to improve model performance. Commonly used activation functions are shown in (2)-(4). That is, the sigmoid function and the tansig function are used in the input and hidden layers, respectively. The purelin function is used as the activation function in the output layer.

[0046]

[0047]

[0048] purelin(x) = x(4)

[0049] For the initialization of the BPNN, the number of training iterations and the training objective are set to 300 and 0.01, respectively. Specifically, the training objective refers to the error between the reference and the BPNN output during training for each sample. Furthermore, the learning rate of the BPNN should be carefully considered. This is because a small learning rate can cause the BPNN to get stuck in local optima, resulting in slow convergence. In contrast, a BPNN with a large learning rate may oscillate around a minimum or even diverge. In the proposed method, the learning rate is set to 0.02 (the learning rate needs to be adjusted based on the different training data). Then, on the training set, the Levenberg-Marquardt method is used to minimize the error between the BPNN's reference and output, aiming to achieve satisfactory training performance for the BPNN. In this method, the Gauss-Newton algorithm is used to iteratively optimize the BPNN parameters. Simultaneously, a decay coefficient is introduced into the BPNN to achieve fast convergence. It is important to note that the decay coefficient will be adjusted according to the BPNN error. When the error meets the requirements, the BPNN model is considered trained. The BPNN training process is as follows... Figure 3 As shown in the figure, after 40 iterations, the errors of the training and test sets are significantly reduced, and the error of BPNN is close to the training target (i.e., 0.01), indicating that BPNN has achieved acceptable training performance.

[0050] Step 5: Model Validation.

[0051] A well-trained model can achieve accurate aging monitoring of bond wires. After model training, the performance of the BPNN for aging monitoring needs to be validated. The model validation results are as follows: Figure 4 As shown, with the help of a well-trained model, the proposed method exhibits satisfactory performance in monitoring the aging of IGBT bond wires.

[0052] The IGBT bond wire aging monitoring method employed in this invention uses the on-state voltage drop during IGBT turn-on as an indicator to monitor bond wire aging, and the on-state voltage drop (OVD) exhibits good sensitivity to bond wire aging. However, while affected by aging, the OVD is also challenged by junction temperature fluctuations and load current dependence. To address this issue, this invention further proposes an IGBT module bond wire aging monitoring method using a backplane junction network (BPNN). The proposed method can effectively solve the problem of severe performance degradation of OVD monitoring of bond wires caused by load current and junction temperature, achieving accurate monitoring of the bond wire aging state.

[0053] This invention enables bond wire aging monitoring solely through on-state voltage drop measurement. Furthermore, by employing a backpropagation neural network (BPNN), it achieves precise identification of IGBT bond wire aging conditions under complex operating conditions. This method is suitable for online aging monitoring of IGBT bond wires in actual converter operation, effectively solving the problem of difficult IGBT bond wire degradation monitoring in real-world operating conditions. The proposed method is characterized by strong applicability, simple operation, and ease of implementation. It can achieve relatively accurate aging parameter identification under different IGBT junction temperatures and load currents, effectively addressing the limitation of existing bond wire aging monitoring methods based on on-state voltage drop, which are affected by junction temperature and current and cannot be used online. This expands the versatility of the method. Essentially, the proposed bond wire aging monitoring method combines traditional aging monitoring methods with machine learning to eliminate interference from factors other than aging, exhibiting strong universality. Its applicability can be further extended to different types of power devices in various power electronic devices.

Claims

1. A method for monitoring the aging of IGBT bond wires based on BPNN, characterized in that, By using the on-state voltage drop (OVD) as a monitoring parameter and combining it with a backpropagation neural network (BPNN) to eliminate the influence of junction temperature fluctuations and load current, accurate information on the bonding wire aging degree of IGBT modules under different operating conditions is provided. Specifically, the following steps are included: Step 1: Data Collection Based on the dual-pulse experimental platform, different experimental conditions were set and experiments were conducted to obtain OVD data under different conditions. The experimental conditions included: load current, number of bond wires stripped, and IGBT chip temperature. The number of bond wires stripped was arranged from least to most to simulate the aging degree of the bond wires from small to large. Step 2: Preprocessing: The OVD data collected from the double-pulse experiment were filtered to reduce the influence of noise, and the features of OVD were extracted from these filtered data. At the same time, the extracted OVD features were normalized to eliminate the influence of dimensionality and eliminate outliers. The max-min normalization method is used to normalize the extracted features within the range [0,1], and it is described as follows: in, x i x min and x max These are the normalized features of the training samples, the i-dimensional features of the training samples, and the minimum and maximum values ​​of the training samples, respectively. Step 3: Sample Splitting: With an appropriate ratio of training set to test set, the processing features of OVD are further divided into training set and test set. Here, 80% of the OVD samples are used for training set, while the test set includes 20% of the OVD samples. Step 4: Model Training First, the architecture and initialization of the BPNN are executed; the number of neurons in the input layer is the dimension of the sample features, while the number of neurons in the hidden layer is determined based on the test results of the BPNN; the activation function shows high importance in the BPNN, and different types of activation functions will be used in the BPNN, as shown in (2)-(4); that is, the sigmoid function and the tansig function are used in the input layer and the hidden layer, respectively; while the purelin function is used as the activation function in the output layer; purelin(x) = x(4) The initialization of the BPNN mainly focuses on setting the number of training iterations and the training target to 300 and 0.01 respectively; the learning rate is set to 0.

02. In the training set, the Levenberg-Marquardt method is used to minimize the error between the reference and output of the BPNN. The Gauss-Newton algorithm is used to iteratively optimize the parameters of the BPNN. At the same time, a decay coefficient is introduced into the BPNN to achieve fast convergence. Step 5: Model Validation The model test set is used to verify the model's effectiveness. The model's output results are compared with the test set results. If the error in the comparison results is within an acceptable range, the model is considered to have a good monitoring effect.

Images