We just talked about heat, why monitor temperature? In fact, the most fragile part of our module is the bonding wire. What makes this bonding wire fragile?

It's because the temperature fluctuates repeatedly. Those who make modules know that these materials are different, and their repeated thermal expansion rates are different. We need to monitor its bonding wire, which is a passive method, while what we just talked about is an active thermal management method. There are several faults in passive mode, one is that the binding wire falls off, and the lead wire breaks. What mechanism do we use to obtain this? The fundamental principle of this method is to collect temperature sensitivity coefficients such as resistance and voltage. The module on the left is a half bridge module, which is a model of IGBT. The same applies to silicon carbide. It's just that its pressure drop and other parameters are slightly different, the principle is the same. We have obtained all its parasitic parameters and detected its voltage on both sides. We can see that if a point falls off at the binding line, the voltage will be slightly higher. The curve is different at each temperature. Finally, we can monitor the change in voltage, obtain the number of points where the binding line falls off, and then know the degree of damage and evaluate whether it should be repaired. The second method is to detect the solder layer. Nowadays, some silicon carbide is sintered with silver, which is a popular process for IGBT.

Here is its model above, and below is the testing tool. After 1300 cycles, the healthy results can be seen to be complete at the back. In about half of the damaged cases, this is the case, and solder aging is visible. The intermediate process involves 1300 cycles, with some solder hotspots occurring more frequently than others. This is the offline method we used when we identified some issues, and we have already removed the module from the system.

The above processing methods are all monitoring. In order to solve these problems, we use technical means to monitor the entire module.

These technological means include installing intelligent sensors and leveraging powerful MCU computing power. We use FPGA to do it, but some people may question whether you can use Infineon chips for calculation? it's fine too. After taking some proactive thermal management measures, the remaining task is to perform some processing on the gate drive, taking into account environmental factors such as humidity, light, pressure, magnetic force, and vibration. Finally, we will develop a driving solution that is more effective than the previous one. Another method is to conduct real-time online evaluation of this device. The original loss model was the cumulative result of the product of current and ESR. The establishment of this model is based on its thermal resistance and loss product plus the ambient temperature. This is a conventional model that many engineers can calculate and directly derive from the data manual. But there is a problem, this is all normal, you can obtain stable values, dynamic ones are not easy to evaluate, and the cumulative results are not like this. Where are the differences in cumulative results reflected? One is that ESR will change, as can be seen from the middle graph, ESR will accumulate over a long period of time and increase. We need to improve its lifespan model, which should be related to the previous thermal models, showing an exponential relationship.

In addition, its damage is the result of cumulative damage, not a constant result over a certain period of time. When engineers choose devices, they usually have a habit of using a constant module, such as 600A, for 500A. At this time, we calculate its loss, which is actually inaccurate because it is constantly changing. We need to improve it. The result of the improvement is to convert variable stress into constant stress throughout the entire lifecycle. We can see its curve from the results of probability statistics.

The curve graph is obtained through Weibull fitting, and at the bottom, we can see the latter half of a power device curve. After fitting, the curve looks like this: how many lifetimes can still exist? This is a means. Silicon carbide high-frequency switches can also bring many problems, such as crosstalk, which can cause some issues. We adjust the main power circuit through other means, such as segmented driving and variable resistance driving. These driving chips include many technological research achievements, such as changing the gate driving resistance. We can achieve it by only dividing it into 2-3 segments, which may not necessarily require so many. The process of oscillation is well known to everyone, and it is precisely because of its small parasitic parameters that we can switch at high frequencies. However, its oscillation during high-frequency switching can affect its lifespan.

The entire intelligent drive adopts a multi information fusion health management concept, which involves data collection, status monitoring, health assessment, and fault prediction, which is provided to maintenance engineers or other after-sales departments. Whether we need to replace this module or wait until the end of its lifespan is like knowing that someone is leaving, we should be prepared. How did we obtain this data? Through human-machine interface and other system interfaces, we have embedded this system into the controller developed by J-Link Electronics. This data can be downloaded offline or exported online through CAN interface.

On this basis, we further investigate and have already involved the reliability of the entire drive system. At this point, we will use some popular techniques, such as convolutional neural networks, digital encoding, and other neural network methods. Detect motor faults and other malfunctions through a giant magnetic detection method. This is a testing platform for motors. On this platform, we process the leakage magnetic signal through Fourier transform to obtain data, and then extract parameters and feature values through convolutional neural network methods. We train the model to predict the current fault and the extent of damage. Another technical approach is to use multi physics field simulation. On the left is to use MATLAB, which is a system level method. The input current and voltage parameters depend on MATLAB and Pspice, reaching the module level. The top is the switch loss, and the bottom is the conduction loss. Finally, a three-dimensional thermal model at the chip level is obtained.

Another approach is that many engineers have various simulation tools, ranging from multi physics fields to single board level, system level, and chip level. Some people feel that the simulation is not accurate, but there is a problem because there is no closed loop. Imitation is imitation, but there is no closed loop. Many structural engineers have experience, so when we simulate it, there is a significant difference between the estimated heat and the actual measured temperature. Why? There is no closed loop. The repeated results of theory to practice and practice to theory are the key to making our simulations more precise and practical. The silicon carbide model on the left is processed using a neural network, which is repeatedly combined with MATLAB and Pspice to form a closed-loop process. This operating condition input can be obtained using Pspice to access the IGBT dynamic database, and then we repeatedly close the loop to obtain a result that is close to reality. It's difficult to be completely accurate. For example, when this model is used for electromagnetic simulation, it's not as accurate in the high frequency range because the parameters change and simulation needs to be combined with reality.