A neural network-based method for monitoring the state of a DC-DC converter

Abstract

The application discloses a DC-DC converter state monitoring method based on a neural network and belongs to the technical field of power generation, power transformation or power distribution. The method uses Laplace inverse transformation to solve a circuit model, uses a Foster thermal network to convert device loss into device temperature, and modifies the value of a device thermal parameter according to the temperature to form temperature feedback. The output obtained through simulation of an electrothermal model is used as the input of a neural network, and the aging parameter in the model is used as the output of the neural network to train the neural network. The experimental output is used as the input of the trained neural network to obtain the aging parameter, and finally, the value of the aging parameter is fed back to the electrothermal model to correct the inaccurate aging parameter. The application uses simulation data instead of experimental data to train the neural network, thereby greatly reducing the time cost of training the neural network while ensuring the monitoring accuracy.

Description

TECHNICAL FIELD

[0001] The application discloses a neural network-based DC-DC converter state monitoring method, relates to power electronic reliability technology, and belongs to the technical field of power generation, power transformation or power distribution. BACKGROUND

[0002] DC-DC converters are widely used in many fields such as renewable energy systems, industrial manufacturing, aerospace and electric vehicles. The increasingly complex operating conditions increase the probability of system failure, affecting the normal operation of the system and putting forward more stringent requirements for the reliability of the system. According to investigations, about 51% of the failure events of power converters are caused by capacitors and power semiconductors, and aging failure is the main cause of failure of these devices. Aging failure refers to the performance degradation of devices in long-term power cycles. This failure is often long-term and difficult to detect, and it is difficult to prevent in advance. Once this failure occurs, a large amount of manpower and material resources will be wasted for maintenance. Therefore, it is of great significance to monitor the state of capacitors, power semiconductors and other devices in DC-DC converters, and to replace the devices that fail due to aging in a timely manner.

[0003] Common DC-DC converter state monitoring methods mainly fall into three categories: experience-based state monitoring methods, model-based state monitoring methods and data-based state monitoring methods. The experience-based state monitoring method judges whether the device is damaged according to the past experience, and it is obvious that the accuracy of this method is low and the application occasions are single. The model-based state monitoring method mainly includes adding a measurement circuit and injecting a measurement signal. This kind of method needs to change the original working mode of the system, although the monitoring accuracy is high, but it belongs to the invasive state monitoring. The data-based state monitoring method directly estimates the aging parameters of the device according to the electrical output of the system which is easy to measure. For example, the state monitoring method based on digital twinning constantly learns and approximates the physical entity through the digital twinning model, and finally considers the aging parameters in the digital twinning model as the aging parameters of the physical entity. Although the digital twinning method belongs to non-invasive state monitoring, its monitoring accuracy is low, and the calculation amount is large and the calculation efficiency is low.

[0004] The present application aims to propose a neural network-based DC-DC converter state monitoring method to overcome the above-mentioned defects. SUMMARY

[0005] The application aims to solve the problems of the prior art and provides a DC-DC converter state monitoring method based on a neural network.

[0006] The application adopts the following technical scheme to solve the technical problems:

[0007] The application provides a DC-DC converter state monitoring method based on a neural network, which comprises the following three parts:

[0008] The first part is to establish an electro-thermal modeling of the DC-DC converter. The electro-thermal model comprises a circuit model of the DC-DC converter and a thermal model of a self-heating device. The circuit model is solved by using Laplace inverse transformation, the power loss of the self-heating device is converted into a self-heating junction temperature by using a Foster thermal network, and the value of a device thermal sensitive parameter is modified according to the self-heating junction temperature to form temperature feedback, that is, the aging parameter of the DC-DC converter is corrected according to the self-heating junction temperature. The device thermal sensitive parameter comprises a capacitance value of a capacitor, a resistance value of an equivalent series resistance of the capacitor and a resistance value of a conduction resistance of a power switch tube.

[0009] The second part is to train the neural network. The electrical output obtained by simulating the electro-thermal model of the DC-DC converter is used as the input of the neural network, and the aging parameter of the DC-DC converter is used as the output of the neural network. The neural network is trained to obtain a corresponding mathematical relationship between the electrical output of the DC-DC converter and the aging parameter of the DC-DC converter.

[0010] The third part is to apply the neural network. The experimentally measured electrical output of the DC-DC converter is used as the input of the trained neural network, a set of DC-DC converter aging parameters is predicted by the neural network, the capacitor voltage experimental data is calculated, the set of DC-DC converter aging parameters predicted by the neural network is substituted into the electro-thermal model of the DC-DC converter to obtain capacitor voltage simulation data, when the difference between the capacitor voltage experimental data and the capacitor voltage simulation data exceeds a threshold value, a set of DC-DC converter aging parameters is re-predicted by the neural network, and the process of calculating the capacitor voltage experimental data, the capacitor voltage simulation data and the difference between the capacitor voltage experimental data and the capacitor voltage simulation data is repeated until the difference between the capacitor voltage experimental data and the capacitor voltage simulation data meets the threshold value requirement. The value of the set of aging parameters finally output by the neural network is fed back to the electro-thermal model to correct the inaccurate aging parameter.

[0011] As a further optimization scheme for the state monitoring method of DC-DC converter based on neural networks, the thermal model of MOSFETs or diodes in self-heating devices is a four-layer Foster thermal network model. This is achieved through the differential equations of the four-layer Foster thermal network model. Perform an inverse Laplace transform to convert the sum of power losses of the self-heating device into the junction temperature of the self-heating device. in, R represents the temperature difference across the heat capacity of the i-th layer of the Foster thermal network, where i = 1, 2, 3, 4. i C i Let P be the thermal resistance and thermal capacity of the i-th layer of the Foster thermal network. loss For the power loss of a MOSFET or diode, T c T is the case temperature of the MOSFET or diode. j This refers to the junction temperature of a MOSFET or diode. The thermal model of the capacitor in a self-heating device is a first-order Foster thermal network model. The capacitor case temperature is calculated by performing a Laplace transform on the first-order Foster thermal network model. T c_C (t) represents the temperature of the capacitor casing at time t, where T is the temperature of the capacitor casing. a For ambient temperature, P loss_C For capacitor power loss, R th With C th The thermal resistance and heat capacity are given by the first-order Foster thermal network model.

[0012] As a further optimization of the state monitoring method for DC-DC converters based on neural networks, the electrical output of the DC-DC converter includes inductor current and output voltage.

[0013] As a further optimization of the DC-DC converter state monitoring method based on neural networks, the expression for calculating the capacitor voltage experimental data based on a set of DC-DC converter aging parameters predicted by the neural network is as follows: Among them, v o For the output voltage, v C Let i be the capacitor voltage. L R is the inductor current. C R is the equivalent series resistance of the capacitor, and R is the output load.

[0014] The present invention, by adopting the above technical solution, has the following beneficial effects:

[0015] (1) Performance advantage: Compared with the traditional state monitoring method, the present application belongs to the non-invasive DC-DC converter state monitoring technology. Firstly, the sum of the power loss of the self-heating device is converted into the DC-DC converter electrothermal model of the junction temperature of the self-heating device, the aging parameters are simulated through the electrothermal model, and then the neural network of the corresponding mathematical relationship between the electrical output of the DC-DC converter and the aging parameters of the DC-DC converter is trained and fitted with the simulation data of the aging parameters. The aging parameters are accurately corrected through the neural network. Compared with the traditional data-driven monitoring technology, the monitoring accuracy of the present application is higher and the calculation efficiency is high, and the value of the aging parameter can be quickly estimated.

[0016] (2) Easy to integrate: The present application does not need to add hardware measurement circuit to the system, and does not need to inject measurement signal. Once the neural network is trained, it can be integrated into the system microprocessor to monitor the system health status in real time.

[0017] (3) Cost advantage: The data required for training the neural network is all derived from the simulation results of the electrothermal model, which avoids the economic cost and time cost loss caused by a large number of aging experiments, reduces the time cost of training the neural network while ensuring the monitoring accuracy, and improves the efficiency. BRIEF DESCRIPTION OF DRAWINGS

[0018] Figure 1 is the overall block diagram of the DC-DC converter state monitoring method based on neural network proposed by the present application.

[0019] Figure 2 is the electrothermal modeling block diagram of the DC-DC converter of the present application.

[0020] Fig. 3(a) and Fig. 3(b) are schematic diagrams of the complex frequency domain model in the on state and the off state of the switch tube in the Buck circuit, respectively.

[0021] Figure 4 is the schematic diagram of the Foster thermal network model of the present application.

[0022] Figure 5 is the capacitor temperature rise curve measured in the experiment of the present application. DETAILED DESCRIPTION

[0023] The present application proposes a DC-DC converter state monitoring method based on neural network. In order to make the purpose, technical scheme and effect of the present application more clear and explicit, the present application will be further described in detail below with reference to the accompanying drawings.

[0024] In view of the problems existing in the existing state monitoring method, the present application proposes a state monitoring method based on neural network as shown in Figure 1 The parts of the method will be described below.

[0025] 1. Electric-thermal modeling

[0026] Take Buck converter (working in CCM mode) as an example, as shown in Figure 2 , the electric-thermal modeling of DC-DC converter can be divided into circuit modeling and device thermal modeling.

[0027] Circuit modeling is shown in Fig. 3(a) and Fig. 3(b). According to the on and off states of the switch, the main power circuit can be divided into two modes. For each mode, complex frequency domain transformation is carried out. According to KCL, KVL and Ohm's law, the expressions of I L (s), V C (s) and V o (s) can be solved, and then the corresponding i L (t), v C (t) and v o (t) can be obtained by inverse Laplace transform. Compared with numerical solution of differential equations such as Runge-Kutta, using inverse Laplace transform to solve circuit model can skip the non-steady state stage, thereby greatly improving the calculation efficiency.

[0028] The parameter D is introduced. When the switch is on, D is 1; when the switch is off, D is 0. Therefore, the differential equation model of Buck converter can be obtained as follows:

[0029]

[0030] In formula (1), i L is the inductor current, v C is the voltage across the capacitor, v o is the output voltage; R dson , R L , R C are the MOSFET on-resistance, inductor equivalent series resistance, and capacitor equivalent series resistance, respectively; V in is the input voltage, V f is the forward voltage drop of the diode, and R is the output load; D represents the switching state of the MOSFET. The above differential equation can be solved by inverse Laplace transform.

[0031] After completing the circuit modeling, the overall electric-thermal modeling of the system is shown in Figure 2 . In actual operation, MOSFET, diode, inductor and capacitor will generate losses, and these losses will be converted into heat and released, which is manifested as the rise of junction temperature. The change of temperature will also affect the MOSFET on-resistance R dson , the forward voltage drop of the diode V f , the capacitance C, the capacitor equivalent series resistance R C , the inductance L, and the inductor equivalent series resistance RL the value of the inductance L and the resistance R L The MOSFET, diode and capacitor are less affected by temperature, so the present application only considers the electro-thermal modeling of the MOSFET, diode and capacitor. After the device loss is calculated according to the circuit model solution, the device loss needs to be converted into device temperature. The common methods mainly include finite element method and thermal network method, but the finite element method needs to know the material parameters of the device, and these parameters are generally kept secret by the manufacturer, and the calculation efficiency of the finite element method is very low. The thermal network method can be divided into Cauer thermal network and Foster thermal network, as shown in Figure 4 . Similarly, to calculate the Cauer thermal network, the material parameters of the device also need to be known. The parameters of the Foster thermal network can be fitted by only the transient thermal impedance curve, and the manufacturer generally provides the transient thermal impedance curve in the data manual. The present application selects the Foster thermal network to express the process of converting the loss into temperature.

[0032] The equation of the transient thermal impedance curve of the Foster thermal network corresponding to the MOSFET or diode can be approximately written in the form of an exponential function, as shown in the following formula:

[0033]

[0034] In formula (2), Z thJC is the transient thermal impedance from the inside of the MOSFET or diode to the shell, R i and C i are the thermal resistance and thermal capacity of the i-th layer of the Foster thermal network, as shown in Figure 4 . The fitting accuracy of the transient thermal impedance curve of the device is different when the number of layers N in the model is different, and generally four layers of Foster can meet the accuracy requirement. The values of the thermal resistance and thermal capacity of each layer can be obtained by fitting the transient thermal impedance curve in MATLAB using the above formula. The differential equation set of four layers of Foster is:

[0035]

[0036] In formula (3), ΔT is the temperature difference across the i-th layer of thermal capacity, P loss is the power loss of the MOSFET or diode. The differential equation can also be solved by Laplace inverse transform, and the final calculated junction temperature T j of the MOSFET or diode can be determined by formula (3):

[0037]

[0038] In formula (4), T cFor MOSFET or diode case temperature, it is worth noting that the transient thermal impedance curve provided by the data sheet of MOSFET and diode is generally the transient thermal impedance from the inside of the device to the case of the device, so the device case temperature T c is a fixed value, so that the accurate self-heating device junction temperature can be calculated.

[0039] In the circuit, the inductance and the capacitance also generate the electro-thermal effect during the operation of the power converter, but since the inductance and the series resistance are less affected by the temperature, the present application only considers the electro-thermal modeling of the capacitance, and since the temperature change of the capacitance is not large, the thermal network only considers the first order. Figure 5 The capacitance temperature rise curve is measured in the experiment, the first order Foster thermal network parameters are obtained by fitting, and the capacitance case temperature T c_C :

[0040]

[0041] In formula (5), T c_C (t) is the capacitance case temperature at time t, T a is the ambient temperature, P loss_C is the power loss of the capacitance, R th is the equivalent series resistance of the capacitance, and C th is the capacitance value. C The values of the capacitance value C and the equivalent series resistance R

[0042] After the self-heating device junction temperature is obtained, the parameters (R dson , V f , C, R C , etc.) of the self-heating device can be corrected according to the parameter-temperature curve provided by the device data sheet.

[0043] 2. Training of neural network

[0044] Artificial neural networks (ANNs) are also simply called neural networks (NNs) or connection models, which are a kind of algorithmic mathematical model that simulates the behavior characteristics of animal neural networks and performs distributed parallel information processing. Such a network relies on the complexity of the system to adjust the relationship between the mutual connection of a large number of nodes inside, so as to achieve the purpose of processing information. Based on the excellent processing ability of neural networks for data information, the present application will use the electrical output of the neural network analysis learning system to estimate the value of the aging parameter.

[0045] For example Figure 1As shown, when other system parameters are known, based on the aging parameter R... dson C, R C The value of i can be determined by the electrothermal model to determine the corresponding electrical output i. L v o The task of condition monitoring is to estimate the values ​​of aging parameters through electrical outputs. This invention uses a neural network to learn the one-to-one correspondence between the electrical outputs of the electrothermal model and the aging parameters. When the electrothermal model matches the electrothermal characteristics of the physical entity, the neural network trained with simulation data can be used to estimate the aging parameters of the physical entity.

[0046] The specific steps for training a neural network are as follows:

[0047] We selected a five-layer feedforward neural network, with 50 neurons in each layer, and trained it using a convolutional neural network (CNN).

[0048] The data required to train the neural network comes from simulation data of the electrothermal model. Taking the Buck circuit as an example, the device aging parameters (C, R) are used. C R dson Dividing the data into ten equal parts within a reasonable numerical range yields 1000 parameter combinations. These parameters are then calculated using an electrothermal model to obtain 1000 sets of i... L With v o , change i L With v o As inputs to the neural network, C and R C R dson As the output of the neural network, out of 1000 sets of input and output data, 700 sets are used as training data, 150 sets are used as validation data, and 150 sets are used as test data; then the neural network is trained in MATLAB.

[0049] 3. Applications of Neural Networks

[0050] like Figure 2 As shown in the circuit model, in the Buck circuit, the capacitor voltage v C With output voltage v o The relationship is:

[0051]

[0052] In a Buck circuit, the capacitor used to suppress output voltage ripple is typically an electrolytic capacitor, and the equivalent series resistance R of an electrolytic capacitor... C The output voltage waveform is relatively large, which means it depends primarily on the inductor current i. L The waveform, and the capacitor voltage v C The information contained therein is overwritten, so the inductor current i L With output voltage vo The change of the capacitance value C is not sensitive, resulting in low monitoring accuracy of the capacitance value C.

[0053] As shown in Figure 1 To solve this problem, the present application uses a trained neural network and an electro-thermal model to correct the estimated capacitance value C. Experimental data i Lm And v om The trained neural network can obtain a set of aging parameters C, R C , R dson , R C , i Lm , v om Substitute into equation (6) to estimate the experimental capacitance voltage v Cm Then substitute the set of aging parameters C, R C , R dson into the electro-thermal model to calculate the corresponding simulation capacitance voltage v Cc Compare v Cm and v Cc If the difference between the two is large, a new set of aging parameters is obtained until the difference between v Cm and v Cc is less than a certain value, at which point it is considered that the estimated value of the capacitance value C is accurate.

[0054] The above examples only illustrate the technical idea of the present application and cannot limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical solution falls within the protection scope of the present application.

Claims

1. A method for monitoring the state of a DC-DC converter based on a neural network, characterized in that, An electrothermal model of the DC-DC converter is established, comprising a circuit model and a thermal model of the self-heating devices. The thermal model of the self-heating devices converts the sum of power losses of the self-heating devices into their junction temperatures. The aging parameters of the DC-DC converter are then corrected based on these junction temperatures. These aging parameters include: capacitance, equivalent series resistance of the capacitor, and on-resistance of the power switch. The thermal model of the MOSFETs or diodes in the self-heating devices is a four-layer Foster thermal network model. The model is further refined using a system of differential equations from the four-layer Foster thermal network model. Perform an inverse Laplace transform to convert the sum of power losses of the self-heating device into the junction temperature of the self-heating device. , R represents the temperature difference across the heat capacity of the i-th layer of the Foster thermal network, where i = 1, 2, 3, 4. i C i Let the thermal resistance and heat capacity of the i-th layer of the Foster thermal network be given. For the power loss of a MOSFET or diode, T c T is the case temperature of the MOSFET or diode. j This refers to the junction temperature of the MOSFET or diode; the thermal model of the capacitor in the self-heating device is a first-order Foster thermal network model. The capacitor case temperature is calculated by performing a Laplace transform on the first-order Foster thermal network model. , Let T be the temperature of the capacitor casing at time t. a For ambient temperature, P loss_C For capacitor power loss, R th With C th The thermal resistance and heat capacity of a first-order Foster thermal network model; Training the neural network: Using the electrical output obtained from the electrothermal model simulation of the DC-DC converter as input data and the aging parameters of the DC-DC converter as output data, train a neural network to obtain the mathematical relationship between the electrical output of the DC-DC converter and the aging parameters of the DC-DC converter. Neural network application: The experimentally measured electrical output of the DC-DC converter is substituted into a neural network that fits the mathematical relationship between the electrical output of the DC-DC converter and the aging parameters of the DC-DC converter. Based on a set of DC-DC converter aging parameters predicted by the neural network, experimental data of the capacitor voltage is calculated. The set of DC-DC converter aging parameters predicted by the neural network is substituted into the electrothermal model of the DC-DC converter to obtain simulated data of the capacitor voltage. When the difference between the experimental data and the simulated data of the capacitor voltage exceeds a threshold, a new set of DC-DC converter aging parameters is predicted and output by the neural network. The process of calculating the experimental data, simulated data, and the difference between the experimental data and the simulated data of the capacitor voltage is repeated until the difference between the experimental data and the simulated data of the capacitor voltage meets the threshold requirement. Finally, the set of DC-DC converter aging parameters output by the neural network is fed back into the electrothermal model of the DC-DC converter to correct the aging parameters.

2. The method for monitoring the state of a DC-DC converter based on a neural network according to claim 1, characterized in that, The electrical outputs of the DC-DC converter include inductor current and output voltage.

3. The method for monitoring the state of a DC-DC converter based on a neural network according to claim 1, characterized in that, The expression for calculating the capacitor voltage experimental data based on a set of DC-DC converter aging parameters predicted by the neural network is as follows: ,in, For output voltage, This is the capacitor voltage. For inductor current, This is the equivalent series resistance of the capacitor. For output load.

Images