CN114330197B - Parameter extraction method of IGBT numerical model based on convolutional neural network - Google Patents

Abstract

The invention relates to the technical field of IGBT process parameters, and discloses a method for extracting parameters of an IGBT numerical model based on a convolutional neural network. , constitute the original data set, expand, standardize and normalize the original data set to obtain a training data set that can be recognized by the convolutional neural network. The convolutional neural network uses the training data set to learn the relationship between the characteristic curve and the process parameters. According to the relationship, the IGBT numerical model parameters are extracted to extract the convolutional neural network. Finally, the waveform extracted from the actual circuit is used as the input value, and the IGBT numerical model parameters are used to extract the convolutional neural network to calculate and solve to obtain the process parameters. The IGBT numerical model parameter extraction method based on the convolutional neural network of the present invention effectively solves the problems of difficulty in experimental extraction of IGBT process parameters and difficulty in accurate construction.

Description

Parameter extraction method of IGBT numerical model based on convolutional neural network

technical field

The invention relates to the technical field of IGBT process parameters, in particular to a method for extracting parameters of an IGBT numerical model based on a convolutional neural network.

Background technique

With the improvement of semiconductor production process, the generation of insulated gate bipolar transistor (IGBT) products has been replaced faster and the product line has become more abundant. Four generations of IGBT modules with trench gate structure have evolved to the sixth and seventh generations of IGBT modules with fine trench gates suitable for new energy vehicles. The complexity of the product process also brings challenges to numerical software modeling and chip process inversion. In order to establish a more accurate numerical model, it is necessary to extract these process parameters reasonably and effectively.

The IGBT process parameters refer to the distribution information of the doping concentration of the device collector side, the emitter layer and the field stop layer with the doping depth. Compared with other parameters, such as the emitter-side parameters, its location and concentration can be easily extracted by formulas, and these process parameters are usually difficult to obtain or accurately reversely analyzed. Using the SRP (extended resistance test) method can roughly measure the actual parameter distribution of some devices, but its accuracy is limited and often has a large error with the actual. Process parameters also determine the core performance advantages of the device, which involve the core secrets of manufacturers and will not be disclosed to general researchers. The variation trend and sensitivity of some parameters can be approximately deduced by the method of formula derivation, but the actual doping distribution of the device is not uniform, and there are many parameter variables. In some methods, the influence of each distribution parameter change on the output characteristic change cannot be directly measured. This makes the method of exploiting formula sensitivity no longer applicable. Correspondingly, in the numerical model parameter calibration, it is often necessary to average the value within the value range of each process parameter, and accumulate a large amount of simulation data for analysis and comparison, which greatly increases the difficulty of parameter extraction.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a method for extracting parameters of IGBT numerical model based on convolutional neural network to solve the problems of difficulty in experimental extraction of IGBT process parameters and difficult to construct accurately.

In order to achieve the above purpose, the method for extracting parameters of IGBT numerical model based on convolutional neural network involved in the present invention includes the following steps:

A) build a numerical model based on TCAD numerical simulation software, and the numerical model includes the cell size structure parameters of the IGBT and the process parameters to be solved;

B) use TCAD numerical simulation software to screen and assign the process parameters, generate a process parameter value netlist, use TCAD numerical simulation software to obtain the characteristic curve output under each group of processes, the characteristic curve and the process parameter value netlist form the original dataset;

C) the original data set is subjected to data expansion, standardization and normalization to obtain a training data set that can be identified by the convolutional neural network;

D) the convolutional neural network uses the training data set to learn the relationship between the characteristic curve and the process parameters, and generates the IGBT numerical model parameter extraction convolutional neural network;

E) Using the waveform extracted from the actual circuit as the input value, using the IGBT numerical model parameters to extract the convolutional neural network to calculate and solve, and obtain the process parameters.

Preferably, in the step A), the cell size and structure parameters are observed and extracted by using an observation tool to observe and extract the chip cross-section sample, and the process parameters include the peak collector concentration Pcollector, the collector junction depth Hemit, and the concentration at the collector junction. Pjuntion, field stop concentration Nbuff and field stop junction depth Hcut.

Preferably, in the step B), the process parameters are screened and assigned, including the following steps:

B1) determine the value range of the process parameter by means of simulation;

B2) The process parameters are valued by the method of taking values at equal intervals and adding interpolation.

Preferably, in the step B1), use TCAD numerical simulation software to screen and assign process parameters, by solving the on-voltage drop Vces of the device, the breakdown voltage Vcesat, and the device loss Eoff, and meet the following conditions:

Vces>1800V

Vcesat=2±0.15V

Eoff=1500±200mJ

Among them, the environmental conditions for calculating the breakdown voltage Vcesat are the gate voltage Vg=15V, the collector current is 3600A, the ambient temperature is 25°C, the environmental conditions for calculating the device loss Eoff are the driving resistance 1 ohm, the device emits collector current Ic= The shutdown is completed under the condition of 3600A, and the ambient temperature is 25 degrees Celsius. On the premise that the above three conditions are met at the same time, the value range of the process parameters is obtained.

Preferably, in the step B2), firstly take values at equal intervals based on the value range, calculate the output characteristic curve of the parameter under a total of k values, and extract the saturated on-voltage of the device under each value through the characteristic curve drop and turn-off loss, where, under the values of two adjacent parameters, there are interpolation ΔV _cesat and interpolation ΔE _off , and the calculated sensitivity k _i =[(ΔVcesat/Vcesat) ² +(ΔEoff/Eoff) ² ] ^{1/ 2} /(Δm _i /m _i ), where i represents the i-th group of parameters, m _i represents the specific value of any parameter in the process parameters, Δm _i =m _i -m _i-1 represents the two most adjacent parameters of the parameter Take the difference between the values and calculate the k sensitivity averages k'. When the sensitivity ki>k', insert the mean point m' _i =(m _i +m _i+1 )/ 2.

Preferably, in the step B), in the step B), the characteristic curve includes the relationship curve between the device collector current Ic and the collector-emitter voltage Vce, the transient turn-off collector-emitter voltage Vce with time Time variation curve and the collector current Ict variation curve with time at transient turn-off.

Preferably, in the step C), the original data set is subjected to data expansion, standardization and normalization, including the following steps:

C1) Using the means of parity interval sampling, according to the data parity of the original data set, after the even sequence of the original data set is sampled at equal intervals, perform equal interval resampling on its adjacent odd bits to obtain the same The output data of the data length realizes the data expansion of the limited number of samples;

C2) Arrange all data in the original data set after data expansion into a data chain with a length of 512 bits, including three groups of two-dimensional column vectors [Vce,Ic],[Time,Ict],[Time,Vce];

C3) Use the normalization strategy for the data in the data chain, and find the maximum value of this type of data in the 5000 groups of data chains for three groups of vectors, respectively, denoted as Icmax, Ictmax, Vcemax, using Ic=Ic/Icmax, Ict= Ictmax, Vce=Vcemax, the normalized data is obtained, and its range is within (0, 1), and then the sequence of the data chain is disrupted and the end-to-end splicing of the arrays in the data chain is completed, and finally a complete input data chain is obtained.

Preferably, in the step D), the convolutional neural network adopts a keras network architecture and has a four-layer structure, including a first convolutional layer and a pooling layer, a second convolutional layer and a pooling layer, and flattened layer and fully connected layer, the input layer of the convolutional neural network is a data chain containing several groups of IGBT characteristic curves, and the output layer is the value of the process parameters.

Preferably, in the step D), the training process of the convolutional neural network is as follows: the input data chain is brought into the first convolution layer, the length of the data chain is 5000 at this time, and the data chain is not expanded by convolution, The depth information is 1, which is counted as (5000, 1), which is convolved by a filter composed of 64 neurons, and activated by the ELU function. The ELU function can be expressed as: f(x)={x(x>0 ); (e ^x -1)(x≤0)}, the data structure is (2500, 64), the information dimension is expanded, and then the first pooling layer is connected, the pooling scaling factor is 5, and the length of the data is further Zoom out to get the data structure as (500, 64), then enter the second convolution layer, the data becomes (250, 128), enter the second pooling layer, the data structure becomes (25, 256), the data at this time After the flattening layer processing, all dimensions are flattened to obtain the data structure (25*256, 1), and then through the full connection layer, the final data structure obtained is (5, 1), and a data structure containing 5 output information is obtained at this time. one-dimensional parameters.

Preferably, in the step E), the actual circuit includes a voltage source V _DC , a power supply terminal capacitor C _DC , a stray inductance Ls at the line connection, a discharge IGBT T1 and an IGBT T2 to be tested, wherein the voltage source VDC is connected through the line The stray inductance Ls is connected to the collector of the discharge IGBT T1, the emitter of the discharge IGBT T1 is connected to the collector of the IGBT T2 to be tested, and the emitter of the IGBT T2 to be tested is connected to the negative electrode of the voltage source V _DC , In the IGBT T2 circuit to be tested, V _GE1 is the gate drive voltage, which can control the turn-on and turn-off time of the IGBT in the circuit, so that the device can be turned off at a specific current, and the device is turned off transient collector-emitter is obtained by means of circuit port testing. Voltage Vce, collector current Ic with time change curve, collector current Ic with collector-emitter voltage Vce change curve, using the step C) preprocessing method, obtain the reorganization curve to obtain the input data link, bring into the training completed IGBT The numerical model parameters are extracted from the neural network, and the process parameters to be solved are calculated.

Compared with the prior art, the present invention has the following advantages:

1. The process parameter extraction method based on neural network training can avoid the traditional method based on SPR experiment, save the time and cost of model construction, and provide an alternative for some units that lack professional experimental equipment or conditions to extract process parameters. Program;

2. The neural network method can fully mobilize the advantages of TCAD software, and the process parameter extraction accuracy is improved through the learning and training process of the data set, so that the model is more in line with the typical characteristics of the actual device;

3. Compared with the method of extracting parameters using experimental curves and approximate formulas, this method can better adapt to the complex process parameters on the collector side of the device, the model accuracy is better, and it can also better adapt to the needs of numerical models;

4. Provide an implementable method for process optimization and improvement of semiconductor manufacturers. The improvement and optimization of parameters in device production can also be realized through a small amount of experiments and numerical model simulation, thereby saving development time.

Description of drawings

Fig. 1 is the schematic diagram of the chip cell structure in the IGBT numerical model parameter extraction method based on convolutional neural network of the present invention;

2 is a reference diagram of the device structure and process parameters of the IGBT in the present invention;

3 is a circuit diagram of a dynamic IGBT on-off test platform in the present invention;

4 is a schematic diagram of data expansion in the present invention;

Fig. 5 is the standardization normalization and random sorting schematic diagram of data in the present invention;

Fig. 6 is the relation diagram between the model training result of the process parameter value that uses neural network model training in the present invention and the number of times of training;

Fig. 7 is the relation diagram between the model training result and the training times of the process parameter value that uses simulation training to provide in the present invention;

Fig. 8 is the static characteristic comparison diagram of simulation result and actual experiment result in the present invention;

Fig. 9 is the dynamic characteristic comparison diagram of simulation result and actual experiment result under a working condition in the present invention;

FIG. 10 is a comparison diagram of the dynamic characteristics between the simulation results and the actual experimental results under another working condition in the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

A method for extracting parameters of an IGBT numerical model based on a convolutional neural network, comprising the following steps:

A) Build a numerical model based on TCAD numerical simulation software. The numerical model includes the cell size structure parameters of the IGBT and the process parameters to be solved. The process parameters include the peak collector concentration Pcollector, the collector junction depth Hemit, and the collector junction concentration Pjuntion, The field cutoff layer concentration Nbuff, the field cutoff layer junction depth Hcut, the cell size and structure parameters are observed and extracted by using observation tools to observe and extract the chip cross-section sample. In this embodiment, the chip cross-section sample is observed by scanning electron microscope. The doping position of the chip was determined experimentally, and the geometric model of the corresponding IGBT cell area was constructed by using the TCAD numerical simulation software. Finally, the cell structure of the chip was shown in Figure 1. Combined with the observation of the chip electron microscope and the results of the dye-junction experiment, Establish the cell size structure parameters of the device;

B) Use TCAD numerical simulation software to screen and assign process parameters, generate a netlist for the value of process parameters, and use TCAD numerical simulation software to obtain the output characteristic curves of each group of processes. The characteristic curves include device collector current Ic and collector-emitter voltage. The relationship curve between Vce, the time curve of the transient turn-off collector-emitter voltage Vce, and the time curve of the collector current Ict at the time of transient turn-off, the latter two characteristic curves are the transient characteristic curves of the device, so The corresponding circuit is shown in Figure 3. The characteristic curve and the process parameter value netlist constitute the original data set. Among them, the process parameters are screened and assigned, including the following steps:

B1) Determine the value range of the process parameters by means of simulation. Refer to Figure 2 for the device structure and process parameters of the IGBT, and determine the range of values by means of simulation. Screening and assignment is performed by solving the on-voltage drop Vces, breakdown voltage Vcesat, and device loss Eoff of the device, and the following conditions are met:

Vces>1800V

Vcesat=2±0.15V

Eoff=1500±200mJ

Among them, the environmental conditions for calculating the breakdown voltage Vcesat are the gate voltage Vg=15V, the collector current is 3600A, the ambient temperature is 25°C, the environmental conditions for calculating the device loss Eoff are the driving resistance 1 ohm, the device emits collector current Ic= The shutdown is completed under the condition of 3600A, and the ambient temperature is 25 degrees Celsius. On the premise that the above three conditions are met at the same time, five process parameter value ranges are obtained, as shown in Table 1;

Table 1 IGBT process parameter description and its range

Chinese name corresponding symbol Ranges unit Collector peak concentration Pcollector 1e17～2e18 /cm³ collector junction depth Hemit 0.6-2 um Concentration at collector junction Pjuntion 1e14～1e15 /cm³ Field Stop Layer Concentration Nbuff 8e15～1.8e16 /cm3 Field stop layer junction depth Hcut 6～30 um

B2) Use the method of taking values at equal intervals and adding interpolation to take values for the process parameters. First, take values at equal intervals based on the value range, calculate a total of k values, and output the characteristic curve for this parameter, and extract each parameter from the characteristic curve. The saturation turn-on voltage drop and turn-off loss of the device under the value of , where, under the values of two adjacent parameters, there are interpolation ΔV _cesat and interpolation ΔE _off , and the sensitivity k _i =[(ΔVcesat/Vcesat) ² +( ΔEoff/Eoff) ² ] ^1/2 /(Δm _i /m _i ), where i represents the i-th group of parameters, _mi represents the specific value of any parameter in the process parameters, and Δm _i =m _i -m _i-1 represents The difference between the two most adjacent values of the parameter is calculated to calculate k sensitivity average values k'. When the sensitivity ki>k', insert the mean point m' in these two parameters _i = (m _i +m _i+1 )/2;

C) The original data set is subjected to data expansion, standardization and normalization to obtain a training data set that can be recognized by the convolutional neural network. The specific steps are as follows:

C1) Using the method of odd-even interval sampling, according to the data parity of the original data set, after the even-numbered sequence of the original data set is sampled at equal intervals, the adjacent odd-numbered bits are resampled at equal intervals to obtain the output of the same data length data to achieve data expansion for a limited number of samples;

C2) Arrange all data in the original data set after data expansion into a data chain with a length of 512 bits, including three groups of two-dimensional column vectors [Vce,Ic],[Time,Ict],[Time,Vce];

C3) Use the normalization strategy for the data in the data chain, and find the maximum value of this type of data in the 5000 groups of data chains for three groups of vectors, respectively, denoted as Icmax, Ictmax, Vcemax, using Ic=Ic/Icmax, Ict= Ictmax, Vce=Vcemax, the normalized data is obtained, and its range is within (0, 1), and then the sequence of the data chain is scrambled and the end-to-end splicing of the arrays in the data chain is completed, and a complete input data chain is finally obtained;

Taking the curve of transient turn-off collector-emitter voltage Vce in the original data set as an example, 512 data points are extracted from the original data set in the form of sampling at equal intervals, and the ordinal numbers of all sampling points in the original data are For even-numbered points, resample the odd-numbered points to the right of each even-numbered point to double the original data expansion, as shown in Figure 4;

Then, the expanded and doubled data is arranged according to the standard data chain format. Each group of data chains contains three sets of two-dimensional column vectors [Vce,Ic],[Time,Ict],[Time,Vce], and the length of the column vector Both are 512, and the number of data chains is 5,000. Then, normalize the standardized data to find the maximum value of the three sets of vectors in the 5,000 sets of data chains, and record them as Icmax, Ictmax, Vcemax. Using Ic=Ic/Icmax, Ict=Ictmax, Vce=Vcemax, the normalized data is obtained, and its range is within (0, 1), and then the original data chain is processed out of order, and the end-to-end splicing is completed, Finally, a complete input data chain is obtained. At the same time, the process parameter value netlist is spliced in the disordered order, and the output data chain is obtained. The whole process is shown in Figure 5. The input and output data chain finally constitutes the training data. set;

D) The convolutional neural network uses the training data set to learn the relationship between the characteristic curve and the process parameters, and generates the IGBT numerical model parameters to extract the convolutional neural network. The convolutional neural network adopts the keras network architecture and has a four-layer structure, including the first Convolutional layer and pooling layer, second convolutional layer and pooling layer, flattening layer and fully connected layer, the input layer of the convolutional neural network is a data chain containing several groups of IGBT characteristic curves, and the output layer is the process parameters The training process of the convolutional neural network is: the input data chain is brought into the first convolutional layer, the length of the data chain is 5000 at this time, the data chain has not been expanded by convolution, and the depth information is 1, which is counted as ( 5000, 1), which is convolved by a filter composed of 64 neurons, and activated by the ELU function. The ELU function can be expressed as: f(x)={x(x>0); (e ^x -1) (x≤0)}, the data structure is (2500, 64), the information dimension is expanded, and then the first pooling layer is connected, the pooling scaling factor is 5, the length of the data is further reduced, and the data structure is (500 , 64), then enter the second convolution layer, the data becomes (250, 128), enter the second pooling layer, the data structure becomes (25, 256), the data at this time is processed by the flattening layer, and all The dimension is flattened to obtain the data structure (25*256, 1), and then through the fully connected layer, the final data structure is (5, 1), and a one-dimensional parameter containing 5 output information is obtained at this time;

In the single learning process of the convolutional neural network, the input data chain in Figure 5 is used to obtain the output data chain under the convolutional neural network. Since the data chain contains five parameters, the length of the output data chain in the training data set is is 5000, so i is the serial number of the parameter, k is the serial number of the group of parameters in the data chain, each element of the output data chain obtained by using the convolutional neural network is m _i,k , and the process parameter value netlist is composed of The output data chain is counted as Mi _,k , and β is the error cost function, β _k =0.2∑[(mi,k-Mi,k)/Mi,k] ² i=1,2,3,4,5 , β=∑β _k , k=1,2,...5000, through the iterative strategy, change the weight value in the convolutional neural network to make the error cost function optimal, if the error function is still not improved after 20 iterative adjustments , the convolution operation is stopped, and the training is considered to meet the requirements. Figures 6 and 7 show the results after the training network processes 1000 data, the process parameter values obtained by the neural network model training and the process parameter values given by the simulation training. From the comparison between the two, it can be seen that the neural network can better simulate the corresponding relationship, and in the case of limited samples, its learning effect is better, and the result error is small;

E) Use the waveform extracted from the actual circuit as the input value, use the IGBT numerical model parameters to extract the convolutional neural network for calculation and solution, and obtain the process parameters. The actual circuit includes the voltage source V _DC , the power supply terminal capacitor C _DC , and the stray inductance Ls at the line connection , the discharge IGBT T1 and the IGBT T2 to be tested, in which the voltage source VDC is connected to the collector of the discharge IGBT T1 through the stray inductance Ls at the line connection, and the emitter of the discharge IGBT T1 is connected to the collector of the IGBT T2 to be tested, and then by The emitter of the IGBT T2 to be tested is connected to the negative electrode of the voltage source V _DC . In the circuit of the IGBT T2 to be tested, V _GE1 is the gate driving voltage, which can control the turn-on and turn-off time of the IGBT in the circuit, so that the device can be turned off at a specific current. The circuit port test method is used to obtain the device turn-off transient collector-emitter voltage Vce, the collector current Ic changing curve with time, the collector current Ic changing curve with the collector-emitter voltage Vce, using step C) preprocessing method to obtain the reorganization The curve is obtained from the input data chain and brought into the trained IGBT numerical model parameter extraction neural network to calculate the process parameters to be solved.

In order to further verify the accuracy of the process parameters, the experimental waveforms and simulation waveforms of the model under the process parameters were compared under multiple working conditions. The final effect is shown below with an example. The content of Table 2 shows the comparison of the results of the traditional search method and the neural network parameter extraction method.

Table 2 Comparison of the output results of the traditional method and the neural network method

Figures 8 to 10 are the comparison diagrams of the final results, of which Figure 8 is the comparison of the static characteristics of the data sheet and the device simulation curve, and Figures 9 and 10 are the comparison of the dynamic turn-off characteristics under different operating conditions. Well, the accuracy requirement can be met, and the correctness of the present invention is verified.

In the present invention, the method of using the neural network can improve the process parameter extraction method, and provide a new method for fitting some process parameters that are difficult to determine. The model can improve the accuracy and efficiency of model extraction. In addition to the above-mentioned examples, the present invention can also be used for the adjustment of some process parameters in the production process of semiconductor devices such as IGBTs. During the debugging process, a process netlist can be constructed near the process parameters to be adjusted, and the neural network can be trained by using the process netlist to output simulation results. . Finally, use the target output curve to find the optimal input process parameters.

Claims (7)

1. a kind of IGBT numerical model parameter extraction method based on convolutional neural network, is characterized in that: comprise the steps: A) build a numerical model based on TCAD numerical simulation software, and the numerical model includes the cell size structure parameters of the IGBT and the process parameters to be solved; B) use TCAD numerical simulation software to screen and assign the process parameters, generate a process parameter value netlist, use TCAD numerical simulation software to obtain the characteristic curve output under each group of processes, the characteristic curve and the process parameter value netlist The original data set is formed, and the process parameters are screened and assigned, including the following steps: B1) Determine the value range of the process parameters by the method of simulation, use TCAD numerical simulation software to screen and assign the process parameters, and solve the on-voltage drop Vces of the device, breakdown voltage Vcesat, and device loss Eoff, and satisfy the following condition: Vces>1800V Vcesat=2±0.15V Eoff=1500±200mJ Among them, the environmental conditions for calculating the breakdown voltage Vcesat are the gate voltage Vg=15V, the collector current is 3600A, the ambient temperature is 25°C, the environmental conditions for calculating the device loss Eoff are the driving resistance 1 ohm, the device emits collector current Ic= The shutdown is completed under the condition of 3600A, and the ambient temperature is 25 degrees Celsius. On the premise that the above three conditions are met at the same time, the value range of the process parameters is obtained; B2) adopt the method of taking values at equal intervals and adding interpolation to take values for the process parameters, first take values at equal intervals based on the value range, calculate a total of k values, and output the characteristic curve of the parameter, and extract each value through the characteristic curve The saturation turn-on voltage drop and turn-off loss of the device under these values, where, under the adjacent two parameter values, there are interpolation ΔV _cesat and interpolation ΔE _off , and the sensitivity k _i =[(ΔVcesat/Vcesat) ² +(ΔEoff/Eoff) ² ] ^1/2 /(Δm _i /m _i ), where i represents the i-th group of parameters, _mi represents the specific value of any one of the process parameters, Δm _i =m _i -m _{i- 1} represents the difference between the two most adjacent values of the parameter, and calculates k sensitivity averages k'. When the sensitivity ki>k', insert the mean point m' in these two parameters _i = (m _i +m _i+1 )/2; C) the original data set is subjected to data expansion, standardization and normalization to obtain a training data set that can be identified by the convolutional neural network; D) the convolutional neural network uses the training data set to learn the relationship between the characteristic curve and the process parameters, and generates the IGBT numerical model parameter extraction convolutional neural network; E) Using the waveform extracted from the actual circuit as the input value, using the IGBT numerical model parameters to extract the convolutional neural network to calculate and solve, and obtain the process parameters. 2. The IGBT numerical model parameter extraction method based on convolutional neural network according to claim 1, is characterized in that: in described step A), described cell size structure parameter utilizes observation tool to observe and extract chip cross-section sample, The process parameters include the collector peak concentration Pcollector, the collector junction depth Hemit, the collector junction concentration Pjuntion, the field stop layer concentration Nbuff and the field stop layer junction depth Hcut. 3. The IGBT numerical model parameter extraction method based on convolutional neural network according to claim 1, wherein: in the step B), in the step B), the characteristic curve comprises the device collector current Ic and the The relationship curve between the collector-emitter voltage Vce, the transient-off collector-emitter voltage Vce vs. time Time curve, and the transient-off collector current Ict vs. time Time curve. 4. The method for extracting parameters of IGBT numerical model based on convolutional neural network according to claim 3, characterized in that: in the step C), the original data set is subjected to data expansion, standardization and normalization and comprises the following steps : C1) Using the means of parity interval sampling, according to the data parity of the original data set, after the even sequence of the original data set is sampled at equal intervals, perform equal interval resampling on its adjacent odd bits to obtain the same The output data of the data length realizes the data expansion of the limited number of samples; C2) Arrange all data in the original data set after data expansion into a data chain with a length of 512 bits, including three groups of two-dimensional column vectors [Vce,Ic],[Time,Ict],[Time,Vce]; C3) Use the normalization strategy for the data in the data chain, and find the maximum value of this type of data in the 5000 groups of data chains for three groups of vectors, respectively, denoted as Icmax, Ictmax, Vcemax, using Ic=Ic/Icmax, Ict= Ictmax, Vce=Vcemax, the normalized data is obtained, and its range is within (0, 1), and then the sequence of the data chain is disrupted and the end-to-end splicing of the arrays in the data chain is completed, and finally a complete input data chain is obtained. 5. The IGBT numerical model parameter extraction method based on convolutional neural network according to claim 1, is characterized in that: in described step D), described convolutional neural network adopts keras network architecture, and there is a four-layer structure, including The first convolutional layer and the pooling layer, the second convolutional layer and the pooling layer, the flattening layer and the fully connected layer, the input layer of the convolutional neural network is a data chain containing several groups of IGBT characteristic curves, and the output The layer is the value of the process parameter. 6. The IGBT numerical model parameter extraction method based on convolutional neural network according to claim 5, wherein: in the step D), the training process of the convolutional neural network is: the input data link is brought into the first In the convolutional layer, the length of the data chain is 5000 at this time, the data chain is not expanded by convolution, the depth information is 1, which is counted as (5000, 1), and the filter composed of 64 neurons is used for convolution operation. The ELU function is activated, and the ELU function can be expressed as: f(x)={x(x>0); (e ^x -1)(x≤0)}, the obtained data structure is (2500, 64), and the information dimension is expanded , and then connect the first pooling layer, the pooling scaling factor is 5, and the length of the data is further reduced to obtain a data structure of (500, 64), and then enter the second convolution layer, the data becomes (250, 128) , enter the second pooling layer, the data structure becomes (25, 256), the data at this time is processed by the flattening layer, and all dimensions are flattened to obtain the data structure (25*256, 1), and then go through the fully connected layer , the final data structure is (5, 1), and a one-dimensional parameter containing 5 output information is obtained at this time. 7. The IGBT numerical model parameter extraction method based on convolutional neural network according to claim 1, wherein: in the step E), the actual circuit comprises a voltage source V _DC , a power supply terminal capacitor C _DC , a line connection The stray inductance Ls, the discharge IGBT T1 and the IGBT T2 to be tested, where the voltage source VDC is connected to the collector of the discharge IGBT T1 through the stray inductance Ls at the line connection, and the emitter of the discharge IGBT T1 is connected to the IGBT T2 to be tested. The collector is connected to the negative electrode of the voltage source V _DC by the emitter of the IGBT T2 to be tested. In the circuit of the IGBT T2 to be tested, V _GE1 is the gate driving voltage, which can control the on-off time of the IGBT in the circuit, so that the device can operate in The turn-off is completed at a specific current, and the device turn-off transient collector-emitter voltage Vce is obtained by means of the circuit port test, the collector current Ic varies with time, and the collector current Ic varies with the collector-emitter voltage Vce. Step C) preprocessing method is used to obtain the shuffling curve to obtain the input data chain, which is brought into the trained IGBT numerical model parameter extraction neural network to calculate the process parameters to be solved.

Images