CN114330197B - Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network - Google Patents
Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network Download PDFInfo
- Publication number
- CN114330197B CN114330197B CN202210250735.4A CN202210250735A CN114330197B CN 114330197 B CN114330197 B CN 114330197B CN 202210250735 A CN202210250735 A CN 202210250735A CN 114330197 B CN114330197 B CN 114330197B
- Authority
- CN
- China
- Prior art keywords
- igbt
- data
- neural network
- convolutional neural
- parameters
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
- Testing Or Measuring Of Semiconductors Or The Like (AREA)
Abstract
The invention relates to the technical field of IGBT (insulated gate bipolar transistor) process parameters, and discloses an IGBT numerical model parameter extraction method based on a convolutional neural network, which comprises the following steps of: the method comprises the steps of constructing a numerical model by using TCAD numerical simulation software, generating a process parameter value net list and an output characteristic curve to form an original data set, performing data expansion, standardization and normalization on the original data set to obtain a training data set which can be identified by a convolutional neural network, learning the relation between the characteristic curve and the process parameters by using the training data set by the convolutional neural network, generating IGBT numerical model parameters to extract the convolutional neural network, and finally extracting the convolutional neural network by using waveforms extracted by an actual circuit as input values and extracting the convolutional neural network by using the IGBT numerical model parameters to calculate and solve to obtain the process parameters. The method for extracting the parameters of the IGBT numerical model based on the convolutional neural network effectively solves the problems that the extraction of an IGBT process parameter experiment is difficult and the IGBT process parameter experiment is difficult to construct accurately.
Description
Technical Field
The invention relates to the technical field of IGBT (insulated gate bipolar transistor) process parameters, in particular to an IGBT numerical model parameter extraction method based on a convolutional neural network.
Background
With the improvement of semiconductor production technology, Insulated Gate Bipolar Transistor (IGBT) products are replaced more quickly, and product lines are more abundant, for example, from the original PT type IGBT module with low power to the IGBT module with a trench gate structure of the fourth generation which is commonly used for high power at present, to the IGBT module with the fine trench gate of the sixth generation and the seventh generation which are currently suitable for new energy vehicles, and the like. The complexity of the product process also brings challenges to the numerical software modeling and the chip process, and the process parameters need to be reasonably and effectively extracted to establish a more accurate numerical model.
The IGBT process parameters refer to distribution information of doping concentration of the collector side, the emission layer and the field stop layer of the device along with doping depth. The location and concentration of these process parameters are easily formulated compared to other parameters, such as emitter-side parameters, which are often difficult to obtain or accurately analyzed in reverse. The method of using SRP (extended resistance test) can roughly measure the actual parameter distribution of some devices, but its precision is limited, and the error is often larger than the actual one. The process parameters also determine the core performance advantages of the device, relate to the core secrets of the manufacturer and are not disclosed to general researchers. The method of formula derivation can be used for approximately deriving the variation trend and the variation sensitivity of certain parameters, but the actual doping distribution of the device is not uniform, the parameter variables are more, the value variation of a certain parameter is often determined by a plurality of distribution quantities, and the influence of the variation of each distribution parameter on the variation of the output characteristic cannot be directly measured in the existing means. This results in that the method using formula sensitivity is no longer applicable. Correspondingly, in the parameter calibration of the numerical model, the average value is often required to be taken within the value range of each process parameter, and a large amount of simulation data is accumulated for analysis and comparison, so that the difficulty in parameter extraction is greatly increased.
Disclosure of Invention
The invention aims to provide a method for extracting parameters of an IGBT numerical model based on a convolutional neural network aiming at the defects of the technology, and effectively solves the problems that the extraction of an IGBT process parameter experiment is difficult and the IGBT process parameter experiment is difficult to construct accurately.
In order to achieve the above purpose, the invention relates to a method for extracting parameters of an IGBT numerical model based on a convolutional neural network, which comprises the following steps:
A) constructing a numerical model based on TCAD numerical simulation software, wherein the numerical model comprises cell size structure parameters of the IGBT and technological parameters to be solved;
B) screening and assigning the process parameters by using TCAD numerical simulation software to generate a process parameter value netlist, obtaining a characteristic curve output under each group of processes by using the TCAD numerical simulation software, wherein the characteristic curve and the process parameter value netlist form an original data set;
C) expanding, standardizing and normalizing the original data set to obtain a training data set for the convolutional neural network to identify;
D) the convolutional neural network learns the relation between the characteristic curve and the process parameters by utilizing a training data set, and generates IGBT numerical model parameters to extract the convolutional neural network;
E) And (3) extracting the waveform extracted by the actual circuit as an input value, and extracting the convolutional neural network by using the parameters of the IGBT numerical model to calculate and solve to obtain process parameters.
Preferably, in the step a), the cell size structure parameters are observed and extracted from the chip cross-section sample by using an observation tool, and the process parameters include a collector peak concentration Pcollector, a collector junction depth Hemit, a collector junction concentration Pjuntion, a field cut-off layer concentration Nbuff and a field cut-off layer junction depth Hcut.
Preferably, in the step B), the screening and assigning values to the process parameters includes the following steps:
B1) determining the value range of the process parameters by a simulation method;
B2) and adopting a method of equal interval value taking and interpolation to take values of the process parameters.
Preferably, in the step B1), TCAD numerical simulation software is used to perform screening assignment on the process parameters, and by solving the conduction voltage drop Vces, breakdown voltage Vcesat, and device loss Eoff of the device, the following conditions are satisfied:
Vces>1800V
Vcesat＝2±0.15V
Eoff＝1500±200mJ
the environment condition for calculating the breakdown voltage Vcesat is that the gate voltage Vg is 15V, the collector current is 3600A, the environment temperature is 25 ℃, the environment condition for calculating the loss Eoff of the device is that the driving resistor is 1 ohm, the device is turned off under the condition that the emission collector current Ic is 3600A, the environment temperature is 25 ℃, and the value range of the process parameter is obtained on the premise that the three conditions are met.
Preferably, in the step B2), the values are taken at equal intervals based on the value range, a characteristic curve is output by the parameter under k values, and the saturation on-state voltage drop and the turn-off loss of the device under each value are extracted through the characteristic curve, wherein interpolation Δ V exists between two adjacent values of the parameter cesat And interpolated Δ E off Calculating the sensitivity k i ＝[(ΔVcesat/Vcesat) 2 +(ΔEoff/Eoff) 2 ] 1/2 /(Δm i /m i ) Where i denotes the i-th set of parameters, m i Indicating a specific value, Δ m, of any one of the process parameters i ＝m i -m i-1 Representing the difference between two most adjacent values of the parameter, and calculating k average values of sensitivity k' when the sensitivity ki>k ', an average value point m ' is inserted into the two parameters ' i ＝(m i +m i+1 )/2。
Preferably, in step B), the characteristic curves include a relation curve between the device collector current Ic and the collector-emitter voltage Vce, a Time variation curve of the transient off collector-emitter voltage Vce, and a Time variation curve of the transient off collector current Ict.
Preferably, in the step C), the data expansion, normalization and normalization of the raw data set comprises the following steps:
C1) adopting a means of odd-even interval sampling, according to the data parity of the original data set, after equal interval sampling is carried out on an even sequence of the original data set, carrying out equal interval resampling on adjacent odd bits of the original data set, obtaining output data with the same data length, and realizing data expansion of limited number of samples;
C2) Arranging all data in an original data set after data expansion into a data chain with the length of 512 bits, wherein the data chain comprises three groups of two-dimensional column vectors of [ Vce, Ic ], [ Time, Ict ], [ Time, Vce ];
C3) and (2) adopting a normalization strategy for data in the data chain, respectively finding the maximum values of the three groups of vectors in 5000 groups of data chains, respectively recording the maximum values as Icmax, Icmax and Vcemax, obtaining normalized data with the ranges of (0 and 1) by using Ic/Icmax, Ict as Icmax and Vce as Vcemax, and then sequentially disordering the data chains, completing the head and tail splicing of the arrays in the data chains, and finally obtaining a complete input data chain.
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, a flattening layer and a full-connection layer, an input layer of the convolutional neural network is a data chain including a plurality of groups of IGBT characteristic curves, and an output layer is a value of a process parameter.
Preferably, in the step D), the training process of the convolutional neural network is: the input data chain is brought into a first convolution layer, the length of the data chain is 5000 at this time, the data chain is not subjected to convolution expansion, the depth information is 1, the value is (5000, 1), the convolution operation is carried out by a filter consisting of 64 neurons, an ELU function is adopted for activation, and the ELU function can be expressed as: (x) { x (x > 0); (e) x -1) (x is less than or equal to 0), obtaining a data structure (2500, 64), expanding information dimensions, then connecting a first pooling layer, wherein a pooling scaling factor is 5, further reducing the length of the data to obtain the data structure (500, 64), then entering a second convolution layer, changing the data into (250, 128), entering a second pooling layer, changing the data structure into (25, 256), flattening all dimensions through flattening processing of the data at the moment to obtain the data structure (25 x 256, 1), then passing through a full connection layer, finally obtaining the data structure (5, 1), and then obtaining a one-dimensional parameter containing 5 output information.
Preferably, in said step E), said actual circuit comprises a voltage source V DC Capacitor C at power supply end DC The device comprises a line junction stray inductance Ls, a discharging IGBT T1 and an IGBT T2 to be tested, wherein a voltage source VDC is connected to a collector of the discharging IGBT T1 through the line junction stray inductance Ls, an emitter of the discharging IGBT T1 is connected to a collector of the IGBT T2 to be tested, and the emitter of the IGBT T2 to be tested is connected with the voltage source V DC Negative pole of (2), V in IGBT T2 circuit to be tested GE1 And C) preprocessing means in the step C) is adopted to obtain a reorganization curve to obtain an input data chain, the reorganization curve is substituted into the trained IGBT numerical model parameter extraction neural network, and the process parameters to be solved are calculated.
Compared with the prior art, the invention has the following advantages:
1. the process parameter extraction method based on neural network training can avoid adopting the traditional method based on SPR experiment, save the time and cost of model construction, and provide an alternative scheme for extracting process parameters for units lacking specialized experimental equipment or conditions;
2. the advantages of TCAD software can be fully adjusted by adopting a neural network method, and the extraction precision of process parameters is improved through the learning training process of a data set, so that the model is more in line with the typical characteristics of an actual device;
3. compared with a method for extracting parameters by using an experimental curve and an approximate formula, the method can better adapt to complex process parameters on the collector side of the device, has better model precision and can better adapt to the requirements of a numerical model;
4. the method provides an implementable method for the process optimization improvement of semiconductor production enterprises, improves and optimizes parameters in device production, and can also be realized by a small amount of experiments and numerical model simulation, thereby saving development time.
Drawings
FIG. 1 is a schematic diagram of a chip cell structure in the method for extracting parameters of an IGBT numerical model based on a convolutional neural network;
FIG. 2 is a reference diagram of the device structure and process parameters of the IGBT of the invention;
FIG. 3 is a circuit diagram of a dynamic IGBT turn-on and turn-off test platform according to the invention;
FIG. 4 is a schematic diagram of data expansion according to the present invention;
FIG. 5 is a diagram illustrating normalized normalization and random ordering of data in accordance with the present invention;
FIG. 6 is a graph of the relationship between the model training results and the training times for process parameter values obtained using neural network model training in the present invention;
FIG. 7 is a graph showing the relationship between the training results and the number of times of training of a model using values of process parameters given by simulation training in accordance with the present invention;
FIG. 8 is a comparison graph of the static characteristics of simulation results and actual experimental results in the present invention;
FIG. 9 is a diagram showing the dynamic characteristics of the simulation result and the actual experiment result under one condition of the present invention;
fig. 10 is a comparison graph of the simulation result and the actual experiment result dynamic characteristics under another working condition in the present invention.
Detailed Description
The invention is described in further detail below with reference to the figures and the specific embodiments.
An IGBT numerical model parameter extraction method based on a convolutional neural network comprises the following steps:
A) the method comprises the steps of constructing a numerical model based on TCAD numerical simulation software, wherein the numerical model comprises cell size structural parameters of an IGBT and technological parameters to be solved, the technological parameters comprise a collector peak value concentration Pcollector, a collector junction depth Hemit, a collector junction concentration Pjunction, a field stop layer concentration Nbaff and a field stop layer junction depth Hcut, and the cell size structural parameters are observed and extracted on a chip section sample by using an observation tool;
B) The method comprises the following steps of screening and assigning process parameters by using TCAD numerical simulation software to generate a process parameter value netlist, obtaining characteristic curves output under each group of processes by using the TCAD numerical simulation software, wherein the characteristic curves comprise a relation curve between collector current Ic and collector-emitter voltage Vce of a device, a Time variation curve of transient off collector-emitter voltage Vce along with Time and a Time variation curve of collector current Ict along with Time during transient off, the last two characteristic curves are transient characteristic curves of the device, a corresponding circuit is shown in figure 3, and the characteristic curves and the process parameter value netlist form an original data set, wherein the process parameters are screened and assigned, and the method comprises the following steps:
B1) the value range of the process parameters is determined by a simulation method, the device structure and the process parameters of the IGBT refer to the graph 2, and the value range is determined by the simulation method, and the specific implementation method comprises the following steps: 5 process parameters are screened and assigned by using TCAD numerical simulation software, and the following conditions are met by solving the conduction voltage drop Vces, the breakdown voltage Vcesat and the device loss Eoff of the device:
Vces>1800V
Vcesat＝2±0.15V
Eoff＝1500±200mJ
the environment condition for calculating the breakdown voltage Vcesat is that the gate voltage Vg is 15V, the collector current is 3600A, the environment temperature is 25 ℃, the environment condition for calculating the device loss Eoff is that the driving resistor is 1 ohm, the device is turned off under the condition that the emission collector current Ic is 3600A, the environment temperature is 25 ℃, and on the premise that the three conditions are met, 5 process parameter value ranges are obtained, as shown in table 1;
TABLE 1 IGBT Process parameter description and ranges thereof
| Name of Chinese | Corresponding symbol | Value range | Unit of | |
| Peak concentration of collector | Pcollector | 1e17～2e18 | /cm 3 | |
| Junction depth of collector | Hemit | 0.6-2 | um | |
| Concentration at collector junction | Pjuntion | 1e14～1e15 | /cm 3 | |
| Concentration of field stop layer | Nbuff | 8e15～1.8e16 | /cm3 | |
| Junction depth of field | Hcut | 6～30 | um |
B2) By taking at equal intervalsThe value-adding interpolation method is used for valuing a process parameter, firstly, equal-interval value taking is carried out based on a value range, a characteristic curve is output by the parameter under k values in total, the saturation on-state voltage drop and the turn-off loss of a device under each value are extracted through the characteristic curve, wherein interpolation delta V exists under two adjacent parameter values cesat And interpolated Δ E off Calculating the sensitivity k i ＝[(ΔVcesat/Vcesat) 2 +(ΔEoff/Eoff) 2 ] 1/2 /(Δm i /m i ) Where i denotes the i-th set of parameters, m i Indicating a specific value, Δ m, of any one of the process parameters i ＝m i -m i-1 Representing the difference between two most adjacent values of the parameter, and calculating k average values of sensitivity k' when the sensitivity ki>k ', an average value point m ' is inserted into the two parameters ' i ＝(m i +m i+1 )/2；
C) The method comprises the following steps of performing data expansion, standardization and normalization on an original data set to obtain a training data set for convolutional neural network identification, wherein the training data set comprises the following specific steps:
C1) adopting a parity interval sampling means, according to the data parity of the original data set, after equal interval sampling is carried out on an even sequence of the original data set, carrying out equal interval resampling on adjacent odd bits to obtain output data with the same data length, and realizing data expansion of limited number of samples;
C2) Arranging all data in an original data set after data expansion into a data chain with the length of 512 bits, wherein the data chain comprises three groups of two-dimensional column vectors of [ Vce, Ic ], [ Time, Ict ], [ Time, Vce ];
C3) adopting a normalization strategy for data in a data chain, respectively finding the maximum values of the three groups of vectors in 5000 groups of data chains, respectively recording the maximum values as Icmax, Icmax and Vcemax, obtaining normalized data with the ranges of Ic/Icmax, Ict as Icmax and Vce as Vcemax, and then sequentially disordering the data chains and completing the head and tail splicing of arrays in the data chains to finally obtain a complete input data chain;
taking a curve of transient turn-off collector-emitter voltage Vce changing with time in an original data set as an example, adopting an equal-interval sampling form, extracting 512 data points in the original data set, wherein ordinal numbers of all sampling points in the original data are even points, and resampling at the odd points on the right side of each even point to realize expansion and doubling of the original data, as shown in FIG. 4;
and then arranging the data expanded and doubled according to a standard data chain format, wherein each group of data chain comprises three groups of two-dimensional column vectors of [ Vce, Ic ], [ Time, Ict ], [ Time, Vce ], the length of each column vector is 512, the number of the data chains is 5000, then, carrying out normalization processing on the normalized data, and respectively finding the maximum values of the data of the three groups of vectors in 5000 groups of data chains, and respectively marking the maximum values as Icmax, Ictmax and Vcemax. Obtaining normalized data with a range of (0, 1) by using Ic (Ic/Icmax), Ict (Ictmax) and Vce (Vcemax), performing disorder processing on an original data chain, completing head-to-tail splicing, and finally obtaining a complete input data chain, meanwhile, performing head-to-tail splicing on a process parameter value netlist according to a disorder sequence to obtain an output data chain, wherein the whole process is as shown in fig. 5, and the input and output data chain finally forms a training data set;
D) The convolutional neural network utilizes the training data set to learn the relation between characteristic curve and the process parameter, generate IGBT numerical model parameter and draw convolutional neural network, convolutional neural network adopts keras network architecture, there is the four-layer structure, including first convolutional layer and pooling layer, second convolutional layer and pooling layer, flattening layer and full tie layer, convolutional neural network's input layer is the data link that includes a plurality of groups IGBT characteristic curve, the output layer is the value of process parameter, convolutional neural network's training process does: the input data chain is brought into a first convolution layer, the length of the data chain is 5000 at this time, the data chain is not subjected to convolution expansion, the depth information is 1, the value is (5000, 1), the convolution operation is carried out by a filter consisting of 64 neurons, an ELU function is adopted for activation, and the ELU function can be expressed as: (x) { x (x > 0); (e) x -1) (x ≦ 0), resulting in a data structure of (2500, 64), information dimension extension, then connection to the first pooling layer, pooling scaling factor of 5, further reduction of data length, resulting in a data structure of (500,64) then, entering a second convolution layer, changing data into (250, 128), entering a second pooling layer, changing a data structure into (25, 256), flattening all dimensions by using the data subjected to flattening processing at the moment to obtain a data structure (25 x 256, 1), then passing through a full connection layer to finally obtain a data structure of (5, 1), and obtaining a one-dimensional parameter containing 5 output information at the moment;
Under the single learning process of the convolutional neural network, the output data chain under the convolutional neural network is obtained by using the input data chain belonging to the convolutional neural network in fig. 5, because the data chain comprises five parameters, and the length of the output data chain in the training data set is 5000, i is the serial number of the parameter, k is the serial number of the group of parameters in the data chain, and each element of the output data chain obtained by using the convolutional neural network is m i,k The output data chain formed by the process parameter value netlist is counted as M i,k The calculated beta is an error cost function, beta k ＝0.2∑[(mi，k-Mi,k)/Mi,k] 2 i＝1,2,3,4,5，β＝∑β k If the error function is not improved after the iterative adjustment for 20 times, the convolution operation is stopped, and the training is considered to meet the requirement, fig. 6 and 7 show the result after the training network processes 1000 data, and the comparison between the process parameter values obtained by the neural network model training and the process parameter values given by the simulation training shows that the neural network can better simulate the corresponding relation, and under the condition of limited samples, the learning effect is better, and the result error is smaller;
E) the waveform extracted by an actual circuit is used as an input value, the convolution neural network is extracted by using IGBT numerical model parameters for calculation and solving to obtain process parameters, and the actual circuit comprises a voltage source V DC Capacitor C at power supply end DC A line junction stray inductor Ls, a discharge IGBT T1 and an IGBT T2 to be tested, wherein a voltage source VDC is connected to a collector of the discharge IGBT T1 through the line junction stray inductor Ls, is connected to a collector of the IGBT T2 to be tested through an emitter of the discharge IGBT T1, and is connected with a voltage source V through an emitter of the IGBT T2 to be tested DC Negative electrode of, to be measuredV in IGBT T2 circuit GE1 The method comprises the steps that for grid driving voltage, the turn-on and turn-off time of an IGBT in a circuit can be controlled, so that a device can be turned off under specific current, the turn-off transient collector and emitter voltage Vce of the device is obtained by means of circuit port testing, the collector current Ic changes along with the time, the collector current Ic changes along with the collector and emitter voltage Vce, an input data chain is obtained by adopting a preprocessing means of step C), and the input data chain is brought into a trained IGBT numerical model parameter extraction neural network to calculate technological parameters to be solved.
In order to further verify the accuracy of the process parameters, the experimental waveforms and the simulated waveforms of the model under the process parameters are compared under multiple working conditions. The final effect is shown in the following example, and the content in table 2 shows the comparison of the results of the conventional searching method and the neural network parameter extraction method,
TABLE 2 comparison of output results of the conventional method and the neural network method
Fig. 8 to fig. 10 are final result comparison graphs, where fig. 8 is a comparison of static characteristic results of a data manual and a device simulation curve, and fig. 9 and fig. 10 are comparison of dynamic turn-off characteristics under different working conditions, and as a result, it is found that the consistency is good, the accuracy requirement can be met, and the correctness of the present invention is verified.
In the invention, the method for extracting the process parameters can be improved by utilizing the neural network method, a new method is provided for fitting part of process parameters which are difficult to determine, the method can also be used for simplifying part of complex process procedures in the actual process, a more accurate model is obtained, and the accuracy and the efficiency of model extraction are improved. In addition to the above examples, the present invention can also be used for adjusting and debugging partial process parameters in the production process of semiconductor devices such as IGBTs, etc., and can construct a process netlist near the process parameters to be adjusted, and train a neural network by using the process netlist to output a simulation result. And finally, searching for the optimal input process parameters by using the target output curve.
Claims (7)
1. A method for extracting parameters of an IGBT numerical model based on a convolutional neural network is characterized by comprising the following steps: the method comprises the following steps:
A) Constructing a numerical model based on TCAD numerical simulation software, wherein the numerical model comprises cell size structure parameters of the IGBT and technological parameters to be solved;
B) screening and assigning the process parameters by using TCAD numerical simulation software to generate a process parameter value netlist, obtaining a characteristic curve output under each group of processes by using the TCAD numerical simulation software, forming an original data set by the characteristic curve and the process parameter value netlist, and screening and assigning the process parameters, wherein the method comprises the following steps of:
B1) determining the value range of the process parameters by a simulation method, screening and assigning the process parameters by using TCAD numerical simulation software, solving the conduction voltage drop Vces, the breakdown voltage Vcesat and the device loss Eoff of the device, and meeting the following conditions:
Vces>1800V
Vcesat＝2±0.15V
Eoff＝1500±200mJ
the environment condition for calculating the breakdown voltage Vcesat is that the gate voltage Vg is 15V, the collector current is 3600A, the environment temperature is 25 ℃, the environment condition for calculating the loss Eoff of the device is that the driving resistor is 1 ohm, the device is turned off under the condition that the emission collector current Ic is 3600A, the environment temperature is 25 ℃, and the value range of the process parameter is obtained on the premise that the three conditions are met;
B2) The process parameters are valued by adopting a method of equal interval value adding interpolation, firstly, equal interval value is carried out based on a value range, a characteristic curve is output by the parameters under k values in total, the saturation conduction voltage drop and the turn-off loss of the device under each value are extracted through the characteristic curve, wherein under two adjacent parameter values, the interpolation existsΔV cesat And interpolated Δ E off Calculating the sensitivity k i ＝[(ΔVcesat/Vcesat) 2 +(ΔEoff/Eoff) 2 ] 1/2 /(Δm i /m i ) Where i denotes the i-th set of parameters, m i Indicating a specific value, Δ m, of any one of the process parameters i ＝m i -m i-1 Representing the difference between two most adjacent values of the parameter, and calculating k average values of sensitivity k' when the sensitivity ki>k ', an average value point m ' is inserted into the two parameters ' i ＝(m i +m i+1 )/2；
C) Expanding, standardizing and normalizing the original data set to obtain a training data set for the convolutional neural network to identify;
D) the convolutional neural network learns the relation between the characteristic curve and the process parameters by utilizing a training data set, and generates IGBT numerical model parameters to extract the convolutional neural network;
E) and (3) taking the waveform extracted by the actual circuit as an input value, extracting the convolutional neural network by using IGBT (insulated gate bipolar transistor) numerical model parameters, and calculating and solving to obtain process parameters.
2. The convolutional neural network-based IGBT numerical model parameter extraction method of claim 1, characterized in that: in the step A), the cell size structure parameters are observed and extracted from a chip section sample by using an observation tool, and the process parameters comprise collector peak concentration Pcollector, collector junction depth Hemit, collector junction concentration Pjunction, field cut-off layer concentration Nbuff and field cut-off layer junction depth Hcut.
3. The convolutional neural network-based IGBT numerical model parameter extraction method as claimed in claim 1, characterized in that: in step B), the characteristic curves include a relation curve between the collector current Ic and the collector-emitter voltage Vce of the device, a Time variation curve of the transient off collector-emitter voltage Vce, and a Time variation curve of the collector current Ict at the Time of transient off.
4. The convolutional neural network-based IGBT numerical model parameter extraction method as claimed in claim 3, characterized in that: in the step C), the step of performing data expansion, normalization and normalization on the original data set comprises the following steps:
C1) adopting a means of odd-even interval sampling, according to the data parity of the original data set, after equal interval sampling is carried out on an even sequence of the original data set, carrying out equal interval resampling on adjacent odd bits of the original data set, obtaining output data with the same data length, and realizing data expansion of limited number of samples;
C2) Arranging all data in the original data set after data expansion into a data chain with the length of 512 bits, wherein the data chain comprises three groups of two-dimensional column vectors of [ Vce, Ic ], [ Time, Ict ], [ Time, Vce ];
C3) and (2) adopting a normalization strategy for data in the data chain, respectively finding the maximum values of the three groups of vectors in 5000 groups of data chains, respectively recording the maximum values as Icmax, Icmax and Vcemax, obtaining normalized data with the ranges of (0 and 1) by using Ic/Icmax, Ict as Icmax and Vce as Vcemax, and then sequentially disordering the data chains, completing the head and tail splicing of the arrays in the data chains, and finally obtaining a complete input data chain.
5. The convolutional neural network-based IGBT numerical model parameter extraction method of claim 1, characterized in that: 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, a flattening layer and a full-link layer, wherein an input layer of the convolutional neural network is a data chain including a plurality of groups of IGBT characteristic curves, and an output layer is a value of a process parameter.
6. The convolutional neural network-based IGBT numerical model parameter extraction method as claimed in claim 5, characterized in that: in the step D), the training process of the convolutional neural network is: 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 subjected to convolution expansion and is deep Degree information is 1, and is counted as (5000, 1), convolution operation is carried out by a filter consisting of 64 neurons, activation is carried out by adopting an ELU function, and the ELU function can be expressed as: (x) { x (x > 0); (e) x -1) (x is less than or equal to 0), obtaining a data structure (2500, 64), expanding information dimensions, then connecting a first pooling layer, wherein a pooling scaling factor is 5, further reducing the length of the data to obtain the data structure (500, 64), then entering a second convolution layer, changing the data into (250, 128), entering a second pooling layer, changing the data structure into (25, 256), flattening all dimensions through flattening processing of the data at the moment to obtain the data structure (25 x 256, 1), then passing through a full connection layer, finally obtaining the data structure (5, 1), and then obtaining a one-dimensional parameter containing 5 output information.
7. The convolutional neural network-based IGBT numerical model parameter extraction method as claimed in claim 1, characterized in that: in said step E), said actual circuit comprises a voltage source V DC Capacitor C of power supply terminal DC A line junction stray inductor Ls, a discharge IGBT T1 and an IGBT T2 to be tested, wherein a voltage source VDC is connected to a collector of the discharge IGBT T1 through the line junction stray inductor Ls, is connected to a collector of the IGBT T2 to be tested by an emitter of the discharge IGBT T1, and is connected to the voltage source V by an emitter of the IGBT T2 to be tested DC Negative pole of (2), V in IGBT T2 circuit to be tested GE1 And C) preprocessing means in the step C) is adopted to obtain a reorganization curve to obtain an input data chain, the reorganization curve is substituted into the trained IGBT numerical model parameter extraction neural network, and the process parameters to be solved are calculated.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210250735.4A CN114330197B (en) | 2022-03-15 | 2022-03-15 | Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202210250735.4A CN114330197B (en) | 2022-03-15 | 2022-03-15 | Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN114330197A CN114330197A (en) | 2022-04-12 |
| CN114330197B true CN114330197B (en) | 2022-07-29 |
Family
ID=81033596
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202210250735.4A Active CN114330197B (en) | 2022-03-15 | 2022-03-15 | Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN114330197B (en) |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN105763121A (en) * | 2016-03-03 | 2016-07-13 | 湖南大学 | Synchronous electric main shaft acceleration strong magnetic control method for variable-load superhigh-speed grinding |
| CN106202590A (en) * | 2015-04-29 | 2016-12-07 | 国网智能电网研究院 | IGBT module switching transients model parameter acquisition methods and method for establishing model |
| CN110502805A (en) * | 2019-07-31 | 2019-11-26 | 中国人民解放军海军工程大学 | IGBT physical model statistic property extracting method |
Family Cites Families (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20200210824A1 (en) * | 2018-12-28 | 2020-07-02 | Utopus Insights, Inc. | Scalable system and method for forecasting wind turbine failure with varying lead time windows |
| CN113935413A (en) * | 2021-10-14 | 2022-01-14 | 国网浙江省电力有限公司金华供电公司 | Distribution network wave recording file waveform identification method based on convolutional neural network |
-
2022
- 2022-03-15 CN CN202210250735.4A patent/CN114330197B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106202590A (en) * | 2015-04-29 | 2016-12-07 | 国网智能电网研究院 | IGBT module switching transients model parameter acquisition methods and method for establishing model |
| CN105763121A (en) * | 2016-03-03 | 2016-07-13 | 湖南大学 | Synchronous electric main shaft acceleration strong magnetic control method for variable-load superhigh-speed grinding |
| CN110502805A (en) * | 2019-07-31 | 2019-11-26 | 中国人民解放军海军工程大学 | IGBT physical model statistic property extracting method |
Non-Patent Citations (2)
| Title |
|---|
| An Improved Physical Model of Power Diode in the Rectifier Circuit;Zenan Shi 等;《 IEEE Journal of Emerging and Selected Topics in Power Electronics》;20210625;第1203-1218页 * |
| IGBT 机理建模及其基于神经网络的参数辨识方法;孙跃 等;《西南交通大学学报》;20151215;第1143-1149+1163页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN114330197A (en) | 2022-04-12 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN108573225B (en) | Partial discharge signal pattern recognition method and system | |
| CN111368454B (en) | SiC MOSFET SPICE model establishment method based on bare chip packaging structure | |
| CN110502805A (en) | IGBT physical model statistic property extracting method | |
| CN110456217B (en) | MMC fault positioning method based on WPD-FOA-LSSVM dual model | |
| CN110108992A (en) | Based on cable partial discharge fault recognition method, system and the medium for improving random forests algorithm | |
| CN103105571A (en) | Simulated measurement method of current characteristics of insulated gate bipolar transistor | |
| CN110726898A (en) | Power distribution network fault type identification method | |
| CN113536661A (en) | TFET device structure optimization and performance prediction method based on neural network | |
| CN113850154A (en) | Inverter IGBT (insulated Gate Bipolar transistor) micro fault feature extraction method based on multi-modal data | |
| CN114330197B (en) | Method for extracting parameters of IGBT (insulated Gate Bipolar transistor) numerical model based on convolutional neural network | |
| CN112685958B (en) | SiC MOSFET blocking voltage determination method based on neural network | |
| CN114564906A (en) | SiC MOSFET simulation modeling method and system | |
| CN107622167B (en) | Collector current soft measurement method for grid control device | |
| CN100561488C (en) | The modeling method of metal-oxide-semiconductor resistance | |
| CN113189513A (en) | Ripple-based redundant power supply current sharing state identification method | |
| CN112036010A (en) | Photovoltaic system dynamic process hybrid equivalent modeling method based on data driving | |
| CN114066005B (en) | CNN network-based silicon carbide diode breakdown voltage prediction method | |
| CN116152629A (en) | Macadimia nut detection and identification method based on Faster R-CNN | |
| CN113497588B (en) | Method and device for testing electrical performance of solar cell and solar cell module | |
| CN112101393A (en) | Wind power plant fan clustering method and device | |
| CN106781502B (en) | A kind of fuel-engined vehicle road conditions recognition methods based on vector quantization training pattern | |
| CN115238620A (en) | Composite algorithm aided simulation IC design method combining response surface with support vector regression | |
| CN114035120A (en) | Three-level inverter open-circuit fault diagnosis method and system based on improved CNN | |
| CN113537327A (en) | Non-invasive load identification method and system based on Alexnet neural network and color coding | |
| CN112213562A (en) | Method for measuring and calculating internal resistance of grid electrode of power semiconductor device |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |