CN110780188A - IGBT drive circuit fault diagnosis method - Google Patents
IGBT drive circuit fault diagnosis method Download PDFInfo
- Publication number
- CN110780188A CN110780188A CN201910993456.5A CN201910993456A CN110780188A CN 110780188 A CN110780188 A CN 110780188A CN 201910993456 A CN201910993456 A CN 201910993456A CN 110780188 A CN110780188 A CN 110780188A
- Authority
- CN
- China
- Prior art keywords
- layer
- fault diagnosis
- signal
- neural network
- driving circuit
- 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.)
- Pending
Links
Images
Classifications
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/28—Testing of electronic circuits, e.g. by signal tracer
- G01R31/316—Testing of analog circuits
-
- G—PHYSICS
- G01—MEASURING; TESTING
- G01R—MEASURING ELECTRIC VARIABLES; MEASURING MAGNETIC VARIABLES
- G01R31/00—Arrangements for testing electric properties; Arrangements for locating electric faults; Arrangements for electrical testing characterised by what is being tested not provided for elsewhere
- G01R31/28—Testing of electronic circuits, e.g. by signal tracer
- G01R31/2832—Specific tests of electronic circuits not provided for elsewhere
- G01R31/2836—Fault-finding or characterising
- G01R31/2846—Fault-finding or characterising using hard- or software simulation or using knowledge-based systems, e.g. expert systems, artificial intelligence or interactive algorithms
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- General Physics & Mathematics (AREA)
- General Engineering & Computer Science (AREA)
- Evolutionary Computation (AREA)
- Health & Medical Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Biophysics (AREA)
- Biomedical Technology (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Life Sciences & Earth Sciences (AREA)
- Medical Informatics (AREA)
- Tests Of Electronic Circuits (AREA)
Abstract
A fault diagnosis method for an IGBT driving circuit comprises the following steps: (1) performing Monte Carlo analysis on the IGBT driving circuit by using PSpice software, and acquiring a time domain response signal of the circuit, namely acquiring an output voltage signal of the driving circuit; (2) performing wavelet transformation on the acquired voltage signal, calculating the energy of a wavelet coefficient as a characteristic parameter, wherein the set of all the characteristic parameters is sample data; (3) establishing a convolutional neural network model based on sample data and training the convolutional neural network model; (4) and taking the established convolutional neural network fault diagnosis model as a classifier to diagnose the faults of the drive circuit of the IGBT. The method for extracting the fault signal features is superior to other methods, and the fault mode classifier has higher recognition rate and adaptivity.
Description
Technical Field
The invention relates to the field of machine learning and electronic circuit engineering, in particular to a fault diagnosis method for an IGBT (insulated gate bipolar transistor) driving circuit.
Background
In recent years, since the IGBT integrates the advantages of the power MOSFET and the bipolar transistor, the IGBT has not only high-speed switching and voltage driving characteristics, but also low saturation voltage characteristics and the ability to easily realize a larger current, so that the IGBT becomes a device of great interest in the field of power electronics and has been rapidly developed. With the wide use of the IGBT, various faults occur in the drive circuit of the IGBT, the performance of equipment is seriously influenced, and even the equipment cannot normally work, so in order to ensure the working efficiency, the research on the fault diagnosis of the IGBT drive circuit needs to be strengthened, and the IGBT drive circuit has important engineering and economic significance for developing power electronic equipment.
At present, few researches are made on a fault diagnosis method of an IGBT driving circuit, and the IGBT driving circuit belongs to one of analog circuits, so that the existing fault diagnosis method of the analog circuit can be used for reference. For the fault diagnosis of an analog circuit, researchers have adopted an analog circuit fault diagnosis method for extracting fault characteristics by wavelet decomposition and fuzzy clustering, however, the fault diagnosis method based on fuzzy clustering needs to adopt expert experience knowledge for determining fuzzy membership parameters and has no learning capability. The prediction method based on the LS-SVM can track the variation trend of the performance parameters of the fault characteristics to realize fault prediction, but is only limited to fault diagnosis of a direct current circuit and has limitation. According to the analog circuit fault diagnosis method based on the support vector machine, the fault identification capability and the system fault diagnosis speed are greatly improved, but the diagnosis effect depends on the selection of the kernel function and the kernel function parameters. Therefore, in order to realize the fault diagnosis of the IGBT driving circuit, it is necessary to select an appropriate feature extraction method and an efficient classification method.
Disclosure of Invention
The technical problem to be solved by the invention is to overcome the defects in the prior art and provide a fault diagnosis method for an IGBT driving circuit, so that the learning capability of a fault mode classifier is improved, and the efficiency and the precision of fault mode diagnosis are improved.
The method comprises the steps of firstly, extracting fault characteristic data by using a decomposition and reconstruction algorithm in wavelet transform, generating a sample set, then establishing a convolutional neural network classification model, dividing the sample set obtained in the previous step into a training part and a testing part, and utilizing training sample data to carry out convolutional neural network training and learning so as to realize the construction of the IGBT drive circuit fault mode classifier.
The purpose of the invention is realized by the following scheme:
a fault diagnosis method for an IGBT driving circuit comprises the following steps:
(1) performing Monte Carlo analysis on the IGBT driving circuit by using PSpice software, and acquiring a time domain response signal of the circuit, namely acquiring an output voltage signal of the driving circuit; the tolerance ranges of fault elements, namely resistance and capacitance, in the circuit are +/-5 percent;
(2) performing wavelet transformation on the acquired voltage signal, calculating the energy of a wavelet coefficient as a characteristic parameter, wherein the set of all the characteristic parameters is sample data;
(3) establishing a convolutional neural network fault diagnosis model based on sample data, and performing training and diagnosis result verification on the convolutional neural network fault diagnosis model;
(4) and taking the established convolutional neural network fault diagnosis model as a classifier to diagnose the faults of the IGBT driving circuit.
Furthermore, in the step (1), the IGBT driving circuit to be tested only has one input end and one output end, the input end is excited by adopting rectangular pulses, and the output end samples voltage signals.
Further, in the step (2), the acquired voltage signal is subjected to wavelet transform, and the steps are as follows:
and (2-a) performing wavelet decomposition on the voltage signal, and obtaining a high-frequency coefficient and a low-frequency coefficient from a component signal obtained by decomposition, wherein the high-frequency coefficient corresponds to a detail signal, and the low-frequency coefficient corresponds to an approximation signal. Noise is detected from the high frequency coefficients, and the low frequency coefficients identify the different frequencies of the component signals.
And (2-b) reconstructing a target signal by using the denoised high-frequency coefficient and the approximated low-frequency coefficient.
The wavelet transform is a time-scale (time-frequency) Analysis method of signals, has the characteristic of Multi-resolution Analysis (Multi-resolution Analysis), has the capability of representing the local characteristics of the signals in two time-frequency domains, and is a time-frequency local Analysis method with a fixed window size, a changeable shape and changeable time window and frequency window. The wavelet transform has a higher frequency resolution and a lower time resolution in the low frequency portion and a higher time resolution and a lower frequency resolution in the high frequency portion, and is suitable for extracting feature vectors of signals.
The basic idea of wavelet transform is:
search for small waves or "wavelets" and use their integer translations and dyadic scaling to generate the entire L
2(R) space, assuming f ∈ L
2(R) is the actual signal to be processed, the measured signal f
jIs f in the scale space V
jOf (1) or (ii). V
jIs V
j+1So that there is V
jAt V
j+1Of orthogonal complement W
j,j→∞:
The above-described spatial orthogonal decomposition process can be passed on, thus for any V
jCan be represented as
Thus, f can be adjusted
jThe multi-resolution is expressed as
f
j=f
j-1+w
j-1
=f
j-2+w
j-2+w
j-1
=…
=f
j0+w
j0+w
j0+1+…+w
j-1;
In the same way, w
jIs f
jAt f
j+1Of orthogonal complement
Wherein the content of the first and second substances,
wherein k is equal to Z, c
l,kAnd phi
l,kAre respectively f
lScale coefficients and wavelet coefficients of (d), likewise
l,kIs w
lScale factor of, #
l,kIs w
lThe wavelet coefficients of (a).
And (3) decomposition: f. of
j→f
j0,w
j0…,w
j-1;
And (3) reconstruction: f. of
j0,w
j0…,w
j-1→f
j。
In the step (2-a), the wavelet decomposition comprises the following specific steps:
(2-a.1) initialization: firstly, according to the actual signal f, determining the approximate scale space V
j. Then select f
j∈V
jSo that f
jIs f to V
jIs best approximated, i.e.
f
j=P
jf denotes f
jIs f to V
jThe best approximation of; phi is a
j,
kIs f
jThe wavelet coefficients of (a); scale factor c
j,
k=<f,φ
j,k>;
Theorem: suppose { V
jJ ∈ Z } is a multi-resolution analysis generated by the scaling function δ, and δ has tight support. If f ∈ L
2(R) is a continuous function, then when j is sufficiently large,
c
j,k≈mf(k/2
j);
c
j,kis f
jThe scale factor of (c);
wherein the content of the first and second substances,
(2-a.2) iteration: using Mallat decomposition algorithm to decompose f
jDecomposition into f
j=f
j0+w
j0+w
j0+1+…+w
j-1;
(2-a.3) termination: the above-mentioned iterative process needs to be carried out to an approximation level V that meets the preset requirements
j0This level is contingent on the actual situation. The simplest is to decompose all the way to level 0. Although in practice multiple decompositions are generally not required. For example, singularity detection may only require decomposition at levels 1 to 2. The termination point is determined by the particular job to be completed.
In the step (2-b), reconstructing the target signal by the de-noised high-frequency coefficient and the approximated low-frequency coefficient, specifically:
according to the scale coefficient c after drying
j0,k(c
j0,kHave the same meaning as
C when l is j0
j0,k) And a scale factor d after drying
l,k(d
l,kHave the same meaning as
D in (1)
l,k) J0, j0+1, … j-1, c being obtained stepwise using the Mallat reconstruction algorithm
l,kJ0+1, …, j-1, j, and then according to the approximated wavelet coefficient phi
l,kAnd psi
l,kJ0+1, …, j-1, j, using
And
thereby obtaining f
j0,w
j0…,w
j-1Finally, a target reconstruction signal f after drying is obtained
j0+w
j0+w
j0+1+…+w
j-1=f
j。
In the step (2), the energy of the wavelet coefficient is calculated as the characteristic parameter, and the wavelet coefficient comprises f
lIs small wave coefficient phi
l,kAnd w
lWavelet coefficient psi
l,k。
Further, in the step (3), training and checking the diagnosis result of the convolutional neural network fault diagnosis model are performed, and the specific steps are as follows:
(a) dividing the sample data into two parts, wherein one part is a training sample and the other part is a test sample; initializing parameters of a convolutional neural network, inputting the classified training samples, training and learning the convolutional neural network, and initializing weights;
(b) extracting samples, namely extracting 1 sample X from the training samples and giving a target output vector of the sample X;
(c) sequentially calculating from the front layer to the rear layer to obtain the output of the convolutional neural network, then sequentially calculating error terms of each layer in the reverse direction (namely from the rear layer to the front layer) until the error is lower than a set threshold value, and finally obtaining the ideal expected output;
(d) and after a series of training and learning are carried out on the model, diagnosing and monitoring by taking the other group of test sample data as input, and verifying the diagnosis effect of the model by using the finally obtained simulation result.
A convolutional neural network fault diagnosis model is established, in the field of mode classification, the convolutional neural network can avoid complex characteristic quantity preprocessing in the early stage, and an output result can be obtained by directly inputting original data. The method for establishing the convolutional neural network fault diagnosis model in the step (3) comprises the following steps:
(a) the convolutional neural network fault diagnosis model consists of an input layer, a convolutional layer, a sampling layer, a full link layer and an output layer, wherein the convolutional layer is used for carrying out convolution on data input by the input layer and then outputting the data; meanwhile, the sampling layer is used as the input of the sampling layer and is output after being subjected to a sampling method; by analogy, the final input can be obtained.
(b) The convolution process is performed by determining the size of the filter, using a convolution kernel K
abFor a feature x of the filter of the previous layer
aWeighted to obtain x
a*K
abThen to x
a*K
abThe sum is then biased as shown below.
Wherein the content of the first and second substances,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; m
bA filter corresponding to neuron b;
the b-th corresponding weight of the q-th neuron a is obtained; b is
qIs the only offset for layer q.
(c) The sampling layer adopts mean-posing (namely, only averaging the characteristic points in the neighborhood) to average all the values in the filter; then multiplying the sampled information by trainable parameters, adding trainable bias, and obtaining the output of the neuron in the sampling layer by the result obtained by the calculation of an activation function (adopting sigmoid function), wherein the output formula of the neuron in the sampling layer is as follows
Wherein, β
qFor the trainable parameters of the q-th level,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; m
bA filter corresponding to neuron b; b is
qIs the only offset for layer q.
(d) For the full link layer, the calculation is directly carried out by adopting a method of a multilayer artificial neural network, and the output formula of the neuron of the full link layer is as follows:
wherein, w
baThe weight value from the node b of the q-th layer to the node i of the q + 1-th layer is obtained;
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; b is
qIs the only offset for layer q.
Further, in the step (4), when the failure of the IGBT drive circuit is diagnosed, the obtained diagnosis result is a diagnosis accuracy.
The convolutional neural network is generally used for image recognition and classification, and the output of the IGBT driving circuit is a non-linear voltage signal, so that no technician proposes to use the convolutional neural network for fault diagnosis of the IGBT driving circuit. Meanwhile, the IGBT driving circuit belongs to one of analog circuits, and generally, a neural network, an SVM (support vector machine), and other methods are applied to fault diagnosis thereof as a classifier. The convolutional neural network belongs to a deep neural network, and solves the problems that after the network hierarchy becomes deep due to the fact that the deep neural network cannot be calculated, network parameters are increased, so that a serious overfitting phenomenon is caused, and in the training process, when the deep neural network is spread by using a BP algorithm, the gradient is rapidly reduced, so that the previous network cannot be trained, and the network is difficult to converge. The method is different from the method that the characteristic data extracted by wavelet transformation is used as the input of a convolutional neural network when the image is classified and identified from the prior art. Meanwhile, at present, the study on the convolutional neural network itself is not deep enough, and although the convolutional neural network has excellent effect, the convolutional neural network still is a black box for technicians. It would be a considerable task to clarify the construction of this black box and thus to better improve it.
The invention has the following advantages:
(1) the provided IGBT driving circuit fault diagnosis concept. The tolerance ranges of the fault elements in the circuit, namely the resistance and the capacitance, are +/-5%, the tolerance ranges are smaller than the tolerance ranges of the fault elements in fault diagnosis of other analog circuits, and identification and diagnosis are more difficult to perform. By using the invention, the fault mode classifier has higher identification rate and adaptability.
(2) The invention extracts the fault characteristic vector by utilizing wavelet decomposition and reconstruction, and the decomposition and reconstruction algorithm is simple and clear and has high calculation speed.
(3) The method utilizes the convolutional neural network to establish the fault diagnosis model, has good convergence, self-learning capability, parallel capability and fault tolerance capability, and can be more quickly and accurately applied to the aspect of fault identification.
Drawings
FIG. 1 is a flow chart of the IGBT drive circuit fault diagnosis method of the invention;
fig. 2 is an IGBT driving circuit diagram.
Detailed Description
The invention is further illustrated with reference to the following figures and examples.
Referring to fig. 1, the invention comprises 4 steps, and step 1 acquires a time domain response signal of the IGBT driving circuit to be tested. And 2, performing wavelet transformation on the acquired fault response signal (namely the time domain response signal acquired in the step 1), and calculating the energy of a wavelet coefficient as a characteristic parameter, wherein the set of all the characteristic parameters is sample data. In the present embodiment, 3-layer wavelet transform is specifically performed, and 3-dimensional wavelet coefficient energy is obtained as a characteristic parameter. And 3, establishing a convolutional neural network fault diagnosis model, training and diagnosing the convolutional neural network fault diagnosis model, wherein the convolutional neural network fault diagnosis model after training and verifying the diagnosis result is a fault classifier. And 4, taking the established convolutional neural network fault diagnosis model as a classifier, diagnosing the faults of the IGBT driving circuit, and outputting the diagnosis result of the test data by using the convolutional neural network fault diagnosis model.
In the step 1, Monte Carlo analysis is carried out on the IGBT drive circuit to obtain a time domain response signal, the IGBT drive circuit to be tested only has one input end and one output end, the input end is excited by rectangular pulses, and the output end samples voltage signals.
The method comprises the following steps of performing wavelet transformation on the acquired voltage signals:
(a) the method comprises the steps of firstly carrying out three-layer decomposition on a signal, obtaining high-frequency and low-frequency coefficients from a component signal obtained by decomposition, wherein the high-frequency coefficients correspond to detail signals, and the low-frequency coefficients correspond to approximation signals. Noise is detected from the high frequency coefficients, and the low frequency coefficients identify the different frequencies of the component signals.
(b) And reconstructing the target signal by using the denoised high-frequency coefficient and the approximated low-frequency coefficient.
Search for small waves or "wavelets" and use their integer translations and dyadic scaling to generate the entire L
2(R) space, assuming f ∈ L
2(R) is the actual signal to be processed, the measured signal f
jIs f in the scale space V
jOf (1) or (ii). V
jIs V
j+1So that there is V
jAt V
j+1Of orthogonal complement W
j,j→∞:
The above-described spatial orthogonal decomposition process can be passed on, thus for any V
jCan be represented as
Thus, f can be adjusted
jThe multi-resolution is expressed as
f
j=f
j-1+w
j-1
=f
j-2+w
j-2+w
j-1
=…
=f
j0+w
j0+w
j0+1+…+w
j-
1;
In the same way, w
jIs f
jAt f
j+1Of orthogonal complement
Wherein the content of the first and second substances,
wherein k is equal to Z, c
l,kAnd phi
l,kAre respectively f
lScale coefficients and wavelet coefficients, likewise, d
l,kIs w
lScale factor of, #
l,kIs w
lWavelet coefficient of
And (3) decomposition: f. of
j→f
j0,w
j0…,w
j-
1;
And (3) reconstruction: f. of
j0,w
j0…,w
j-1→f
j。
The wavelet decomposition comprises the following specific steps:
initialization: firstly, according to the actual signal f, determining the approximate scale space V
j. Then select f
j∈V
jSo that f
jIs f to V
jIs best approximated, i.e.
f
j=P
jf denotes f
jIs f to V
jThe best approximation of; phi is a
j,
kIs f
jThe wavelet coefficients of (a); scale factor c
j,
k=<f,φ
j,k>;
Theorem: suppose { V
jJ ∈ Z } is a multi-resolution analysis generated by the scaling function δ, and δ has tight support. If f ∈ L
2(R) is a continuous function, then when j is sufficiently large,
c
j,k≈mf(k/2
j);
c
j,kis f
jThe scale factor of (c);
wherein the content of the first and second substances,
iteration: using Mallat decomposition algorithm to decompose f
jIs decomposed into
f
j=f
j0+w
j0+w
j0+1+…+w
j-1;
And (4) terminating: the above-mentioned iterative process needs to be carried out to an approximation level V that meets the preset requirements
j0This level is contingent on the actual situation.
Reconstructing a target signal by using the denoised high-frequency coefficient and the approximated low-frequency coefficient, which specifically comprises the following steps: according to the scale coefficient c after drying
j0,k(c
j0,kHave the same meaning as
C when l is j0
j0,k) Andscale factor d after drying
l,k(d
l,kHave the same meaning as
D in (1)
l,k) J0, j0+1, … j-1, c being obtained stepwise using the Mallat reconstruction algorithm
l,kJ0+1, …, j-1, j, and then according to the approximated wavelet coefficient phi
l,kAnd psi
l,kJ0+1, …, j-1, j, using
And
thereby obtaining f
j0,w
j0…,w
j-1Finally, a target reconstruction signal f after drying is obtained
j0+w
j0+w
j0+1+…+w
j-1=f
j。
In step 3, training and checking the diagnosis result of the convolutional neural network fault diagnosis model, specifically comprising the following steps:
(a) dividing the sample data into two parts, wherein one part is a training sample and the other part is a test sample; initializing parameters of a convolutional neural network, inputting the classified training samples, and performing convolutional neural network training learning to initialize weights, namely initializing all weights into 1 smaller random number;
(b) extracting samples, namely extracting 1 sample X from the training set and giving a target output vector of the sample X;
(c) sequentially calculating from the front layer to the rear layer to obtain the output of the convolutional neural network, then sequentially calculating error terms of each layer in the reverse direction (namely from the rear layer to the front layer) until the error is lower than a set threshold value, and finally obtaining the ideal expected output;
(d) and after a series of training and learning are carried out on the model, diagnosing and monitoring by taking the other group of test sample data as input, and verifying the diagnosis effect of the model by using the finally obtained simulation result.
The method for establishing the convolutional neural network in the step (3) comprises the following steps:
(a) the convolutional neural network consists of an input layer, a convolutional layer, a sampling layer, a full link layer and an output layer, wherein the convolutional layer is used for carrying out convolution on data input by the input layer and then outputting the data; meanwhile, the sampling layer is used as the input of the sampling layer and is output after being subjected to a sampling method; by analogy, the final input can be obtained.
(b) The convolution process is performed by determining the size of the filter, using a convolution kernel K
abFor a feature x of the filter of the previous layer
aWeighted to obtain x
a*K
abThen to x
a*K
abThe sum is then biased as shown below.
Wherein the content of the first and second substances,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; m
bA filter corresponding to neuron b;
the b-th corresponding weight of the q-th neuron a is obtained; b is
qIs the only offset for layer q.
(c) The sampling layer adopts mean-posing (namely, only averaging the characteristic points in the neighborhood) to average all the values in the filter; then multiplying the sampled information by trainable parameters, adding trainable bias, and obtaining the output of the neuron in the sampling layer by the result obtained by the calculation of an activation function (adopting sigmoid function), wherein the output formula of the neuron in the sampling layer is as follows
Wherein, β
qFor the trainable parameters of the q-th level,
filters for the upper layer, i.e. layer q-1One of the features of; (x) is sigmoid function; m
bA filter corresponding to neuron b; b is
qIs the only offset for layer q. .
(d) For the full link layer, the calculation is directly carried out by adopting a method of a multilayer artificial neural network, and the output formula of the neuron of the full link layer is as follows:
wherein, w
baIs the weight of the node b of the q-th layer to the node i of the q + 1-th layer,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; b is
qIs the only offset for layer q.
In the step (4), when the fault of the IGBT drive circuit is diagnosed in the step (4), the obtained diagnosis result is a diagnosis accuracy.
In order to demonstrate the procedures and performance of the fault diagnosis method for the IGBT driving circuit according to the present invention, an example is described herein.
Fig. 2 shows an IGBT driver circuit. Taking this circuit as an example, the whole flow of the IGBT driving circuit fault diagnosis method proposed by the present invention is shown, where the excitation source is a rectangular pulse signal with a V1-6V, a pulse width PW-40 μ s, and a period PER-50 μ s, and the fault time domain response signal is obtained by subtracting the voltages at two ends of the IGBT. The tolerance ranges of the resistor and the capacitor are set to be +/-5% and +/-5%, respectively. Selecting normal, C1 ↓, C1 ↓, C2 ↓, C2 ↓, R1 ↓, R1 ↓, R2 ↓, R2 ↓, R3 ↓, R3 ↓, R4 ↓, R4 ↓, R5 ↓, R5 ↓, R6 ↓, R6 ↓, R7 ↓, R7 ↓, R8 ↓, R8 ↓21common fault categories, wherein ↓and ↓indicatethat the fault value is higher and lower than the nominal value respectively, and normal indicates normal. Table 1 gives the fault code, fault class, nominal value and fault value of the circuit element. And respectively acquiring 100-dimensional feature vectors for each fault category, dividing the data according to the proportion of 50/100 and 50/100, respectively training a convolutional neural network model, and testing the accuracy of the fault diagnosis model.
Table 1 fault code, fault category, nominal value and fault value.
Fault code | Class of failure | Nominal value | Fault value |
F0 | Normal | - | - |
F1 | C1↑ | 100nF | 105nF |
F2 | C1↓ | 100nF | 95nF |
F3 | C2↑ | 100nF | 105nF |
F4 | C2↓ | 100nF | 95nF |
F5 | R1↑ | 470Ω | 493.5Ω |
F6 | R1↓ | 470Ω | 446.5Ω |
F7 | R2↑ | 470Ω | 493.5Ω |
F8 | R2↓ | 470Ω | 446.5Ω |
F9 | R3↑ | 1kΩ | 1050Ω |
F10 | R3↓ | 1kΩ | 950Ω |
F11 | R4↑ | 330Ω | 346.5Ω |
F12 | R4↓ | 330Ω | 313.5Ω |
F13 | R5↑ | 10Ω | 10.5Ω |
F14 | R5↓ | 10Ω | 9.5Ω |
F15 | R6↑ | 0.1Ω | 0.105Ω |
F16 | R6↓ | 0.1Ω | 0.095Ω |
F17 | R7↑ | 0.1Ω | 0.105Ω |
F18 | R7↓ | 0.1Ω | 0.095Ω |
F19 | R8↑ | 8Ω | 8.4Ω |
F20 | R8↓ | 8Ω | 7.6Ω |
Modeling, training and learning are carried out by combining training samples obtained by wavelet transformation until the system completely reaches the expected target; the test sample failure data obtained from the experiment herein is then further subjected to simulation analysis, wherein a portion of the test data is shown in table 2. The model accuracy rate is 100% in the states of Normal, C1 ↓, C1 ↓, R1 ↓, R2 ↓, R2 ↓, R3 ↓, R4 ↓, R4 ↓, R5 ↓, R6 ↓, R7 ↓, and R7 ↓; (ii) a Under the state of R6 ↓, the model accuracy rate is 98%; under the states of C2 ↓andR 5 ↓, the model accuracy rate is 97%; in the states of R3 ↓andR 8 ↓, the accuracy rate of the model is 96%; in the states of R1 ↓andR 8 ↓, the model accuracy was 95%. The modeling training learning is tested in combination with the results of table 2, and partial fault diagnosis results are shown in table 2.
TABLE 2 partial diagnosis results for each failure category
Fault code | Class of failure | Accuracy rate |
F0 | Normal | 100% |
F1 | C1↑ | 100% |
F2 | C1↓ | 100% |
F3 | C2↑ | 100% |
F4 | C2↓ | 97% |
F5 | R1↑ | 95% |
F6 | R1↓ | 100% |
F7 | R2↑ | 100% |
F8 | R2↓ | 100% |
F9 | R3↑ | 96% |
F10 | R3↓ | 100% |
F11 | R4↑ | 100% |
F12 | R4↓ | 100% |
F13 | R5↑ | 100% |
F14 | R5↓ | 97% |
F15 | R6↑ | 100% |
F16 | R6↓ | 98% |
F17 | R7↑ | 100% |
F18 | R7↓ | 100% |
F19 | R8↑ | 95% |
F20 | R8↓ | 96% |
Through statistics, the diagnosis accuracy of the test sample reaches 98.5%. The fault diagnosis result of the IGBT driving circuit by using the wavelet transformation and the convolution neural network is basically consistent with the actual condition, and the accurate positioning of the fault position can be realized. The method is used for demonstrating the superiority of the IGBT driving circuit fault diagnosis method combining wavelet transformation and a convolutional neural network. The comparison test of the fault diagnosis rate of the fault diagnosis method of the power electronic circuit based on the LS-SVM and the fault diagnosis method of the analog circuit based on the support vector machine is carried out, and the test results are shown in the table 3. It can be seen that the fault diagnosis rate of the method of the invention is significantly higher than that of the two methods.
TABLE 3 comparative experimental results
By means of | Diagnosis rate/% |
The invention provides a method | 98.5 |
LS-SVM | 97.1 |
Support vector machine | 96.2 |
Claims (6)
1. The IGBT driving circuit fault diagnosis method is characterized by comprising the following steps:
(1) performing Monte Carlo analysis on the IGBT driving circuit by using PSpice software, and acquiring a time domain response signal of the circuit, namely acquiring an output voltage signal of the driving circuit; the tolerance ranges of fault elements, namely resistance and capacitance, in the circuit are +/-5 percent;
(2) performing wavelet transformation on the acquired voltage signal, calculating the energy of a wavelet coefficient as a characteristic parameter, wherein the set of all the characteristic parameters is sample data;
(3) establishing a convolutional neural network fault diagnosis model based on sample data, and performing training and diagnosis result verification on the convolutional neural network fault diagnosis model;
(4) and taking the established convolutional neural network fault diagnosis model as a classifier to diagnose the faults of the IGBT driving circuit.
2. The method for diagnosing the fault of the IGBT driving circuit according to claim 1, wherein in the step (1), the IGBT driving circuit to be tested only has one input end and one output end, the input end is excited by adopting rectangular pulses, and the output end samples voltage signals.
3. The IGBT driving circuit fault diagnosis method according to claim 1 or 2, wherein in the step (2), the acquired voltage signal is subjected to wavelet transform, and the steps are:
performing wavelet decomposition on the voltage signal, and obtaining a high-frequency coefficient and a low-frequency coefficient from a component signal obtained by decomposition, wherein the high-frequency coefficient corresponds to a detail signal, and the low-frequency coefficient corresponds to an approximation signal; detecting noise from the high frequency coefficients, and identifying different frequencies of each component signal from the low frequency coefficients;
and (2-b) reconstructing a target signal by using the denoised high-frequency coefficient and the approximated low-frequency coefficient.
4. The method for diagnosing the faults of the IGBT driving circuit according to claim 3, wherein the wavelet transform is a time-scale or time-frequency analysis method of signals, has the characteristic of multi-resolution analysis, has the capability of representing the local characteristics of the signals in two time-frequency domains, and is a time-frequency localization analysis method with a fixed window size, a changeable shape and changeable time window and frequency window; the wavelet transform has high frequency resolution and low time resolution in the low frequency part and high time resolution and low frequency resolution in the high frequency part, and is suitable for extracting the feature vector of the signal;
search for small waves or "wavelets" and use their integer translations and dyadic scaling to generate the entire L
2(R) space, assuming f ∈ L
2(R) is the actual signal to be processed, the measured signal f
jIs f in the scale space V
j(ii) an approximation of (d); v
jIs V
j+1So that there is V
jAt V
j+1Of orthogonal complement W
j,j→∞:
The above-described spatial orthogonal decomposition process can be passed on, thus for any V
jCan be represented as
Thus, f can be adjusted
jThe multi-resolution is expressed as
f
j=f
j-1+w
j-1
=f
j-2+w
j-2+w
j-1
=…
=f
j0+w
j0+w
j0+1+…+w
j-1;
In the same way, w
jIs f
jAt f
j+1Of orthogonal complement
Wherein the content of the first and second substances,
wherein k is equal to Z, c
l,kAnd phi
l,kAre respectively f
lScale coefficients and wavelet coefficients of (d), likewise
l,kIs w
lScale factor of, #
l,kIs w
lThe wavelet coefficients of (a);
and (3) decomposition: f. of
j→f
j0,w
j0…,w
j-1;
And (3) reconstruction: f. of
j0,w
j0…,w
j-1→f
j。
5. The IGBT drive circuit fault diagnosis method according to claim 4,
the wavelet decomposition comprises the following specific steps:
(2-a.1) initialization: firstly, according to the actual signal f, determining the approximate scale space V
j(ii) a Then select f
j∈V
jSo that f
jIs f to V
jIs best approximated, i.e.
f
j=P
jf denotes f
jIs f to V
jThe best approximation of; phi is a
j,kIs f
jThe wavelet coefficients of (a); coefficient of scale
c
j,k=<f,φ
j,k>;
Theorem: suppose { V
jJ ∈ Z } is a multi-resolution analysis generated by the scale function δ, and δ has tight support; if f ∈ L
2(R) is a continuous function, then when j is sufficiently large,
c
j,k≈mf(k/2
j);
c
j,kis f
jThe scale factor of (c);
wherein the content of the first and second substances,
(2-a.2) iteration: using Mallat decomposition algorithm to decompose f
jIs decomposed into
f
j=f
j0+w
j0+w
j0+1+…+w
j-1;
(2-a.3) termination: the above-mentioned iterative process needs to be carried out to an approximation level V that meets the preset requirements
j0This isA level is contingent on practice;
in the step (2-b), reconstructing the target signal by the de-noised high-frequency coefficient and the approximated low-frequency coefficient, specifically:
according to the de-noised scale coefficient c
j0,kAnd de-noised scale coefficient d
l,kJ0, j0+1, … j-1, c being obtained stepwise using the Mallat reconstruction algorithm
l,kJ0+1, …, j-1, j, and then according to the approximated wavelet coefficient phi
l,kAnd psi
l,kJ0+1, …, j-1, j, using
And
thereby obtaining f
j0,w
j0…,w
j-1Finally, a de-noised target reconstruction signal f is obtained
j0+w
j0+w
j0+1+…+w
j-1=f
j。
6. The IGBT driving circuit fault diagnosis method according to claim 1, wherein the method for establishing the convolutional neural network fault diagnosis model in the step (3) is as follows:
(a) the convolutional neural network fault diagnosis model consists of an input layer, a convolutional layer, a sampling layer, a full link layer and an output layer, wherein the convolutional layer is used for carrying out convolution on data input by the input layer and then outputting the data; meanwhile, the sampling layer is used as the input of the sampling layer and is output after being subjected to a sampling method; and so on, the final input can be obtained;
(b) the convolution process is performed by determining the size of the filter, using a convolution kernel K
abFor a feature x of the filter of the previous layer
aWeighted to obtain x
a*K
abThen to x
a*K
abAfter summing, biasing is performed as shown in the following equation
Wherein the content of the first and second substances,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; m
bA filter corresponding to neuron b;
the b-th corresponding weight of the q-th neuron a is obtained; b is
qIs the only offset for layer q;
(c) the sampling layer calculates the average value of all values in the filter in a mode of only calculating the average of the feature points in the neighborhood; then multiplying the sampled information by trainable parameters, adding trainable bias, and obtaining the result by activating function calculation, the output formula of the sampling layer neuron is as follows
Wherein, β
qFor the trainable parameters of the q-th level,
a feature in the filter of the previous layer, i.e., layer q-1; (x) is sigmoid function; m
bA filter corresponding to neuron b; b is
qIs the only offset for layer q;
(d) for the full link layer, the calculation is directly carried out by adopting a method of a multilayer artificial neural network, and the output formula of the neuron of the full link layer is as follows:
wherein, w
baThe weight value from the node b of the q-th layer to the node i of the q + 1-th layer is obtained;
one of the filters of the previous layer, i.e. the q-1 th layerCharacteristic; (x) is sigmoid function; b is
qIs the only offset for layer q.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910993456.5A CN110780188A (en) | 2019-10-18 | 2019-10-18 | IGBT drive circuit fault diagnosis method |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201910993456.5A CN110780188A (en) | 2019-10-18 | 2019-10-18 | IGBT drive circuit fault diagnosis method |
Publications (1)
Publication Number | Publication Date |
---|---|
CN110780188A true CN110780188A (en) | 2020-02-11 |
Family
ID=69385906
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201910993456.5A Pending CN110780188A (en) | 2019-10-18 | 2019-10-18 | IGBT drive circuit fault diagnosis method |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN110780188A (en) |
Cited By (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111398787A (en) * | 2020-04-17 | 2020-07-10 | 安徽理工大学 | Fault diagnosis method for three-phase voltage type PWM (pulse-width modulation) rectification circuit under complex working condition |
CN112986784A (en) * | 2021-04-21 | 2021-06-18 | 国网江西省电力有限公司电力科学研究院 | Abnormity identification method and device for high-power welding type IGBT module |
CN113408815A (en) * | 2021-07-02 | 2021-09-17 | 湘潭大学 | Deep learning-based traction load ultra-short-term prediction method |
CN113533945A (en) * | 2021-06-30 | 2021-10-22 | 桂林电子科技大学 | Analog circuit fault diagnosis method based on two-dimensional convolutional neural network |
CN114325352A (en) * | 2022-01-04 | 2022-04-12 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
CN114454951A (en) * | 2021-12-30 | 2022-05-10 | 南京航空航天大学 | Dual-motor steer-by-wire system and convolutional neural network fault-tolerant control method thereof |
CN115078952A (en) * | 2022-08-19 | 2022-09-20 | 江苏东海半导体股份有限公司 | IGBT driving fault detection method and system |
Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102636742A (en) * | 2012-05-15 | 2012-08-15 | 长沙河野电气科技有限公司 | Large-scale analogue circuit fault diagnosis method based on wavelet neural network |
CN105841961A (en) * | 2016-03-29 | 2016-08-10 | 中国石油大学(华东) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network |
US20180129911A1 (en) * | 2016-11-04 | 2018-05-10 | Case Western Reserve University | Histomorphometric classifier to predict cardiac failure from whole-slide hematoxylin and eosin stained images |
CN108178037A (en) * | 2017-12-30 | 2018-06-19 | 武汉大学 | A kind of elevator faults recognition methods based on convolutional neural networks |
CN109389104A (en) * | 2018-11-30 | 2019-02-26 | 浙江碳银互联网科技有限公司 | A kind of family photovoltaic plant fault of converter prediction technique |
CN109444667A (en) * | 2018-12-17 | 2019-03-08 | 国网山东省电力公司电力科学研究院 | Power distribution network initial failure classification method and device based on convolutional neural networks |
KR101984730B1 (en) * | 2018-10-23 | 2019-06-03 | (주) 글루시스 | Automatic predicting system for server failure and automatic predicting method for server failure |
-
2019
- 2019-10-18 CN CN201910993456.5A patent/CN110780188A/en active Pending
Patent Citations (7)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN102636742A (en) * | 2012-05-15 | 2012-08-15 | 长沙河野电气科技有限公司 | Large-scale analogue circuit fault diagnosis method based on wavelet neural network |
CN105841961A (en) * | 2016-03-29 | 2016-08-10 | 中国石油大学(华东) | Bearing fault diagnosis method based on Morlet wavelet transformation and convolutional neural network |
US20180129911A1 (en) * | 2016-11-04 | 2018-05-10 | Case Western Reserve University | Histomorphometric classifier to predict cardiac failure from whole-slide hematoxylin and eosin stained images |
CN108178037A (en) * | 2017-12-30 | 2018-06-19 | 武汉大学 | A kind of elevator faults recognition methods based on convolutional neural networks |
KR101984730B1 (en) * | 2018-10-23 | 2019-06-03 | (주) 글루시스 | Automatic predicting system for server failure and automatic predicting method for server failure |
CN109389104A (en) * | 2018-11-30 | 2019-02-26 | 浙江碳银互联网科技有限公司 | A kind of family photovoltaic plant fault of converter prediction technique |
CN109444667A (en) * | 2018-12-17 | 2019-03-08 | 国网山东省电力公司电力科学研究院 | Power distribution network initial failure classification method and device based on convolutional neural networks |
Non-Patent Citations (6)
Title |
---|
HE Y G 等: "Incipient Fault Diagnosis Method for IGBT Drive Circuit Based on Improved SAE", 《IEEE ACCESS》 * |
WANG D H 等: "Transmission Line Fault Diagnosis Based on Wavelet Packet Analysis and Convolutional Neural Network", 《2018 5TH IEEE INTERNATIONAL CONFERENCE ON CLOUD COMPUTING AND INTELLIGENCE SYSTEMS (CCIS)》 * |
章华 等: "基于卷积神经网络的煤岩识别研究", 《绥化学院学报》 * |
蒋英春: "《小波分析基本原理》", 31 December 2012, 天津大学出版社 * |
许天周 等: "《近代数学基础》", 31 May 2009, 北京理工大学出版社 * |
谢涛: "基于神经网络和小波变换的模拟电路故障诊断理论与方法", 《中国优秀硕士学位论文全文数据库 信息科技辑》 * |
Cited By (11)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN111398787A (en) * | 2020-04-17 | 2020-07-10 | 安徽理工大学 | Fault diagnosis method for three-phase voltage type PWM (pulse-width modulation) rectification circuit under complex working condition |
CN111398787B (en) * | 2020-04-17 | 2022-09-16 | 安徽理工大学 | Fault diagnosis method for three-phase voltage type PWM (pulse-width modulation) rectification circuit under complex working condition |
CN112986784A (en) * | 2021-04-21 | 2021-06-18 | 国网江西省电力有限公司电力科学研究院 | Abnormity identification method and device for high-power welding type IGBT module |
CN113533945A (en) * | 2021-06-30 | 2021-10-22 | 桂林电子科技大学 | Analog circuit fault diagnosis method based on two-dimensional convolutional neural network |
CN113408815A (en) * | 2021-07-02 | 2021-09-17 | 湘潭大学 | Deep learning-based traction load ultra-short-term prediction method |
CN114454951A (en) * | 2021-12-30 | 2022-05-10 | 南京航空航天大学 | Dual-motor steer-by-wire system and convolutional neural network fault-tolerant control method thereof |
CN114454951B (en) * | 2021-12-30 | 2022-12-06 | 南京航空航天大学 | Dual-motor steer-by-wire system and convolutional neural network fault-tolerant control method thereof |
CN114325352A (en) * | 2022-01-04 | 2022-04-12 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
CN114325352B (en) * | 2022-01-04 | 2023-04-18 | 电子科技大学 | Analog filter circuit fault diagnosis method based on empirical wavelet transform |
CN115078952A (en) * | 2022-08-19 | 2022-09-20 | 江苏东海半导体股份有限公司 | IGBT driving fault detection method and system |
CN115078952B (en) * | 2022-08-19 | 2022-12-23 | 江苏东海半导体股份有限公司 | IGBT driving fault detection method and system |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN110780188A (en) | IGBT drive circuit fault diagnosis method | |
CN105548862B (en) | A kind of analog-circuit fault diagnosis method based on broad sense multi-kernel support vector machine | |
WO2019090878A1 (en) | Analog circuit fault diagnosis method based on vector-valued regularized kernel function approximation | |
CN109597043B (en) | Radar signal identification method based on quantum particle swarm convolutional neural network | |
He et al. | Wavelet neural network approach for fault diagnosis of analogue circuits | |
He et al. | A rough membership neural network approach for fault classification in transmission lines | |
CN101477172B (en) | Analogue circuit fault diagnosis method based on neural network | |
Kashyap et al. | Classification of power system faults using wavelet transforms and probabilistic neural networks | |
CN107894564A (en) | A kind of analog-circuit fault diagnosis method based on intersection wavelet character | |
Ma et al. | Capturing group variability using IVA: a simulation study and graph-theoretical analysis | |
CN110346736B (en) | NPC three-level inverter fault diagnosis method based on improved treelet transformation | |
CN104198924A (en) | Novel analog circuit early fault diagnosis method | |
Tan et al. | Analog fault diagnosis using S-transform preprocessor and a QNN classifier | |
CN104793124A (en) | Switched circuit fault diagnosis method based on wavelet transformation and ICA (independent component analysis) feature extraction | |
Sohl-Dickstein et al. | Hamiltonian annealed importance sampling for partition function estimation | |
Moezi et al. | Simultaneous fault localization and detection of analog circuits using deep learning approach | |
CN114692758A (en) | Power communication fault analysis method and device, terminal equipment and medium | |
Pan et al. | Test generation for linear time-invariant analog circuits | |
CN113111731A (en) | Deep neural network black box countermeasure sample generation method and system based on channel measurement information | |
Pan et al. | Learning approach based DC arc fault location classification in DC microgrids | |
CN113112039B (en) | Active power distribution system initial fault identification method based on time-frequency memory recurrent neural network | |
CN102445650B (en) | Blind signal separation algorithm-based circuit fault diagnosis method | |
CN110458071A (en) | A kind of fiber-optic vibration signal characteristic abstraction and classification method based on DWT-DFPA-GBDT | |
CN109782158A (en) | A kind of Analog circuit diagnosis method based on multiclass classification | |
Sridhar et al. | Detection and classification of power quality disturbances in the supply to induction motor using wavelet transform and neural networks |
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 | ||
RJ01 | Rejection of invention patent application after publication | ||
RJ01 | Rejection of invention patent application after publication |
Application publication date: 20200211 |