Method for Extracting Micro Fault Features of Inverter IGBT Based on Multimodal Data

Abstract

The present invention relates to a method for extracting micro-fault characteristics of IGBTs in an inverter based on multi-modal data, comprising the following steps: 1) Establish an equivalent second-order system for the transient mode of a two-level traction inverter, and identify the parameters of the equivalent second-order system expression; 2) Perform modal division on the inverter output voltage data through an Elman neural network; 3) Extract the corresponding micro-fault characteristic parameters in each transient mode, including voltage overshoot σ, peak time t p and the absolute value of voltage slope Δ; 4) Extract the voltage steady-state value V in each steady-state mode stab ; 5) Perform fault detection based on the extracted characteristic parameters of IGBT micro-faults. Compared with the prior art, the present invention has the advantages of being easy to implement, cost-saving, reducing noise interference, improving the accuracy of modal classification, accurately extracting and fault detecting, etc.

Description

Technical Field

[0001] The present invention relates to the field of inverter fault detection, and particularly to a method for extracting micro fault features of IGBTs in an inverter based on multi-modal data. Background Art

[0002] During the long-term operation of IGBTs in common two-level traction inverters, due to the cumulative effects of voltage and current, the interference of line electromagnetic waves, and the thermal stress generated at the interface due to different thermal expansion coefficients of various materials inside the module, micro faults such as bond wire breakage and solder layer fatigue will occur. If these micro faults are not detected in a timely and effective manner and allowed to develop, they will eventually evolve into significant faults such as short circuits or open circuits, which will have an adverse impact on the operation of the system. Compared with significant faults, IGBT micro faults have characteristics such as small fault characteristic values, slow fault development, and easy masking of fault characteristics by unknown disturbances and noises, making their detection more difficult. Therefore, selecting appropriate fault characteristic variables that can reflect IGBT micro faults and accurately extracting fault characteristic parameters from these fault characteristic variables play a decisive role in accurately detecting micro faults.

[0003] In recent years, experts and scholars at home and abroad have conducted extensive and in-depth research on IGBT micro fault detection methods, and have determined some fault characteristic variables that can clearly characterize IGBT micro faults, including the collector-emitter saturation conduction voltage drop V CE(on) , gate voltage V GE , turn-on threshold voltage V GE(th) , turn-off time t off , turn-on time t on , junction-case thermal resistance R th , junction temperature T j etc. However, for these known fault characteristic variables, corresponding sensors need to be installed on the IGBT module inside the inverter to obtain them, which not only increases the complexity of the system but also improves the sampling difficulty and is not convenient for timely obtaining fault characteristic information. Summary of the Invention

[0004] The purpose of the present invention is to overcome the defects of the above-mentioned existing technologies and provide a method for extracting micro fault features of IGBTs in an inverter based on multi-modal data.

[0005] The purpose of the present invention can be achieved by the following technical solutions:

[0006] A method for extracting micro fault features of IGBTs in an inverter based on multi-modal data, which is used to extract characteristic parameters that can reflect micro faults of IGBTs in the inverter for subsequent fault detection, includes the following steps:

[0007] 1) Establish an equivalent second - order system for the transient mode of a two - level traction inverter, and identify the parameters of the equivalent second - order system expression;

[0008] 2) Divide the inverter output voltage data into modes through an Elman neural network;

[0009] 3) Extract the corresponding micro - fault characteristic parameters in each transient mode, including voltage overshoot σ, peak time t p and the absolute value of voltage slope Δ;

[0010] 4) Extract the voltage steady - state value V in each steady - state mode stab ;

[0011] 5) Conduct fault detection based on the extracted characteristic parameters of IGBT micro - faults.

[0012] The specific steps of step 1) include the following steps:

[0013] 101) Construct the transfer function of the equivalent second - order system for the transient mode of the two - level traction inverter, and its expression is:

[0014]

[0015] where c1 and c2 are both parameters of the transfer function, and s1 and s2 are the roots of the characteristic equation of the transfer function G(s) respectively;

[0016] 102) Obtain the corresponding step response when a step signal with a set sampling interval is used as the system input;

[0017] 103) Substitute the step response into the second - order difference equation to solve. After solving the parameters of the second - order difference equation, calculate and obtain the parameters of the transfer function and the roots of the characteristic equation. The expression of the second - order difference equation is:

[0018] g(t)+a1g(t + Δt)+a2g(t + 2Δt)=0

[0019] where a1 and a2 are the parameters of the second - order difference equation, g(t), g(t + Δt), and g(t + 2Δt) are the step responses at sampling times t, t + Δt, and t + 2Δt respectively, and Δt is the sampling interval.

[0020] The specific steps of step 2) include the following steps:

[0021] 201) Divide a batch of original working mode data of the inverter, that is, the original output voltage data, into s time slices to obtain a data set X i composed of s data tuples x set , and the data tuple As a classification characteristic parameter of the inverter operating mode, where is the average value of the original operating mode data;

[0022] 202) Combine the data sets of all n batches to form a training sample set Sample_Set, and use the training sample set Sample_Set as the input data set of the Elman neural network for training;

[0023] 203) Perform mode division on the inverter output voltage data according to the trained Elman neural network.

[0024] In step 201) described above, each data set X set contains at least one complete operating mode switching cycle, including 6 transient modes and 6 steady-state modes.

[0025] Step 3) described above includes two processes: transient parameter calculation and weight fusion. That is, for the voltage transient mode, first calculate the overshoot and peak time in the original output voltage data and the parameter identification system data respectively, then fuse and average the two overshoots and peak times according to the weights, and finally obtain the absolute value of the voltage slope Δ through combined calculation.

[0026] Step 3) specifically includes the following steps:

[0027] 301) Extract the first k large voltage values in the original output voltage data to form a maximum value set V max , and calculate the overshoot σ Dataset of the original data accordingly. Then there is:

[0028]

[0029] where avg(V max ) is the average value of the maximum value set, and V norm is the reference level corresponding to the transient mode;

[0030] Form a peak time set with the times corresponding to the voltage values in the maximum value set V max , and calculate the average of the times in the peak time set. Then the peak time extracted from the original output voltage data is:

[0031] t p_Dataset = avg(t max ) - t0

[0032] where t p_Dataset is the peak time, avg(t max ) is the average of the times in the peak time set, and t0 is the transient start time;

[0033] 302) Obtain the damping ratio ξ based on the identified equivalent second-order system transfer function, and calculate the system overshoot σ System and the peak time t p_System , then there is:

[0034]

[0035]

[0036] where ω n is the undamped natural oscillation frequency;

[0037] 303) Determine the fusion weight α of the original data information and the fusion weight β of the identified system information respectively, and obtain the fused voltage overshoot σ and peak time t according to the fusion weights p , then there is:

[0038] σ = avg(σ Dataset , σ System ) = α' × σ Dataset + β' × σ System

[0039] t p = avg(t p_Dataset , t p_System ) = α' × t p_Dataset + β' × t p_System

[0040]

[0041] where α' and β' are the normalized fusion weights respectively;

[0042] 304) Calculate the absolute value Δ of the transient modal voltage slope through non-linear combination, then there is:

[0043]

[0044] In the said step 303), the expression of the fusion weight α of the original data information is:

[0045]

[0046]

[0047]

[0048] where size(Dataset) is the size of the actual data volume, k is the total number of large voltage values selected from the original output voltage data, t i , t i+1 respectively represent the moments corresponding to the i-th and i + 1-th large voltage values.

[0049] In the aforesaid step 303), the expression of the fusion weight β for identifying system information is as follows:

[0050] β = e -L

[0051] where e is the natural logarithm, and L is a loss function reflecting the main error of parameters in the identification system.

[0052] The aforesaid step 4) is specifically as follows:

[0053] Obtain the intersection coordinates of the output characteristic curves of the IGBT module at different temperatures. Then, the abscissa value of the intersection point is the IGBT saturation conduction voltage drop that is not affected by the junction temperature. This IGBT saturation conduction voltage drop that is not affected by the junction temperature is only related to the minor fault state of the IGBT module. Make the IGBT module in the inverter work at this IGBT saturation conduction voltage drop that is not affected by the junction temperature, and collect the steady-state voltage value V corresponding to the output voltage of the inverter stab .

[0054] The extraction system for implementing the extraction method for the minor fault characteristics of the IGBT of the inverter includes:

[0055] Inverter equivalent second-order system parameter identification module: used to construct an equivalent second-order system for the transient mode of the two-level traction inverter and perform parameter identification;

[0056] Inverter operating mode division module: used to perform mode division on the output voltage data of the inverter through an Elman neural network;

[0057] Transient feature extraction module and steady-state feature extraction module: used to extract the fault characteristics of each transient mode and steady-state mode respectively, including voltage overshoot σ, peak time t p and the absolute value of voltage slope Δ and voltage steady-state value V stab .

[0058] Compared with the prior art, the present invention has the following advantages:

[0059] (1) At the output level of the inverter system, the present invention directly collects the three-phase voltages at the output end of the inverter as fault feature variables, and processes the three-phase voltage data to obtain corresponding transient and steady-state fault feature parameters, avoiding installing sensors inside the inverter, which is easier to implement in engineering and more cost-saving.

[0060] (2) The present invention proposes a feature parameter extraction method that combines the identification features obtained by a system parameter identification method with the original data features, which not only reduces the interference of data noise but also avoids the influence of parameter identification errors on the feature extraction effect.

[0061] (3) The method for dividing the operating modes of an inverter based on an Elman neural network proposed by the present invention combines the operating mode data and historical data, greatly improving the accuracy of mode classification and solving the problem of misclassification of modes.

[0062] (4) The present invention proposes an effective method for extracting the saturation conduction voltage drop under a specific collector current, realizing the precise extraction of the characteristics of small steady-state faults of the inverter voltage type. Description of the Drawings

[0063] Figure 1 is the system structure block diagram of the present invention.

[0064] Figure 2 is the main circuit topology of a two-level inverter and the internal structure of a non-ideal IGBT.

[0065] Figure 3 is the step response output waveform and the display diagram of four fault characteristic parameters in the switching case of a two-level inverter.

[0066] Figure 4 is the schematic diagram for dividing the operating modes of the single-phase output phase voltage of a two-level inverter under square wave control.

[0067] Figure 5 is the schematic diagram for obtaining the training data of an Elman neural network.

[0068] Figure 6 is the flow chart for extracting the characteristic parameters of small faults in the voltage transient mode.

[0069] Figure 7 is the output characteristic curve of the IGBT module at different temperatures.

[0070] Figure 8 is the classification result of the operating modes of an Elman neural network.

[0071] Figure 9 is the comparison diagram of the three-dimensional distribution of characteristic parameters in the normal working state and the small fault state.

[0072] Figure 10 is the comparison diagram of the two-dimensional distribution of characteristic parameters in the normal working state and the small fault state. Detailed Embodiment

[0073] The present invention provides a method for extracting micro - fault characteristics of IGBTs in an inverter based on multi - modal data. From the perspective of the inverter system, by analyzing the influence of IGBT micro - faults on the output port characteristics of the traction inverter, the inverter output voltage that can reflect the IGBT micro - faults is selected as the micro - fault characteristic variable. Then, according to the change law of the inverter output voltage signal before and after the fault, the transient signal characteristic parameters and steady - state signal characteristic parameters of the voltage are extracted, and these characteristic parameters provide a basic guarantee for subsequent IGBT micro - fault detection.

[0074] As Figure 1 shown, the present invention proposes and designs a system for extracting micro - fault characteristics of IGBTs in an inverter based on multi - modal data. The system structure mainly includes an inverter equivalent second - order system parameter identification module, an inverter operating mode division module, a transient feature extraction module, and a steady - state feature extraction module. This system mainly collects and processes the three - phase output voltage signals of a two - level traction inverter, and extracts four characteristic parameters that can reflect the IGBT micro - faults in the inverter.

[0075] The research object of the present invention is a two - level traction inverter composed of IGBT modules. Existing research shows that in order to effectively detect the micro - faults of IGBT modules, it is necessary to fully consider the distribution of stray parameters during the operation of the modules. By studying the reliability of IGBTs, it can be known that the occurrence of micro - faults will cause changes in the stray parameters inside the IGBT module. Therefore, the two - level traction inverter, which is the research object of the present invention, is composed of six non - ideal IGBT modules considering stray parameters, and its topological structure and the internal structure of the non - ideal IGBT module are as Figure 2 shown. Further analyzing the equivalent circuit model of the non - ideal IGBT module, it can be seen that when it is in the switching operation state, the inverter system can be equivalent to a second - order system. The change in stray parameters caused by IGBT micro - faults will also change the output parameters of the step response of the equivalent second - order system, mainly reflected in the overshoot, peak time of the transient response in the step output response, and the absolute value of the voltage slope calculated by combining the two, and at the same time, it will also cause a change in the amplitude of the steady - state response. Therefore, the present invention processes the collected inverter output voltage signal and extracts the transient characteristic parameter overshoot σ, peak time t p and the absolute value of the voltage slope Δ, as well as the steady - state characteristic parameter voltage steady - state value V stab , and the oscillation waveform of the second - order system and the four proposed fault characteristic parameters are as Figure 3 shown.

[0076] At the same time, for the convenience of description, the inverter in the present invention adopts a square - wave control method, and the corresponding waveform of the inverter output phase voltage is as Figure 4As shown, it can be seen that within one period of the output voltage of the two-level inverter, there are 6 different steady-state operating states and 6 different transient oscillation states. Therefore, it is necessary to accurately divide these 12 different operating modes first, and then the corresponding transient fault characteristic parameters and steady-state fault characteristic parameters can be extracted in each operating mode.

[0077] Based on the above system, the present invention further proposes a method for extracting micro fault characteristics of IGBTs in an inverter based on multi-modal data, including the following steps:

[0078] (1) Establish an equivalent second-order parameter identification system for the inverter and solve the equivalent second-order system expression.

[0079] By identifying the second-order system parameters corresponding to the inverter, they can be used as the input parameters of the mode division module to achieve mode classification, and the overshoot and peak time of the transient characteristic parameters can also be calculated according to the identified parameters. The specific steps are as follows:

[0080] (101) Any second-order system can be represented by a second-order difference equation, that is:

[0081] g(t)+a1g(t + Δt)+a2g(t + 2Δt) = 0 (1)

[0082] Where a1 and a2 are parameters related to the actual second-order system.

[0083] Assume that the sampling interval is Δt. According to the transfer form of the difference equation, the time is delayed by Δt in sequence, and n equations can be written:

[0084]

[0085] Write the above formula in matrix form:

[0086]

[0087] Y = θX

[0088] Using the least squares method, a set of estimated values can be found to minimize the sum of squared errors:

[0089]

[0090] where L is the loss function, which takes the partial derivative with respect to the parameter matrix and finally the following can be obtained:

[0091]

[0092] In this way, the second-order system parameters a1 and a2 can be obtained.

[0093] (102) Similarly, assuming that the roots of the characteristic equation of the transfer function G(s) of the equivalent second-order system of the inverter transient mode are denoted as s1 and s2, then the second-order system can be expressed as:

[0094]

[0095] When the system input is a step signal u(t), the corresponding output is the step response g(t), which can be expressed as:

[0096]

[0097] Therefore, with a sampling interval of Δt, the step responses at times t + Δt, t + 2Δt, …, t + nΔt can be expressed as:

[0098]

[0099] (103) Substituting equations (3) and (4) into equation (1), we get:

[0100]

[0101] For equation (5) to hold, the value inside the square brackets should be 0, that is:

[0102]

[0103] Let Then the solution of the above equation can be written as:

[0104]

[0105] Correspondingly, we can obtain:

[0106] s1 = lnx1 / Δt, s2 = lnx2 / Δt

[0107] Substituting x1, x2 and t = 0 into equations (3) and (4), we get:

[0108]

[0109] Solving this equation system jointly, we can obtain c1 and c2.

[0110] Therefore, by finding the parameters a1 and a2 of the second-order system, we can further obtain the parameters s1, s2, c1, and c2 of the second-order system corresponding to equation (2).

[0111] (2) Based on the Elman neural network, complete the modal division of the inverter output voltage data.

[0112] When the two-level inverter is dividing the working mode, the timing characteristics of mode switching are a very significant feature. The appearance of a certain mode may only occur after a fixed working mode. The Elman neural network is a typical feedback network that can completely store the neuron information of the previous moment and can solve the timing characteristics problem well. The Elman neural network takes advantage of the fact that the neurons in the successive layer directly record the output results of the hidden layer at the previous moment, further imposes sequential restrictions on the modal classification results, and adaptively establishes a model based on the data timing characteristics. The specific steps are as follows:

[0113] (201) After the second-order system parameter identification in step (1), the original sampled data can use the identified parameters a1 and a2 to describe the distribution of some data, which can be used as the input parameters of the Elman neural network. However, after the parameter identification, the significant differences between some modes are eliminated, and it is impossible to distinguish between the transient mode d2 and the transient mode d6. The d2 and d6 data in the original data are distributed at different voltage levels, and the original working mode data can be averaged. The difference between the d2 and d6 modes can be reflected by comparing the average values. Stable at the 4 reference levels of the inverter, while the average value of the transient mode data Stable between the various reference levels, so the average The steady-state modes and transient modes can also be distinguished well.

[0114] (202) Based on the analysis of the above classification characteristics, the classification characteristic parameters of the inverter working mode are selected as At the same time, the same batch of data to be processed is divided into n time slices to obtain n data tuples x i Composed as a network training processing unit X set and use this as the input data set of the Elman neural network.

[0115] (203) Before using the Elman neural network to perform modal classification on the collected data, the neural network is trained first. Figure 5 As shown in the figure, after the original data is divided into time slices, parameter identification and mean extraction, the data tuple x is obtained. m If a batch of processed data contains s time slices, s data tuples are obtained, and these s data constitute the network training processing unit X set After the above processing, a total of n batches of data form a training sample set Sample_Set, which contains n network training processing units X set , each X set It contains at least one complete working mode switching cycle (6 transient states and 6 steady states), and each Xset Various training-validation experiments can be carried out without considering their timing characteristics, enabling the model to achieve the best matching effect for the data. At the same time, k-fold cross-validation is used to further improve the classification effect.

[0116] (3) According to the six transient modes obtained by the division in step (2), corresponding fault characteristic parameters - voltage overshoot σ, peak time t p and absolute value of voltage slope Δ are extracted in the transient modes.

[0117] To reduce the error of the fault characteristic parameters while reflecting the actual fault characteristics, this process is divided into two main processes: transient parameter calculation and weight fusion. That is, for the voltage transient mode, the overshoot and peak time are calculated in the original data and the parameter identification system respectively, and then the two overshoots and peak times are fused and averaged according to the weights. Finally, the absolute value of voltage slope Δ is calculated through the combination of the two. The specific steps are as follows:

[0118] (301) According to the definition, the signal overshoot represents the ratio of the maximum overshoot of the signal step response to the system stable value, and the peak time represents the time corresponding to the voltage peak minus the transient start time. To reduce the influence of interference, the first k largest voltage values of the original voltage data are taken The value of k is determined by the actual data volume, forming the maximum value set V max , and the overshoot σ of the original data can be calculated according to formula (6) Dataset .

[0119]

[0120] Similarly, the moments corresponding to the first k largest voltage values of the voltage data are selected to form the peak time set, and the mean value of these moments is calculated to smooth these data. Thus, the peak time extracted from the original data is:

[0121] t p_Dataset = avg(t max ) - t0 (7)

[0122] (302) Calculate the parameters s1, s2, c1, c2 according to a1 and a2 obtained in step (1), and obtain the standard form of the second-order system transfer function:

[0123]

[0124] The damping ratio ξ of the second-order system can be calculated from equation (8), and the overshoot σ of the system can be calculated using ξ System :

[0125]

[0126] Similarly, the peak time of the second-order system can be calculated based on Equation (10) as follows:

[0127]

[0128] (303) Determine the fusion ratio of the parameters calculated from the original data and the parameters obtained from the identification system, respectively.

[0129] (3031) The parameter deviation in the original data mainly comes from various disturbances. The most ideal situation for extracting the maximum k value is that k values are continuous, and the time differences obtained in sequence according to the index are all the sampling interval Δt, and the product of the differences is (Δt) k-1 , which is the minimum time difference product; the worse situation for extracting the maximum k value is that k values are scattered in each segment of a working mode period, and the product of the time differences obtained in sequence according to the index number is the largest, that is size(Dataset) is the size of the actual data volume.

[0130] It is hoped that in the most ideal situation, the highest score can be obtained for this data set, that is, let:

[0131] At the same time, when the sampling points are the most scattered, the lowest score can be obtained, that is, let:

[0132]

[0133] Mapping the lowest and highest score intervals to the [0,1] interval, the fusion weight α of the original data information is:

[0134]

[0135] According to the above formula, in the most ideal case, α = 1, and when the data is the most scattered, α = 0.

[0136] (3032) The main error of the parameters in the identification system comes from the difference between the identification system and the actual system, and the loss function L exactly reflects the magnitude of this difference. When the fitting effect of the original data and the identification system is good and the data is completely matched, the minimum value of the loss function L is 0; when the difference between the original data and the identification system is extremely large and the identification result does not converge within a limited number of iterations, the loss function L may be very large, even approaching ∞.

[0137] It is hoped that the credibility of the overshoot information in the identification system can be reflected through the fusion weight. The reciprocal of the loss function of the identification system can be taken as the system score. The better the data fitting, the closer the system score is to +∞; the worse the data fitting, the closer the system score is to 0. At the same time, the score is also mapped to the [0,1] interval, which can be achieved by using the negative exponential function with e as the base. Therefore, the fusion weight β of the identification system information is:

[0138] β = e -L

[0139] Normalize α and β to obtain α' and β':

[0140]

[0141] Therefore, the overshoot σ after fusion is:[[]]

[0142] σ = avg(σ Dataset , σ System ) = α' × σ Dataset + β' × σ System

[0143] The peak time t after fusion p is:[[]]

[0144] t p = avg(t p_Dataset , t p_System ) = α' × t p_Dataset + β' × t p_System

[0145] (304) After extracting the system overshoot σ and peak time t p , the absolute value of the transient modal voltage slope is obtained through the non - linear combination of the two:[[]]

[0146]

[0147] Although the voltage slope feature is calculated by combining other fault features, it does not include the non - linear mapping of data when adopting the diagnostic method, so it can bring new discriminant information for fault diagnosis.[[]]

[0148] To sum up, the extraction process of the fault feature parameters of the voltage transient process is as Figure 6 shown.[[]]

[0149] (4) Extract the voltage steady - state value in the voltage steady - state mode.[[]]

[0150] A minor fault in the IGBT of the inverter will cause a change in the saturation conduction voltage drop V CE(on) of the IGBT, which will further affect the change of the inverter output voltage steady - state value V stab . And V CE(on) is not only affected by the minor fault of the IGBT module, but also affected by the junction temperature of the IGBT module.[[]] Figure 7 The output characteristic curves of the IGBT module at different temperatures are shown as follows. To eliminate the influence of different junction temperatures on V CE(on) , it is necessary to determine Figure 7 the intersection coordinates (V alter , I alter of the output characteristic curves of the IGBT module at different temperatures shown as follows), the abscissa value of this intersection point corresponds to a specific collector current I c corresponding V CE(on) , which is a constant value and is not affected by temperature changes. Therefore, for the IGBT module, to determine the V CE(on) that is not affected by temperature changes, the specific steps are as follows:

[0151] (401) The IGBT can be equivalently regarded as a PiN diode and a MOSFET in series. V CE(on) is determined by the voltage drop competition between the diode and the MOSFET. When I c is small, V CE(on) is mainly determined by the diode; when I c is large, V CE(on) is mainly determined by the MOSFET, and the obtained intersection point is the critical point of the competition between the two. Among them, when the diode is dominant, the relationship between the IGBT current and the voltage across its two ends is:

[0152]

[0153] In Equation (11), V PiN is the voltage drop of the PiN diode, K is the Boltzmann constant, T is the absolute temperature of the IGBT, q is the electric charge quantity, D a is the bipolar diffusion coefficient, n i is the carrier concentration, W N is the width of the N-base region, L a is the diffusion length, F(·) represents the functional relationship, and J C is the collector current density.

[0154] (402) When the MOSFET dominates V CE(on) , the relationship between the IGBT current and the voltage across its two ends is:

[0155]

[0156] In Equation (12), V MOSFET is the voltage drop across the MOSFET channel when operating in the linear region, μ ni is the channel carrier mobility, C OX is the metal oxide layer capacitance, V G is the gate voltage, V TH is the threshold voltage, p is the cell size (I C = J C pZ, where Z is the length orthogonal to the cross-section of the IGBT structure), and L CH is the channel diffusion length.

[0157] (403) The on-state voltage drop of the IGBT is the sum of the voltage of the PiN part and the voltage of the MOSFET part, that is:

[0158] V CE(on) = V PiN + V MOSFET

[0159] When different currents flow through the IGBT, the dominant structure of the voltage is different, and the conduction voltage drop can be approximately expressed only by V PiN or only by V MOSFET When the voltage dominant structure changes, the critical point value is (V alter , I alter ), that is:

[0160]

[0161] To facilitate the solution of the critical point, the inverse function of the above formula is obtained:

[0162]

[0163] To solve the critical point, take the first derivative of Equation (14) with respect to V CE(on) as the variable, and the following can be obtained:

[0164]

[0165] Continue to take the second derivative to get:

[0166]

[0167] Only need to determine the intersection of the second derivative of J C and the value of 0. The corresponding V CE(on) of the intersection is the IGBT saturation conduction voltage drop that is not affected by the junction temperature and is only related to the micro-fault state of the IGBT module. Let the IGBT module in the inverter work at this collector saturation conduction voltage drop, and collect the corresponding steady-state voltage value V stab of the inverter output voltage, that is, it is not affected by temperature changes and can be used as a characteristic parameter to reflect micro-faults.

[0168] (5) Micro-fault detection based on transient modal characteristics and steady-state modal characteristics

[0169] According to the above steps, the transient characteristic parameters overshoot, peak time and absolute value of voltage slope of 6 transient modes of the inverter output voltage, and the steady-state characteristic parameter steady-state voltage value of 6 steady-state modes can be extracted. If it is necessary to detect whether a certain switch tube in the inverter has a micro-fault, then select the transient and steady-state modes corresponding to the switching working state of the switch tube, calculate 4 characteristic parameters, and use the PCA method for fault detection:

[0170] The PCA method is used to further calculate and process the four characteristic parameters to obtain two statistics of PCA, the Hotelling T 2 statistic and the SPE (Squared Prediction Error) statistic. By comparing these two statistics with the statistical control limits under normal operating conditions, the following detection results can be obtained:

[0171] ①Both SPE and T 2 do not exceed their respective control limits, indicating that the switching transistor has not failed;

[0172] ②If either SPE or T 2 exceeds its control limit, it can be determined that the switching transistor has a minor fault.

[0173] To verify the effectiveness of the present invention, a corresponding simulation model was built according to the two-level traction inverter shown in Figure 2 . Tables 1 and 2 show the non-ideal IGBT simulation parameters and the main circuit simulation parameters of the inverter respectively. The three-phase output voltage of the inverter was collected and processed according to the method proposed in the present invention to obtain the verification results.

[0174] Table 1 Non-ideal IGBT simulation parameters

[0175]

[0176] Table 2 Main circuit simulation parameters of the two-level inverter

[0177]

[0178] To verify the effectiveness of the mode division strategy, the output phase voltage data of the inverter was collected, and the sampling interval was set to 0.01 μs. The identification parameters and data mean were obtained according to steps (1) and (2), and then Figure 5 were used to form the input variables of the working mode division model. On this basis, the Elman neural network model was trained and k-fold cross-validation was used. Taking k = 3, the working mode division results of the a-phase voltage data are shown in Figure 8 . The calculated model accuracy is 0.99. It can be seen that the Elman neural network can effectively and accurately divide the 12 working modes of the two-level inverter.

[0179] To verify the effectiveness of the extracted fault characteristic parameters, the inverter A-phase voltage data under normal conditions and bond wire breakage faults were collected respectively. Assuming that the bond wire breakage fault occurs in the VT3 module as shown in Figure 1 , at this time, the emitter stray parameter R of the VT3 module EIt changes from 0.136 Ω to 0.204 Ω. Select the data of 100 cycles among them as samples, and calculate the transient characteristic parameters and steady-state characteristic parameters of the d3 and s3 modes in each cycle according to the proposed method. The average values of the four fault characteristic parameters are shown in Table 3. It can be seen that although the change amplitude is very small, after the occurrence of a minor fault, the four extracted characteristic parameters have all changed correspondingly compared with the normal state, that is, the overshoot becomes smaller, the peak time becomes larger, and the steady-state voltage value becomes larger, which is in line with the theoretical analysis results.

[0180] Table 3 Extraction results of minor fault characteristic parameters of the inverter

[0181]

[0182] To further verify the effectiveness of the extracted fault characteristic parameters, select the overshoot, peak time, and voltage steady-state value as the three-dimensional coordinate variables. The comparison results of the fault characteristic values in the normal working state and the minor fault state formed thereby are as Figure 9 shown. Among them, the blue circles represent the normal working state, and the red triangles represent the minor fault state. It can be seen from the figure that the characteristic parameters in the normal working state and the minor fault state are clustered in different spaces respectively, and the boundary between the two states is clear, without any mixing situation. Thus, it shows that the fault characteristic parameters extracted by the method proposed in the present invention can effectively distinguish the normal and fault states.

[0183] Similarly, in order to more specifically reflect the effectiveness of the extracted fault characteristic parameters, select the overshoot and peak time as the two-dimensional coordinate variables. The comparison results of the fault characteristic values in the normal working state and the minor fault state formed thereby are as Figure 10 shown. In this plane, the characteristic parameters in the normal working state and the minor fault state are also clustered in two different regions. Although there are several sample overlaps in the overshoot characteristic values, there is still a very clear boundary. Combining Figure 9 with the distribution status of the characteristic parameters, it can also be obtained that after a minor fault occurs in the IGBT, the overshoot of the characteristic parameters becomes smaller, the peak time becomes larger, and the steady-state voltage value becomes larger, which is consistent with the conclusion obtained in Table 3 and in line with the theoretical analysis results.

[0184] In summary, the method proposed in the present invention can accurately and effectively extract four fault characteristic parameters, and based on this, it can relatively clearly distinguish the minor fault state and the normal working state, providing good characteristic inputs for subsequent fault detection.

Claims

1. A method for extracting micro-fault features of IGBTs in inverters based on multi-modal data, which is used to extract characteristic parameters that can reflect the micro-faults of IGBTs in inverters for subsequent fault detection, is characterized in that, It includes the following steps: 1) Establish an equivalent second-order system for the transient mode of a two-level traction inverter, and identify the parameters of the equivalent second-order system expression; 2) Perform modal division on the inverter output voltage data through an Elman neural network; 3) Extract the corresponding micro fault feature parameters in each transient mode, including voltage overshoot σ, peak time t p and the absolute value of voltage slope Δ; 4) Extract the voltage steady-state value V in each steady-state mode stab ; 5) Conduct fault detection based on the extracted characteristic parameters of the IGBT micro-fault; The specific steps of step 2) include the following steps: (201) Divide the original operating mode data of a batch of inverters, that is, the original output voltage data, into s time slices to obtain a data set X i composed of s data tuples x set , and the said data tuple is used as the classification feature parameter of the inverter operating mode, where is the average value of the original operating mode data; 202) Combine all n batches of data sets to form a training sample set Sample_Set, and use the training sample set Sample_Set as the input data set of the Elman neural network for training; 203) Perform modal division on the inverter output voltage data according to the trained Elman neural network; Step 3) includes two processes: transient parameter calculation and weight fusion. That is, for the voltage transient mode, first calculate the overshoot and peak time in the original output voltage data and the parameter identification system data respectively, then fuse and average the two overshoots and peak times according to the weights, and finally obtain the absolute value of the voltage slope Δ through combined calculation; The specific content of step 4) is: Obtain the intersection coordinates of the output characteristic curves of the IGBT module at different temperatures. Then, the abscissa value of the intersection point is the IGBT saturation conduction voltage drop that is not affected by the junction temperature. This IGBT saturation conduction voltage drop that is not affected by the junction temperature is only related to the minor fault state of the IGBT module. Make the IGBT module in the inverter work at this IGBT saturation conduction voltage drop that is not affected by the junction temperature, and collect the steady-state voltage value V corresponding to the output voltage of the inverter stab .

2. The method for extracting the micro-fault characteristics of the inverter IGBT based on multi-modal data according to claim 1, wherein The specific steps of step 1) include the following steps: 101) Construct the transfer function of the equivalent second-order system for the transient mode of the two-level traction inverter, and its expression is: where c1 and c2 are both parameters of the transfer function, and s1 and s2 are the roots of the characteristic equation of the transfer function G(s) respectively; 102) Obtain the corresponding step response when a step signal with a set sampling interval is used as the system input; 103) Substitute the step response into the second-order difference equation to solve. After obtaining the parameters of the second-order difference equation, calculate and obtain the parameters of the transfer function and the roots of the characteristic equation. The expression of the second-order difference equation is: g(t)+a1g(t + Δt)+a2g(t + 2Δt) = 0 where a1 and a2 are the parameters of the second-order difference equation, g(t), g(t + Δt), and g(t + 2Δt) are the step responses at the sampling times of t, t + Δt, and t + 2Δt respectively, and Δt is the sampling interval.

3. A method for extracting micro-fault characteristics of an inverter IGBT based on multi-modal data according to claim 1, characterized in that, In the described step 201), each data set X set includes at least one complete working mode switching cycle, including 6 transient modes and 6 steady-state modes.

4. A method for extracting the micro - fault characteristics of an inverter IGBT based on multi - modal data according to claim 1, characterized in that, The specific steps of step 3) include the following steps: 301) Extract the first k large voltage values from the original output voltage data to form a maximum value set V max , and calculate the overshoot σ of the original data based on this Dataset , then there is: where avg(V max ) is the average value of the maximum value set, and V norm is the reference level corresponding to the transient mode; The set of maximum values V max The moments corresponding to each voltage value in it form a set of peak moments. By calculating the mean of the moments in the set of peak moments, the peak time extracted from the original output voltage data is: t p_Dataset = avg(t max ) - t0 where t p_Dataset is the peak time, avg(t max ) is the mean time of the peak time set, and t0 is the transient start time; (302) Obtain the damping ratio ξ based on the identified equivalent second-order system transfer function, and calculate the system overshoot σ System and the peak time t p_System , then we have: Among them, ω n is the undamped natural vibration frequency; 303) Determine the fusion weights α of the original data information and the fusion weight β of the identification system information respectively, and obtain the fused voltage overshoot σ and peak time t according to the fusion weights p , then we have: σ = avg(σ Dataset , σ System ) = α' × σ Dataset + β' × σ System t p = avg(t p_Dataset , t p_System ) = α' × t p_Dataset + β' × t p_System where α' and β' are the fused weights after normalization respectively; 304) Obtain the absolute value of the transient mode voltage slope Δ through non-linear combination calculation, then there is:

5. A method for extracting micro - fault features of an inverter IGBT based on multi - modal data according to claim 4, characterized in that, In step 303), the expression of the fused weight α of the original data information is: Among them, size(Dataset) is the actual data volume size, k is the total number of large voltage values selected from the original output voltage data, and t i , t i+1 respectively represent the moments corresponding to the i-th and (i + 1)-th large voltage values.

6. A method for extracting the micro-fault characteristics of an inverter IGBT based on multi-modal data according to claim 4, characterized in that In step 303), the expression of the fused weight β of the identification system information is: β＝e -L where e is the natural logarithm, and L is the loss function reflecting the main errors of the parameters in the identification system.

Images