TWI408387B - Inverter fault diagnosis method - Google Patents

Abstract

An inverter fault diagnosis method is disclosed for detecting each power transistor failure of inverter main circuit. The method includes steps as following: Simulating many characteristic values represent to each of the power transistors is fault, and generating a fault matter-element model thereof. Using an extension neural network algorithm step flow and the fault matter-element model to establish a weighting matrix. Using the fault matter-element model and the weighting matrix to establish a fault diagnosis system. Inputting a data into the fault diagnosis system, and executing a recognizing step flow of the detecting system to determine which transistor of the inverter is fault.

Description

Inverter fault detection method

The present invention relates to a method of detecting a fault in a frequency converter, and more particularly to a method of detecting a multi-level inverter fault.

In recent years, the rapid development of industry relies on motors to drive loads and create automation in the industry. The speed control of the motor is achieved by using a frequency converter, which converts direct current (DC) to alternating current (AC) output, so when the output AC frequency changes, the motor speed will change. However, the traditional two-level inverter is limited by the high voltage and high power requirements, so in order to overcome this shortcoming, many experts are working on multi-level inverters. (Multi-level inverter) architecture and application. The multi-level inverter is composed of a plurality of power semiconductors in a serial and parallel manner. The advantage is that it can improve the power quality due to the voltage line waveform of the inverter output line to ground. Many similar Step waveforms are embedded, causing the waveform to become a trapezoidal voltage waveform, making it closer to the Sinusoidal waveform, thus reducing Harmonics components and having low loss energy and improvement. Electromagnetic interference (EMI) and other advantages. In general, the more power transistors of a multi-level inverter, the lower the voltage change rate (dv/dt) at the output, and the closer the output voltage waveform is to a sine wave. The effective reduction of the withstand voltage of the power crystal and the power rating of the component will extend the service life of the power crystal switch, so the multi-level inverter circuit architecture is widely used in many places. However, multi-level inverters also have some disadvantages. The higher the number of layers, the more the number of power crystals required, which increases the cost of the circuit and makes the circuit difficult to control, which in turn leads to a more prone to failure when the inverter is running. Therefore, in order to avoid damage to the machine caused by the fault of the inverter, the fault detection of the inverter is quite worthy of research.

Please refer to FIG. 1 , which is a schematic structural diagram of a conventional neutral clamp type three-level inverter. Neutral point clamped (NPC) frequency converters are the simplest and most controllable among the various types of inverter types. Three-level inverters are used in practical applications. After a period of use, problems such as aging or damage of components will occur, and motor drive system failures are prone to occur. However, it is an extremely complicated and awkward thing to confirm which power transistor is faulty.

One aspect of the present invention provides a method for detecting a fault of a frequency converter, which is applied to detecting a plurality of power transistors of a main circuit to quickly identify which power transistor in the frequency converter is faulty.

A fault detection method for a frequency converter according to an embodiment of the present invention includes at least the following steps: simulating a plurality of characteristic values of the plurality of power transistor faults to establish a fault matter element model. A weight-value matrix is established by using an extension-like neural network learning process and the above-mentioned fault object model. A judgment system is established by using the fault matter metamodel and the weight value matrix. And inputting a piece of data to be tested, and using the judging system to perform an identification process to determine one of the faulty power transistors represented by the data to be tested. Therefore, the present embodiment can establish a judgment system in advance, and the judgment system mainly has a fault matter element model and an extension type neural network to quickly identify which power transistor fault in the inverter.

Please refer to FIG. 2, which is a flow chart showing the steps of the inverter fault detecting method according to an embodiment of the present invention. The inverter fault detection method of the present embodiment includes at least the following steps: First, as shown in step 110, a plurality of characteristic values of the plurality of power transistor faults are simulated to establish a fault matter element model. Then, as shown in step 120, a weight-value matrix is established by using an extension-type neural network learning process and the above-described fault object model. Next, as shown in step 130, a fault determination system is established using the faulty matter model and the weight value matrix. Finally, as shown in step 140, a piece of data to be tested is input, and an identification process is performed by the fault judging system to determine one of the faulty power transistors represented by the data to be tested. Therefore, the present embodiment can establish a fault judgment system in advance, and the judgment system mainly has a fault matter element model and an extension type neural network to quickly identify which power transistor fault in the frequency converter. The specific operation principle is explained as follows: Generally speaking, the types of power crystal switch failures in the inverter can be roughly divided into three categories: short-circuit fault, open-circuit fault and trigger signal. Trigger signal fault, etc. Among them, the short circuit fault is that the voltage across the switch is too large and exceeds the rated voltage of the switch, causing the power crystal to be broken down; the open circuit fault means that there is no trigger signal in each power crystal to make it act; and the trigger signal fault refers to the power crystal The drive circuit cannot generate a normal trigger pulse to the corresponding switch, which will cause the switch to not work properly. Therefore, in the present embodiment, a simulation environment of a three-level inverter of a neutral point clamp type is established by using a PSIM power electronic analog circuit software, and a single time and any switching fault (Switch fault) is studied. Through simulation analysis, it can be observed that if the inverter is in a normal state, the measured waveform will exhibit three phase balancing. For example, when the operating frequency of the inverter is 60 Hz and none of them In the case of a power crystal failure, the output waveforms of the phase voltages and their relative spectra are shown in Figures 3A to 3C and 4A to 4C, respectively. It can be seen from the 3A to 3C drawings that the voltage waveforms of the respective phases have the same magnitude and shape and the phase difference between them is 120°, which is an inherent characteristic of the three-phase balance. Therefore, some characteristics will change once any power crystal in the frequency converter fails. For example, when the switch S11 in the inverter architecture is set to a fault, it can be seen that the output voltage waveforms of the phases in the inverter are as shown in FIGS. 5A-5C. It can be observed from Fig. 5A that the phase voltage V _uo of the u arm is distorted, and the phase voltages V _vo and V _{wo of the} other two arms (ie, the v arm and the w arm) and the waveform of the inverter during normal operation. No different. The 6A-6C diagram is the voltage waveform of each phase when the switch S22 in the inverter fails, and the phase voltage V _{vo of} the 6B phase diagram is obviously different from the waveform of the inverter under normal operation. The 7A-7C diagram shows the three-phase phase voltage waveform (Voltage waveform) measured when the switch S24 is faulty. It can be observed that the output phase voltage V _vo waveform of the v arm is obviously distorted, and the phase voltage waveforms of other arms are observed. are not affected; FIG. 8A ~ 8C and the first three-phase voltage waveforms when compared with the fault switch S33, which was observed from the figure are three-phase voltage waveforms will be affected, particularly in the phase voltage waveform V w _wo arm of the most serious .

Therefore, it can be seen from the above analysis that once the inverter fails, it can be observed that the spectrum of its output phase voltage will be abnormal. For example, when the operating frequency of the inverter is 60 Hz and the switch S33 is faulty, each phase can be obtained. The frequency spectrum is as shown in Figures 9A to 9C. If the spectrum of the 4A to 4C non-fault is compared with each other, a characteristic spectrum capable of representing a power crystal failure can be found. In the present invention, the spectrum at ( m _f -1) and ( m _f +3) is selected as the characteristic spectrum, where m _{f is} defined as a frequency modulation index, and the mathematical expression can be expressed as

Where f _tri is the frequency of the carrier waveform; and f _sin is the sine wave frequency (ie the operating frequency of the frequency converter). After the simulation analysis, the data related to each power crystal fault can be obtained, and the fault detection system of the inverter is established by using the extension-like neural network algorithm as the fault location of the three-level inverter.

Next, the fault element matter element model established in the present embodiment is explained as follows: the matter element is a basic element for describing a thing in the extension theory, and is represented by the symbol “ R ”, which is the name of the given thing. N (Name), wherein c (characteristic) and the other corresponding to a characteristic value of ν (value) is constituted, it may represent element of this equation was as follows:

R =( N , c , ν ) ......(2)

In addition, according to the definition of the matter element, the relationship between the characteristics of the three basic elements and the values corresponding to the features can be mathematically expressed as:

ν = c ( N ) ......(3)

Therefore, the second form of the matter element can be converted into

R =( N , c , c ( N )) ......(4)

To understand more about the relationship between the basic elements of matter elements, you can use the matter space as shown in Figure 10 to represent the value of the object, its characteristics and its characteristics in the space coordinates x , y , The z- axis is represented and the characteristics of its combined transformation can be displayed. In addition, in the extension matter-element theory, if the matter element has multiple features, it can be m features c ₁ , c ₂ ,..., c _m and m corresponding feature values v ₁ , v ₂ ,..., v _m are respectively represented, so the relationship between them can be expressed by mathematics.

From the fifth formula, R is referred to as an m -dimensional matter element, and R _k = (N, c _k , v _k ) (k = 1, 2, ..., m) is represented as each of the sub-object elements of R. Therefore, Equation 5 can be rewritten as R = (N, C, V) , where matrix C = [c ₁ , c ₂ , ..., c _m ] ^T , and matrix V = [v ₁ , v ₂ ,. ..,v _m ] ^T . Therefore, in the form of multidimensional matter elements, it is possible to describe anything in real life. For example, the inverter in the AC motor drive system is a straight/AC converter. If the value of the event is expressed in the form of a meta-element, it can be written as follows:

Next, in order to calculate the correlation degree between the data to be tested and the faulty matter metamodel established by each fault characteristic value according to the present embodiment, that is, which power transistor fault should be classified as the data to be tested. . The present embodiment refers to extension mathematics to calculate the degree of relevance, and the principle thereof is explained as follows: classical mathematics uses a two-bit logical value as the identification for distinguishing whether a thing belongs to the set, and is represented by {1, 0}. , that is, when a certain thing is "1", it is judged to belong to this set; when it is "0", it is judged that it does not belong to this set, and thus it is impossible to classify clearly for the fuzzy zone. For this reason, the fuzzy set is a degree to which a certain thing belongs to the set, and is represented by continuous data of 0 to 1, and the closer to "1", the higher the degree of belonging to the set. As for the extension theory, the degree range of the fuzzy set mathematics [0, 1] is extended to the extension set of the continuous multi-value (-∞~+∞) range. For the item to be identified, the real function is found in the range of (-∞~+∞) by the correlation function method as the basis for the degree of a certain feature. Therefore, a positive value indicates that the item belongs to the positive strength of the feature; a negative value indicates that the item does not belong to the negative strength of the feature, where “0” is a characteristic that belongs to the feature at the same time or does not belong to the feature. .

In extension mathematics, the extension set definition can be expressed in the following mathematical form: suppose W is a universe of discourse and has any element a . , there will be a real number

Corresponding to it. If To represent a set of extensions in the domain W, the mathematical relationship can be expressed as:

And in the eighth formula, The correlation function is γ= K(a) , and K(a) represents a about Correlation degree. Therefore, the K(a) range is classified as follows:

The 9th formula is Positive region; the tenth formula is Negative region; while the 11th is The criticality. In addition, if ,then ,Simultaneously . Figure 11 is the range of each region of the extension set, and can be observed from the figure, if assumed (or a ₂ ), element a will be classified into positive and negative domains.

In the extension set, the correlation function is used as a way to solve the contradictory problem. Therefore, in order to establish the correlation function on the real axis, we must first introduce the concept of "Distance" and "Rank value".

The distance system is the distance between a certain point and the interval on the real domain in the extension set, and is defined as follows: assuming that a is any point on the real domain (-∞~+∞), and any interval A ₁ =< va ₁ , va ₂ > also belongs to the real domain, so the distance between point a and interval A ₁ can be represented by the 12th formula, while the 12A and 12B diagrams show the meaning of the distance, and the 12A and 12B diagrams are the point a in the interval A, respectively. ₁ is related to the interval A ₁ .

In addition, in addition to considering the relevance of points and intervals in real problems, it is necessary to consider the relationship between intervals and intervals, points and two intervals. Therefore, the relationship between a point and two intervals is represented by a position value, and is defined as follows: Let A ₁ = < va ₁ , va ₂ >, A ₂ = < va ₃ , va ₄ > belong to the second interval of the real domain, respectively. interval a ₁ and a ₂ in the section, the section position and the value of a point a ₁ and a ₂ zone can be expressed as:

Among them, Fig. 13A and Fig. 13B show the meaning of D(a, A ₁ , A ₂ ) .

The basic formula of the correlation function is as follows: Let A ₁ = < va ₁ , va ₂ >, A ₂ = < va ₃ , va ₄ >, And without a common endpoint, the correlation function between point a and the two intervals A ₁ , A ₂ can be expressed as:

And has the following properties:

(1) And a ≠ va ₁ , ;

(2) a = va ₁ or ;

(3) , And _{a ≠ va 1, va 2,} va 3, ;

(4) a = va ₃ or ;

(5) And a ≠ va ₃ , .

Among them, if the correlation function can reach the maximum value at a = 0.5 (va ₁ + va ₂ ) , this function is called the elementary correlation function, as shown in Fig. 14. Among them, it is observable that when K(a) <-1, it means that point a is outside the interval A ₂ (=< va ₃ , va ₄ >); when K(a) > 0 means that point a is in the interval A ₁ (=< va _1, va within _{2>); -1 <K (} a) <0 indicates a point falls within the extension, if the conditions for a conversion point, will be classified as ₁ point to a section A. In addition, the elementary correlation function is one of the special functions of extension mathematics.

Next, the present embodiment uses an extension-like neural algorithm for fault detection of a three-level frequency converter. This theory includes the correlation function value operation of the extension theory, the parallel computing power of the neural network and its learning. Features such as (Learn), Recall, and Generalize are therefore ideal for characterizing objects. Figure 15 shows the proposed extension-like neural network architecture, which includes an input layer, a calculation layer, and an output layer. The process of this architecture sends a feature sample (Sample) to the input layer, and is set by the weighting parameter range in the calculation layer to generate an image corresponding to the feature sample. In the calculus of this architecture, the weight value Is the upper limit of the classical domain of the input feature; Is the classical domain lower limit of the input feature, and and It is the maximum and minimum values of the weight values between the mth input endpoint and the gth output endpoint, respectively. In addition, from the output of Figure 15, each feature sample input of the architecture can be found, and only one output endpoint is changing, so this result can be used as the identification result of the input feature sample. If the result of the discrimination is not as expected, the original weight value will be adjusted and learned until the required target value is reached. In addition, the computing mode of the extension-like neural network can be divided into two types: learning mode and arithmetic mode. First, the supervised learning calculus program for extension-type neural networks is described as follows:

Step 1: Using the extension matter metamodel, set the link weights of the input and output nodes as:

In the formula 15, c _{m is} expressed as the mth feature of the category name Ng , and And was based on a classical field is characterized in clusters g c _m, and y is the total number of classified output terminal of the cluster. Among them, the classical domain range of the extensional neural theory (the invention sets the classical domain as the range of the corresponding magnitude of the fault feature) can be obtained by the following mathematical formula:

among them, Represents the input learning materials of the extension-like neural network.

Step 2: Calculate the weight center of each cluster, represented by the symbol W _{cen , g} , so it can be expressed mathematically as follows:

W _cen,g ={ w _{cen , g 1} , w _{cen , g 2} ,..., w _{cen , gm} ,...} ......(18)

g =1,2,..., y ; m =1,2,..., q ......(20)

And the first feature weight center in the cluster data g is represented by the symbol w _{cen , g 1} .

Step 3: b is read into the category number of learning samples s pen (Learning sample), which may be expressed as follows:

Where y is the total number of category categories.

Step 4: Using the Extension Distance (ED) formula, as shown in Equation 22, the distance between the learning sample and each cluster can be calculated as:

In the 22nd The s pen learning sample whose category number is b and whose characteristic is m ; and W _{cen , gm} is the center of the weight value between the mth input end point and the gth output end point. Figure 16 shows a schematic diagram of the extension distance (ED), which is used to describe the distance between point p and the interval < w ^L , w ^H >. In addition, it can be seen from Fig. 16 that when the range < w ^L , w ^H > is changed, the extension distance will be similarly changed due to the sensitivity characteristics. Generally speaking, when the range of the feature classic domain is large, it indicates that the required data is relatively fuzzy and causes a low sensitivity extension distance; conversely, if the feature classic domain range is narrowed, the data demand will be clearer and the sensitivity will be improved. For extension distance applications. Step 5: After the operation, the new cluster of learning materials can be obtained as g ^* and available. . If g ^* = b , it will jump directly to step 7; otherwise, proceed to step 6.

Step 6: Adjust and update the weight values of the b-th and g ^* -th clusters as follows:

(1) Update the weight center value of the b-th and g ^* -th clusters.

(2) Update the weight values of the b-th and g ^* -th clusters.

In the 23rd to 26th formulas, the representative meanings of the variables are defined as follows: η: the learning rate of the extension-type neural network.

: The new weight center value of the feature m and the category number b (after learning).

: The feature weight m and the category number is the old weight center value of b (before learning).

: The new weight center value of the feature m and the cluster number g ^* (after learning).

: Feature m and cluster number is the old weight center value of g ^* (before learning).

: The weight limit upper limit and lower limit (new value) of the feature m and the category number b .

: is the upper and lower weight (old value) of the weight m and the category number b .

: The upper and lower weights (new values) of the weight of the feature m and the cluster number g ^* .

: is the upper and lower weight (old value) of the weight m and the cluster number is g ^* .

The weight value adjustment process of step 6 can be illustrated in FIGS. 17A and 17B. In Figure 17A, it means that the learning sample p _sm should belong to cluster B , but the calculation result of the extension distance formula is So that the sample p _sm is classified as cluster A. However, if the weight value is adjusted, the new extension distance is obtained. As shown in Figure 17B, it can be learned that this sample is classified as Cluster B. In addition, in the learning process of the weight value of step 6, the weight value adjustment is performed only for the clusters b and g ^{* of} one category, and the weight values of the other clusters are not affected. Therefore, the extension-type neural network has faster computing speed and higher adaptability than other artificial intelligence (AI) with supervised learning rules, that is, there is a new sample input. When you are able to quickly find the corresponding output value.

Step 7: Repeat the calculation procedure of steps 3 to 6 until all the samples have been classified, and then end a learning batch (Epoch).

Step 8: When all of the cluster classification procedures have reached a converged state, or E _r is the total error rate reaches the target value calculation program is stopped, if the operation is not complete it returns to step 3.

When the extension-type neural network completes the learning process, the cluster type can be identified and classified by the new data volume, so the identification procedure is as follows:

Step 1: Read in the weight value matrix of the extension class neural network.

Step 2: Calculate the initial weight center values for the different clusters.

Step 3: Read the sample to be tested

P _t ={ p _{t 1} ,p _{t 2} ,...,p _tq } (27)

Step 4: Using the extension distance formula of Equation 22, calculate the distance between the test sample and the cluster after learning.

Step 5: Find the value g ^* value so that In addition, the corresponding output node value is set to 1, to determine which cluster category the test sample belongs to.

Step 6: When all test samples are classified, stop the operation program, otherwise skip to step 3.

From the simulation results of the above-mentioned inverter faults, it can be known that the fault category of the inverter and its fault characteristics must be related. Therefore, by the form of the extension matter element, the relationship between each other when the failure occurs can be written as:

R =( N , c , v )=( F _n , FI _n , v ) ......(28)

Where R is the fault element of the frequency converter; N is the fault category of the frequency converter and is represented by the symbol F _n ( n =0,1,...); c is the characteristic of the fault and is symbolized FI _n ( n = 1, 2, ...); and v is the magnitude corresponding to the fault feature. The invention is based on the failure discussion of the frequency converter, and the simulation analysis is performed separately when the power crystal is faultless and the 12 kinds of switches (S11~S34) are faulted, so the 13 fault categories are respectively represented by the following symbols.

F ₀ : The power crystal is not faulty. F ₇ : Power crystal S23 is faulty.

F ₁ : Power crystal S11 is faulty. F ₈ : Power crystal S24 is faulty.

F ₂ : Power crystal S12 is faulty. F ₉ : Power crystal S31 is faulty.

F ₃ : Power crystal S13 is faulty. F ₁₀ : Power crystal S32 is faulty.

F ₄ : Power crystal S14 is faulty. F ₁₁ : Power crystal S33 is faulty.

F ₅ : Power crystal S21 is faulty. F ₁₂ : Power crystal S34 is faulty.

F ₆ : Power crystal S22 is faulty.

The fault characteristics can be expressed separately

FI ₁ and FI ₂ : a characteristic spectrum of the u-arm phase voltage v _uo .

FI ₃ and FI ₄ : a characteristic spectrum of the v-arm phase voltage v _vo .

FI ₅ and FI ₆ : a characteristic spectrum of the w arm phase voltage v _wo .

The proposed three-level inverter fault detection system uses the PSIM software to simulate the fault of each single power crystal switch separately, and selects the operating frequency range of the inverter to be 30 Hz to 90 Hz. In addition, in order to increase the sensitivity of the determination of the characteristic spectrum, the two characteristic spectra are processed in the 29th and 30th equations respectively, and are characterized as the fault type of the inverter.

FI ₁ = FI ₃ = FI ₅ =0.5[( m _f -1)+( m _f +3)] (29)

FI ₂ = FI ₄ = FI ₆ = 0.5[( m _f -1)-( m _f +3)] (30)

In the embodiment, the operating frequency of the frequency converter is in the range of 30 Hz to 90 Hz, and 25 different frequencies and spectral characteristic values under different switching faults are extracted, and a total of 325 data are used as learning materials of the extension type neural network, and established. Three-level inverter fault detection system. In order to verify the effectiveness of the fault detection system of the proposed frequency converter, a total of 468 fault data measured by the simulation are used as the test data of the inverter fault detection system, and the identification results are shown in Table 1. From the table, the feasibility of the inverter fault detection system established by the extension-like neural network can be found.

In addition, Tables 2 to 4 are the spectrum data of each phase voltage when the operating frequency is 35 Hz, 55 Hz and 75 Hz. The test data of Tables 2 to 4 are input to the established inverter fault detection system, and the identification results of Tables 5 to 7 can be obtained. It is found in the table that each fault data can be correctly identified. For example, from Table 6, the ED _{F 2} =2.81985 of the serial number 2 test data is the minimum extension distance value, which is known to belong to the power crystal switch S11 fault, and in the same data, ED _{F 9} = 18.045 is the maximum value. , indicating that this test data has the lowest probability of failure of switch S24. In the same way, the identification results of other pen test data can be analyzed.

Although the present invention has been disclosed in the above embodiments, it is not intended to limit the present invention, and the present invention can be modified and modified without departing from the spirit and scope of the present invention. The scope is subject to the definition of the scope of the patent application attached.

S11～S34. . . Power transistor

110~140. . . step

The above and other objects, features, advantages and embodiments of the present invention will become more apparent and understood.

The first figure is a schematic diagram of the structure of a conventional neutral-clamped three-level inverter.

Fig. 2 is a flow chart showing the steps of a method for detecting a fault of a frequency converter according to an embodiment of the present invention.

Figures 3A to 3C are schematic diagrams of three-phase phase voltage waveforms in a normal state.

The 4A-4C diagram is a schematic diagram of the three-phase phase voltage spectrum under normal conditions.

The 5A to 5C diagram is a schematic diagram of the waveform of the three-phase phase voltage when the switch S11 is faulty.

6A to 6C are schematic diagrams of three-phase phase voltage waveforms when the switch S22 is faulty.

7A to 7C are schematic diagrams of three-phase phase voltage waveforms when the switch S24 fails.

The 8A-8C diagram is a schematic diagram of the three-phase phase voltage waveform when the switch S33 is faulty.

The 9A-9C diagram is a schematic diagram of the three-phase phase voltage spectrum when the switch S33 fails.

Figure 10 is a schematic diagram of the matter space.

Figure 11 is a schematic diagram of the extent of each region of the extension set.

Figure 12A is a schematic diagram of the distance relationship between one point a of the extension set and the interval A ₁ .

Figure 12B is a schematic diagram of the distance relationship between one point a of the extension set and the interval A ₁ .

Figure 13A is a schematic diagram showing the positional relationship between point a and interval A ₁ and interval A ₂ of the extension set, which is the right position value of point a .

Figure 13B is a schematic diagram showing the positional relationship between point a and section A ₁ and section A ₂ of the extension set, which is the left position value of point a .

Figure 14 is a schematic diagram of the elementary correlation function.

Figure 15 is a schematic diagram of the structure of the extension-like neural network architecture.

Figure 16 is a schematic diagram of the extension distance (ED).

FIG. 17A is a schematic diagram of the weight adjustment of the present embodiment, which is before the weight adjustment.

FIG. 17B is a schematic diagram of the weight adjustment of the present embodiment, which is shown after the weight adjustment.

110~140. . . step

Claims (4)

A fault detection method for a frequency converter is used for detecting a fault of any one of the power transistors in the main circuit of the frequency converter, and at least comprising: simulating a characteristic spectrum of the output phase voltage of the inverter when any one of the power transistors is faulty, to establish a fault a matter-element model in which the amplitude of the spectrum at the operating frequency of the inverter output phase voltages ( m _f -1) and ( m _f +3) times is taken as the characteristic spectrum, where m _{f is} the pulse width modulation of the frequency converter Change the frequency modulation index and meet the following conditions: , where f _tri is the triangular carrier frequency of the pulse width modulation of the frequency converter, f _sin is the sine wave frequency of the pulse width modulation of the frequency converter; using an extension type neural network learning process and the fault matter element model Establishing a weight value matrix; establishing a fault judgment system by using the fault matter element model and the weight value matrix; and inputting a data to be tested, and using the fault judgment system to perform an identification process to determine the data to be tested Represents one of the faulty power transistors. The method for detecting a fault of a frequency converter according to claim 1, wherein the learning process of the extension type neural network comprises: establishing an input layer to receive the plurality of learning materials; and establishing a calculation layer to determine the fault element model according to the fault The plurality of learning materials generate a plurality of weight values, thereby establishing the weight value matrix; and establishing an output layer to output a signal representing a faulty power transistor. The method for detecting a fault of a frequency converter according to claim 2, wherein the step of generating the plurality of weight values according to the faulty matter element model and the plurality of learning materials comprises: reading the plurality of learning materials; The model generates a plurality of clusters; calculating an initial weight center of each of the clusters; calculating a plurality of extension distances of each of the plurality of learning materials and the plurality of initial weight centers; and determining a minimum of the plurality of extension distances The value produces the plurality of weight values. The fault detection method of the frequency converter according to claim 3, the identification process includes: reading the data to be tested; reading the weight value matrix to obtain the plurality of clusters and the plurality of weight values; according to the plurality of a weight value, calculating a plurality of weight center values of the plurality of clusters; calculating a plurality of distance values of the test data and the plurality of weight center values; and finding a minimum distance value to summarize the data to be tested Cluster and output.