A fault diagnosis method for IGBT switch tubes and current sensors of a three-phase alternating current motor driving circuit

Abstract

The application provides a fault diagnosis method for IGBT switching tubes and sensors of a three-phase alternating current motor driving circuit, and steps of the method comprise the following steps: S1, defining fault modes and labels of an inverter; S2, transforming three-phase currents measured in a fault state through Park transformation to obtain two-phase rotating coordinate system d axis currents and q axis currents; S3, using FFT Fourier transformation on the obtained q axis currents to obtain fault characteristic data in a frequency domain; and S4, using a TCN time convolution neural model to perform fault diagnosis according to the extracted fault characteristic data in the frequency domain. IGBT The application provides a new simple and convenient Park sensor fault diagnosis method, which transforms obtained three-phase alternating current signals to obtain q axis direct current signal samples, Park the transformed current signals not only eliminate sinusoidal fundamental wave components, which is helpful to feature signal extraction, but also select q axis currents as feature extraction samples to reduce a data set.

Description

Technical Field

[0001] This invention relates to the technical field of fault diagnosis, and more particularly to a fault diagnosis method for IGBT switching transistors and current sensors in a three-phase AC motor drive circuit. Background Technology

[0002] With the rapid development of the power industry and the continuous advancement of power electronics technology, inverters are being used in an increasing number of industries. As a crucial component of power electronic systems, inverter circuit failures can lead to system malfunctions, performance degradation, and in severe cases, system shutdown. Therefore, accurate and rapid fault diagnosis after an inverter malfunctions is of paramount importance.

[0003] Inverter fault modes are mainly divided into short-circuit faults and open-circuit faults. Short-circuit faults in the power switch are generally caused by overvoltage, overheating, or incorrect commands. Once this occurs, an overcurrent may be generated within a very short time, damaging the components. Therefore, in the circuit design process, a fast-acting fuse is added to quickly cut off the current and convert the short-circuit fault into an open-circuit fault. Unlike short-circuit faults, when an open-circuit fault occurs in the switching device, although the system can still operate normally for a short time, if it is not addressed for a long time, it may lead to secondary faults in the system, causing system shutdown and resulting in huge economic losses.

[0004] In inverters, besides IGBT (Instrument Block Transistor) failures, related sensor failures are also a common type of fault. Compared to open-circuit faults in power switches, sensor failures not only distort the measurement signal but also quickly generate overcurrent, leading to more serious damage to other components. Measurements show that current sensor faults can be categorized as: disconnection faults, jamming faults, and gain faults.

[0005] Over the decades of research on three-phase inverter fault diagnosis, diagnostic methods can generally be divided into two main categories: those based on mathematical models and those based on data-driven approaches. Mathematical model methods primarily include observer methods, parameter estimation techniques, and equivalent space methods. For example, the paper "Current Sensor Fault-Tolerant Control for Encoderless IPMSM Drives Based on Current Space Vector Error Reconstruction" establishes a single-phase current sensor-based encoderless method for fault diagnosis and fault-tolerant control of permanent magnet synchronous motors based on a sliding mode observer. The paper "Open Circuit Fault Diagnosis of Midpoint Clamped Three-Level Grid-Connected Inverters Based on Adaptive Sliding Mode Observer" establishes an adaptive sliding mode observer for open circuit fault diagnosis of three-level grid-connected inverters. Data-driven methods are mainly divided into signal processing methods, knowledge-based methods, and artificial intelligence techniques. Signal processing methods have been widely used in fault diagnosis and can generally be divided into voltage signal-based and current signal-based methods. As can be seen from the above, fault diagnosis of IGBT switching devices and sensors in inverters is generally performed separately, neglecting to consider the coupling relationship between the two types of faults. This can lead to misdiagnosis of switching device faults when sensors malfunction. A method that can simultaneously diagnose faults in both switching devices and sensors would avoid such misdiagnosis.

[0006] To address the aforementioned issues, numerous scholars have conducted research in recent years. One such method, proposed in "A Robust Observer-Based Method for IGBTs and Current Sensors Fault Diagnosis in Voltage-Source Inverters of PMSM Drives," diagnosed IGBT and current sensor faults in PMSM drives. This method distinguishes between IGBT and sensor faults by verifying whether the sum of the three-phase currents is zero, and then locates the fault by obtaining a residual threshold. However, this method cannot differentiate between fault types. Another method, "A New Approach for Current Sensor Fault Diagnosis in PMSG Drives for Wind Energy Conversion Systems," proposes a different approach from traditional state-transformer-based current sensor fault diagnosis methods. This method requires a system model and corresponding parameters, using only measured phase currents. However, this method cannot distinguish between specific fault types of sensors and faulty switches. The paper "An Online Data-Driven Method for Simultaneous Diagnosis of IGBT and Current Sensor Fault of Three-Phase PWM Inverter in Induction Motor Drives" proposes an online data-driven method for diagnosing IGBT open-circuit faults and current sensor faults in inverters. However, it is difficult to find the optimal hidden layer weights and output layer weights required for training the RVFL network in this diagnostic method. Summary of the Invention

[0007] To address the technical problem of diagnosing and locating simultaneous faults in IGBT switching transistors and current sensors in three-phase AC motors, this invention proposes a fault diagnosis method for IGBT switching transistors and current sensors in three-phase AC motor drive circuits. This method reduces the amount of data, improves noise immunity, and highlights fault characteristics, facilitating accurate fault detection and diagnosis.

[0008] To achieve the above objectives, the technical solution of the present invention is implemented as follows: a fault diagnosis method for IGBT switching transistors and sensors in a three-phase AC motor drive circuit, comprising the following steps:

[0009] S1: Define the inverter's fault modes and labels;

[0010] S2: The measured three-phase current under fault conditions is transformed by Park transformation to obtain the d-axis current and q-axis current in a two-phase rotating coordinate system;

[0011] S3: Use FFT Fourier transform on the obtained q-axis current to obtain fault characteristic data in the frequency domain;

[0012] S4: Using the TCN temporal convolutional neural model, fault diagnosis is performed based on the extracted fault features in the frequency domain.

[0013] The method for defining the fault mode and tag of the inverter in step S1 is as follows: The three-phase inverter diagnoses single-tube and dual-tube faults, and the fault type is:

[0014]

[0015] Current sensor faults are generally classified into three types: jamming fault, gain fault, and disconnection fault. The definitions of jamming fault, gain fault, and disconnection fault for current sensors are as follows:

[0016]

[0017] Where i(t) is the output current, i n (t) represents the output current under normal operating conditions, C1 is a constant, C2 is a constant less than 0.8 or greater than 1.2, and t1 is the time when the fault occurs;

[0018] There are 21 types of IGBT faults and 9 types of current sensor faults, which are represented by different numbers:

[0019]

[0020]

[0021] The method for obtaining the d-axis and q-axis currents in the two-phase rotating coordinate system described in step S2 is as follows:

[0022] The process of converting the measured three-phase current in the natural coordinate system abc to the two-phase current in the two-phase stationary coordinate system αβ is called the Clark transformation.

[0023] The Clark transform formula is:

[0024]

[0025] Among them, i α with i β These represent the currents in the two-phase stationary coordinate systems along the α and β axes, respectively. a i b with i c These are the corresponding three-phase currents;

[0026] Obtained using Clark transform The calculation yields:

[0027]

[0028] Among them, i d with i q These are the d-axis and q-axis currents of the corresponding two-phase rotating coordinate systems, respectively, and θ is the angle between the d-axis and the α-axis.

[0029] By combining the formulas for the current sensor in case of jamming fault and the current sensor in case of gain fault, the two-phase currents in the d and q axes can be obtained directly from the natural coordinate system abc.

[0030] The combined formula of the jamming fault current sensor formula and the gain fault current sensor formula is as follows:

[0031]

[0032] The method using FFT Fourier transform described in step S3 is as follows:

[0033] Based on the q-axis variable, the DFT is defined as:

[0034]

[0035] in, Let x(n) be the time-domain input signal, and let X(k) be the relative amplitude of the N discrete frequency points corresponding to x(n). The FFT is obtained by decomposing the N-point DFT into two N / 2-point DFTs, i.e., (N / 2)² + (N / 2)² = N / 2, and continuing to decompose.

[0036] Decompose x(n) into the sum of two sequences: even numbers and odd numbers.

[0037] x(n) = x1(n) + x2(n)

[0038] FFT is defined as:

[0039]

[0040] Based on the scaling property of the rotation factor, further calculation yields:

[0041]

[0042] Where x1(n) and x2(n) correspond to the even and odd sequences of x(n) respectively, and X1(k) and X2(k) are the DFTs of x1(n) and x2(n) respectively. For the corresponding even number x1(n) or odd number x2(n);

[0043] After being transformed to the frequency domain by FFT, the mean, standard deviation, skewness, and root mean square are used as fault characteristics in the frequency domain.

[0044] The method for fault diagnosis based on fault feature data extracted in the frequency domain described in step S4 is as follows: the fault feature data obtained in step S3 is classified and numbered according to the fault type, and then the numbered fault features are input into the TCN temporal convolutional neural model. The TCN temporal convolutional neural model is used to diagnose faults based on the fault feature data extracted in the frequency domain and to classify the faults.

[0045] The TCN temporal convolutional neural model described in step S4 includes at least one residual module. The residual module comprises a first dilated causal convolutional layer, a second dilated causal convolutional layer, a first normalization, a second normalization, a first activation function, a second activation function, a first regularization, a second regularization, and a 1*1 convolutional layer. The input of the first dilated causal convolutional layer is connected to the input of the convolutional layer. The output of the first dilated causal convolutional layer, the first normalization, the first activation function, the first regularization, the second dilated causal convolutional layer, the second normalization, the second activation function, and the second regularization are connected in sequence. The output of the second regularization is connected to the output of the convolutional layer.

[0046] Both the first and second activation functions are ReLU activation functions, and both the first and second regularizations are droupout regularizations.

[0047] The method for fault diagnosis based on fault feature data extracted in the frequency domain, as described in step S4, is as follows:

[0048] Fault characteristics should be used as input data into the model:

[0049] o = Activation(x + F(x));

[0050] In the above formula, Activation is the activation function, where x is the input data and F(x) is the output after the convolutional layer;

[0051]

[0052]

[0053] in, This is the output of the i-th residual module. This is the output of the (i-1)th residual module. This is represented as a batch normalization process, where γ and β are the training parameters for batch normalization, and a l This represents the output after batch normalization of the l-th layer and the activation function f. Data for batch normalization.

[0054] The method for classifying faults is as follows: using the Softmax function to classify faults based on their characteristics.

[0055] The beneficial effects of this invention are as follows: This application proposes a novel and simple method for fault diagnosis of IGBTs and sensors. This method obtains a q-axis DC current signal sample by performing a Park transform on the acquired three-phase AC current signal. The Park-transformed current signal not only eliminates the sinusoidal fundamental component, which is helpful for feature signal extraction, but also reduces the dataset by selecting the q-axis current as the feature extraction sample. Finally, the feature is extracted by utilizing the salient feature extraction capability of FFT. Using FFT for fault feature extraction not only improves the calculation speed but also enhances the signal's anti-interference capability by converting the signal to the frequency domain. For fault diagnosis, a TCN temporal convolutional neural network is used, which allows for efficient parallel computation, ensures the same input and output length, and does not prematurely leak future data, to identify and classify fault types. By combining causal convolution and dilated convolution to form a dilated causal convolutional layer, not only is the input and output length the same, but the output result is also only affected by previously occurring elements. By using residual block stacking connections, cross-layer connections can be achieved for information transmission in the network. This method provides good diagnostic results for open-circuit faults in switching transistors, stuck faults in sensors, gain faults, and disconnection faults in three-phase inverters. Compared with other methods, this method has better performance in fault diagnosis. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 This is a flowchart of the method of the present invention.

[0058] Figure 2 This is a diagram of the residual module of the invention.

[0059] Figure 3 This is a TCN network structure.

[0060] Figure 4 The curve shows the recognition accuracy of the TCN model in the experiment.

[0061] Figure 5 This is the model loss curve used in the experimental verification.

[0062] Figure 6 The following are comparison charts showing the experimental verification results: (a) shows the results for the training set; (b) shows the results for the test set. Detailed Implementation

[0063] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0064] like Figure 1 As shown, the fault diagnosis method for the IGBT switching transistor and current sensor in a three-phase AC motor drive circuit includes the following steps:

[0065] S1: Define the inverter's fault modes and labels.

[0066] The structural topology of a three-phase SPWM pulse width modulation inverter is as follows: Figure 1 As shown, each phase of the three-phase inverter has two power switches with anti-parallel diodes and a current sensor.

[0067] When switch T1 experiences an open-circuit fault, even if a conduction signal is applied to T1, the upper bridge arm cannot conduct properly due to T1's open-circuit state, resulting in no positive voltage output in phase A. Only when switch T4 is turned on will phase A have a negative voltage output. Therefore, from the moment T1 fails, phase A only has a negative current output, with no positive current. If both switches T1 and T4 fail simultaneously, phase A will have no output current.

[0068] In three-phase inverters, short-circuit faults are typically converted to open-circuit faults using fuses. Faults involving more than two switching transistors are very rare, and the ideal repair method is component replacement. Therefore, this invention only diagnoses single-transistor and dual-transistor faults. Fault types are shown in Table 1 below:

[0069] Table 1 IGBT Fault Types

[0070]

[0071] To meet safety requirements, inverters typically contain three current sensors. Based on measurement results, current sensor faults are generally categorized into three types: jamming, gain fault, and disconnection fault. When a sensor experiences a jamming fault, it can not only cause current distortion but may also lead to circuit overload, potentially causing fires, short circuits, and other safety hazards. A gain fault can prevent the system from accurately identifying the current magnitude, leading to system malfunction. A drift fault can cause the sensor output signal to gradually deviate from the true value, making accurate system control impossible. The definitions of jamming, gain fault, and disconnection faults for current sensors are as follows:

[0072]

[0073] Where i(t) is the output current, i n (t) represents the output current under normal operating conditions, C1 represents a constant, C2 represents a constant less than 0.8 or greater than 1.2, and t1 represents the time when the fault occurs.

[0074] There are 21 types of faults in the IGBT switching transistor and 9 types in the current sensor. Fault diagnosis of the inverter is essentially the identification of its operating status.

[0075] Different inverter faults result in different output currents. To accurately diagnose and locate faults, this application uses different numbers to represent IGBT faults and current sensor faults, as shown in Table 2. This includes not only normal operation, single IGBT faults, and dual IGBT faults, but also sensor stagnation, gain issues, and disconnection faults.

[0076] Table 2 Inverter Fault Labels

[0077]

[0078] S2: The measured three-phase current under fault conditions is transformed by Park transformation to obtain the d-axis and q-axis currents in a two-phase rotating coordinate system.

[0079] The Parker transformation is a coordinate transformation method for transforming three-phase currents from the abc coordinate system to the d and q axes. The principle is as follows: First, the measured three-phase currents in the natural coordinate system abc are transformed into two-phase currents in the two-phase stationary coordinate system αβ. This process is called the Clarker transformation.

[0080] The Clark transform formula is:

[0081]

[0082] Among them, i α with i βThese represent the currents in the two-phase stationary coordinate systems along the α and β axes, respectively. a i b with i c These are the corresponding three-phase currents.

[0083] The essence of the Park transformation is to transform the stationary αβ coordinate system into a rotating d, q coordinate system, using the Clark transformation. The calculation yields:

[0084]

[0085] Among them, i d with i q Let ...

[0086] By combining the formulas for the current sensor in case of jamming fault and the current sensor in case of gain fault, we can obtain the two-phase current formulas directly in the natural coordinate system abc, which are given by the formulas for the d and q axes.

[0087]

[0088] The collected three-phase current is converted into two-phase current by Park transformation, which reduces the data set, effectively monitors changes in AC quantity, and eliminates the sinusoidal fundamental component, which helps to extract fault characteristics.

[0089] S3: Use FFT Fourier transform on the obtained q-axis current to obtain fault characteristic data in the frequency domain.

[0090] The Park transform aligns the d-axis with the α-axis. Since the q-axis current exhibits different characteristics under different fault conditions, only the q-axis current is used for fault feature extraction. This significantly reduces the dataset size and computation time. The FFT, a fast algorithm adapted from the Discrete Fourier Transform (DFT), transforms the signal from the time domain to the frequency domain, improving its anti-interference capability, effectively extracting feature information, and removing white noise and high-frequency components.

[0091] DFT is defined as:

[0092]

[0093] The Faster FFT (FFT) decomposes an N-point Discrete Fourier Transform (DFT) into two N / 2-point DFTs. This halves the total computation of the two N / 2-point DFTs, i.e., (N / 2)² + (N / 2)² = N / 2. This decomposition can be continued, further subdividing the N / 2 points into N / 4-point DFTs, and so on. The more values ​​being computed, the greater the reduction in computational cost.

[0094] FFT is defined as:

[0095]

[0096] Where x1(n) and x2(n) correspond to the even and odd sequences of x(n) respectively, and X1(k) and X2(k) are the DFTs of x1(n) and x2(n) respectively. For the corresponding even number x1(n) or odd number x2(n);

[0097] After converting to the frequency domain using FFT, the fault characteristics in the frequency domain are obtained using formulas such as mean, standard deviation, skewness, and root mean square.

[0098] average value:

[0099] Standard deviation:

[0100] Skewness:

[0101] S4: Perform fault diagnosis based on the fault features extracted in the frequency domain.

[0102] Specifically, the fault feature data obtained from S3 is numbered and then input into the TCN temporal convolutional neural model. The TCN temporal convolutional neural model is used to perform fault diagnosis based on the fault feature data extracted in the frequency domain.

[0103] Temporal Convolutional Neural Networks (TCNs) were chosen for fault diagnosis. The TCN model has advantages such as efficient parallel computation, equal length of input and output sequences, and no gradient explosion or gradient vanishing.

[0104] Unlike traditional convolutional neural networks, to ensure that the input and output sequences are of equal length and that future data is not prematurely leaked, temporal convolutional neural networks need to combine 1D-FCN with causal convolution as dilated causal convolution. For example... Figure 2 As shown, the model employs a multi-layered stacked residual block approach instead of the previous layer-by-layer connections. Each residual block comprises a first dilated causal convolutional layer, a second dilated causal convolutional layer, a first normalization, a second normalization, a first activation function, a second activation function, a first regularization, a second regularization, and a 1x1 convolutional layer. The input of the first dilated causal convolutional layer is connected to the input of the convolutional layer. The output of the first dilated causal convolutional layer, the first normalization, the first activation function, the first regularization, the second dilated causal convolutional layer, the second normalization, the second activation function, and the second regularization are sequentially connected. The output of the second regularization is connected to the output of the convolutional layer.

[0105] By utilizing residual modules, the output is made relevant only to previously occurring factors without losing information. Stacking residual blocks ensures that the input and output have the same number of channels and addresses the vanishing and exploding gradient problems in the model. Adding ReLU activation and droupout regularization to the residual blocks simplifies computational complexity and reduces overfitting. Fault features should be used as input data to the model.

[0106] o = Activation(x + F(x));

[0107] In the above formula, Activation is the activation function, where x is the input data, F(x) is the output after the convolutional layer, and the residual module is as follows: Figure 2 As shown.

[0108]

[0109] in, This is the output of the i-th residual module. This is the output of the (i-1)th residual module. This is represented as a batch normalization process, where γ and β are the training parameters for batch normalization. l This represents the output after batch normalization of the l-th layer and the activation function f.

[0110] This method simplifies fault feature extraction, improves noise immunity, and highlights fault characteristics, facilitating accurate fault detection and diagnosis. The simplified feature extraction and enhanced noise immunity are achieved through the Park transform and FFT transform components. These are inherent characteristics of the Park transform and FFT transform themselves.

[0111] This application utilizes MATLAB software to program the TCN fault diagnosis model. The program flowchart is as follows: Figure 3 As shown.

[0112] like Figure 4 The diagram shows the network structure of the TCN model. This TCN model consists of six stacked residual modules. Since the ReLU activation function not only reduces the computational cost of TCN but also enables fast model convergence, it is used in the module connections. A Softmax function is added to the final output layer of the TCN model to classify faults. The Adam algorithm is used to optimize gradients during the training process of the TCN model. For determining the consistency between the model's input and output objectives, this application selects the cross-entropy loss function.

[0113] The ReLU activation function is as follows:

[0114]

[0115] Cross-entropy loss function:

[0116]

[0117] Wherein, P(y i Therefore, it represents the probability distribution of the sample dataset. This represents the probability distribution of the prediction dataset.

[0118] Fault features have already been extracted in the previous Park transform and FFT transform. This final step in the model diagnosis uses the Softmax function to filter out the maximum probability for each category corresponding to each input data point. The results are shown in Table 2. In Table 2, each fault is encoded according to its type and location. The type and category determined by the final Softmax function can be used to determine the location and type of the fault. Finally, the cross-entropy loss function is used to evaluate the accuracy of the model's diagnosis.

[0119] A three-phase inverter circuit model was built and samples were collected to verify the method proposed in this application. When collecting samples, 2000 sampling points were collected every 0.02 seconds as one sample. A total of 60 samples were collected for each operating state, for a total of 480 samples. After the collected samples were shuffled, they were divided into training set and test set in a ratio of 7:3.

[0120] Figure 5 The curves show the model accuracy and loss rate during training with the TCN model. It can be seen that the recognition accuracy and loss rate of the TCN model on the training set tend to stabilize after 40 iterations. The recognition accuracy eventually remains at 100%, while the loss function eventually approaches zero.

[0121] By inputting the training set and test set into the already trained and saved TCN model, we can obtain the following results: Figure 5 , Figure 6 The test results showed that the accuracy of the test set reached 100%, and the accuracy of the training set reached 99.375%. This indicates that the model can not only accurately identify the trained data, but also identify device faults well in unseen data, demonstrating good generalization ability. At the same time, according to the confusion matrix results, it can be seen that the TCN model can accurately identify switch tube faults or sensor faults and the location of each fault type, proving its accuracy and effectiveness in the diagnosis of switch tube and sensor faults.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for fault diagnosis of IGBT switching transistors and sensors in a three-phase AC motor drive circuit, characterized in that the steps are as follows: include: S1: Define the inverter's fault modes and labels; S2: The measured three-phase currents under fault conditions are transformed sequentially through Clark transformation and Park transformation to obtain the d-axis current and q-axis current in a two-phase rotating coordinate system; S3: Use FFT Fourier transform on the obtained q-axis current to obtain fault characteristic data in the frequency domain; The fault characteristics in the frequency domain include mean, standard deviation, skewness, and root mean square. S4: Using the TCN temporal convolutional neural model, fault diagnosis is performed based on the extracted fault features in the frequency domain. The TCN temporal convolutional neural model described in step S4 includes at least one residual module. The residual module comprises a first dilated causal convolutional layer, a second dilated causal convolutional layer, a first normalization, a second normalization, a first activation function, a second activation function, a first regularization, a second regularization, and a 1*1 convolutional layer. The input of the first dilated causal convolutional layer is connected to the input of the convolutional layer. The output of the first dilated causal convolutional layer, the first normalization, the first activation function, the first regularization, the second dilated causal convolutional layer, the second normalization, the second activation function, and the second regularization are connected in sequence. The output of the second regularization is connected to the output of the convolutional layer. The TCN temporal convolutional neural model stacks 6 layers of residual modules. The first and second activation functions are both ReLU activation functions, and the first and second regularizations are both droupout regularizations. There are 21 types of faults in the IGBT switching transistor and 9 types of faults in the current sensor, which are represented by different numbers. The faults of the current sensor include jamming fault, gain fault, and disconnection fault.

2. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 1, characterized in that: The method for defining the fault modes and tags of the inverter in step S1 is as follows: the inverter diagnoses single-transistor and dual-transistor faults, and the fault types are: ； Current sensor faults are generally classified into three types: jamming fault, gain fault, and disconnection fault. The definitions of jamming fault, gain fault, and disconnection fault for current sensors are as follows: ； ； ； Where i(t) is the output current, i n (t) represents the output current under normal operating conditions, C1 is a constant, C2 is a constant less than 0.8 or greater than 1.2, and t1 is the time when the fault occurs; There are 21 types of IGBT faults and 9 types of current sensor faults, which are represented by different numbers: 。 3. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 2, characterized in that: The method for obtaining the d-axis and q-axis currents in the two-phase rotating coordinate system described in step S2 is as follows: The process of converting the measured three-phase current in the natural coordinate system abc to the two-phase current in the two-phase stationary coordinate system αβ is called the Clark transformation. The Clark transform formula is: ； in, and Let be the currents in the stationary coordinate systems along the α and β axes, respectively. , and These are the corresponding three-phase currents; Obtained using Clark transform The calculation yields: ； in, and These are the d-axis and q-axis currents of the corresponding two-phase rotating coordinate systems, respectively. The angle between the d-axis and the α-axis; By combining the formulas for the current sensor in case of jamming fault and the current sensor in case of gain fault, the two-phase currents in the d and q axes can be obtained directly from the natural coordinate system abc.

4. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 3, characterized in that: The combined formula of the jamming fault current sensor formula and the gain fault current sensor formula is as follows: 。 5. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 3 or 4, characterized in that: The method using FFT Fourier transform described in step S3 is as follows: Based on the q-axis variable, the DFT is defined as: ； in, For rotation factor, The input time-domain signal, It corresponds to The relative amplitudes of N discrete frequency points; FFT is achieved by decomposing the N-point DFT into two N / 2-point DFTs, i.e., (N / 2)² + (N / 2)² = N / 2, and continuing to decompose; Will It can be decomposed into the sum of two sequences: even numbers and odd numbers. ； FFT is defined as: ； Based on the scaling property of the rotation factor, further calculation yields: ； in, and They correspond to as Two sequences, even and odd numbers, and They are respectively and Point DFT, For the corresponding even number or odd number Rotation factor at time; After being transformed to the frequency domain by FFT, the mean, standard deviation, skewness, and root mean square are used as fault characteristics in the frequency domain.

6. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 5, characterized in that: The method for fault diagnosis based on fault feature data extracted in the frequency domain described in step S4 is as follows: the fault feature data obtained in step S3 is classified and numbered according to the fault type, and then the numbered fault features are input into the TCN temporal convolutional neural model. The TCN temporal convolutional neural model is used to diagnose faults based on the fault feature data extracted in the frequency domain and to classify the faults.

7. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 6, characterized in that: Both the first and second activation functions are ReLU activation functions, and both the first and second regularizations are droupout regularizations.

8. The fault diagnosis method for the IGBT switching transistor and sensor in the three-phase AC motor drive circuit according to claim 7, characterized in that: The method for fault diagnosis based on fault feature data extracted in the frequency domain, as described in step S4, is as follows: Fault characteristics should be used as input data into the model: ； In the above formula, Activation is the activation function, where For input data, This is the output after passing through the convolutional layer; ； ； ； in, This is the output of the i-th residual module. This is the output of the (i-1)th residual module. This is represented as a batch normalization process. β are batch-normalized training parameters. This represents the output after batch normalization of the l-th layer and the activation function f. For batch normalization of data.

9. The fault diagnosis method for the IGBT switching transistor and sensor of the three-phase AC motor drive circuit according to any one of claims 6-8, characterized in that: The method for classifying faults is as follows: using the Softmax function to classify faults based on their characteristics.

Images