A method for diagnosing a fault of a multi-level inverter power device

Abstract

The application provides a kind of multilevel inverter power device fault diagnosis method, it is characterized in that, using feedforward convolution-pooling network and auxiliary neural network to the operating mode of each phase of multilevel inverter is identified, obtains diagnostic result, wherein, the weight and bias parameter learned by auxiliary neural network is migrated to the convolution kernel of feedforward convolution-pooling network, to share the learning ability of auxiliary neural network, make feedforward convolution-pooling network have the feature identification of the operating mode of multilevel inverter.Using the diagnostic method proposed in the application, the inherent problems of traditional circuit model method and data-driven method, i.e., the contradiction between model dependency and diagnostic speed, are effectively improved.The diagnostic network proposed in the application has high diagnostic precision while reducing the demand for training samples and speeding up signal processing by sharing learning ability and simplifying convolutional neural network structure, which is helpful for online fault diagnosis.

Description

TECHNICAL FIELD

[0001] The application relates to a multi-level inverter power device fault diagnosis method based on a shared learning type neural network and belongs to the technical field of power electronic equipment fault diagnosis. BACKGROUND

[0002] The multi-level inverter has been widely concerned and applied in industrial production due to the advantages of high efficiency, good harmonic characteristics and low power device stress. However, compared with the traditional two-level inverter, the multi-level inverter contains more power semiconductor devices, Figure 1 A typical topology of a multi-level inverter is a T-type three-level inverter. According to an industrial statistical report [1], the power semiconductor switch is one of the most vulnerable components in the power electronic system, and the multi-level inverter usually increases the number of power switches to improve the number of output levels, so that the fault will have a great threat to the reliable operation of the inverter. As a common form of power semiconductor devices, the open circuit fault will not cause transient mutation of electrical quantities in a short time, so that the fault characteristics are difficult to identify. However, if it is not handled in time, the open circuit fault will lead to distortion of the inverter output and increase of electrical stress, cause secondary damage to other parts of the circuit, and eventually lead to catastrophic failure. Therefore, it is of great significance to design a fast and reliable power semiconductor switch open circuit fault diagnosis method to ensure system safety.

[0003] At present, the fault diagnosis methods for power semiconductor switches mainly include the circuit model-based diagnosis method and the data-driven diagnosis method. The former accurately models the target circuit, analyzes the voltage and current signals collected by the sensor using the circuit model, evaluates the analysis results, and deduces the type and position of the fault according to the corresponding diagnosis rules, such as the observed current residual method (patent application number 201310743597.4), the output voltage average method [2], the voltage residual method between bridge arms [3], and the line voltage envelope method (patent application number 201610286526.X). Figure 2 The implementation principle diagram of this method is shown. When the circuit topology is different, the circuit model and the diagnosis rules to be established are different. With the increase of the number of power switches, the fault modes of the circuit increase, which makes the design of the logic framework of the diagnosis method more complex. In addition, this method is sensitive to circuit parameters. The above problems limit the practical application of the circuit model method.

[0004] Data-driven diagnostic methods do not require detailed analysis of the inverter's working mechanism or model. Instead, they assess the system's operating status by extracting feature information from the collected data. Examples include artificial neural networks (patent application number 201610438466.9, [4]), extreme learning machines [5], random forests [6], and vector machines (patent application number 201410005022.7). Artificial neural networks (ANNs), as a nonlinear model with multiple inputs and multiple outputs, have shown good performance in simulating the characteristic relationship between the input and output of inverters and have received widespread attention and application. Their diagnostic principle diagram is shown in the figure below. Figure 3 As shown. Many studies have also combined signal analysis strategies with ANN to further improve diagnostic accuracy, such as the ANN diagnostic model based on Fourier transform [7] and the ANN diagnostic method based on discrete wavelet transform [8]. However, the above methods usually require extracting fault features from a large amount of data, which is computationally intensive. This not only places high demands on the performance of the circuit controller, but also prolongs the diagnostic time, making it difficult to achieve rapid online diagnosis.

[0005] References:

[0006] [1] S.Yang, A.Bryant, P.Mawby, D.Xiang, L.Ran, and P.Tavner, "An industry-based survey of reliability in power electronic converters," IEEETrans.Ind.Appl., vol.47, no.3, pp.1441–1451, May / Jun.2011.

[0007] [2] Z.Li, H.Ma, Z.Bai, Y.Wang, and B.Wang, "Fast transistor open-circuitfaults diagnosis in grid-tied three-phase VSIs based on average bridge armpole-to-pole voltages and error-adaptive thresholds," IEEE Trans.PowerElectron., vol.33, no.9, pp.8040–8051, Sep.2018.

[0008] [3] L. M. A. Caseiro and A. M. S. Mendes, “Real-time IGBT open-circuit fault diagnosis in three-level neutral-point-clamped voltage-source rectifiers based on instant voltage error,” IEEE Trans. Ind. Electron., vol. 62, no. 3, pp. 1669-1678, Mar. 2015.

[0009] [4] S. S. Moosavi, A. N’Diaye, A. Djerdir, et al. Artificial neural network based fault diagnosis in the AC-DC converter of the power supply of series hybrid electric vehicle. IET Electr. Syst. Transp., vol. 6, no. 2, pp. 96-106, Jun. 2016.

[0010] [5] Xia Y, Xu Y. A transferrable data-driven method for IGBT open-circuit fault diagnosis in three-phase inverters [J]. IEEE Trans. Power Electron., 2021, 36(12): 13478-13488.

[0011] [6] Kou L, Liu C, Cai G W, et al. Fault diagnosis for open-circuit faults in NPC inverter based on knowledge-driven and data-driven approaches [J] IET Power Electron., 2020, 13(6): 1236-1245.

[0012] [7] Han P, He X, Ren H, et al. Fault diagnosis and system reconfiguration strategy of single-phase three level neutral-point-clamped cascaded inverter [J]. IEEE Transactions on Industry Applications, 2019, 55(4): 3863-3876.

[0013] [8] Zhang H, An Q, Hu Z, et al. Fault detection wavelet fractal method of circuit of three-phase bridge rectifier [C]. 2010 International Conference on Intelligent System Design and Engineering Application. IEEE, 2010: 725-729.

[0014] [9] Gong W, Chen H, Zhang Z, et al. A data-driven-based fault diagnosis approach for electrical power DC-DC inverter by using modified convolutional neural network with global average pooling and 2-D feature image [J]. IEEE Access, 2020, 8: 73677-73697.

[0015]

[10] M. Lin, Q. Chen, and S. C. Yan, “Network in network,” in Proc. Int. Conf. Learn. Represent., 2014, pp. 1-10. SUMMARY

[0016] The technical problem this invention aims to solve is that while circuit model-based diagnostic methods offer good accuracy and speed, they require establishing precise models and parameters of the inverter system. Furthermore, they necessitate remodeling for different topologies, resulting in poor versatility. Data-driven methods, on the other hand, have lower accuracy requirements for circuit models and parameters but require extensive data computation and analysis, leading to longer diagnostic times and higher computational resource consumption, making them unsuitable for online fault monitoring of inverter systems.

[0017] To address the aforementioned technical problems, the present invention provides a method for diagnosing power device faults in a multi-level inverter. The method is characterized by using a feedforward convolutional-pooling network and an auxiliary neural network to identify the operating modes of each phase of the multi-level inverter and obtain diagnostic results. Specifically, the weights and bias parameters learned by the auxiliary neural network are transferred to the convolution kernels of the feedforward convolutional-pooling network to share the learning capabilities of the auxiliary neural network, thereby enabling the feedforward convolutional-pooling network to identify the characteristics of the multi-level inverter's operating modes.

[0018] Preferably, the data input to the feedforward convolutional-pooling network and the auxiliary neural network is h, which is obtained by reconstructing the three-phase one-dimensional time-series signal transformation based on sensor acquisition. d ×3s d Two-dimensional signal matrix M fm .

[0019] Preferably, the two-dimensional signal matrix M is obtained using the following method. fm :

[0020] Step 1: Use sensors to acquire the three-phase output current signal i of the multilevel inverter. X After performing Park transform, the current amplitude is normalized. Then, the normalized three-phase current signal is attenuated and accelerated to obtain a three-phase one-dimensional timing diagnostic signal, where X = A, B, and C represent phases A, B, and C, respectively. The three-phase one-dimensional timing diagnostic signal is a one-dimensional timing signal sequence corresponding to the sampling time. The one-dimensional timing diagnostic signal of phase X is represented as {i X.N3 [1], i X.N3 [2],...,i X.N3 [n],...},i X.N3 [n] represents the diagnostic signal of phase X at the nth sampling time;

[0021] Step 2: Set the size to 1×N DW A sliding detection window is used to collect one-dimensional timing diagnostic signals for each phase. Therefore, the sliding detection window of the three phases of the multi-level inverter collects 3N signals each time. DW A diagnostic signal;

[0022] Step 3, reconstructing the collected three-phase one-dimensional time sequence signal into a two-dimensional signal matrix M according to the following formula d ×3s d fm fm The input feature map of the feedforward convolution-pooling network and the auxiliary neural network is:

[0023]

[0024] wherein k = 1: h d , l = 1: s d , and h d · s d = N DW ; x k，l represents the data in the kth row and the lth column of the two-dimensional signal matrix M fm .

[0025] Preferably, the auxiliary neural network comprises an input layer, a hidden layer and an output layer, the number of neurons r of the input layer satisfies the relationship r = 3s fm with the size of the two-dimensional signal matrix M d ; after the input feature map is input into the input layer, the neurons of the hidden layer and the output layer process the input feature map respectively, and then the input feature map is input into the Softmax layer for classification to obtain the final running mode classification result, wherein the processing of the hidden layer and the output layer on the signal is respectively represented by the following formula.

[0026]

[0027]

[0028] wherein z j represents the output information of the jth hidden layer neuron; f hid () is the activation function of the hidden layer; y j represents the information received by the jth hidden layer neuron; is the weight value of the information transmitted from the nth input layer neuron to the jth hidden layer neuron; x n represents the information received by the nth input layer neuron at a specific moment; a j is the bias of the neuron receiving information; o l represents the output information of the lth output layer neuron; f out () is the activation function of the output layer; is the weight value of the information transmitted from the kth hidden layer neuron to the lth output layer neuron; b l represents the bias of the output layer neuron receiving information.

[0029] ​​Preferably, the feedforward convolution-pooling network comprises two convolution layers and one global pooling layer, the two convolution layers are defined as convolution layer one and convolution layer two respectively, wherein:

[0030] The convolution layer can directly process the input feature map, and the information in the N DW sampling periods is included in the receptive field analysis, and the characteristic information of the multi-level inverter operating state in this time period is comprehensively extracted. The convolution operation process is shown in the following formula:

[0031]

[0032] In the formula, z l represents the information matrix of the lth output feature map; f conv () is the activation function of the convolution layer; w l，k represents the weight matrix of the kth channel in the lth group of convolution kernels; represents the convolution operator; x k represents the information matrix of the kth channel in the input feature map; β l represents the signal bias of the lth output feature map of the convolution layer;

[0033] All features processed by the convolution layer are dimensionally compressed by the global pooling layer, and finally the compressed information is classified by using the Softmax function to obtain the probability result of the input feature map belonging to each operating mode, indicating the operating state of the multi-level inverter.

[0034] Preferably, the shared condition for the feedforward convolution-pooling network to share the learning ability of the auxiliary neural network is:

[0035]

[0036]

[0037] In the formula: represents the jth convolution kernel array of the convolution layer one; represents the lth convolution kernel array of the convolution layer two; represents the weight array of the jth neuron of the hidden layer of the auxiliary network; represents the weight array of the lth neuron of the output layer of the auxiliary network; represents the signal bias of the convolution layer one; represents the signal bias of the convolution layer two; p is the number of neurons of the hidden layer of the auxiliary neural network; f convI represents the activation function of the convolution layer one; f convII represents the activation function of the convolution layer two.

[0038] Preferably, the convolution layer adopts a multi-channel convolution layer, each set of convolution kernel parameters can be shared by all channels, each channel respectively receives the diagnostic signal collected by each phase of the multi-level inverter in the sliding detection window, and the diagnostic signal is filtered to contain features after being processed by the convolution kernel.

[0039] The diagnostic method disclosed in the present application effectively improves the inherent problems of the traditional circuit model method and the data-driven method, i.e., the contradiction between model dependency and diagnostic speed. First, the method disclosed in the present application learns the operating characteristics of the circuit based on a neural network and training data, without analyzing the circuit model and parameters, without additional sensors, and is easy to implement. Compared with the traditional neural network diagnostic method, the diagnostic network disclosed in the present application has high diagnostic accuracy while reducing the demand for training samples and speeding up signal processing, which is helpful for online fault diagnosis. To verify the performance of the method of the present application, the current signals of a T-type three-level inverter are trained and tested. The results show that the diagnostic model disclosed in the present application can achieve better diagnostic performance and effectively improve the reliability of the multi-level inverter. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 Fig. 1 is a schematic diagram of a three-phase T-type three-level inverter topology;

[0041] Figure 2 Fig. 4 is a schematic diagram of a diagnostic method based on a circuit model;

[0042] Figure 3 Fig. 5 is a schematic diagram of a data-driven diagnostic method;

[0043] Figure 4 Fig. 6 illustrates a neural network with shared feature learning capability;

[0044] Figure 5 Fig. 7 is a schematic diagram of the diagnostic method in the present application;

[0045] Figure 6 Fig. 8 illustrates the principle of forming a feature map based on multi-signal sequence reconstruction;

[0046] Figure 7 Fig. 9 is a schematic diagram of convolution operation;

[0047] Figure 8 Fig. 10 is a confusion matrix diagram based on the diagnostic results of the training data. DETAILED DESCRIPTION

[0048] The application will be further described in connection with specific embodiments. It should be understood that these embodiments are only used to illustrate the application and not used to limit the scope of the application. Furthermore, it should be understood that those skilled in the art can make various modifications or changes to the application after reading the content of the application, and these equivalent forms also fall within the scope defined by the appended claims.

[0049] The application designs a neural network architecture for fault diagnosis of a multi-level inverter, as shown in Figure 4 , which consists of a feedforward convolution-pooling network and an auxiliary neural network, and can realize diagnosis and positioning of open-circuit faults of power switches of the inverter. The principle of the fault diagnosis method based on the model is as shown in Figure 5 , and the implementation steps are as follows:

[0050] Step 1, current signal preprocessing

[0051] First, the three-phase output current signals i X (X=A, B, C) of the inverter in normal and fault operation modes are collected by using sensors, and park transformation is performed, and the current amplitude is normalized, as shown in equations (1) and (2).

[0052]

[0053]

[0054] In equations (1) and (2), i d represents the d-axis component of the three-phase current in the park coordinate system; i q represents the q-axis component of the three-phase current in the park coordinate system; ω is the angular velocity of the three-phase sinusoidal current; φ x represents the initial phase angle of phase X, X=A, B, C; i X.N [n] represents the normalized X-phase current value at the nth sampling time; i X [n] represents the X-phase current value at the nth sampling time; i d [n] represents the value of the current component i d at the nth sampling time; i q [n] represents the value of the current component i q at the nth sampling time.

[0055] Then, the normalized three-phase current signals are subjected to an attenuation acceleration process as shown in equation (3). The current signals after the above preprocessing are used as three-phase one-dimensional time sequence diagnosis signals i A.N3 , i B.N3 , i C.N3 .

[0056]

[0057] In equation (3): i X.N3 [n] represents the diagnostic signal of phase X at the nth sampling time.

[0058] Step 2: Diagnostic signal sequence reconstruction

[0059] This invention proposes a multi-signal sequence reconstruction method that converts a three-phase one-dimensional time-series diagnostic signal into a two-dimensional signal matrix, such as... Figure 6 As shown. The three-phase one-dimensional timing diagnostic signal of the inverter is a one-dimensional timing signal sequence corresponding to the sampling time, represented as {i X.N3 [1], i X.N3 [2],...,i X.N3 [n], ...}, where n represents the nth sampling time. First, by setting a sliding window, the one-dimensional time-series diagnostic signals of each phase are collected in batches, for example, using a window with a size of 1×N. DW The sliding detection window collects one-dimensional timing diagnostic signals for each phase, and slides N times on the one-dimensional timing diagnostic signal each time. DW To update the collected signal, the sliding window of the inverter's three phases collects 3N signals each time. DW A diagnostic signal. Then, according to formula (4), the collected three-phase one-dimensional time-series signal is transformed and reconstructed into an h. d ×3s d Two-dimensional signal matrix M fm Also known as the input feature map of a convolutional network.

[0060]

[0061] In equation (4): k = 1: h d , l = 1: s d And satisfy h d ·s d =N DW ;x k，l Represents a two-dimensional signal matrix M fm The data in the k-th row and l-th column.

[0062] Step 3: Training the auxiliary neural network

[0063] The two-dimensional signal matrix M fm The data in each column vector is sent to... Figure 4 The auxiliary neural network is trained on each neuron in its input layer. The auxiliary neural network is a simple three-layer artificial neural network, where the number of neurons r in the input layer is related to the two-dimensional signal matrix M. fm The dimensional relationship satisfies r = 3s dThe input data are processed by neurons of the hidden layer and the output layer respectively, and then input into the Softmax layer for classification.

[0064]

[0065]

[0066] In the formula (6) and (7), z j represents the output information of the jth hidden layer neuron; f hid () is an activation function of the hidden layer; y j represents the information received by the jth hidden layer neuron; is a weight value of the information transmitted from the nth input layer neuron to the jth hidden layer neuron; x n represents the information received by the nth input layer neuron at a specific moment; a j is a bias of the neuron receiving information; o l represents the output information of the lth output layer neuron; f out () is an activation function of the output layer; is a weight value of the information transmitted from the kth hidden layer neuron to the lth output layer neuron; b l represents a bias of the output layer neuron receiving information; n, j, k, l = 1, 2,...

[0067] The training of the auxiliary neural network is based on the evaluation of the error between the expected output and the actual output of the input signal, and the parameters (weight values and the bias a j , b l ) of the auxiliary neural network are continuously adjusted to minimize the error, that is, the input diagnostic signal can correctly correspond to the corresponding operation mode. The process uses a traditional cross-entropy loss function to calculate the output error of the neural network, and based on the error back propagation algorithm (BP algorithm), the weight values and bias parameters of each layer are corrected and updated by differential bias[9].

[0068] Step 4, construction of the convolution-pooling network and migration of the learning parameters

[0069] The convolution-pooling network designed in the present application is shown in Figure 4 , which is formed by the combination of two convolution layers and one global pooling layer. The convolution layer can directly process the two-dimensional signal matrix M fm , and analyze the information in the N DW sampling periods in the receptive field, and comprehensively extract the feature information of the inverter operation state in the time period. The convolution operation process is shown in formula (8).

[0070]

[0071] In formula (8), z l represents the information matrix of the lth output feature map; f conv is the activation function of the convolution layer; w l，k represents the weight matrix of the kth channel in the lth group of convolution kernels; represents the convolution operator; x k represents the information matrix of the kth channel in the input feature map; β l represents the signal bias of the lth output feature map of the convolution layer.

[0072] Figure 7 is the information processing process of a three-channel convolution layer, and each group of convolution kernel parameters can be shared by all channels. Each channel receives the diagnostic signals collected by the inverter phases through a sliding window, and filters the features contained therein through the convolution kernel. All features processed by the convolution layer are compressed in dimension by using a global pooling layer

[10] , that is, each row vector in the feature map is directly averaged. The pooling operation is shown in formula (9). Finally, the compressed information is classified by using a Softmax function to obtain the probability result of the input signal belonging to each working mode, indicating the running state of the inverter.

[0073]

[0074] In formula (9), o n，j represents the information at position (n, j) in the pooled feature map; h p and s p represent the height and width of the pooling window; z k，l represents the data on the kth row and lth column in the feature map obtained after the convolution layer processing.

[0075] The neural network shared learning method proposed in the present application is to migrate the weight and bias parameters learned by the auxiliary neural network to the convolution kernel of the convolution-pooling network, to share the learning ability of the auxiliary neural network, so that the convolution-pooling network also has the feature recognition of the inverter running mode. According to the information processing process of the auxiliary neural network and the convolution network in formulas (6)-(8), the sharing conditions of the learning ability can be summarized as follows:

[0076]

[0077]

[0078] In formulas (10) and (11), z represents the jth convolution kernel array of the convolution layer I; an array of the lth convolution kernel of the convolution layer II; an array of the weight of the jth neuron of the hidden layer of the auxiliary network; an array of the weight of the lth neuron of the output layer of the auxiliary network; a signal bias of the convolution layer I; a signal bias of the convolution layer II; r, p and g are respectively the number of neurons of the input layer, the hidden layer and the output layer of the auxiliary neural network; f convI an activation function of the convolution layer I; f convII an activation function of the convolution layer II.

[0079] To verify the performance of the diagnostic method, the method is applied to Figure 1 the three-phase T-type three-level inverter operation mode detection (including one normal mode and 12 fault modes of corresponding switching tubes) as shown in the figure. Considering the symmetry of the three-phase topology of the inverter, only the diagnostic results of the A phase are analyzed. Figure 4 The neural network model parameters are shown in the following table.

[0080]

[0081] First, the learning ability of the neural network model is trained. The three-phase current of the T-type three-level inverter is collected, the sampling interval is 10 μs, 800 labeled signals are collected for each phase under each working mode, and the three-phase current signal is used as a training sample. 800 training samples can form 50 feature maps. The training set of all working modes contains a total of 10400 samples. Based on the training set samples, the diagnostic signal is obtained and the auxiliary neural network is trained. The trained learning parameters are migrated to the convolution-pooling network, and the operation mode of the A phase of the T-type three-level inverter is identified. The diagnostic results are represented by a confusion matrix as shown in the figure. Figure 8 Although this method has a small amount of misdiagnosis in identifying the fault of the T A3 tube, the diagnostic accuracy can reach 99.7%.

[0082] Then, the performance of the neural network model proposed in the patent is compared with that of the traditional neural network model. The diagnostic accuracy of the inverter in normal and fault conditions is as follows.

[0083]

[0084] ANN in the table is traditional artificial neural network, whose parameters and structure are the same as auxiliary neural network, and CPN is a feedforward convolution-pooling network sharing learning ability. Comparing model I and II with model III and IV, it can be seen that only normalizing the collected current signal, the diagnostic accuracy of network model is low, while using the current preprocessing method proposed in the patent to obtain the diagnostic signal i X.N3 , the diagnostic accuracy of network model can be significantly increased. At the same time, comparing model II and model IV, it can be seen that by sharing learning ability, convolution-pooling network can effectively improve the recognition accuracy of system operation mode compared with traditional neural network.

[0085] Finally, the fault detection time of the model is tested. The simulation of T A1 tube fault triggers 20 times on average in a current fundamental period (based on the fundamental phase angle of phase A current), and the fault occurrence time and the corresponding fault diagnosis time are as shown in the following table.

[0086]

[0087] Since the T A1 tube fault directly affects the operation of the inverter A phase in the positive half cycle of the current, the diagnosis time is shorter when the fault triggers in the positive half cycle, and the shortest diagnosis time is 2.2 ms. Since the fault triggers in the negative half cycle does not affect the operation of the inverter, the diagnosis time is longer, and the longest time is 14.6 ms.

Claims

1. A method of diagnosing a fault in a multi-level inverter power device, characterized by, The operating modes of each phase of a multi-level inverter are identified using a feedforward convolutional-pooling network and an auxiliary neural network to obtain diagnostic results. Specifically, the weights and bias parameters learned by the auxiliary neural network are transferred to the convolutional kernels of the feedforward convolutional-pooling network to share its learning capabilities, enabling the feedforward convolutional-pooling network to identify the characteristics of the multi-level inverter's operating modes. The sharing condition for the feedforward convolutional-pooling network to share the learning capabilities of the auxiliary neural network is as follows: In the formula: , This represents the first convolutional layer. An array of convolutional kernels; , This represents the second convolutional layer. An array of convolutional kernels; , The hidden layer of the auxiliary network is represented by the first... The weight array of one neuron; , This represents the weight array of the l-th neuron in the output layer of the auxiliary network; This indicates the signal bias of convolutional layer one; The signal bias of convolutional layer 2 is represented by p; the number of neurons in the hidden layer of the auxiliary neural network is represented by f. convI f represents the activation function of convolutional layer one; convII This represents the activation function of the second convolutional layer.

2. The method of claim 1, wherein the method further comprises: The data input into the feedforward convolution-pooling network and the auxiliary neural network is a h d ×3s d two-dimensional signal matrix transformed and reconstructed from a three-phase one-dimensional time-series signal collected based on a sensor 3. A method of fault diagnosis of a multi-level inverter power device according to claim 2, characterized in that, The two-dimensional signal matrix is obtained by the following method : Step 1, collecting three-phase output current signals i of the multi-level inverter by using a sensor X , and performing park transformation and then normalizing the current amplitude, then performing attenuation and acceleration processing on the normalized three-phase current signals to obtain three-phase one-dimensional time sequence diagnostic signals, wherein X = A, B, C respectively represent phase A, phase B and phase C, the three-phase one-dimensional time sequence diagnostic signals are one-dimensional time sequence signal sequences corresponding to sampling time, and the one-dimensional time sequence diagnostic signal of phase X is represented as , represents the diagnostic signal of phase X at the nth sampling time ; Step 2, setting a sliding detection window with a size of 1xN DW , collecting one-dimensional time sequence diagnostic signals of each phase by using the sliding detection window, then the sliding detection window of the three-phase of the multi-level inverter collects 3N DW diagnostic signals each time; Step 3, converting and reconstructing the collected three-phase one-dimensional time sequence signals into a two-dimensional signal matrix with a size of h d x 3s d according to the following formula , the two-dimensional signal matrix is an input feature map of the feedforward convolution-pooling network and the auxiliary neural network: , wherein: k = 1:h d , l = 1:s d , and h d s d =N DW ; x k,l represents data in the kth row and the lth column of the two-dimensional signal matrix .

4. The method of claim 3, wherein the method further comprises: The auxiliary neural network includes an input layer, a hidden layer, and an output layer. The number of neurons r in the input layer is related to the two-dimensional signal matrix. The size relationship satisfies r=3sd; after the input feature map is input to the input layer, it is processed by neurons in the hidden layer and the output layer respectively, and then input to the Softmax layer for classification to obtain the final running pattern classification result. The signal processing of the hidden layer and the output layer is represented by the following formulas: , In the formula: z j f represents the output information of the j-th hidden layer neuron; hid () represents the activation function of the hidden layer; y j This represents the information received by the j-th hidden layer neuron; For information from the first The weights passed from one input layer neuron to the j-th hidden layer neuron; x n This represents the information received by the nth input layer neuron at a specific time; a j The bias for neurons to receive information; o l f represents the output information of the l-th output layer neuron; out () is the activation function of the output layer; Information is transmitted from the k-th hidden layer neuron to the k-th hidden layer neuron. The weights of each output layer neuron; b l This indicates the bias of the information received by the output layer neurons.

5. A method of fault diagnosis of a multi-level inverter power device as claimed in claim 4, characterized in that, The feedforward convolution-pooling network comprises two convolution layers and a global pooling layer, the two convolution layers are defined as a first convolution layer and a second convolution layer respectively, wherein the convolution layers are capable of directly processing the input feature map, and information in N DW sampling periods is included in the receptive field analysis, and characteristic information of the operating state of the multi-level inverter in the corresponding time period is comprehensively extracted, and the convolution operation process is shown in the following formula: , wherein: represents the information matrix of the lth output feature map; f conv () is an activation function of the convolution layer; represents the weight matrix of the kth channel in the lth group of convolution kernels; represents a convolution operator; represents the information matrix of the kth channel in the input feature map; represents a signal bias of the lth output feature map of the convolution layer; all features processed by the convolution layer are dimensionally compressed by a global pooling layer, and finally the compressed information is classified by a Softmax function to obtain probability results of the input feature map belonging to each operating mode, thereby indicating the operating state of the multi-level inverter.

6. The method of claim 3, wherein the method further comprises: The convolution layer adopts a multi-channel convolution layer, each set of convolution kernel parameters can be shared by all channels, each channel respectively receives the diagnostic signals collected by each phase of the multi-level inverter in the sliding detection window, and the diagnostic signals are filtered to contain features through the processing of the convolution kernel.

Images