A three-level ANPC converter fault diagnosis method and system based on a double-flow network

Abstract

The application discloses a kind of three-level ANPC converter fault diagnosis method and system based on double-flow network, method steps include, three-phase ac signal during the fault transient of converter is collected, and is spliced into one-dimensional basic data sequence;Parallelly execute windowed short-time Fourier transform and synchronous compression wavelet transform, extract low-frequency continuous harmonic distortion and high-frequency transient impact characteristics respectively in combination with nonlinear logarithmic scale enhancement technology, generate double-state time-frequency spectrum with physical complementarity;Double-state spectrum is fed into weight-independent double-flow convolutional neural network, after local feature independent extraction and element-by-element addition deep layer coupling, output single-tube open-circuit fault positioning result.The application realizes the spatiotemporal two-dimensional complementary decoupling of fault characteristics, significantly improves the diagnostic accuracy and generalization ability of converter under extreme conditions such as data deficiency and noise interference.

Description

Technical Field

[0001] This invention relates to the field of power electronic converter detection and control, specifically to a fault diagnosis method and system for a three-level ANPC converter based on a dual-current network. Background Technology

[0002] Three-level ANPC converters have gained widespread attention in renewable energy grid-connected and grid-forming applications due to their advantages such as high output power quality and uniform switching loss distribution. However, with the increase in the number of power devices, IGBT single-transistor open-circuit faults caused by thermal stress aging and bond wire detachment are highly insidious. If not diagnosed and treated in time, continuous asymmetrical operation can lead to cascading system failures.

[0003] Existing data-driven diagnostic methods heavily rely on large-scale fault samples. In practical engineering, it is difficult to obtain a large amount of fault data for converters. When the system faces "small sample" operating conditions, the traditional feature extraction methods based on one-dimensional time-series signals have limited representation capacity and are unable to characterize the spatiotemporal two-dimensional coupled texture of low-frequency continuous evolution and high-frequency transient impacts. This leads to neural network models being prone to "inter-class aliasing" and overfitting, resulting in a significant decrease in generalization ability.

[0004] Therefore, in order to solve the problems of low fault diagnosis accuracy and poor feature decoupling capability in existing technologies under small sample conditions, this invention provides a fault diagnosis method and system for three-level ANPC converters based on a dual-stream network. Summary of the Invention

[0005] In view of the shortcomings of the existing technology, the purpose of this invention is to provide a fault diagnosis method and system for a three-level ANPC converter based on a dual-stream network.

[0006] To achieve the above objectives, the present invention provides the following technical solution: A fault diagnosis method for a three-level ANPC converter based on a dual-stream network includes: The signal acquisition and sequence reconstruction steps involve acquiring the three-phase AC signals during the fault transient of the three-level ANPC converter and reconstructing the sequence to build a one-dimensional basic data sequence that includes cross-phase linkage evolution information. The STFT continuous evolution spectrum construction steps involve performing a windowed short-time Fourier transform on the one-dimensional basic data sequence and combining it with a nonlinear logarithmic scaling transform to generate an STFT continuous evolution energy spectrum characterizing the continuous distortion of low-frequency harmonics. The SWT transient impact map construction steps involve performing continuous wavelet transform and synchronous compression rearrangement in parallel on the one-dimensional basic data sequence, and combining nonlinear logarithmic scaling compression enhancement to generate an SWT transient impact feature map characterizing high-frequency transient jumps. The feature fusion and fault location steps involve inputting the STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum into a weighted, independently weighted, pre-defined dual-stream convolutional neural network for feature extraction. The features are then deeply coupled using an element-by-element addition strategy to determine the converter's operating status and perform multi-class identification. The single-tube open-circuit fault location result of the ANPC converter is obtained through the classifier output.

[0007] As a further improvement of the present invention, the construction of the one-dimensional basic data sequence includes: real-time acquisition of the three-phase voltage and current signals output from the AC side of the three-level ANPC converter controlled by the virtual synchronous generator; extracting signal data windows including the fault occurrence time and the transient response process of the system fault to generate a three-phase AC voltage timing signal; and splicing the three-phase AC voltage timing signal end to end according to the A, B, C phase sequence to obtain the one-dimensional basic data sequence.

[0008] As a further improvement of the present invention, the generation of the STFT continuous evolution energy spectrum includes introducing a sliding window mechanism into the one-dimensional basic data sequence, and obtaining a one-dimensional windowed data frame sequence by truncating and windowing through a Hamming window, performing a discrete short-time Fourier transform on the one-dimensional windowed data frame sequence to extract the low-frequency harmonic continuous distortion features induced by open-circuit faults, and performing a nonlinear logarithmic scaling transform on the low-frequency harmonic continuous distortion features to generate the STFT continuous evolution energy spectrum.

[0009] As a further improvement of the present invention, the nonlinear logarithmic scaling transformation includes performing amplitude transformation on the feature matrix calculated by the discrete short-time Fourier transform, converting the complex frequency domain features into real features characterizing the physical energy intensity, introducing a nonlinear scaling compression operator to perform dynamic range compression on the real features, and mapping the compressed feature data to a time-frequency decoupled two-dimensional discrete rectangular coordinate grid to generate an STFT continuous evolution energy spectrum.

[0010] As a further improvement of the present invention, the generation of the SWT transient impact feature map includes: performing continuous wavelet transform on the one-dimensional basic data sequence to obtain a wavelet coefficient matrix; estimating the true instantaneous frequency of the signal based on the phase derivative of the wavelet coefficients; introducing a synchronous compression operator; using the true instantaneous frequency as the energy convergence center; rearranging the spatial positions of the wavelet coefficients in the scale domain to accumulate and converge the divergent transient energy; and performing nonlinear logarithmic scaling compression enhancement on the rearranged wavelet coefficient matrix to output an SWT transient impact energy map with steep time-frequency energy lines.

[0011] As a further improvement of the present invention, the nonlinear logarithmic scale compression enhancement includes: extracting the rearranged coefficient matrix, performing an amplitude transformation on the coefficient matrix to convert it into real-valued features characterizing physical energy intensity, introducing a nonlinear scale enhancement operator to perform dynamic range series compression and span adjustment on the real-valued features, enhancing transient impact details and tuning the contrast of feature distribution at the decibel scale, and mapping the enhanced real-valued features onto a two-dimensional discrete rectangular coordinate grid decoupled from the time axis and the scale axis to generate an SWT transient impact feature map.

[0012] As a further improvement of the present invention, the dual-stream convolutional neural network includes two feature extraction branches with equivalent topology and independent weights, which uniformly convert the STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum interpolation into RGB three-channel feature tensors of the same dimension, and input them to the two feature extraction branches respectively.

[0013] As a further improvement of the present invention, the element-wise addition strategy includes coupling the feature vector output by the dual-stream convolutional neural network in the feature fusion layer, inputting the coupled feature vector into a Softmax classifier with 19 output nodes, calculating the probability distribution, and outputting the specific single-tube open-circuit fault location results for the 18 power devices in the three phases of the ANPC converter under normal operating conditions.

[0014] A fault diagnosis system for a three-level ANPC converter based on a dual-stream network includes: The signal acquisition and reconstruction module is used to acquire the three-phase AC signals during the fault transient of the three-level ANPC converter at high frequency and splice them into a one-dimensional basic data sequence according to the phase sequence. The continuous evolution feature reconstruction module is used to perform windowed short-time Fourier transform and combine it with nonlinear logarithmic scaling transform to generate STFT continuous evolution energy spectrum. The transient impact feature reconstruction module is used to perform continuous wavelet transform and synchronous compression rearrangement in parallel, and combined with nonlinear logarithmic scaling compression enhancement to generate SWT transient impact feature maps. The TSCNN feature fusion and localization module is used to input the dual-state map into a weighted dual-stream convolutional neural network to extract features independently. It adopts an element-wise addition strategy for deep coupling and outputs the localization result through a Softmax classifier.

[0015] The beneficial effects of this invention are: it aims to improve the fault diagnosis generalization ability and noise resistance robustness of converters under "small sample" data-limited operating conditions. By using Short Time Fourier Transform (STFT) and Synchronous Compressed Wavelet Transform (SWT), the one-dimensional signal is mapped to a two-dimensional physical complementary spectrum that characterizes low-frequency continuous evolution and high-frequency transient impact, effectively avoiding mutual interference of heterogeneous features in the shallow layer of the network.

[0016] In terms of feature extraction and classification, this method abandons the single-input network approach and proposes a two-stream convolutional neural network (TSCNN) architecture. It combines nonlinear logarithmic scaling enhancement techniques with an element-wise additive feature fusion strategy to effectively suppress intra-class divergence and inter-class aliasing in high-dimensional space. This method improves accuracy under the constraint of very few samples in extreme operating conditions and is suitable for engineering applications where obtaining massive amounts of fault data is difficult. Attached Figure Description

[0017] Figure 1 This invention is based on a fault diagnosis flowchart for dual-stream networks.

[0018] Figure 2 This is a schematic diagram of a dual-stream convolutional neural network according to an embodiment of the present invention.

[0019] Figure 3 This is the time-frequency spectrum of the continuous evolution of STFT based on a logarithmic scale under normal operating conditions in an embodiment of the present invention.

[0020] Figure 4 This is the time-frequency spectrum of the continuous evolution of STFT based on a logarithmic scale under the open-circuit fault condition of T1 tube in an embodiment of the present invention.

[0021] Figure 5 This is the time-frequency spectrum of the continuous evolution of STFT based on a logarithmic scale under the open-circuit fault condition of T2 tube in this embodiment of the invention.

[0022] Figure 6 This is the SWT transient time-frequency spectrum based on linear amplitude under normal operating conditions in this embodiment of the invention.

[0023] Figure 7 This is the SWT transient time-frequency spectrum based on linear amplitude under the open-circuit fault condition of tube T1 in this embodiment of the invention.

[0024] Figure 8 This is the SWT transient time-frequency spectrum based on linear amplitude under the open-circuit fault condition of tube T2 in this embodiment of the invention.

[0025] Figure 9 This is the SWT transient time-frequency spectrum of nonlinear reconstruction under normal operating conditions in an embodiment of the present invention.

[0026] Figure 10 This is the transient time-frequency spectrum of the nonlinear reconstructed SWT under the open-circuit fault condition of tube T1 in this embodiment of the invention.

[0027] Figure 11 This is the transient time-frequency spectrum of the nonlinear reconstructed SWT under the open-circuit fault condition of the T2 tube in this embodiment of the invention.

[0028] Figure 12This is a bar chart comparing multi-dimensional evaluation indicators under standard operating conditions in embodiments of the present invention.

[0029] Figure 13 This is a bar chart comparing the multidimensional evaluation indicators under constrained working conditions in this application.

[0030] Figure 14 This is a comparative diagram of multi-dimensional evaluation indicators under extreme working conditions in this application. Detailed Implementation

[0031] The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Identical components are denoted by the same reference numerals. It should be noted that the terms "front," "rear," "left," "right," "upper," and "lower" used in the following description refer to directions in the accompanying drawings, and the terms "bottom surface," "top surface," "inner," and "outer" refer to directions toward or away from the geometric center of a specific component, respectively.

[0032] The first embodiment of this invention proposes a fault diagnosis method for three-level ANPC converters under small sample operating conditions. Its core lies in upscaling the one-dimensional time-series fault signal into two two-dimensional feature maps with physical spatial positive complementarity: STFT (extracting continuous evolution of low-frequency harmonics) and SWT (extracting high-frequency transient impulse jumps). After nonlinear logarithmic enhancement to overcome energy masking effects, these maps are input into a weighted independent dual-stream convolutional neural network (TSCNN), where deep feature coupling is achieved through an element-wise addition strategy. This method overcomes the limitations of traditional one-dimensional feature representation capacity and the susceptibility to inter-class aliasing in the model. It improves the converter's fault location accuracy and degradation robustness under extreme operating conditions with scarce data and electromagnetic noise. The method of this embodiment is as follows: Figures 1 to 2 As shown, it includes: The signal acquisition and sequence reconstruction steps involve real-time high-frequency acquisition of the three-phase AC signals during the fault transient of the three-level ANPC converter, followed by sequence reconstruction to construct a one-dimensional basic data sequence that includes cross-phase linkage evolution information.

[0033] Specifically, the construction of the one-dimensional basic data sequence includes real-time acquisition of the three-phase voltage and current signals output from the AC side of a three-level ANPC converter controlled by a virtual synchronous generator, and extracting signal data windows including the fault occurrence time and the system fault transient response process to generate a three-phase AC voltage timing signal. The three-phase AC voltage timing signal is then concatenated end-to-end according to the A, B, and C phase sequence to obtain the one-dimensional basic data sequence x[n]. To fully explore the spatial coupling information in the constrained samples and expand the feature capacity of a single two-dimensional spectrum, a joint voltage sequence spliced ​​end-to-end in the order of three phases A, B, and C is used as the basis for analysis. Assuming the preprocessed ANPC three-phase spliced ​​voltage signal is x[n], the parameter configuration when performing a discrete short-time Fourier transform on the one-dimensional basic data sequence x[n] directly affects the time-frequency resolution and characterization accuracy of the three-level ANPC converter in terms of low-frequency harmonic continuous distortion characteristics. Because this invention employs a strategy of splicing the transient response signals of three phases A, B, and C end-to-end to form a one-dimensional basic data sequence x[n], this sequence not only contains single-phase time evolution but also implicitly contains cross-phase linkage evolution information at the splicing breakpoints. Its mathematical expression for the discrete short-time Fourier transform is: ; Where N is the number of points in the Discrete Fourier Transform (DFT), used to define the number of discrete points on the frequency axis (vertical axis) of the generated STFT feature matrix. Combined with the discrete frequency index (0 < k < N-1), it divides the continuous spectrum of the fault signal into N parallel frequency domain observation channels. The number of points N in the DFT determines the discrete sampling density of the frequency axis. In this embodiment, the recommended value range for N is 256 < N < 1024. When a single open-circuit fault occurs in a three-level ANPC converter, the induced distortion is mainly concentrated in the fundamental frequency, low-order characteristic harmonics such as the 2nd, 3rd, and 5th harmonics, and the DC component. If N is too small (less than 256), the frequency resolution will be too low, causing spectral aliasing of similar low-order harmonics in the frequency domain, making it impossible to accurately extract the unique low-frequency features of different devices when open-circuited. If N is too large (greater than 1024), although the frequency domain resolution is improved, it will significantly increase the computational overhead of the dual-stream convolutional neural network front-end data preparation and cause the time domain resolution to diverge. Therefore, a value within this range is usually chosen to achieve a better frequency domain resolution boundary in the characterization of low-frequency fault features of the converter, ensuring that the continuous evolution of low frequencies induced by open-circuit faults can be accurately discretized and mapped. R is the sliding step size of adjacent time windows, which refers to the number of discrete points the window function slides on the time axis in two adjacent STFT calculations. In this embodiment, the value of R is directly determined by the window length L and the preset overlap rate, satisfying R = L × (1 - Overlap). The overlap rate is set to 50% ≤ Overlap ≤ 75%, preferably 50%. When an ANPC converter experiences a transient open-circuit fault, the fundamental frequency distortion caused by the fault is a continuously evolving process. If the step size R is set too large (i.e., the overlap rate is too low), the sliding window will miss the phase connection information before and after the breakpoint at the moment the fault occurs, resulting in gaps and missed detections in the time-domain features. If the step size R is set too small (i.e., the overlap rate is too high), it will lead to extremely high data redundancy between adjacent time-domain frames, causing not only a sharp expansion of the feature matrix size but also introducing noise interference into the subsequent neural network. By setting an overlap rate of 50%, the continuity in the time domain can be ensured while suppressing spectral leakage to the greatest extent, so that the time frame index m (where the upper limit of m is determined by the total length of the one-dimensional sequence and the step size R) can accurately and without omission index every tiny time step of the fault transient response. m is the time frame index, which refers to the sequence number of the discrete time frame obtained after the one-dimensional basic data sequence is truncated and framed by the sliding window. Each specific index value m corresponds to a specific time slice (i.e., the m-th data frame) in the original three-phase spliced ​​signal. By incrementing m, the timing trajectory of the converter evolving from normal operating conditions to transient open-circuit fault conditions can be tracked in time steps. In the final generated two-dimensional STFT continuous evolution energy spectrum, the time frame index m is explicitly mapped to the horizontal axis (time axis) of the image.By arranging the spectrum of each frame corresponding to m in parallel columns, the three-phase voltage distortion, which was originally aliased in the time domain, is reconstructed into low-frequency harmonic horizontal stripes that evolve continuously over time. This provides a two-dimensional spatial geometric topology with clear time evolution characteristics for the two-stream convolutional neural network, enabling it to accurately capture the time-series boundary points of the fault occurrence and subsequent transient response. w[n-mR] is a real-number window function that slides along the time axis. The window length L of this function should be less than or equal to the number of discrete Fourier transform points N. In this embodiment, a Hamming window is preferred, and the window length L is set to be the same as N. The AC signal output from the three-level ANPC converter is accompanied by complex switching noise in both steady-state and transient states. If a traditional rectangular window is used, due to its slow sidelobe attenuation, severe spectral leakage will occur when performing frame cutting of the first and last splicing sequence, resulting in blurred edges of low-frequency stripes in the fault spectrum. The Hamming window exhibits excellent nonlinear attenuation of sidelobes, suppressing the first sidelobe to approximately -43dB. It smooths the boundary truncation abrupt changes introduced by phase differences in the spliced ​​sequence, highlighting the true and essential low-frequency harmonic continuous distortion characteristics caused by open-circuit faults. k is a discrete frequency index, which, together with the time frame index m, constitutes the two-dimensional spatial coordinate system of the final generated feature matrix X[m,k]. In this embodiment, by traversing all m and k, the mathematical expression transforms the one-dimensional time-series voltage signal, which was originally impossible to decouple in the shallow layers of the network, into a complex matrix of size M×N. The row index m of this complex matrix maps to the evolution trajectory in the time dimension, and the column index k maps to the energy distribution in the frequency dimension. This explicit time-frequency dual-drive mapping mechanism directly provides the basic data structure required for subsequent nonlinear logarithmic scaling, eliminating the risk of inter-class feature aliasing caused by the input one-dimensional signal in neural networks under traditional small-sample conditions.

[0034] By extracting and splicing the three-phase transient response signals end to end, the cross-phase distortion characteristics caused by asymmetric faults can be integrated into a single time series, effectively expanding the information carrying capacity of a single spectrum and laying the foundation for full state characterization under small samples.

[0035] The steps for constructing the STFT continuous evolution spectrum include performing a windowed short-time Fourier transform on the one-dimensional basic data sequence and combining it with a nonlinear logarithmic scaling transform to generate an STFT continuous evolution energy spectrum characterizing the continuous distortion of low-frequency harmonics, as shown below. Figures 3 to 5As shown, the generation of the STFT continuous evolution energy spectrum includes: introducing a sliding window mechanism into the one-dimensional basic data sequence, and obtaining a one-dimensional windowed data frame sequence by truncating and windowing using a Hamming window; performing a discrete short-time Fourier transform on the one-dimensional windowed data frame sequence to extract the low-frequency harmonic continuous distortion features induced by open-circuit faults; and performing a nonlinear logarithmic scaling transform on the low-frequency harmonic continuous distortion features to generate the STFT continuous evolution energy spectrum, which is as follows: In the continuous evolution feature map reconstruction strategy based on Short-Time Fourier Transform (STFT), a sliding window mechanism is introduced into the spliced ​​one-dimensional sequence x[n]. A Hamming window with smooth edges approaching zero is used for truncation. The window length L=256 and the sliding step size R=128. At the system sampling frequency f s At 15kHz, the temporal resolution window width can be derived as Δt: ; This parameter configuration, close to the fundamental period (i.e., 20ms), ensures that the spectrum has sufficient frequency resolution, enabling the extraction of continuously evolving harmonic components in the three-phase voltage splicing signal caused by faults in different sub-modules, and the calculation of the discrete STFT matrix.

[0036] Specifically, in this embodiment, the nonlinear logarithmic scaling transformation includes: performing amplitude transformation on the feature matrix calculated by the discrete short-time Fourier transform to convert the complex frequency domain features into real features characterizing physical energy intensity; introducing a nonlinear scaling compression operator to dynamically compress the real features; and mapping the compressed feature data onto a time-frequency decoupled two-dimensional discrete rectangular coordinate grid to generate an STFT continuously evolving energy spectrum. The formula for the nonlinear logarithmic scaling transformation based on the short-time Fourier transform is: ; First, an absolute value operator is applied to the feature matrix (complex coefficients) X[m,k], i.e., the magnitude is calculated. The three-phase AC signals acquired by the three-level ANPC converter under normal or fault transient conditions, after discrete short-time Fourier transform, have their original coefficients X[m,k] mathematically represented as complex tensors containing real and imaginary parts. Since the subsequent two-stream convolutional neural network cannot directly process the discrete grid of the complex domain when extracting two-dimensional spatial features, the complex frequency domain features are losslessly transformed into real amplitude scalars that can directly characterize physical energy intensity by constructing the absolute value operator |X[m,k]| as the square root of the sum of the squares of the real and imaginary coefficients. This provides a unique, computer vision-compliant underlying data source for the subsequent construction of time-frequency image formats.

[0037] The core of the formula construction lies in introducing a commonly used logarithmic operator with a base of 10 and combining it with a coefficient of 20 for decibel (dB) mapping. In a three-level ANPC converter system controlled by a virtual synchronous generator, the voltage or current amplitude of the system's fundamental frequency (50Hz) is extremely large under normal operating conditions. When a power device, such as the upper arm T1 tube of phase A, experiences a weak single-tube open-circuit fault, the absolute amplitude of the low-frequency characteristic distortion harmonics it causes, such as the second and third harmonics, is often only 0.1% to 1% of the fundamental frequency amplitude. If the absolute amplitude |X[m,k]| is directly mapped to two-dimensional image pixels, due to the huge gap of several orders of magnitude between the fundamental frequency energy and the fault transient harmonic energy in absolute value, the high-energy fundamental frequency will produce a severe feature masking effect, making the fault feature stripes almost invisible in visual texture, and the network is prone to inter-class aliasing. By introducing 20log 10 The operator utilizes the nonlinear slope characteristics of the logarithmic function—strongly compressing large values ​​and amplifying small values ​​exponentially—to forcibly reduce the infinitely wide absolute energy range proportionally and map it to a finite decibel scale space. This not only suppresses strong fundamental wave energy by peak clipping but also significantly enhances the contrast of weak transient harmonics induced by open-circuit faults, transforming low-frequency continuous distortion characteristics into clearly visible horizontal fringe topology in the time-frequency grid.

[0038] S obtained after the above logarithmic scaling transformation STFT The [m,k] matrix contains real decibel values ​​at each time-frequency grid point. In this embodiment, this matrix is ​​used for full-matrix contrast linear normalization, linearly mapping values ​​in the -120dB to 40dB range to a standard integer pixel value range of 0 to 255. Through this step, the decibel scalar is completely reconstructed into a pixel grayscale matrix of a two-dimensional image. Subsequently, it is converted into an RGB three-channel feature tensor using pseudo-color encoding or interpolation, and directly used as the standard input to the first topologically symmetric branch of a two-stream convolutional neural network, completing a perfect cross-modal transition from the transient signal of the physical converter to the semantic features of AI space.

[0039] The SWT transient impact map construction steps involve performing continuous wavelet transform and synchronous compression rearrangement in parallel on the one-dimensional basic data sequence, and combining this with nonlinear logarithmic scaling compression enhancement to generate an SWT transient impact feature map characterizing high-frequency transient jumps.

[0040] Specifically, in this embodiment, the generation of the SWT transient impact feature map includes feeding the one-dimensional basic data sequence x[n] into a continuous wavelet transform operator, and performing inner product projection using preset wavelet basis functions under different scale factors and translation factors to obtain a complex-domain original wavelet coefficient matrix. Due to the physical constraints of the Heisenberg uncertainty principle, the high-frequency sudden jump characteristics generated by the three-level ANPC converter at the moment of failure appear in the original wavelet coefficient matrix as an energy halo morphology that diverges along the scale axis and has blurred morphological boundaries, which cannot be directly and accurately recognized by a two-stream convolutional neural network. To eliminate energy divergence in scale space and achieve engineering tuning of time-frequency convergence, the spatial location transfer in this embodiment includes: using the phase derivatives of adjacent grid points in the original wavelet coefficient matrix to calculate the true instantaneous frequency corresponding to the current time-series location online; constructing a synchronous compression operator; traversing each discrete grid in the original wavelet coefficient matrix; and reorganizing and transferring the original wavelet coefficients that diverge on the scale axis but whose corresponding true instantaneous frequencies are the same as the convergence center line in memory according to their instantaneous frequency orientation, directly adding or mapping them to the corresponding instantaneous frequency ridge grid points. For other scale grid points that deviate from the instantaneous frequency ridge line and do not contain fault characteristics, their wavelet coefficient values ​​are forcibly cleared to zero or subjected to low-energy suppression. Through this centripetal rearrangement, the divergent transient energy is forcibly compressed and converged on the time-frequency surface, forming a high-energy transient jump line with a steep physical boundary, and extracting the high-frequency sudden jump characteristics of the fault instant, such as... Figures 6 to 8 As shown, since the parallel branches of deep learning networks cannot directly process data streams containing complex numbers, this embodiment performs a real-value transformation on the rearranged complex coefficient matrix. The real and imaginary coefficients of each grid point are extracted, and by performing a square root operation (amplitude transformation), the complex coefficients with phase are converted into a real scalar feature matrix that can directly characterize the intensity of high-frequency transient impacts. This provides a unique underlying data source for subsequently generating spectral images that conform to computer vision standards. Nonlinear logarithmic scaling enhancement and pixel normalization reconstruction are performed on the real scalar feature matrix. A common logarithmic operator with a base of 10 is introduced, along with a decibel conversion gain coefficient, to perform dynamic range series compression on the real scalar matrix. This filters out high-frequency random switching noise background introduced by the normal power switching of the converter online, while nonlinearly amplifying the weak transient jump characteristics induced by single-tube open circuits at the decibel scale. Figures 9 to 11 As shown, the compressed and tuned feature data are sequentially mapped onto a two-dimensional discrete Cartesian coordinate grid decoupled from the discretized time-shift index (horizontal axis) and the discretized scale index (vertical axis). This completely reconstructs the one-dimensional time-series signal into an SWT transient impact feature map with a standard pixel range.

[0041] in, Figures 6 to 8It directly generates SWT transient time-frequency spectra based on linear amplitudes, which results in feature loss and energy masking. Figures 9 to 11 The masking effect of the fundamental wave energy was effectively suppressed by introducing nonlinear logarithmic scale compression enhancement technology, and the originally weak high-frequency transient characteristics were significantly amplified and clearly aggregated into a bright vertical ridge pattern.

[0042] In this embodiment, the transient impulse feature map reconstruction strategy based on Synchronous Compressed Wavelet Transform (SWT) estimates the true instantaneous frequency of the signal at any time-frequency point [a,b] using the phase derivative of the wavelet coefficients, which can be expressed as: ; In the formula, a is the scaling factor, which controls the scaling of the mother wavelet; b is the time shift factor; ψ * N is the complex conjugate of the mother wavelet function. s The signal length is given. In a three-level ANPC converter controlled by a virtual synchronous generator, the transient response signal caused by an open-circuit fault in a device manifests as an extremely short jump on the time axis. If a traditional damped oscillatory wavelet is used, due to its wide time-domain support range, it is prone to generating false edge oscillation textures at the breakpoints of the spliced ​​sequence. Therefore, in this embodiment, the mother wavelet function is preferably an analytical Morse wavelet with optimal time-frequency localization performance. It can accurately fit the step-type abrupt changes in the three-phase AC current and neutral point voltage when a single tube is open-circuited, ensuring that the high-frequency energy can be completely captured instantaneously without time-domain spread when the complex conjugate operator performs inner product projection. Therefore, ψ is selected. * The signal length N s The number of points in the one-dimensional basic data sequence after the previous steps is determined. The scale factor 'a' determines the frequency range of wavelet detection. In this embodiment, the recommended discrete value range of 'a' is configured as follows: 10 < N per octave. octave <32 Uniform discretization values ​​(preferably N) octave =16), with a total scale configuration of 64 < A total <128. This value range ensures that the scaling factor a can cover the microsecond-level transient components when the converter experiences high-frequency sudden changes without omission. The time shift factor b, as the time axis discrete index, advances step by step in the time domain with a single sampling point as the step size, and the value of b ranges from 0 to N. s -1.

[0043] In this embodiment, a synchronous compression operator is introduced to rearrange and focus the wavelet coefficients in the scale domain, mapping and accumulating the divergent energy to the instantaneous frequency center ω, resulting in: ; In the formula, the summation condition is |ω-ω x (a i b)|≤ ω / 2 (i.e., the instantaneous frequency falls within a specific frequency band) (ω inside), ( a) i is the discrete scale step size. Through the above mechanism, the transient jump characteristics caused by faults in the three-phase voltage splicing signal are clustered into time-frequency ridges in the two-dimensional spectrum. This formula establishes a nonlinear mapping from wavelet time-scale space to synchronously compressed time-frequency space. The formula introduces... This is the energy-normalized tuning gain coefficient. Since the energy density carried by the continuous wavelet transform varies at different scales—high energy at small scales and low energy diverging at large scales—direct rearrangement would cause high-frequency features to overwhelm low-frequency features. By introducing a gain coefficient of this specific power, combined with the discrete scale step size… Performing a cumulative summation mathematically eliminates the energy imbalance caused by scale stretching. This allows the magnitude of the rearranged characteristic matrix along the frequency axis ω at any time shift factor b to objectively and proportionally reflect the transient physical impact intensity of a single tube opening in the converter. |ω-ω x (a i b)|≤ ω / 2 is the absolute physical constraint controlling the spatial location transport of wavelet coefficients. It represents the condition that ω / 2 only applies when a certain discrete scale a... i The instantaneous frequency ω is calculated from the phase derivative of the local wavelet coefficients at a given point. x (a i b) When the absolute difference between the current preset frequency center ω and the current frequency center ω falls within the set target bandwidth Δω, the energy W at that scale grid point is... x (a i Only those (b) are eligible to be stripped and moved to the grid point at the center ω of that frequency.

[0044] In practical converter engineering applications, the value of the specific bandwidth Δω directly determines the concentration of the spectrum. In this embodiment, its value range is strictly configured as Δω = ω. s / N f , where ω s The high-frequency sampling rate of the system transient signal (within the range of 10kHz ≤ ω) s ≤50kHz), N f The desired number of discrete points on the frequency axis (within the range of 128 ≤ N). f512). If the value of Δω is too large, the summation constraint will be relaxed, which will cause different frequency characteristic components to alias and merge again, making it impossible to form a sharp shape on the time-frequency surface; if the value of Δω is too small, the rearrangement path will be too narrow, and some fault transient edge energy will be filtered out. Therefore, by using this adaptive proportional frequency band limitation, it is possible to forcibly map and accumulate the divergent energy to the most realistic physical instantaneous frequency center ω in each frame advanced by the time shift factor b.

[0045] The iterative execution of this formula completely eliminates switching modulation interference noise during the steady-state operation of the converter. It precisely compresses and reshapes the divergent energy induced by the fault instant in the three-phase voltage and current spliced ​​signals into a bright, steep, and spatially topologically strong vertical time-frequency ridge in a two-dimensional Cartesian network graph. This not only completely prevents gradient distortion of samples during network training but also provides a high-frequency spatial semantic feature source with high spatial resolution and deterministic texture boundaries for the second parallel branch of the subsequent two-stream convolutional neural network.

[0046] Specifically, in this embodiment, the nonlinear logarithmic scale compression enhancement includes: extracting the rearranged coefficient matrix, performing an amplitude transformation on the coefficient matrix to convert it into real-valued features characterizing physical energy intensity, introducing a nonlinear scale enhancement operator to perform dynamic range series compression and span adjustment on the real-valued features, enhancing transient impact details and tuning the contrast of feature distribution at the decibel scale, and mapping the enhanced real-valued features onto a two-dimensional discrete rectangular coordinate grid decoupled from the time axis and scale axis to generate an SWT transient impact feature map.

[0047] In this embodiment, the formula for nonlinear logarithmic scaling of the SWT feature matrix is:

[0048] In the formula, For the enhanced transient impact feature map, q represents the magnitude of the complex coefficients after synchronous compression and rearrangement, p is the discretization time shift index, and q is the discretization scale index.

[0049] The complex coefficients T after synchronous compression transformation and refocusing in the preceding steps x Although [p,q] eliminates the scale divergence caused by continuous wavelet transform in the time-frequency plane, the three-level ANPC converter, during normal modulation operation, naturally carries extremely weak, discretely grid-like high-frequency modulation switching noise on the AC side and neutral point due to the high-frequency switching action of PWM or SVPWM space vector modulation. If directly based on the linear amplitude |T x[p,q]| When visualized, the widely distributed absolute amplitudes of these switching noises undergo severe nonlinear interleaving with the sudden jump characteristics of the fault. This not only masks the true fault characteristics but also results in a large number of false noise specks in the extracted two-dimensional map background, which easily interferes with the convolutional attention mechanism of the second branch of the two-stream network. By introducing a decibel scaling operator 20log 10 This technology leverages the strong nonlinear attenuation characteristics of the logarithmic function near the negative and zero domains to proportionally suppress extremely low-energy background modulation clutter to a lower decibel baseline. Simultaneously, it nonlinearly stretches and tunes the contrast of high-frequency transient impulses induced by sudden current and voltage jumps during abrupt open-circuit events at the decibel level. This not only eliminates feature masking effects but also forcibly extracts bright, noise-free vertical topographic textures from the time-frequency plane.

[0050] In the specific algorithm engineering construction, in order to prevent T from being affected at certain scale grid points after synchronous compression and rearrangement... x [p,q]=0 causes a computational singularity of negative infinity in logarithmic calculations. In this embodiment, the formula is adaptively extended during actual programming implementation, adding an adaptive scale bias constant to prevent division by zero. The enhanced transient impact feature map... The output amplitude dynamic range is strictly truncated between -90dB and 30dB. By introducing this range and bias, the switching noise peaks during steady-state operation of the converter can be forcibly suppressed to below -90dB, stabilizing the peak energy of the high-frequency vertical ridge at fault moments around 30dB. This not only ensures that the underlying data of the digital signal processor or embedded edge computing chip does not overflow abnormalities, but also gives the final output data sequence a highly standardized span, effectively supporting the robust convergence of subsequent artificial intelligence networks.

[0051] The discretization time shift index p and the discretization scale index q together determine the generated S. SWTThe [p,q] dB matrix represents the two-dimensional spatial resolution. In this embodiment, the entire dB matrix is ​​linearly normalized for contrast using [-90dB, 30dB] as the baseline boundary, linearly mapping the dB values ​​to the standard integer pixel range of [0, 255]. Subsequently, the discretized time-shift index p is mapped to the horizontal axis of the image, and the discretized scale index q is mapped to the vertical axis of the image, thereby completely reconstructing the physical energy fluctuations of the one-dimensional transient sequence in scale space into a high-resolution SWT transient impact feature map with standard color or grayscale density. This map represents the time discontinuity of the fault transient with highly clear and vertically distinct bright lines, and is then interpolated into an RGB three-channel feature tensor of the same dimension, which is directly fed into the second feature extraction branch of the dual-stream convolutional neural network used to represent high-frequency mutation characteristics, thus completely opening up the entire closed loop of engineering flow from power electronic transient response to high-level semantics of computer vision.

[0052] The feature fusion and fault location steps involve inputting the STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum into a weighted, independently weighted, pre-defined dual-stream convolutional neural network for feature extraction. The features are then deeply coupled using an element-by-element addition strategy to determine the converter's operating status and perform multi-class identification. The single-tube open-circuit fault location result of the ANPC converter is obtained through the classifier output.

[0053] Specifically, in this embodiment, the dual-stream convolutional neural network includes two feature extraction branches with equivalent topologies and independent weights. The STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum interpolation are uniformly converted into RGB three-channel feature tensors of the same dimension, and then input to the two feature extraction branches respectively.

[0054] Specifically, in this embodiment, the element-wise addition strategy includes coupling the feature vector output by the dual-stream convolutional neural network in the feature fusion layer, inputting the coupled feature vector into a Softmax classifier with 19 output nodes, calculating the probability distribution, and outputting the specific single-tube open-circuit fault location results for the 18 power devices in the three phases of the ANPC converter under normal operating conditions.

[0055] The STFT continuous evolution energy spectrum and the SWT transient impact feature map generated in the previous steps are input into a preset dual-stream convolutional neural network. To eliminate the geometric asymmetry in the matrix dimension caused by temporal framing and wavelet scale rearrangement, a bicubic interpolation algorithm is first used before feeding them into the network input layer to uniformly scale and convert the STFT continuous evolution energy spectrum and the SWT transient impact feature map into a fixed-size 64×64 RGB three-channel feature tensor with red, green and blue channels.

[0056] The dual-stream convolutional neural network comprises two feature extraction branches with completely identical topologies and independent weights: a first feature extraction branch and a second feature extraction branch. The scaled STFT continuous evolution energy spectrum tensor and the SWT transient impact feature spectrum tensor are input in parallel into the two feature extraction branches.

[0057] During the forward propagation within each branch, the input tensor sequentially passes through two sets of time-frequency morphological feature extraction modules composed of convolutional layers, batch normalization layers, modified linear unit activation layers, and max-pooling layers to deeply mine the higher-order semantics of low-frequency continuous stripe structures and high-frequency abrupt vertical ridges in spatial topology. Subsequently, a flattening operator is used to transform the multidimensional feature map into a one-dimensional first-order tensor, which is then fed into a fully connected network containing two hidden layers. The number of nodes in the two hidden layers is sequentially configured as 64 and 32. Through linear weighting and nonlinear mapping of the fully connected matrix, the output results of the two branches are finally converged, decoupled, and projected into 32-dimensional high-order semantic feature vectors of low-frequency harmonics and 32-dimensional high-order semantic feature vectors of transient impulses, respectively.

[0058] After extracting the high-order features of the two branches, the fully connected layer at the tail of the two feature extraction branches forces the low-frequency harmonic high-order semantic feature vector and the transient impact high-order semantic feature vector to be projected to the same geometric dimension space, realizing point-to-point alignment of cross-modal features in the spatial tensor dimension. Subsequently, an element-wise addition strategy is used in the feature fusion layer to deeply couple and fuse the dual-mode high-order semantic vectors.

[0059] When the element-wise addition strategy is executed, it iterates through the elements at the same index position in the two 32-dimensional vectors and performs a weighted summation. This allows for the lossless superposition of discrete time-domain impact semantics and continuous frequency-domain distortion semantics at the same spatial position, thereby forcibly activating the representational capacity of the fused feature vector and constructing a comprehensive diagnostic feature vector that takes into account both global transients and local mutations.

[0060] The combined feature vector obtained after coupling is directly fed into a Softmax classifier containing 19 output nodes. The Softmax classifier calculates the full probability distribution online based on the input combined feature vector. A one-to-one mapping relationship is established between the 19 output nodes and the physical operating state of the converter. The first output node represents the converter in normal operating condition, and the remaining 18 output nodes sequentially represent the single-tube open-circuit fault conditions of the 18 specific power devices in the three-phase structure of the three-level ANPC converter. By identifying the node with the highest probability in the full probability distribution, the final converter state diagnosis and accurate single-tube open-circuit fault location result are output.

[0061] To comprehensively verify the anti-degradation capability of the present invention under different data scale constraints and the diagnostic robustness under small sample conditions, this embodiment constructs a standard operating condition sample set, a restricted operating condition sample set, and an extreme operating condition sample set for comparative iterative verification.

[0062] The diagram showing the comparison of multi-dimensional evaluation indicators under standard working conditions is as follows: Figure 12 As shown in the diagram, a comparison of multi-dimensional evaluation indicators under constrained working conditions is presented. Figure 13 As shown in the diagram, a comparison of multi-dimensional evaluation indicators under extreme working conditions is presented. Figure 14 As shown.

[0063] The comparison results of multidimensional evaluation indicators under various working conditions show that as the training sample size gradually decreases, the comparison models based on single feature input or traditional feature cascade splicing input exhibit significant nonlinear decay and technical degradation in core indicators such as accuracy, precision, recall and F1 score due to the sparse spatial features and easy overfitting in small sample environments.

[0064] The method proposed in this invention, under extreme constraints of extremely scarce training data, significantly reduces the network's dependence on massive training samples by introducing a time-frequency dual-drive spectral reconstruction process at the front end and achieving lossless complementary conjugate superposition of time-frequency isotopic features using an element-wise addition strategy at the feature fusion layer. Experimental data confirms that this invention can maintain extremely high levels of evaluation indicators across all dimensions without technical degradation even under constraints of very few samples, perfectly verifying the high-dimensional decoupling and accurate discrimination capabilities of dual-state complementary features and dual-stream networks in environments with scarce small sample data.

[0065] The second embodiment of the present invention proposes a fault diagnosis system for a three-level ANPC converter based on a dual-stream network, comprising: The signal acquisition and reconstruction module is used to acquire the three-phase AC signals during the fault transient of the three-level ANPC converter at high frequency and splice them into a one-dimensional basic data sequence according to the phase sequence. This module is physically connected to the AC output terminal and neutral point of the converter. By configuring a high-frequency hardware sampling chip, it captures transient waveforms, including the moment when a single power switch is open, within a millisecond-level transient window. Subsequently, the data reconstruction logic unit eliminates the independent isolation between the three-phase channels, reconstructing them into a one-dimensional pipelined data source containing implicit cross-phase linkage evolution information, providing a basic time series tensor for subsequent time-frequency upscaling.

[0066] The continuous evolution feature reconstruction module performs a windowed short-time Fourier transform (STFT) and combines it with a nonlinear logarithmic scaling transform to generate an STFT continuous evolution energy spectrum. This module is cascaded with the output of the signal acquisition and reconstruction module. Internally, it incorporates a sliding frame division unit and a Hamming window multiplication unit. By smoothing out edge abrupt changes introduced by frame truncation, it discretizes the one-dimensional time-series voltage or current signal into a time-frequency decoupled complex matrix. Subsequently, using an embedded nonlinear logarithmic scaling operator and an adaptive logarithmic bias term, it performs large dynamic range compression on the complex amplitude to suppress high-energy fundamental waves and proportionally amplify weak fault harmonics, ultimately outputting an image reflecting the continuous distortion morphology of low-frequency harmonics.

[0067] The transient impact feature reconstruction module performs continuous wavelet transform and synchronous compression rearrangement in parallel, combined with nonlinear logarithmic scaling compression enhancement, to generate an SWT transient impact feature map. This module is cascaded in parallel with the output of the signal acquisition and reconstruction module. Internally, it includes a scale-space detection unit based on analytic Morse wavelets and a synchronous compression rearrangement control unit based on instantaneous frequency ridges. It uses the phase derivative to estimate the instantaneous frequency online and performs centripetal position transport and high-energy accumulation on the divergent wavelet coefficients deviating from the ridge grid points in memory, thereby compacting the morphological boundaries of high-frequency transient jumps. Finally, it combines decibel scaling transform to filter out high-frequency modulation switching clutter, outputting an image with clear vertical ridge features.

[0068] The TSCNN feature fusion and localization module is used to independently extract features from the dual-state spectrum input weighted dual-stream convolutional neural network, employing an element-wise addition strategy for deep coupling, and outputting the localization result via a Softmax classifier. The two parallel inputs of this module are connected to the outputs of the continuous evolution feature reconstruction module and the transient impact feature reconstruction module, respectively. Internally, it includes a bicubic interpolation scale alignment unit, two topologically symmetric and weighted convolutional feature extraction branches, an isotopic element-wise addition feature fusion layer, and a Softmax multi-classifier with 19 output nodes. It activates the feature representation capacity under small sample conditions by losslessly superimposing continuous frequency domain distortion semantics and discrete time domain impact semantics at the same spatial location, ultimately outputting hardware control signals for single-tube open-circuit fault localization accurate to 18 specific power devices in the three phases of the ANPC converter.

[0069] The foregoing has illustrated and described the basic features, principles, and advantages of the present invention. It should be noted that the present invention is not limited to the above embodiments, but only to some embodiments. Any improvements and additions made without departing from the spirit and scope of the present invention are considered to be within the scope of protection of the present invention.

Claims

1. A fault diagnosis method for a three-level ANPC converter based on a dual-stream network, characterized in that, include: The signal acquisition and sequence reconstruction steps involve acquiring the three-phase AC signals during the fault transient of the three-level ANPC converter and reconstructing the sequence to build a one-dimensional basic data sequence that includes cross-phase linkage evolution information. The STFT continuous evolution spectrum construction steps involve performing a windowed short-time Fourier transform on the one-dimensional basic data sequence and combining it with a nonlinear logarithmic scaling transform to generate an STFT continuous evolution energy spectrum characterizing the continuous distortion of low-frequency harmonics. The SWT transient impact map construction steps involve performing continuous wavelet transform and synchronous compression rearrangement in parallel on the one-dimensional basic data sequence, and combining nonlinear logarithmic scaling compression enhancement to generate an SWT transient impact feature map characterizing high-frequency transient jumps. The feature fusion and fault location steps involve inputting the STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum into a weighted, independently weighted, pre-defined dual-stream convolutional neural network for feature extraction. The features are then deeply coupled using an element-by-element addition strategy to determine the converter's operating status and perform multi-class identification. The single-tube open-circuit fault location result of the ANPC converter is obtained through the classifier output.

2. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 1, characterized in that, The construction of the one-dimensional basic data sequence includes real-time acquisition of three-phase voltage and current signals output from the AC side of a three-level ANPC converter controlled by a virtual synchronous generator, extracting signal data windows including the fault occurrence time and the system fault transient response process to generate a three-phase AC voltage timing signal, and splicing the three-phase AC voltage timing signal end to end according to the A, B, C phase sequence to obtain the one-dimensional basic data sequence.

3. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 1, characterized in that, The generation of the STFT continuous evolution energy spectrum includes introducing a sliding window mechanism into the one-dimensional basic data sequence, and obtaining a one-dimensional windowed data frame sequence by truncating and windowing through a Hamming window. The one-dimensional windowed data frame sequence is then subjected to a discrete short-time Fourier transform to extract the low-frequency harmonic continuous distortion features induced by open-circuit faults. Finally, the low-frequency harmonic continuous distortion features are subjected to a nonlinear logarithmic scaling transform to generate the STFT continuous evolution energy spectrum.

4. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 3, characterized in that, The nonlinear logarithmic scaling transformation includes performing amplitude transformation on the feature matrix calculated by the discrete short-time Fourier transform, converting the complex frequency domain features into real features characterizing the physical energy intensity, introducing a nonlinear scaling compression operator to dynamically compress the real features, and mapping the compressed feature data to a time-frequency decoupled two-dimensional discrete rectangular coordinate grid to generate an STFT continuous evolution energy spectrum.

5. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 1, characterized in that, The generation of the SWT transient impact feature map includes: performing continuous wavelet transform on the one-dimensional basic data sequence to obtain a wavelet coefficient matrix; estimating the true instantaneous frequency of the signal based on the phase derivative of the wavelet coefficients; introducing a synchronous compression operator; using the true instantaneous frequency as the energy convergence center; rearranging the spatial positions of the wavelet coefficients in the scale domain to accumulate and converge the divergent transient energy; and performing nonlinear logarithmic scaling compression enhancement on the rearranged wavelet coefficient matrix to output an SWT transient impact energy map with steep time-frequency energy lines.

6. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 5, characterized in that, The nonlinear logarithmic scale compression enhancement includes extracting the rearranged coefficient matrix, performing an amplitude transformation on the coefficient matrix to convert it into real-valued features characterizing physical energy intensity, introducing a nonlinear scale enhancement operator to perform dynamic range series compression and span adjustment on the real-valued features, enhancing transient impact details and tuning the contrast of feature distribution at the decibel scale, and mapping the enhanced real-valued features onto a two-dimensional discrete rectangular coordinate grid decoupled from the time axis and scale axis to generate an SWT transient impact feature map.

7. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 1, characterized in that, The dual-stream convolutional neural network includes two feature extraction branches with equivalent topologies and independent weights. The STFT continuous evolution energy spectrum and the SWT transient impact feature spectrum interpolation are uniformly converted into RGB three-channel feature tensors of the same dimension, and then input to the two feature extraction branches respectively.

8. The fault diagnosis method for a three-level ANPC converter based on a dual-stream network according to claim 1, characterized in that, The element-wise addition strategy includes coupling the feature vector output by the dual-stream convolutional neural network in the feature fusion layer, inputting the coupled feature vector into a Softmax classifier with 19 output nodes, calculating the probability distribution, and outputting the specific single-tube open-circuit fault location results for the 18 power devices in the three phases of the ANPC converter under normal operating conditions.

9. A fault diagnosis system for a three-level ANPC converter based on a two-stream network, applicable to the fault diagnosis method for a three-level ANPC converter based on a two-stream network as described in any one of claims 1 to 8, characterized in that, include: The signal acquisition and reconstruction module is used to acquire the three-phase AC signals during the fault transient of the three-level ANPC converter at high frequency and splice them into a one-dimensional basic data sequence according to the phase sequence. The continuous evolution feature reconstruction module is used to perform windowed short-time Fourier transform and combine it with nonlinear logarithmic scaling transform to generate STFT continuous evolution energy spectrum. The transient impact feature reconstruction module is used to perform continuous wavelet transform and synchronous compression rearrangement in parallel, and combined with nonlinear logarithmic scaling compression enhancement to generate SWT transient impact feature maps. The TSCNN feature fusion and localization module is used to input the dual-state map into a weighted dual-stream convolutional neural network to extract features independently. It adopts an element-wise addition strategy for deep coupling and outputs the localization result through a Softmax classifier.

Images