A deep parallel fusion multi-level inverter single tube open circuit fault diagnosis method

Abstract

This invention provides a deeply parallel fusion method for diagnosing single-tube open-circuit faults in multi-level inverters. This method is adaptable to different multi-level inverter circuits and effectively solves the problems of low positioning accuracy and slow response speed in diagnosing single-tube open-circuit faults in multi-level inverter circuits under complex operating conditions using existing diagnostic technologies. The proposed method constructs a parallel fusion architecture consisting of two front-end deep learning networks and one back-end deep learning network. The method extracts multi-scale time-frequency features of the fault signal through the front-end parallel networks, outputs probability scores, and merges them column-wise. The back-end deep learning network then performs deep fusion on the combined probability scores. The proposed method exhibits good topology adaptability, significantly shortening inference time while ensuring diagnostic robustness. This enables highly reliable, real-time fault monitoring and protection of multi-level inverters in new energy grid-connected systems.

Description

Technical Field

[0001] This invention belongs to the field of power electronics and intelligent fault diagnosis technology, and relates to a deep parallel integration method for diagnosing open-circuit faults in single-tube multilevel inverters. It is applicable to the status monitoring and reliability assurance of power conversion equipment in grid-connected scenarios of new energy power generation systems (such as wind power and photovoltaic). Background Technology

[0002] With the large-scale grid connection of new energy sources, the stable operation of inverters in multi-level inverter circuits is particularly important during the grid connection process. Most new energy systems operate in areas with extreme weather and harsh working environments, making the stable operation of inverters in the inverter circuit a vulnerable link. Inverter failures directly disrupt the energy balance between the grid and the inverter, leading to equipment damage, grid fluctuations, and even safety accidents. However, existing traditional methods for classifying and diagnosing single-tube open-circuit faults in multi-level inverter circuits suffer from low fault location accuracy, slow diagnostic speed, and poor adaptability, failing to meet the needs of today's more complex new energy grid-connected systems. Therefore, quickly and accurately detecting inverter faults in multi-level inverter circuits to take appropriate measures has become a significant challenge.

[0003] Currently, fault detection methods for inverter transistors in multi-level inverter circuits mainly fall into three categories: model-based, signal-based, and data-driven. Model-based methods primarily utilize observers to model the system and then diagnose faults by analyzing the residual signals between the actual and estimated signals. The disadvantages of model-based methods are their strong dependence on the system model and weak robustness and adaptability. While model-based methods are general, their accuracy depends on the accuracy of the model, making them unsuitable for grid-connected systems with variable loads and complex operating conditions. Signal-based methods primarily utilize measured signals to extract fault information features. By extracting fault features from signals such as current, voltage, and temperature, the state can be directly determined. Signal-based methods do not rely on precise models and are adaptable to the complexity of grid-connected scenarios. Signal processing methods include Fast Fourier Transform, Wavelet Transform, and Empirical Mode Decomposition. The disadvantages of signal-based methods are that most require additional sensors, have poor adaptability, and low accuracy in diagnosing complex faults. Although signal-based methods do not rely on models, they suffer from significant computational burdens and load variations, leading to misdiagnosis and making them unsuitable for today's more complex new energy grid-connected systems. Data-driven approaches primarily train models using deep learning neural networks by combining a large amount of grid-connected fault data. Deep learning excels at extracting fault features and exhibits strong generalization capabilities, making it a current trend in fault diagnosis of inverter tubes in multi-level inverter circuits. However, achieving high accuracy with a single neural network requires a sufficiently deep and complex network model structure. This demands high-performance hardware and significantly increases model training time. Inference time may exceed the millisecond-level response requirements for new energy grid connection, leading to stability issues. Summary of the Invention

[0004] This invention proposes a deep parallel fusion method for diagnosing open-circuit faults in single-transistors of multi-level inverters. It fuses the outputs of two deep learning networks through a third deep learning network, enabling high accuracy and fast diagnosis in diagnosing and locating open-circuit faults in single-transistors of multi-level inverter circuits under complex operating conditions. The steps of using this deep parallel fusion method for diagnosing open-circuit faults in single-transistors of multi-level inverters are as follows:

[0005] Step (1): Setup H-bridge cascade Level inverter circuit, Indicates the number of phases in the inverter circuit and ; This indicates the number of voltage levels in the inverter circuit. It is an odd number greater than 3; H-bridge cascade There are a total of One inverter tube; this invention Mutually The level inverter circuit is not limited to the H-bridge cascade topology; it can also be applied to diode-clamped, flying capacitor, modular, hybrid, and T-type topologies. Mutually Different topologies of level inverter circuits correspond to different numbers of inverter transistors;

[0006] Step (2): Set each pipe to experience an open-circuit fault individually, so that... The sampling frequency is 1 second per second to collect load-side voltage and current data, as well as load-side voltage and current data under normal conditions without inverter tube faults. The duration of the time interval between two samplings; sampling Each cycle duration, The number of sampling periods; H-bridge cascade Level inverter circuit One set of fault data and one set of normal data; in this invention The value can be set to , and etc.; with other parameters remaining unchanged, An excessively large dataset means that each subsequent dataset contains sampling data points with a longer duration, which in turn leads to a longer diagnosis time. If the sample size is too small, each dataset will contain too little information, which will affect the accuracy of diagnosis. Then it can be 1.5 or 2, etc.; The value of will affect the number of samples in the subsequent dataset. Too small a value will result in incomplete sampling data, while too large a value will consume more data memory; furthermore, voltage and current data should be collected simultaneously, not just voltage or current data; the reason is that... H-bridge cascade Level inverter circuits are not linear load circuits. When an inverter tube on the same bridge arm fails, the voltage or current signal on the load side may be the same. In this case, if only voltage or current data is used for subsequent processing and training, it will be impossible to accurately diagnose and locate which tube is faulty. For example, in a single-phase H-bridge cascaded 5-level inverter circuit, when inverter tubes 4 and 5 have a single-tube open-circuit fault, the voltage waveform on the load side is the same. Therefore, analyzing the voltage alone cannot diagnose and locate these two inverter tubes; the current signal is also required. Similarly, in a single-phase H-bridge cascaded 7-level inverter circuit, when inverter tubes 3 and 7 have a single-tube open-circuit fault, the current waveform on the load side is the same. Therefore, analyzing the current alone cannot diagnose and locate these two inverter tubes; the voltage signal is also required.

[0007] Step (3): Preprocess the collected voltage and current data; using the voltage and current data under normal conditions as the reference value, process the data for the first step. The voltage and current data collected during root canal malfunction are compared with a reference value using a differential processing method. This invention does not limit the use of normal voltage and current data as the reference value during preprocessing; voltage and current data from the malfunctioning canal can also be used as the reference value, and even different reference values ​​can be selected for voltage and current data from different canals. The voltage and current data after differential processing can also be multiplied by a certain factor (e.g., ...). , or (etc.) The difference between the amplifying tubes increases the distinguishability;

[0008] The voltage and current data after differential processing are as follows:

[0009]

[0010]

[0011] In the formula, The inverter tube is numbered, with an initial value of 1; The voltage data has undergone differential processing; For the first Voltage data collected during root canal malfunction; Voltage data under normal conditions; The current data has undergone interpolation processing; For the first Current data collected during root canal malfunction; Normal current data; in this step, if the voltage and current data are transformed using Fourier transform... and Alternatively, perform a differentiation operation on the voltage and current data to transform it into... and to replace and Following these steps can lead to confusion during inverter tube diagnosis; for example, in a single-phase H-bridge cascaded 5-level inverter circuit, regardless of whether Fourier transform, differentiation, integration, or a combination thereof is used... and This will affect the diagnosis of inverter tubes 4 and 5; however, this problem can be solved by using differential processing, i.e., the method of this invention.

[0012] Step (4): The voltage and current data after the difference processing are merged with the original voltage and current data in the column direction. After merging, a total of List;

[0013] The data after merging by column direction is as follows:

[0014]

[0015] In the formula, For the first Data merged in column direction when there is a root canal malfunction; The voltage data has undergone differential processing; For the first Voltage data collected during root canal malfunction; The current data has undergone interpolation processing; For the first Current data collected during root canal malfunction;

[0016] Furthermore, the data merging in this step does not require specific ordering of the voltage and current data; they can be merged in any order. If this step is not present, only the following steps are used: replace This can lead to missing voltage and current data, affecting the model's diagnostic accuracy for faulty inverter transistors; for example, in a single-phase H-bridge cascaded 5-level inverter circuit, only using... Replacing this step will cause inverter tubes 4 and 5 to be confused as the same type when performing fault diagnosis and location.

[0017] Step (5): Extract Data and scaled to size size;

[0018] in, for The row number in the code is initialized to 1. Indicates the truncation length, which is an integer greater than 0; It can be configured according to actual needs. Desirable , , wait; Represented as The Middle Arrive at the OK; This refers to the length of the data after scaling. This refers to the width size after scaling the data. and It can be set to 224, 227, 256, and 299, etc., as needed; in this step, The value of determines how much information about faulty inverter tubes is contained in the dataset samples; A larger value means more fault information is contained, but it also means that each dataset sample contains more sampling data points over a longer period, thus leading to a longer diagnosis time; If the value is too small, each dataset sample will contain less fault information, making it difficult to train and achieve high accuracy; therefore, a reasonable value needs to be selected. The value is determined first, and then the size is scaled up to achieve high accuracy while minimizing diagnostic time. Without the size scaling operation, the data size will be too small, making it difficult to train the model properly in subsequent network training.

[0019] Step (6): Input the scaled voltage and current data into... Store and generate corresponding category tags ;

[0020] in, The input data is a 4-dimensional matrix; The output labels for the classification type are a 1D matrix. H-bridge cascade Level inverter circuit has Different classifications;

[0021] Step (7): Repeat steps (5) and (6) to perform the next set of sample data truncation, scaling, and storage operations; until the number of stored samples is reached. achieve ;

[0022] in, Indicates the sliding window step size, which is an integer greater than 0; A positive integer, representing the total number of samples required for each tube; Values ​​can be retrieved ,and It is an integer; Desirable ,and It is an integer; in this invention The value of will affect the sampling duration in step (2). And subsequent training accuracy; If the dataset is too large, the overlap between samples will be too low, resulting in low training accuracy, and at the same time, more data is needed. Also bigger; If the sample size is too small, the overlap between the samples in the dataset will be too high, making it difficult to represent all the voltage and current states of the faulty tube in one cycle with a small number of samples; thus, more samples are needed for training, and more time is needed for model training. If the value is too large, the dataset will become too large, which will require more time to train the model. Taking values ​​that are too small will result in insufficient training samples, leading to low accuracy.

[0023] Step (8): Repeat steps (3), (4), (5), (6) and (7) to perform the second step. Root canal voltage and current data processing, up to the first The root canal sample data has been processed.

[0024] Step (9): Form a size of Input data and corresponding size is Output category labels , and Together they form the dataset;

[0025] Step (10): Shuffle and corresponding And in accordance with The proportions are divided into training set and test set. The size of the training set; in this invention The value can be 7, 8, or 9, etc.

[0026] Step (11): Build a deep learning network 1 model and train network 1 using the training set. The deep learning network 1 model takes an image input as the input layer and a fully connected layer, activation function, and classification layer as the output module. The training input data and training output labels are trained by the deep learning network 1 model to output the probability score of network 1. Any network model that conforms to the principle of taking an image input as the input layer and a fully connected layer, activation function, and classification layer as the output module can be used as the deep learning network 1 model in this step. This includes network models such as SqueezeNet, GoodLeNet, ResNet-50, EfficientNet-b0, DarkNet-53, DarkNet-19, ShuffleNet, Xception, MobileNet-v2, DenseNet-201, ResNet-18, Inception-v3, ResNet-101, VGG-19, VGG-16, AlexNet, Transformer, and transfer learning.

[0027] For example, if the deep learning network 1 model uses the ResNet-18 network, the residual blocks in the ResNet-18 network can enhance gradient flow, alleviate the gradient vanishing problem, and allow the network to learn residuals instead of the original mappings, thus making it easier to optimize.

[0028] The output of the feature map after passing through the residual block is:

[0029]

[0030] In the formula, Input feature map; This is the first convolution; This is the first batch normalization; For activation functions; This is the second convolution; This is the second batch normalization; for Output feature map after residual block;

[0031] The output of the feature map after passing through the residual structure is:

[0032]

[0033] In the formula, Input feature map; It is an identity mapping; when the number of channels remains constant, When the number of channels or the size changes, , for Convolution kernel; for Output feature map after residual block; For activation functions; for The output after passing through the residual structure;

[0034] Step (12): Build a deep learning network 2 model and train the network 2 using the training set. The deep learning network 2 model takes an image input as the input layer and a fully connected layer, activation function, and classification layer as the output module. The training input data and training output labels are trained by the deep learning network 2 model to output the probability score of the network 2 model. Any network model that conforms to the principle of taking an image input as the input layer and a fully connected layer, activation function, and classification layer as the output module can be used as the deep learning network 2 model in this step. This includes network models such as SqueezeNet, GoodLeNet, ResNet-50, EfficientNet-b0, DarkNet-53, DarkNet-19, ShuffleNet, Xception, MobileNet-v2, DenseNet-201, ResNet-18, Inception-v3, ResNet-101, VGG-19, VGG-16, AlexNet, Transformer, and transfer learning.

[0035] For example, if the deep learning network 2 model uses the MobileNet-v2 network, the inverted residual block in the MobileNet-v2 network first expands the low-dimensional input to a high-dimensional space for nonlinear transformation, and then compresses it back to the low-dimensional output, while retaining the information direct path, which can achieve the optimal balance between accuracy and efficiency.

[0036] The output of the feature map after passing through the inverted residual block is:

[0037]

[0038] In the formula, Input feature map; This is the first convolution; This is the first batch normalization; For activation functions; for Depthwise convolution; This is the second batch normalization; For activation functions; This is the third convolution; This is the third batch normalization; for Output feature map after residual block;

[0039] When the number of input channels and the number of output channels are equal, the output of the feature map after passing through the inverted residual structure is:

[0040]

[0041] In the formula, Input feature map; for Output feature map after inverted residual block; for The output after passing through the inverted residual structure;

[0042] When the number of input channels and the number of output channels are not equal, the output of the feature map after passing through the inverted residual structure is:

[0043]

[0044] In the formula, for Output feature map after inverted residual block; for The output after passing through the inverted residual structure;

[0045] Step (13): Construct a deep learning network 3 framework. The deep learning network 3 model uses feature input as the input layer and fully connected layers, activation functions, and classification layers as output modules. Any network model that conforms to the present invention, using feature input as the input layer and fully connected layers, activation functions, and classification layers as output modules, can be used as the deep learning network 3 model in this step. This includes network models such as SqueezeNet, GoodLeNet, ResNet-50, EfficientNet-b0, DarkNet-53, DarkNet-19, ShuffleNet, Xception, MobileNet-v2, DenseNet-201, ResNet-18, Inception-v3, ResNet-101, VGG-19, VGG-16, AlexNet, Transformer, and transfer learning.

[0046] For example, if the deep learning network 3 uses an intermediate deep learning module composed of fully connected layers, ReLU activation function and dropout layer, the deep learning network 3 can automatically learn the optimal fusion weights, abstract and reorganize the input features layer by layer, and further mine more complex and discriminative high-level semantic information from the features identified at the feature front end, and finally improve the model fusion effect.

[0047] The output of the deep fusion module after the fused input feature matrix is:

[0048]

[0049] In the formula, The input feature matrix after fusion; These are the weights of the fully connected layer; For biasing the fully connected layer; For activation functions; This is a random discard operation; for The output after passing through the deep learning module;

[0050] Step (14): Combine the probability scores from steps (11) and (12) in the column direction and use them as input to the deep learning network 3, and train the deep learning network 3 together with the training labels;

[0051] Step (15): Input the test data into the trained deep learning network 1 to obtain the probability score of the test data;

[0052] Step (16): Input the test data into the trained deep learning network 2 to obtain the probability score of the test data;

[0053] Step (17): Combine the probability scores from steps (15) and (16) in the column direction, and then input them into the trained deep learning network 3 to finally obtain the results. H-bridge cascade Diagnostic results of a single open-circuit fault in a level inverter circuit.

[0054] The present invention has the following advantages and effects compared with the prior art:

[0055] (1) Existing deep learning-based fault classification and diagnosis methods improve network accuracy by increasing network depth and width. This requires high-performance hardware and significantly increases model training time. Inference time may exceed the millisecond-level response requirements for new energy grid connection, which can easily lead to stability problems. Generally speaking, the prediction accuracy of a single network is limited. However, this invention fuses the probability scores of the outputs of deep learning network 1 and deep learning network 2 through a third deep learning network from the perspective of network output. This preserves local details while enriching features, thereby improving model accuracy. This invention uses the collected voltage and current data as fault features to train the basic network, and then fuses the probability scores of the outputs of the two single networks. By fusing the results and characteristics of the outputs of the two individual networks, the diagnostic accuracy is improved.

[0056] (2) Generally speaking, more samples mean more accurate training of the deep network. However, this invention converts the fault sample dataset into an image input format to adapt to the network model, achieving efficient training with a small number of data samples. This invention converts voltage and current data into 4D graph matrix data through data preprocessing, improving data utilization. A small amount of sampled data can amplify fault features through data preprocessing, enabling rapid and accurate fault diagnosis and timely decision support for subsequent fault isolation and maintenance.

[0057] (3) Typically, increasing network depth and width, and repeatedly adjusting parameters for repetitive training consumes a significant amount of time and effort for researchers, resulting in unsatisfactory training outcomes. In contrast, this invention employs a method where deep learning networks 1, 2, and 3 are trained independently, reducing the time and effort required by researchers and minimizing the memory burden on computers during synchronous training. Compared to a single, massive, deep, and wide network, this invention first trains networks 1 and 2 separately, then merges the probability scores obtained from training networks 1 and 2 and inputs them into deep learning network 3 for deep fusion training. If the accuracy of the results is low, the simpler network 3 can be run multiple times, allowing it to complete training quickly. This avoids the process of repeatedly training a single network from scratch and reduces the memory burden on computers during synchronous training. Attached Figure Description

[0058] Figure 1 This is a flowchart of the network usage of the method of the present invention.

[0059] Figure 2 This is a flowchart of the data preprocessing method of the present invention.

[0060] Figure 3 This is the overall structure diagram of the ResNet-18 network model of this invention.

[0061] Figure 4 This is the overall structure diagram of the MobileNet-v2 network model of this invention.

[0062] Figure 5 This is a diagram of the overall structure of the deep fusion network model pre-defined by the method of this invention.

[0063] Figure 6 This is a structural diagram of the residual block in the ResNet-18 network model of the present invention.

[0064] Figure 7 This is a structural diagram of the inverted residual block in the MobileNet-v2 network model of the present invention.

[0065] Figure 8 This is a structural diagram of the feature fusion input module of the method of the present invention. Detailed Implementation

[0066] The present invention proposes a method for diagnosing open-circuit faults in a single transistor of a multi-level inverter with deep parallel fusion, which is described in detail below with reference to the accompanying drawings:

[0067] Figure 1 This is a flowchart of the network training and testing process of the present invention. First, the network is trained. The training data and corresponding labels are input into deep learning network 1 for training, resulting in probability scores for network 1. Then, the training data and corresponding labels are input into deep learning network 2 for training, resulting in probability scores for network 2. Next, the probability scores of networks 1 and 2 are merged column-wise and used as input to deep learning network 3, along with the corresponding labels, for training. This completes the training of networks 1, 2, and 3. Then, the network is tested. Test data is input into the trained deep learning network 1 to obtain output probability scores. Then, the test data is input into the trained deep learning network 2 to obtain output probability scores. Finally, the probability scores of networks 1 and 2 are merged column-wise and input into the trained deep learning network 3, resulting in the test results.

[0068] Figure 2 This is a flowchart of the data preprocessing method of the present invention. First, the parameters are initialized. Then begin processing the first... Root canal and setting Next, the first During root canal malfunctions, voltage and current data collected are compared with baseline values ​​using a differential process. The differentially processed data is then merged with the original, undifferentiated voltage and current data column-wise. Then, the data is truncated. Arrive at the The merged data is then scaled to the desired size. Size; then insert 4D data The first in Store data in each location and generate corresponding category labels. ; then renew and judge Is the result greater than If it is not greater than, then return to the truncation point. Arrive at the The process of merging rows and data continues in a loop; if... The result is greater than Then proceed And update Next, make a judgment. Is it greater than If it is not greater than, return to the beginning of the process. Root canal and setting This step; if Greater than This indicates that all data processing is complete, and 4-dimensional input data has been obtained. and the corresponding output tags Finally press The dataset is divided into training and testing sets.

[0069] Figure 3 This is a diagram showing the overall structure of the ResNet-18 network model of this invention. Taking ResNet-18 as an example, the specific content of the ResNet-18 network model is as follows:

[0070] Connect to one image input layer;

[0071] Then add another 7×7 convolutional layer;

[0072] Then add one more batch normalization layer;

[0073] Then add another activation function layer, namely ReLU;

[0074] Then connect to one more max pooling layer;

[0075] It then splits into two branches;

[0076] Branch 1 connects to one residual block;

[0077] After connecting the residual block, branches 1 and 2 are added together and then connected to an activation function, ReLU.

[0078] It then splits into two branches;

[0079] Branch 3 connects to one residual block;

[0080] After connecting the residual block, branches 3 and 4 are added together and then connected to an activation function, ReLU.

[0081] It then splits into two branches;

[0082] Branch 5 connects to one residual block;

[0083] Branch 6 connects to one convolutional layer;

[0084] Then add a batch normalizer after the convolutional layer;

[0085] Branch 5, which connects to the residual block, and branch 6, which connects to the convolutional layer and after batch normalization, are added together and then connected to an activation function, ReLU.

[0086] It then splits into two branches;

[0087] Branch 7 connects to one residual block;

[0088] After connecting the residual block, branches 7 and 8 are added together and then connected to an activation function, ReLU.

[0089] It then splits into two branches;

[0090] Branch 9 connects to one residual block;

[0091] Branch 10 connects to one convolutional layer;

[0092] Then add a batch normalizer after the convolutional layer;

[0093] Branch 9, which connects to the residual block, and branch 10, which connects to the convolutional layer and after batch normalization, are added together and then connected to an activation function, ReLU.

[0094] It then splits into two branches;

[0095] Branch 11 connects to one residual block;

[0096] After connecting the residual block, branches 11 and 12 are added together and then connected to an activation function, ReLU.

[0097] It then splits into two branches;

[0098] Branch 13 connects to one residual block;

[0099] Branch 14 connects to one convolutional layer;

[0100] Then add a batch normalizer after the convolutional layer;

[0101] Branch 13 after accessing the residual block and branch 14 after accessing the convolutional layer and batch normalization are added together and then connected to an activation function, namely ReLU.

[0102] It then splits into two branches;

[0103] Branch 13 connects to one residual block;

[0104] After connecting the residual block, branches 13 and 14 are added together and then connected to an activation function, ReLU.

[0105] Then add one more average pooling layer;

[0106] Add another fully connected layer;

[0107] Then, another activation function, ReLU, is added.

[0108] Finally, a classification layer is added.

[0109] Figure 4 This is a diagram showing the overall structure of the MobileNet-v2 network model of this invention. Taking MobileNet-v2 as an example, the specific content of the MobileNet-v2 network model is as follows:

[0110] Connect to one image input layer;

[0111] Then add two more inverted residual blocks;

[0112] It then splits into two branches;

[0113] Branch 1 connects to an inverted residual block;

[0114] After connecting to the inverted residual block, branches 1 and 2 are added together and then connected to another inverted residual block;

[0115] It then splits into two branches;

[0116] Branch 3 connects to an inverted residual block;

[0117] After connecting the inverted residual block, branches 3 and 4 are added together and then divided into two branches;

[0118] Branch 5 connects to an inverted residual block;

[0119] After connecting to the inverted residual block, branches 5 and 6 are added together and then connected to another inverted residual block;

[0120] It then splits into two branches;

[0121] Branch 7 connects to an inverted residual block;

[0122] After connecting the inverted residual block, branches 7 and 8 are added together and then divided into two branches;

[0123] Branch 9 connects to an inverted residual block;

[0124] After connecting the inverted residual block, branches 9 and 10 are added together and then divided into two branches;

[0125] Branch 11 connects to an inverted residual block;

[0126] After connecting to the inverted residual block, branches 11 and 12 are added together and then connected to another inverted residual block;

[0127] It then splits into two branches;

[0128] Branch 13 connects to an inverted residual block;

[0129] After connecting the inverted residual block, branches 13 and 14 are added together and then divided into two branches;

[0130] Branch 15 connects to an inverted residual block;

[0131] Branches 15 and 16, after being connected to the inverted residual block, are added together and then connected to another inverted residual block;

[0132] It then splits into two branches;

[0133] Branch 17 connects to an inverted residual block;

[0134] After connecting the inverted residual block, branches 17 and 18 are added together and then divided into two branches;

[0135] Branch 19 connects to an inverted residual block;

[0136] After connecting to the inverted residual block, branches 19 and 20 are added together and then connected to another inverted residual block;

[0137] Then add another 1×1 convolutional layer;

[0138] Then add one more batch normalization;

[0139] Then add another activation function, ReLU6;

[0140] Then add one more average pooling layer;

[0141] Add another fully connected layer;

[0142] Then add another activation function, namely Softmax;

[0143] Finally, a classification layer is added.

[0144] Figure 5 This is a diagram of the overall structure of the deep fusion network model pre-defined by the method of this invention. Taking deep learning network 3, which uses fully connected layers, ReLU activation functions, and dropout layers as the intermediate deep learning module, as an example, the specific content of the pre-defined deep fusion network model is as follows:

[0145] Connect to one feature input layer;

[0146] Add another fully connected layer;

[0147] Then add another activation function layer, namely ReLU;

[0148] Connect another discard layer;

[0149] Add another fully connected layer;

[0150] Then add another activation function layer, namely Softmax;

[0151] Finally, a classification layer is added as the output.

[0152] Figure 6 This is a structural diagram of the residual block in the ResNet-18 network model of the present invention. The specific content of the residual block is as follows:

[0153] The input is then fed into a 3×3 convolutional layer.

[0154] Then add one more batch normalization layer;

[0155] Then add another activation function layer, namely ReLU;

[0156] Then add another 3×3 convolutional layer;

[0157] Then add one more batch normalization layer;

[0158] Final output.

[0159] Figure 7 This is a structural diagram of the inverted residual block in the MobileNet-v2 network model of the present invention. The specific content of the inverted residual block is as follows:

[0160] The input is then fed into a 1×1 convolutional layer.

[0161] Then add one more batch normalization layer;

[0162] Then add another activation function layer, namely ReLU6;

[0163] Then add another 3×3 convolutional layer;

[0164] Then add one more batch normalization layer;

[0165] Then add another activation function layer, namely ReLU6;

[0166] Then add another 1×1 convolutional layer;

[0167] Then add one more batch normalization layer;

[0168] Final output.

[0169] Figure 8 This is a structural diagram of the feature fusion input module of the method of the present invention. The specific contents of the feature fusion input module are as follows:

[0170] Network 1 and Network 2 each correspond to two branches, namely branch 1 and branch 2;

[0171] Branch 1 is followed by an activation function layer, namely Softmax, and then outputs a probability score of 1.

[0172] Branch 2 is followed by an activation function layer, namely Softmax, and then outputs a probability score of 2.

[0173] Probability score 1 and probability score 2 are merged in the column direction to form branch 3;

[0174] Branch 3 is followed by a feature input layer;

[0175] Final output.

[0176] The above description is only a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for diagnosing open-circuit faults in a single transistor of a deeply parallel-integrated multi-level inverter, characterized in that, The outputs of two deep learning networks are fused using a third deep learning network to perform feature fusion. This method is used for fault diagnosis and localization of single-transistor open-circuit faults in multi-level inverter circuits under complex operating conditions, achieving high accuracy and fast diagnostic speed. The steps in its application are as follows: Step (1): Setup H-bridge cascade Level inverter circuit, Indicates the number of phases in the inverter circuit and ; This indicates the number of voltage levels in the inverter circuit. It is an odd number greater than 3; H-bridge cascade There are a total of One inverter tube; Step (2): Set each pipe to experience an open-circuit fault individually, so that... The sampling frequency is 1 second per second to collect load-side voltage and current data, as well as load-side voltage and current data under normal conditions without inverter tube faults. The duration of the time interval between two samplings; sampling Each cycle duration, The number of sampling periods; H-bridge cascade Level inverter circuit One set of fault data and one set of normal data; Step (3): Preprocess the collected voltage and current data; using the voltage and current data under normal conditions as the reference value, process the data for the first step. The voltage and current data collected during root canal malfunctions are compared with the baseline values ​​using a differential calculation. The voltage and current data after differential processing are as follows: In the formula, The inverter tube is numbered, with an initial value of 1; The voltage data has undergone differential processing; For the first Voltage data collected during root canal malfunction; Voltage data under normal conditions; The current data has undergone interpolation processing; For the first Current data collected during root canal malfunction; Current data under normal conditions; Step (4): The voltage and current data after the difference processing are merged with the original voltage and current data in the column direction. After merging, a total of List; The data after merging by column direction is as follows: In the formula, For the first Data merged in column direction when there is a root canal malfunction; The voltage data has undergone differential processing; For the first Voltage data collected during root canal malfunction; The current data has undergone interpolation processing; For the first Current data collected during root canal malfunction; Step (5): Extract Data and scaled to size size; in, for The row number in the code is initialized to 1. Indicates the truncation length, which is an integer greater than 0; Represented as The Middle Arrive at the OK; This refers to the length of the data after scaling. This refers to the width size of the data after scaling. Step (6): Input the scaled voltage and current data into... Store and generate corresponding category tags ; in, The input data is a 4-dimensional matrix; The output labels for the classification type are a 1D matrix. H-bridge cascade Level inverter circuit has Different classifications; Step (7): Repeat steps (5) and (6) to perform the next set of sample data truncation, scaling, and storage operations; until the number of stored samples is reached. achieve ; in, Indicates the sliding window step size, which is an integer greater than 0; A positive integer, representing the total number of samples required for each tube; Step (8): Repeat steps (3), (4), (5), (6) and (7) to perform the second step. Root canal voltage and current data processing, up to the first The root canal sample data has been processed. Step (9): Form a size of Input data and corresponding size is Output category labels , and Together they form the dataset; Step (10): Shuffle and corresponding And in accordance with The proportions are divided into training set and test set. This refers to the proportion of the training set. Step (11): Build a deep learning network 1 model and train the network 1 using the training set. The deep learning network 1 model takes the image input as the input layer and the fully connected layer, activation function and classification layer as the output module. The training input data and training output labels are trained by the deep learning network 1 model to output the probability score of network 1. Step (12): Build a deep learning network 2 model and train the network 2 using the training set. The deep learning network 2 model takes the image input as the input layer and the fully connected layer, activation function and classification layer as the output module. The training input data and training output labels are trained by the deep learning network 2 model to output the probability score of the network 2. Step (13): Build a deep learning network 3 framework. The deep learning network 3 model uses feature input as the input layer and fully connected layer, activation function and classification layer as the output module. Step (14): Combine the probability scores from steps (11) and (12) in the column direction and use them as input to the deep learning network 3, and train the deep learning network 3 together with the training labels; Step (15): Input the test data into the trained deep learning network 1 to obtain the probability score of the test data; Step (16): Input the test data into the trained deep learning network 2 to obtain the probability score of the test data; Step (17): Combine the probability scores from steps (15) and (16) in the column direction, and then input them into the trained deep learning network 3 to finally obtain the results. H-bridge cascade Diagnostic results of a single open-circuit fault in a level inverter circuit.

Images