Early fault diagnosis method and system for dc-dc converter based on improved se-resnet

Abstract

The application discloses a DC-DC converter early fault diagnosis method and system based on an improved SE-Resnet, wherein the method comprises the following steps: determining the early fault type of a DC-DC converter and obtaining the corresponding sample data; adopting a Gram angle field to perform two-dimensional image coding on the time domain signal of the DC-DC converter, converting the sample signal of the DC-DC converter into different images, pre-processing the images, obtaining an image data set, and dividing the image data set into a training set, a verification set and a test set; training an early fault diagnosis model based on the image data set and an improved SE-Resnet model, and using the model to complete fault diagnosis classification. The application adopts the improved SE-Resnet model, can effectively extract the weak features of the early faults of key components of the DC-DC converter, thereby identifying the early soft faults caused by the performance degradation of the key components of the DC-DC converter, and has the characteristics of high accuracy and strong generalization ability.

Description

Technical Field

[0001] This application relates to a circuit fault diagnosis method, belonging to the field of fault diagnosis technology, and particularly to an early fault diagnosis method and system for DC-DC converters based on an improved SE-Resnet. Background Technology

[0002] DC-DC converters are widely used in the power systems of aircraft, launch vehicles, missiles, radar antennas, and other equipment. Their failures affect the overall performance of the power system, and their stability and reliability directly impact the equipment's effectiveness. Traditional DC-DC converter fault diagnosis methods include data-driven and model-driven methods. Both rely on manual feature extraction and expert knowledge, and therefore have certain diagnostic limitations.

[0003] Currently, fault diagnosis methods for DC-DC converters mainly focus on soft and hard faults caused by large deviations in component parameters. The output changes caused by these faults are observable. However, some faults are caused by subtle degradation of critical components. These faults have weak characteristics and are called early soft faults, i.e., the minimum parameter variation of the component is within the tolerance range of 5%. Summary of the Invention

[0004] According to one aspect of this application, a fault diagnosis method is provided that can effectively identify subtle early degradation of internal components of a DC-DC converter.

[0005] The method for early fault diagnosis of DC-DC converters based on the improved SE-Resnet includes:

[0006] (1) Determine the early fault type of the DC-DC converter and obtain its corresponding sample data;

[0007] (2) The time-domain signal of the DC-DC converter is encoded in two dimensions using Gram angle field, and the image is preprocessed to obtain an image dataset;

[0008] (3) Based on the image dataset and the improved SE-Resnet model, an early fault diagnosis model is trained;

[0009] (4) Use the early fault diagnosis model to complete the fault diagnosis classification.

[0010] Preferably, the early fault types include at least one of the following: N1-N3 and F1-F7, wherein N1-N3 is a normal set of device parameters that vary within the tolerance range, and F1-F7 is a set of early soft faults where device parameters vary beyond the tolerance range.

[0011] Preferably, the sample data is obtained by using the PSpice software via Monte Carlo analysis.

[0012] Preferably, in step (2), the size of the Gram differential angle field image obtained after encoding the two-dimensional image is 28×28.

[0013] Preferably, the preprocessing includes: converting the image size to 224×224 using data augmentation, and using it as input features for the improved SE-Resnet model.

[0014] Preferably, the image dataset includes: preprocessed images randomly divided into a training set, a test set, and a validation set in a ratio of 6:3:1.

[0015] Preferably, the BasicBlock of each network layer in the improved SE-Resnet model contains 3, 3, 4, and 3 units respectively.

[0016] Preferably, the improved SE-Resnet model is as follows: based on the SE-Resnet model, the Ranger optimizer and h-swish activation function are selected to replace the Adam optimizer and ReLU activation function, the label-smoothed cross-entropy loss function is selected to replace the cross-entropy function, and the hyperparameters of the model are optimized.

[0017] Preferably, the hyperparameters include at least one of the following: learning rate, batch size, Ranger optimizer parameters, and adjustable parameters of the loss function.

[0018] According to another aspect of this application, an early fault diagnosis system for DC-DC converters based on an improved SE-Resnet is provided, the system comprising:

[0019] The data module is used to determine the early fault type of the DC-DC converter and obtain its corresponding sample data;

[0020] The processing module is used to perform two-dimensional image encoding on the time-domain signal of the DC-DC converter using Gram angle field, and to preprocess the image to obtain an image dataset;

[0021] The diagnostic module is used to train an early fault diagnosis model based on the image dataset and the improved SE-Resnet model, and to use the model to complete fault diagnosis and classification.

[0022] The beneficial effects that this application can produce include:

[0023] 1) This application uses Gramian Angular Field (GAF) technology to convert the one-dimensional time-domain signal of the DC-DC converter into a two-dimensional image dataset, which is convenient for input into the Deep Residual Network (ResNet) and effectively improves the fault diagnosis rate and generalization ability of the DC-DC converter.

[0024] 2) This application employs an improved SE-ResNet model, introducing the SE attention mechanism into the traditional ResNet50 network to improve its residual module. It selects the h-swish activation function and the Ranger optimizer to ensure efficient image feature learning and improve model performance. A label-smoothing cross-entropy loss function is chosen to improve image classification accuracy and prevent overfitting. This application achieves high-accuracy fault identification of weak early-stage faults.

[0025] 3) This application has streamlined and optimized the network architecture in the model, changing the number of units in the BasicBlock of each network layer to 3, 3, 4, and 3 respectively, thereby reducing memory consumption and improving the computational power of the diagnostic model.

[0026] 4) This application can classify early soft faults in DC-DC converters, effectively improving the early fault diagnosis performance of power supply systems. This application can simultaneously and accurately identify minor performance degradation within the tolerance range of multiple key components, with a simulation accuracy of up to 99.32% and a hardware experiment accuracy of up to 94.92%, effectively enhancing the reliability of power supply systems. Attached Figure Description

[0027] Figure 1 This is a flowchart illustrating an early fault diagnosis method for DC-DC converters based on an improved SE-Resnet, as described in this application.

[0028] Figure 2 The residual unit of the improved SE-Resnet model in Embodiment 1 of this application;

[0029] Figure 3 This is a network structure diagram of the improved SE-Resnet model in Embodiment 1 of this application;

[0030] Figure 4 The accuracy curves of the training set and validation set of the model in Embodiment 1 of this application are shown.

[0031] Figure 5 The loss curves for the training and validation sets of the model in Embodiment 1 of this application are shown.

[0032] Figure 6This is the fault diagnosis confusion matrix of the improved SE-Resnet model in Embodiment 1 of this application. Detailed Implementation

[0033] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.

[0034] Please see Figure 1 This paper presents a flowchart illustrating an early fault diagnosis method for DC-DC converters based on an improved SE-ResNet. The method includes:

[0035] (1) Determine the early fault types of various components in the DC-DC converter and obtain their corresponding sample data;

[0036] (2) The time-domain signal of the DC-DC converter is encoded in two dimensions using Gram angle field, and the image is preprocessed to obtain an image dataset;

[0037] (3) Based on the image dataset and the improved SE-Resnet model, an early fault diagnosis model is trained;

[0038] (4) Use the early fault diagnosis model to complete the fault diagnosis classification.

[0039] Example 1

[0040] Step (1) specifically includes:

[0041] (11) The tolerance of the components is set to 5%. For the five components, namely inductor, inductor equivalent series resistance, capacitor, inductor equivalent series resistance and MOSFET equivalent resistance, the minimum degradation deviation (tolerance) of the early soft fault identifiable parameters is set to 5%.

[0042] (12) The sample data of the circuit is obtained by Monte Carlo analysis using PSpice software.

[0043] Step (2) specifically includes:

[0044] (21) The one-dimensional time series data of the DC-DC converter is converted into two-dimensional image data using GAF technology, which facilitates fault classification of the image by the deep residual network. GAF is divided into two types of images: Gram summation angle field (GASF) and Gram difference angle field (GADF). The calculation process of GASF and GADF is as follows:

[0045] Let X = {x1, x2, x3, ..., xn} be a one-dimensional time-domain signal with n values ​​from a DC-DC converter. n}, normalize the time series to the range [-1, 1], and obtain

[0046] The normalized time series is represented in polar coordinates as shown in the following formula:

[0047]

[0048] In the formula, t i For timestamps, N represents dividing the unit length of the polar radius coordinate into N equal parts.

[0049] Using trigonometric functions and angle formulas, the correlation between each time point is defined, and GASF and GADF are defined as follows:

[0050]

[0051]

[0052] In the formula, T represents transpose, φ i (i = 1, 2, ..., n) represents the angle in polar coordinates at the i-th moment;

[0053] Using the above calculation formula, the one-dimensional time-domain signals of 10 fault types of DC-DC converters (as shown in Table 1 below) are converted into 28×28 GAF two-dimensional images.

[0054] Table 1. Early Soft Fault Types of Various Components in Buck Circuits

[0055]

[0056] Electrolytic capacitors, inductors, and MOSFETs are inextricably linked to the overall performance degradation of the DC-DC converter. Based on component degradation models and fault mechanism analysis, the degradation of capacitors and inductors shows a decreasing trend, while the degradation of the equivalent series resistance shows an increasing trend. Therefore, fault injection was performed on the Buck simulation circuit in PSpice software, with fault type settings shown in Table 1. N1-N3 represent the normal set of device parameter variations within the tolerance range (5%), while F1-F7 represent the early soft fault set of device parameter variations exceeding the tolerance range (5%).

[0057] (22) Preprocess the 28×28 GAF two-dimensional image dataset using the transform function from torchvision toolbox:

[0058] Use transforms.RandomResizedCrop(size=256, scale=(0.8,1.0)) to randomly crop the image to 256×256;

[0059] Use transforms.CenterCrop(size=224) to crop the image center to 224×224, and transforms.ToTensor() to convert it into a tensor.

[0060] The tensors obtained are normalized using transforms.Normalize([0.485,0.456,0.406], [0.229,0.224,0.225]).

[0061] The improved SE-Resnet model uses the Resnet50 network as its base model and makes improvements on it.

[0062] First, the improved SE-ResNet model adds an SE attention mechanism to the traditional ResNet50 network. The SE module can automatically analyze the characteristics of different channels, model the relationships between them, and assign corresponding weights by analyzing the importance of the channels themselves, introducing a new dimension for channel fusion.

[0063] like Figure 2 As shown, let the three dimensions of the two-dimensional GAF image be H, W, and C, where h and w are the length and width of the tensor, respectively, and c is the number of channels. The SE module first performs a Squeeze operation, using global average pooling to compress the two-dimensional image from [H,W,C] to a feature map of [1,1,C]. The Excitation operation passes through a fully connected layer, reducing the feature map of [1,1,C] to [1,1,C / r]. Weights are assigned according to the different importance of each channel, where r is the scaling parameter. After passing through an h-swish activation function and a fully connected layer, the feature map size is increased to [1,1,C]. Then, an h-sigmoid function normalizes the weight vector. The Scale operation generates a weighted feature map.

[0064] The improved SE-Resnet-50 consists of five stages. The first stage takes a 224×224 GAF image as input, performs 7×7 convolution operations and max pooling to transform the GAF image into a 56×56 size. The second, third, fourth, and fifth stages each consist of four blocks, with each block containing 3, 3, 4, and 3 bottom edges, respectively. Figure 3As shown. Redundant layers lead to redundant features and weights, increasing the computational load of the model, consuming more memory, reducing computational efficiency, and increasing the unnecessary costs of fault diagnosis. This application improves computational efficiency, reduces memory consumption, and ensures improved fault diagnosis classification accuracy by retaining the feature extraction space of the lowest and highest layers and compressing the intermediate layers accordingly.

[0065] The improved SE-Resnet-50 also optimizes the relevant hyperparameters of the network model, and obtains a better diagnostic model through continuous iterative training.

[0066] (31) The Ranger optimizer is chosen to replace the Adam optimizer of the traditional ResNet50, which can dynamically adjust the learning rate in real time during training, thereby achieving a higher convergence speed.

[0067] The Ranger optimizer combines two novel optimizers: RAdam and Lookahead. During model training, two sets of weight values ​​need to be updated: the slow weights and the fast weights. And fast weights θ. The RAdam optimizer employs a dynamic rectifier, adjusting adaptive momentum based on variance. First, the RAdam optimizer is used to update the fast weights θ, achieving dynamic adjustment of the learning rate. The Lookahead optimizer improves the convergence of the inner optimizer, reduces dependence on hyperparameters, and is adaptable to changes in the inner loop optimizer, the number of updates to the fast weights θ, and the slow weights. The learning rate exhibits strong robustness and achieves rapid convergence in experiments.

[0068] (32) The h-swish activation function was chosen to replace the traditional ReLU activation function. This function was first used in the MobileNetV3 network. It has the characteristics of being unbounded at the upper limit, bounded at the lower limit, smooth, and non-monotonic, and is superior to the ReLU activation function in deep models. The traditional ReLU activation function has good nonlinearity when extracting high-dimensional features, but it destroys nonlinearity when extracting low-dimensional features, which has certain limitations in the training process of the model network.

[0069] The h-swish activation function is an improvement on the swish function, and experiments show that it demonstrates its advantages when used in deeper network layers. The calculation process of the h-swish activation function is as follows:

[0070] swish(x) = x·sigmoid(βx)

[0071]

[0072]

[0073] Where β is a hyperparameter, which can be a training parameter or a constant; Swish(x) is the swish activation function; and Sigmoid(x) is the sigmoid activation function.

[0074] (33) The improved SE-Resnet-50 model is trained using the label smoothing cross-entropy loss function to improve the accuracy of image classification and prevent model overfitting.

[0075] Because different fault types in DC-DC converters exhibit extremely high similarity, overconfidence can arise when extracting fault features, potentially leading to incorrect classification. To avoid this overconfidence problem, a label-smoothed cross-entropy loss function is used instead of the traditional cross-entropy loss function. Label smoothing effectively prevents overfitting, and its calculation formula is as follows:

[0076] L c =-∑y true ln y pred

[0077] Label smoothing is applied to the traditional cross-entropy loss function, i.e., a smoothing coefficient ε is added:

[0078] L celoss =-∑[(1-ε)*y true +ε / N]ln y pred = (1-ε)∑y true ln y pred -ε / N∑ln y pred

[0079] In the formula, y pred The predicted classification result; y true The actual classification result is given; N represents the number of fault categories for the DC-DC converter.

[0080] Adjust the model's hyperparameters, such as learning rate, batch size, Ranger optimizer parameters, and adjustable parameters of the loss function, to optimize the network model. Save the model with the highest accuracy on the validation set as the trained model.

[0081] To verify the effectiveness of the improved SE-Resnet-50 model, six ablation experiments were designed to analyze the impact of the SE attention mechanism, h-swish activation function, label smoothness cross-entropy loss function, and Ranger optimizer on the fault diagnosis results of DC-DC converters. The experimental results are shown in Table 2.

[0082] Table 2 Contributions of the ResNet-50 Improvement Module

[0083]

[0084] To verify the effectiveness of SE-Net, the residual units in the traditional ResNet-50 model were replaced with the improved residual units of SE-Net. Adding only the SE attention mechanism improved the fault diagnosis accuracy from 94.50% to 98.47%, demonstrating that the SE attention mechanism can effectively improve the feature extraction performance of the fault diagnosis model and optimize the network.

[0085] To analyze the effectiveness of the h-swish activation function, the ReLU activation function in the SE-Net residual units and ResNet-50 was replaced with h-swish, and the sigmoid activation function was replaced with h-sigmoid. Experimental results show that h-swish can improve accuracy by 0.76% while reducing computational cost, offering better cost-effectiveness.

[0086] To demonstrate the superiority of the Ranger optimizer, the commonly used Adam optimizer was replaced with the Ranger optimizer. Results show that introducing the Ranger optimizer improves accuracy by 0.18%. The Ranger optimizer is a synergistic combination of RAdam and LookAhead, reducing the need for hyperparameter tuning and achieving fast convergence of classification tasks with lower computational cost.

[0087] To verify the impact of the label smoothing regularization loss function on model training, the traditional cross-entropy loss function was replaced with a label smoothing method. Experimental results show that it can not only alleviate the impact of mislabels, but also improve the accuracy by 0.40%.

[0088] The above six ablation experiments show that the model achieves the highest accuracy when SE-Net, h-swish activation function, Ranger optimizer, and label smoothing are used simultaneously. Since some early faults in DC-DC converters are caused by subtle degradation of key components, and these faults have very weak characteristics, achieving an accuracy of 99.32% already demonstrates a significant performance improvement for soft fault diagnosis.

[0089] (34) The image dataset collected in step (2) is randomly divided into training set, test set and validation set in a ratio of 6:3:1. The improved SE-Resnet model is trained using the training set and validation set. The accuracy and loss rate of the training set and validation set are recorded for each training session. The network model with the highest accuracy on the validation set is saved.

[0090] (4) Input the test set into the trained model to complete the early soft fault classification of various key components of the DC-DC converter. The diagnostic results are shown in Table 3. The average accuracy of fault identification is as high as 99.32%.

[0091] Table 3 Experimental Results of Early Fault Identification of DC-DC Converters

[0092]

[0093] like Figure 4 , 5 The figures show the accuracy and loss curves of the proposed method on the training and validation sets, respectively. Both curves indicate that the model converges around 20 iterations without overfitting.

[0094] To more vividly demonstrate the classification performance of this method on the test set, a confusion matrix is ​​introduced to analyze the simulation results. The fault diagnosis results of the improved SE-ResNet model in this application on the test set are as follows: Figure 6 As shown in the figure, there are 59 samples for each fault type, totaling 10 fault types. The numbers on the diagonal of the figure represent the number of correctly identified faults for each type. The figure shows that only in the identification of fault F3 were two GADF diagrams misidentified as N1, and in the identification of fault N2 were one GADF diagram misidentified as F3; the rest were correct.

[0095] The method described in this application is compared with other classic deep learning models, as shown in Table 4. When AlexNet, ResNet-18, GoogleNet, VGG-16, and DenseNet-201 were applied to fault diagnosis of DC-DC converters, their accuracies were 89.83%, 88.98%, 94.41%, 96.27%, and 99.26%, respectively. The accuracy, precision, recall, and F1 score of this application were 99.32%, 99.36%, 99.09%, and 99.11%, respectively, all higher than the other five models. In conclusion, the fault diagnosis accuracy of the improved SE-ResNet model is significantly higher than other models, indicating that it can reliably identify soft faults in DC-DC converters in practical applications.

[0096] Table 4 Experimental results of DC-DC converter fault diagnosis under different deep learning algorithms

[0097]

[0098] This application also provides a diagnostic system corresponding to the above-described fault diagnosis method, the system comprising:

[0099] The data module is used to determine the early fault type of the DC-DC converter and obtain its corresponding sample data;

[0100] The processing module is used to perform two-dimensional image encoding on the time-domain signal of the DC-DC converter using Gram angle field, and to preprocess the image to obtain an image dataset;

[0101] The diagnostic module is used to train an early fault diagnosis model based on the image dataset and the improved SE-Resnet model, and to use the model to complete fault diagnosis and classification.

[0102] The above description is merely a few embodiments of this application and is not intended to limit this application in any way. Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any changes or modifications made by those skilled in the art without departing from the scope of the technical solution of this application using the disclosed technical content are equivalent to equivalent implementation cases and fall within the scope of the technical solution.

Claims

1. A method for early fault diagnosis of DC-DC converters based on an improved SE-ResNet, characterized in that, The method includes: (1) Determine the early fault type of the DC-DC converter. The early fault type includes the normal set N1-N3 and the early soft fault set F1-F7, where N1-N3 is the normal set where the device parameters change within the tolerance range, and F1-F7 is the early soft fault set where the device parameters change beyond the tolerance range. (2) Obtain sample data and use PSpice software to obtain the time domain signal of the DC-DC converter through Monte Carlo analysis; (3) The time-domain signal of the DC-DC converter is encoded in two dimensions using Gram angle field to obtain a Gram differential angle field image with a size of 28×28; (4) The image is preprocessed and its size is converted to 224×224 using data augmentation, which is then used as the input feature of the improved SE-Resnet model; (5) The preprocessed images are randomly divided into training set, test set and validation set in a ratio of 6:3:1; (6) Train an improved SE-Resnet model based on the image training set. The BasicBlock of each network layer of the improved SE-Resnet model contains 3, 3, 4 and 3 units respectively. The h-swish activation function is replaced with the ReLU activation function, the Ranger optimizer is replaced with the Adam optimizer, and the label-smoothed cross-entropy loss function is replaced with the cross-entropy function. (7) Use the trained improved SE-Resnet model to complete the fault diagnosis and classification.

2. The method for early fault diagnosis of DC-DC converters based on improved SE-Resnet according to claim 1, characterized in that, The hyperparameters of the model include at least one of the following: learning rate, batch size, parameters of the Ranger optimizer, and adjustable parameters of the loss function.

3. A DC-DC converter early fault diagnosis system based on an improved SE-Resnet, characterized in that, The system includes: The data module is used to determine the early fault types of the DC-DC converter. The early fault types include the normal set N1-N3 and the early soft fault set F1-F7, where N1-N3 is the normal set where the device parameters change within the tolerance range, and F1-F7 is the early soft fault set where the device parameters change beyond the tolerance range. Sample data is acquired by using PSpice software to obtain the time domain signal of the DC-DC converter through Monte Carlo analysis. The processing module is used to perform two-dimensional image encoding on the time-domain signal of the DC-DC converter using Gram angle field to obtain a Gram differential angle field image with a size of 28×28; the image is preprocessed: the image size is transformed to 224×224 using data augmentation, which is used as the input feature of the improved SE-Resnet model; the preprocessed image is randomly divided into training set, test set and validation set in a ratio of 6:3:

1. The diagnostic module is used to train an improved SE-ResNet model based on the image training set. The BasicBlock of each network layer of the improved SE-ResNet model contains 3, 3, 4, and 3 units respectively, and the h-swish activation function replaces the ReLU activation function, the Ranger optimizer replaces the Adam optimizer, and the label-smoothed cross-entropy loss function replaces the cross-entropy function. The trained improved SE-ResNet model is used to complete fault diagnosis and classification.

Images