Soft Fault Diagnosis for DC–DC Converter Based on Improved ResNet-50

DC-DC converter is the vital part of the power system, and its fault can cause the failure of complex electronic equipment. Therefore, the timely and accurate fault diagnosis of DC-DC converter is particularly important. This paper proposes a DC-DC converter fault diagnosis method based on improved SE-ResNet algorithm. First, in order to extract all information about the signals, we transform the one-dimensional signal into Gramian angular differential field (GADF) images. Then, the Squeeze-and-Excitation (SE) network is added to ResNet-50, combined with h-swish function and label smoothing cross entropy loss function (LSCE) to improve the model performance. In addition, simplify the redundancy layer of the network and achieve lightweight to improve the training efficiency. In the end, the logsoftmax is used for evaluating the effectiveness of the proposed method. The simulation experimental accuracy is up to 99.3220%,which is higher than other five classical deep learning algorithms. The hardware experiment also denotes the engineering practicability of the proposed method.


I. INTRODUCTION
As an important component of power supply, DC-DC converter is widely used in aerospace, communication system, renewable energy and other fields [1]. The key components of DC-DC converter involve capacitors, inductors and MOSFET [2]. The power supply needs to work in complex environment for long periods of time. The external factors such as temperature and humidity can cause component degradation, which can lead to the shutdown of the equipment [3]. The faults of DC-DC converter mainly include two types. The one is hard fault caused by structural damage, and the other is soft fault caused by the components. Therefore, timely and effective fault diagnosis of DC-DC converter is of great significance.
The study on fault diagnosis of DC-DC converter is increasing. The one is model-based method. The fault mechanism is analyzed by means of state observer, parameter evaluation and redundancy analysis. Alexandre et al. proposed an inversion method based on an integrated fuzzy logic The associate editor coordinating the review of this manuscript and approving it for publication was Zhehan Yi . for detecting sudden failure at unknown time [4]. Yao et al. designed a noninvasive method to monitor the capacitor for the power converter [5]. Su et al. proposed an adaptive parameter identification method for PV system to realize parameter estimation and fault detection [6]. The other is knowledgebased method. Zhang et al. used fault mode analysis and fault tree to realize the fault reasoning and location of UAV power system [7]. Wang conducted an expert system for satellite power system [8]. However, these methods have some disadvantages: (i) Expert knowledge is limited, it is difficult to comprehensively analyze the complex fault mechanism; (ii) It requires sensor and diagnostic circuit to obtain information, and the cost is high.
Deep learning algorithms have developed rapidly and have an extensive use in fault diagnosis recently. Compared with traditional methods, deep learning can dig for information to realize the end-to-end fault diagnosis [9]. It doesn't need to extract fault characteristics manually, but uses multi-source massive data to represent the operating state of the system [10]. In recent years, ResNet has developed rapidly in the fields of computer vision, semantic segmentation, text analysis and fault diagnosis [11]. Wu [17].
This paper designs a fault diagnosis method based on improved ResNet-50 for DC-DC converter. First, transform the time series into Gramian angular differential field (GADF) images, which are used as the ResNet input. Then, introduce the Squeeze-and-Excitation (SE) network to ResNet-50 and improve the structure of SE-ResNet. Finally, use logsoftmax classifier to realize fault classification.
Compared with other fault diagnosis methods of DC-DC converters in Table 1. First of all, the number of fault components is different. Most of them only include inductor and capacitor. But the proposed method can realize fault location of multiple components simultaneously. In addition, our parameter deviation can reach 5%, which can be seen as incipient faults. Thirdly, some literatures used machine learning methods to extract features manually, which can spend a lot of time. But the proposed method presents a deep learning method to achieve efficient fault diagnosis.
The innovation points are as follows: (i) A fault diagnosis framework of DC-DC converter is designed. The improved ResNet-50 model achieves lightweight and reduces computing cost.
(ii) To solve the problem of slow convergence during model training, h-swish activation function is introduced. ranger optimizer can further improve the accuracy. In addition, label smoothed cross entropy loss function (LSCE) is used for restraining the problem of overfitting and increasing the generalization ability. (iii) The hardware experiment of buck circuit is added on the basis of simulation experiment. The performance of this method is further verified under noisy environment.
The organization of this paper is as follows: Section II introduces some theories. Section III represents the framework of the improved ResNet-50. The simulation and hardware experiments and compared results are shown in Section IV. Section V concludes the paper.

II. THEORETICAL FOUNDATION A. GADF TRANSFORMATION
In the preprocessing stage, GADF is used to realize the image coding of time series [24]. It transforms the output signal into images, which are further used as the ResNet input [25].
Step1: The core of GADF is to represent the time-domain signals in polar coordinates. First, each output signals of DC-DC converter X ={x 1 , x 2 , x 3 ,. . . , x n }(n=2000) was normalized between [0, 1]: Step2: The normalized signals was mapped to polar coordinates: where φ i is polar angle, t i is the timestamp and N is constant Step3: For DC-DC converter, trigonometric difference was defined: where φ i and φ j respectively denote the polar angle of x i and x j . The GADF matrix was defined:

B. THE RESIDUAL NETWORK
Theoretically, when faced with large datasets, the number of model layers will increase, which can reduce training accuracy. The deep model will increase computation and cause the gradient explosion problem. For this phenomenon, He et al. proposed ResNet based on CNN [26]. The key point of ResNet is the residual unit which is shown in Fig 2 (b), where x and H (x) represent input and output, and F(x) is mapping function. Pass input directly to output by shortcut: This learning process simplifies the network and improves the discriminant ability [27].
Assuming that a residual network consists of L residual modules, its output can be expressed as: x r+1 = f (y r ) where x r x r+1 W r respectively represent the input, output and parameter of the residual network. F(·) is residual mapping and f (·) is ReLU. By the recursive iteration method, the residual unit is as follows:

A. FRAMEWORK OF THE PROPOSED METHOD
At present, the fault diagnosis methods for DC-DC converter mainly include the soft and hard faults. But some faults may be caused by the subtle parameter deviations of the components in the early stage. This type of fault is called the incipient fault, which indicates the minimum parameter change of the component is within the tolerance range. In this paper, the type of fault is the incipient fault, and the fault signal of DC-DC converter is output voltage V o of R L . In practice, DC-DC converter has an widely use in the power supply of aircraft, rocket, missile and radar antenna. Its stability and reliability is very important [28]. However, some traditional fault diagnosis methods have some disadvantages. In addition, in the face of industrial application, noise will affect the effect of fault diagnosis. A novel method is proposed to diagnose the soft faults of DC-DC converter, namely the improved ResNet-50. The overall framework is shown in Fig. 3. The detailed procedure of the improved ResNet-50 are as follows: Step1: Through GADF transformation, the original 1D-signals of DC-DC converter are transformed into 2D-images.
Step2: The improved ResNet-50 model is trained by the training set.
Step3: The model with the lowest loss of the validation set is selected as the best performance.
Step4: The fault diagnosis results of testing set are obtained.

B. IMPROVED SE-ResNet
Traditional CNN has strong feature representation ability, but the convolution kernel size limits the receptive field. So the global expression ability of ResNet-50 is poor and has some information loss [29]. In recent years, combine deep learning with attention mechanism can autonomously adjust weight to adapt the classification tasks [30]. So this paper proposes a method to improve ResNet-50 based on the SE-Net.
The structure of the improved SE-ResNet for fault diagnosis is shown in Fig. 2. First, The SE attention is added into the residual unit to analyze different feature channels, which can enhance useful feature information and compress invalid information. Then, the activation function is replaced with h-swish and h-sigmoid. The model is trained with the ranger optimizer, which can improve the performance of the network. Thirdly, the loss function is replaced with the label smoothed cross entropy loss function to solve the overconfidence in the image classification. Furthermore, the layers of the ResNet-50 are lightened to save memory resources. The above improvements are described as follows:

1) SQUEEZE-AND-EXCITATION NETWORK
ResNet relies on convolution operation and convolution kernel for automatic feature capture, but it is difficult to break through in network performance. In order to synthesize the characteristics of different receptive fields, Hu et al. proposed the SE-Net by adding spatial scale to the network [31]. SE-Net draws on visual attention mechanism and allocates attention according to the importance of channels.
SE-Net improves performance from the spatial dimension. It consists of a number of Squeeze-and-Excitation blocks (SE blocks), which is shown in Fig. 4. SE Block can add to any network to excite channel features. It involves convolution layer, Squeeze layer and Excitation layer. First, the input Iis obtained by convolution operation F tr . They are compressed by Squeeze operation F sq . Then, by the excitation F ex , the weights are obtained by Sigmoid function. The reweight F scale loads the weights of channel into features. The calculation of F ex , F scale is as follows: where δ and σ are respectively ReLU and Sigmoid. W 1 and W 2 are the weights of fully connected layers, u c denotes the c channel in the feature diagram.

2) H-SWISH ACTIVATION FUNCTION
Swish activation function is designed by Google, which solves saturation problem of Sigmoid and has better performance than ReLU [32]. But there are some problems such as much computation and slow convergence. So h-swish is introduced to ensure the lightweight of the model in this paper [33]. In addition, the piecewise linear function h-sigmoid replaces Sigmoid. The calculation h-swish is shown in Eq. (8).
The ReLU in the SE-Net is replaced by h-swish and Sigmoid is replaced by h-sigmoid. ResNet-50 network does the VOLUME 11, 2023  same as above. This operation is useful to reduce the memory and deploy the mobile devices.

3) RANGER OPTIMIZER
The optimizer is used to ensure the fast convergence of the model and obtain the optimal solution of the parameters.
Common optimizers include Adam and SGD. Tong proposed the ranger optimizer which combines RAdam and Lookahead. RAdam can realize the dynamic adjustment of learning rate, and Lookahead can achieve faster convergence with minimal computation. So the ranger optimizer can realize dynamic and stable descent of training gradient. The 81160 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.   Adam optimizer in the SE-Net is replaced by the ranger optimizer to improve the expression performance of the model.

4) LABEL SMOOTHED CROSS ENTROPY LOSS FUNCTION (LSCE)
The signals of DC-DC converter are a little simple, GADF images of different fault types are similar. So it will cause the excessive trust in the training stage. In addition, the label problem of the data will also lead to the overfitting problem. The idea of label smoothed regularization is added to the loss function to reduce the interference of noisy labels and inhibit the overfitting [34].
The cross entropy loss function is represented by The calculation of the label smoothing: Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.  Add the smoothing coefficient ε to L c , the label smoothed cross entropy loss function is: y true ln y pre − ε/N lny pre (11) where y pre is the predicted label; y true is the true label; Nis the number of fault types. It can be seen that this idea smoothens the gap between the maximum predicted and the average value and enhance the generalization ability.

5) RESNET-50 MODEL LIGHTWEIGHT
Considering the multiple interference factors in actual industrial environment, deep learning algorithm is selected for fault diagnosis in this study. The gradient explosion and gradient dispersion occur in traditional networks with increasing depth. However, ResNet has stronger feature expression ability and can solve the problems of overfitting and degradation. Considering the influence of learning ability and model parameters, ResNet-50 is used as the basic model [35].
There are 49 convolutional layers in ResNet-50. First, 7 × 7 average and 3 × 3 maximum pooling are used to reduce the size of GADF images. Then, extract features of images by layer1-layer4. Each layer of traditional ResNet-50 respectively contains 3,4,6,3 residual units. The redundant features and weights will lead to the large computing memory. So ResNet-50 is lightweight, improved layer contains 3,3,4,3 residual units. The shallowest and deepest layer remain as before, and the middle layer is deleted and optimized. Then, the adaptive global averaging pooling is used to get the output.

C. EVALUATION INDICATORS
This paper used four indicators to evaluate the fault diagnosis performance: Accuracy, Precision, Recall and F1 .The values are higher denoting that the algorithm is more successful. The equations of them are as follows: 81162 VOLUME 11, 2023 Authorized licensed use limited to the terms of the applicable license agreement with IEEE. Restrictions apply.
where TP presents positive judgment, FP refers positive misjudgment, TN refers negative judgment and FN refers negative misjudgment. In addition, the confusion matrix is also used for showing fault diagnosis results. It can show how many fault labels the model has predicted correctly.

1) EXPERIMENT SET UP
The buck converter is selected as the research object, which is conducted in PSpice. As is shown in Fig. 5, the DC signal is +6V,the period of PWM is set as 50 s, and the duty ratio is 50%. The tolerance of components are set to 5%.
The capacitors, inductors and MOSFET are related to the degradation of converters. In order to verify the proposed  method and its wide coverage on soft faults, the normal states and soft faults are injected into the buck converter. All soft fault set is shown in Table 2 according to the degradation model of these components, where N1-N3 are fault types within the tolerance, which can be seen as the normal states. F1-F6 are out of tolerance, which can be seen as soft faults. These fault types are caused by arbitrary changes of components, which is closer to the practical application background. The voltage of 19.5 ms-20 ms is selected with 0.5 s sample interval. From Fig. 6, we can see that the output waveforms of N1-F7 soft faults have small difference. 200 samples are collected for 10 faults by Monte Carlo. Then, they are transformed into GADF images whose size is 28 × 28. 2D images of N1-F7 are shown in Fig.7.  The accuracy of the test dataset is 99.3220%. The precision, recall, F1 and accuracy values are given in Table 3. The confusion matrix is represented in Fig. 9. For each fault type, there are 59 images for testing, and the diagonal line represents the correct types. In the F3 fault prediction, 2 GADF images are misjudged as N1. In the N2 fault prediction, 1 GADF image is misjudged as F3, and in the N3 fault prediction, 2 GADF images are misjudged as F7.

3) INFLUENCE OF RESNET-50 IMPROVEMENTS ON FAULT DIAGNOSIS
To explain the impact of the improved ResNet-50, six ablation experiments are used to explain the impact of the SE-Net, the h-swish function, label smoothed cross entropy loss function and ranger on the fault diagnosis.
For the SE-Net, the residual units of ResNet-50 is combined with SE-Net. As is shown in Table 4, the accuracy is 98.47%,which is a rise of 3.97% after adding SE-Net, denoting that SE-Net can improve the feature extraction ability.
For the h-swish function, all ReLU functions are replaced with h-swish, and sigmoid functions are replaced with h-sigmoid. The results show that h-swish can reduce the calculation cost and improve the accuracy by 0.76%.
For the ranger optimizer, replace the most used Adam with the ranger. The ranger increased the accuracy by 0.18%. The ranger optimizer is a combination of RAdam and LookAhead, reducing the need for hyperparameter adjustment and realizing fast convergence.
For the LSCE, the traditional loss function is replaced with label smoothing. The results show that it can alleviate the influence of mislabeling, and improve the accuracy by 0.40%. The six tests indicate that the introduction of the SE-Net, h-swish function, ranger optimizer and label smoothing to the model is effective and reliable.

B. HARDWARE EXPERIMENTAL RESULTS
The hardware experiment of buck converter is shown in Fig. 10. The VICTOR LCR instrument is used to measure the key components. The R C and Care 1.452 and 4.712µF. The L and R L are 1.071mH and 1.423 . The MOSFET is IRFP150 and R on is 36m obtained from datasheet. The composition of the platform is as followed. Agilent DC power supply is +6V. The signal generator provides the PWM to control MOSFET. The frequency of the PWM is 20 kHz, and the duty ratio is 50%. The acquisition of output signal is completed by oscilloscope. The output waveforms of buck circuit collected by oscilloscope are shown in Fig11 where i L is the inductor current, u o is the output voltage, and u d is the forward voltage of diode. It can be seen that the output signals of the actual circuit contain noise.
The proposed method is aimed at soft faults caused by degradation of converters. So the components with different values are selected to constitute the soft faults of the hardware experiment. Soft fault set of HF1-HF10 is shown in Table 5. 200 samples of 10 faults are acquired from the platform and the sample interval is 1ms. Fig. 12 shows that soft faults   Parameter settings: batch-size=32, training epochs=65, learning rate = 0.001. Fig. 14 denotes the model training and validation curves in the hardware experiment. The fault diagnosis results are represented in Table 6. The accuracy is up to 94.9153%. The confusion matrix is shown in Fig. 15. The accuracy of HF1-HF9 is above 93%, and the recognition effect of HF10 is poor. In the HF10 fault prediction, one sample is misjudged as HF9, six samples are misjudged as HF1, six samples are misjudged as HF6, and two samples are misjudged as HF3. We suspect that the GADF images of HF10 are similar to the GADF images of HF1 and HF6, so the model learned similar features. In addition, it may be due to the overfitting of the improved SE-ResNet. The experiment results indicate that the proposed method has good soft fault identification effect in the noisy environment. In conclusion, the improved ResNet-50 has the best performance, indicating that the method can identify soft faults of DC-DC converter reliably.

2) COMPARISON WITH EXISTING LATEST STUDY OF FAULT DIAGNOSIS
The research related to fault diagnosis of DC-DC converter is shown in Table 8 [39]. But these methods both need to extract fault features manually, which leads to low diagnosis efficiency.
Jiang et al. used an online anomaly identification method to forecast the normal range of converters [18]. The accuracy was 91%, but it could not identify degradation and realize the fault location.
Jiang et al. adopted the support vector data description (SVDD) to identify incipient faults [20]. It avoided the inadequacy of the fault samples in the industrial application. Wei et al. proposed optimal fractional wavelet transform method to obtain the accuracy of 99.24% [19]. But it need to extract fault features manually and the cost is large.
In deep learning fields, Jia et al. designed a deep auto encoder model to achieve the transfer diagnosis [40]. The accuracy is 95.34%. Xia combined 1-D convolutional neural network (1DCNN) with long short-term memory (LSTM) to achieve the fault diagnosis [22]. However, it only considered the faults caused by capacitor, without other components.
Take into account the above shortcomings and realize the efficient fault diagnosis in the actual environment, an improved SE-ResNet is designed in this study. Efficient weight allocation is reached by adding the SE-Net to ResNet-50. Then, the accuracy is enhanced by using the h-swish activation function, the ranger optimizer and the LSCE. The test results show that the accuracy can be up to 99.3220%, denoting that the feasibility of soft faults for DC-DC converter.

D. VARIABLE WORKING CONDITION
For the same DC-DC converters, the variable working conditions have different effect on the output. In the actual application, the input voltage as well as the duty cycle might vary based on different load conditions. So the experiment is carried out to validate the proposed method in variable working conditions. Fault types are shown in Table 9. The parameters of the components are the same as Section IV-B.
Parameter settings: batch-size=16, training epochs=55, learning rate = 0.00005. The fault diagnosis results are represented in Table 10. The accuracy is up to 97.7011%. The results show that the proposed method can also handle the variable conditions.

V. CONCLUSION
In this study, a DC-DC converter fault diagnosis method based on improved SE-ResNet is proposed. The conclusions are as follows: (i) The accuracy of the simulation and hardware experiment is up to 99.3220% and 94.9153%, respectively. And the ablation results show that the SE-Net, h-swish, LSCE and ranger optimizer increase the accuracy by 3.97%, 0.76%, 0.40% and 0.18%, respectively. Compared with other methods, the proposed method has a better performance in soft fault diagnosis.
(ii) This study still has some drawbacks. The accuracy of the hardware experiment still needs to be improved. So the robustness and anti-noise ability in practical applications need to be enhanced. Moreover, the fault types are only aimed at components with large output impact, but not includes diodes.
(iii) In the future work, we intend to combine the fault diagnosis method with digital twin technology in order to solve the problem of insufficient fault data in practical applications, such as the secondary power supply of rocket or aircraft.
This method has practical significance for the health management and fault diagnosis of power supply, which is conducive to the stable operation of electronic equipment.