Fault States Diagnosis of Marine Diesel Engine Valve Based on a Modified VGG16 Transfer Learning Method

Cai, Yijie; Xu, Zhe; Wen, Quan; Shi, Jinni; Zhong, Fei; Yang, Xiaojun; Yang, Jianguo; Zhou, Hongdi

doi:https://doi.org/10.1155/2023/1225536

Mathematical Problems in Engineering

On this page

Abstract Introduction Conclusion Data Availability Conflicts of Interest Acknowledgments References Copyright Related Articles

Research Article | Open Access

Volume 2023 | Article ID 1225536 | https://doi.org/10.1155/2023/1225536

Fault States Diagnosis of Marine Diesel Engine Valve Based on a Modified VGG16 Transfer Learning Method

Yijie Cai,^1,2,3Zhe Xu,^1,2Quan Wen,⁴Jinni Shi,⁴Fei Zhong,^1,2Xiaojun Yang,^1,2Jianguo Yang,³and Hongdi Zhou^1,2

Academic Editor: John S. Sakellariou

Received11 Nov 2022

Revised23 Feb 2023

Accepted22 Mar 2023

Published08 May 2023

Abstract

The marine diesel engine is an important power machine for ships. Traditional machine learning methods for diesel engine fault diagnosis usually require a large amount of labeled training data, and the diagnosis performance may decline when encounters vibrational and environmental interference. A transfer learning convolutional neural network model based on VGG16 is introduced for diesel engine valve leakage fault diagnosis. The acquired diesel engine cylinder head vibration signal is first converted to time domain, frequency domain, and wavelet decomposition images. Secondly, the VGG16 deep convolutional neural network is pretrained using the ImageNet dataset. Subsequently, fine tuning the network based on the pretrained basic parameters and image enhancement methods. Finally, the well-trained model is adopted to train and test the target dataset. In addition, the cosine annealing learning rate setting method is used to make the learning rate close to the global optimal solution. Experimental results show that the proposed method has higher accuracy and better robustness against noise with a small sample dataset than traditional methods and deep learning models. This study not only demonstrates a novel view for the diagnosis of marine diesel engine valve leakage, but also provides an applicable diagnosis method for other similar issues.

1. Introduction

Diesel engines are widely used in the fields of ships and vehicles. However, due to the complex mechanical structure of diesel engines, faults frequently occur, including misfires, cylinder collisions, and valve faults. The failure of the marine diesel engine causes the ship to stop working and increases operating costs. It is very necessary to carry out condition monitoring and fault diagnosis techniques to ensure the safe and reliable operation of the diesel engine. Intake and outlet valves are critical parts of accomplishing the combustion process and suffer from high pressure and high temperature working conditions. A valve leakage fault increases the discharge temperature, decreases the output power, and reduces the overall performance of the diesel engine. Since unexpected failures of valves may cause the breakdown of a diesel engine and even lead to significant economic loss, various diagnosis methods for diesel engine valve leakage faults are developed based on vibration signals to improve the reliability of diesel engines [1–3].

During the working process, the periodical shocks generate a vibration signal because of the strike between the valve and valve seat, which can be transmitted to the diesel engine cylinder head surface. However, the vibration signals sampled from the cylinder head surface are easily affected by the internal parts, and the periodical shocks are often submerged in noise, which significantly increases the difficulty of fault diagnosis. Moreover, diesel engine failure is relatively rare and hard to observe. Once a failure occurs, a major accident occurs. In addition, the cost of failure simulation is high, and it is a tough task to achieve failure data for quantitative research. Therefore, valve leakage fault diagnosis is a significant challenge [4].

There are three categories of vibration signal fault diagnosis methods:

The first category is the traditional signal analysis, including time domain, frequency domain, and time-frequency analysis methods, such as variational mode decomposition (VMD) and wavelet packet transform (WPT). These methods extract and compare the features from fault signal to accomplish fault diagnosis [5–10].

The second category is the traditional machine learning algorithm, such as stream learning, support vector machines (SVM), random forests, and back propagation neural networks (BPNN), which can simplify the signal processing process and improve the efficiency of identification compared with the traditional vibration signal analysis method [11–14]. These diagnostic methods can achieve high diagnostic accuracy for diesel engine fault datasets due to their effective feature extraction and reasonable classifier design.

The third category of fault diagnosis is analyzing pictures and extracting graphic features, such as recurrent neural network (RNN), stacked self-encoder (SAE), and 2D convolutional neural network (2D CNN). These methods can learn the inherent features in the picture adaptively, which can greatly improve the accuracy and efficiency of recognition. Khorram et al. [15] proposed an end-to-end fault detection method. A new convolutional long short-term memory recurrent neural network (CRNN) is used as an input to detect bearing faults with the highest accuracy in the shortest possible time. Jiang et al. [16] proposed a one-dimensional convolutional long short-term memory (1D-CLSTM)-based multifactor condition identification method for diesel engine fault diagnosis. Lu et al. [17] proposed a new multisensor information fusion method to achieve fault classification. This method constructs the time-domain vibration signals from multiple sensors at different locations into a rectangular 2D matrix, and then uses an improved 2D CNN to achieve signal classification. Li et al. [18] proposed a supervised model called modified auxiliary classifier GAN (MACGAN) designed with a new framework. The abovementioned studies show that the combination of image and deep learning methods has a high accuracy rate, but improving its application in fault diagnosis is still necessary. Duan et al. [19] proposed a novel multiscale-stacked sparse principal component analysis network (MS-SSPCANet) to predict tool wear. Duan et al. [20] proposed a novel deep learning network named hybrid attention-based parallel deep learning (HABPDL) model to address these problems. Global average pooling (GAP) is applied to reduce superfluous spatial features and increase model interpretability after the CBAM layer. Abovementioned studies show that the combination of image and deep learning methods has a high accuracy rate, but in practical applications, diesel engines continuously work in harsh environments, vibration signals appear nonstationary, and fault features merge into heavy noise. Although the abovementioned intelligent methods have made remarkable achievements, the following inherent limitations still exist:(1)In practical industrial applications, diesel engines can work continuously in extremely harsh working conditions, fault characteristics are often difficult to distinguish from heavy noise. The collected vibration signals are always nonstationary and noisy, and the traditional methods may not be unsuitable. Thus, using additional and suitable signal processing methods to extract fault signals is significant.(2)Traditional deep learning diagnostic methods, such as CNNs and DBNs, may suffer significant performance losses when applied to new diagnostic tasks, or only a small dataset is available, even if the new task is similar to the original task.

Fortunately, transfer learning provides a way to deal with such problems. Massive data can be obtained in laboratory experiments by fault simulation, and thus the model can be trained sufficiently. Then, transfer the well-trained model to other similar tasks.

This method has achieved remarkable success in a number of vision recognition tasks. In the field of fault diagnosis, transfer learning with deep neural networks has been less explored due to the limitation of a domain-specific dataset of sufficient size and a common deep network model.

Yuan et al. [13] proposed a generic intelligent-bearing fault diagnosis system based on AlexNet with transfer learning to automatically identify and classify different bearing faults. Zhao et al. [21] construct a new taxonomy and perform a comprehensive review of UDTL-based IFD according to different tasks. Comparative analysis of some typical methods and datasets reveals some open and essential issues in UDTL-based IFD, which is rarely studied, including transferability of features, the influence of backbones, negative transfer, physical priors, and so on. Chen et al. [22] proposed a transferable CNN to improve the learning of the target task. A transfer learning strategy was used to train a deep model of the target task using a pretrained network. Ma et al. [23] proposed an AlexNet-based transfer learning convolutional neural network (TLCNN) method for bearing fault diagnosis, using a two-dimensional image representation to convert the vibration signal into a two-dimensional time-frequency image. The proposed fault diagnosis model has higher accuracy and better robustness against noise compared to other deep learning methods and traditional methods. Zhang et al. [24] proposed a novel DL framework by applying convolutional neural networks (CNNs) based on the optimization of transfer learning (TL). TL can help the model achieve higher precision with less computational cost by transferring low-level features and fine-tuning high-level layers. Zhang et al. [25] proposed a convolutional neural network (CNN)-based two-layer transfer learning (CTTL) method for fault diagnosis. CTTL changes the process of the transfer learning method from learning the distribution of domains to learning the distribution of fault types in more detail, which will get higher accuracy. Zhao et al. [26] proposed a novel transfer learning framework based on deep multiscale convolutional neural networks (MSCNN). First, a new multiscale module is cleverly built on top of the expanded convolution as a key part to obtain differentiated features through different perceptual domains. Then, in order to further reduce the complexity of the proposed model, a global average pooling technology is adopted to replace the traditional fully-connected layer. Finally, the architecture and weights of the MSCNN pretrained in the source domain are transferred to other different but similar tasks with appropriate fine tuning instead of training a network from scratch.

In this study, a novel VGG16-based transfer learning fault diagnosis method for diesel engine valve leakage is proposed. This method can effectively extract features from images with well robustness against noise in a heavy noise environment, reduces the computational complexity and time-consuming nature of the neural network. This method uses a combination of time domain, frequency domain, and wavelet decomposition of diesel engine cylinder head surface vibration signal to form a dataset. The main contributions of this study are summarized below:(1)For a small sample dataset, a combination of time domain, frequency domain, and wavelet decomposition is used to increase the sample diversity and achieve high accuracy in diesel engine valve leakage fault diagnosis.(2)The proposed method gives reasonable parameter initialization to the target model by a pretraining strategy, and provides a potential tool to train a deep network based diagnosis system fast and efficiently with less overfitting risk. It can improve the model performance and reduce the computational cost.(3)Based on the concept of transfer learning, a pretrained modified VGG16 deep neural network model is used, and the model is modified. The classification of diesel engine valve leakage fault conditions is achieved by a modified VGG16 deep neural network model, which does not require any subjective preprocessing techniques to assist in feature extraction. The result shows that the fault diagnosis model has higher accuracy and better robustness compared with ResNet50, AlexNet, MACGAN, and Inception-v3 methods.

The rest of this study is organized as follows: Section 2 describes the structure of the VGG16 model and the training process of transfer learning. Section 3 describes the preparation process of the datasets. Section 4 presents the experimental contents and details of the datasets. Section 5 provides the conclusions of this research.

2. Transfer Learning and Modified Convolutional Neural Network Models

2.1. Convolutional Neural Network (CNN)

Convolutional neural network (CNN) is one of the representative algorithms of machine learning. The CNN has a powerful automatic feature extraction capability. The basic structure of a typical CNN is shown in Figure 1, which consists of convolutional layers, pooling layers, and a fully connected layer [27].

The convolution operation in Figure 1 is calculated as follows:where input data is denoted as the th eigenvalue of the eigenmaps in the layer of the network, L denotes the convolution kernel size, denotes the weight coefficient, b is the deviation value, and is the activation function.

Pooling is a subsampling method, which aims to reduce the size of the feature map. So that the number of parameters in the fully connected layer is reduced and overfitting is prevented. There are three kinds of pooling function: max polling, mean pooling, and weight pooling. Max pooling is most commonly adopted, which is written as follows:where represents the output feature map of the band pooling layer. and are the subsampling factors and pooling size.

Suppose the set of training input samples is , where the class of the input sample elements is one of the classes in the set , and the corresponding label is . Thus, the probability of judging the input sample as one class in the set of class is . The function of softmax is as follows:where is the correlation between category and the whole classification category, is the normalization function.

2.2. Transfer Learning

Transfer learning is a new machine learning method developed for image recognition, since the training task of a deep learning model is prone to overfitting due to the small sample dataset. Unlike other neural networks, the training set for transfer learning does not necessarily belong to the same class or come from the same physical background. In transfer learning, there is no need to train the model in the target domain from scratch, which can significantly reduce the need for training data and training time in the target domain [28].

Transfer learning involves a domain (domain) and a task (task). The domain D can be expressed as follows:where denotes the feature space and denotes the edge probability distribution; the task can be expressed as follows:where denotes the category space and denotes the prediction model, i.e., which is the conditional probability distribution. The source domain of the existing labeled data and the learning task , the can be expressed as follows:the target domain of unlabeled data and the learning task , the can be expressed as follows:

Transfer learning improves the accuracy of the target domain and reduces the generalization error based on the condition or .

2.3. The VGG16 Convolutional Neural Network Model and Transfer Learning Training Process

VGG16 is a deep CNN model trained by the ImageNet dataset, which can be considered as a universal CNN feature extractor, which has been applied to many image recognition tasks and has achieved astounding performance. Impressive results have been obtained on several image classification datasets using the pretrained VGG16, as well as object detection, action recognition, human pose estimation, image segmentation, optical flow, image captioning, and others [29–31]. Thus, the pretrained VGG16 is applied to diesel engine valve leakage fault diagnosis.

The ImageNet is a dataset that contains 1,000 classification problems, so the VGG16 network has a huge number of classification layer parameters. However, this study does not need so many classifiers, so the classification layer VGG16 network is fine tuned. Although the images of the vibration signal may be different from the images used for pretrain of VGG16, useful features can be extracted similarly as long as the employed neural network is capable of recognizing features [32]. The VGG16 model uses the ReLU function, i.e., the linear correction unit. Compared with sigmoid and tanh function, ReLU will not encounter the problem of gradient disappearance, and the overfitting phenomenon is mitigated. The output of ReLU is as follows:

When is greater than 0, the input of ReLU is equal to the output, when is less than or equal to 0, the output of ReLU is 0, which indicates that the ReLU activation function will turn off the computation of neurons when the output is less than or equal to 0. This operation reduces the computation of neurons, the interaction between networks, and the weight of the model.

The VGG16 convolutional neural network is designed for image extraction and continuous deep refinement. VGG16 has high accuracy in image classification and an enormous computational cost. In practice, researchers transfer the pretrained VGG16 model based on the source domain to solve specific problems in the target domain, which is different from the source domain. The “fine-tuning” technique is adopted to improve pretrained models. The combination of “transfer and fine tuning” is adopted to modify VGG16 for diesel engine valve leakage diagnosis, as presented in Figure 2.

Due to the fact that the source domain is not very similar to the target domain data, this study freezes the weights of the first 7 layers in the pretrained model, and then retrains the later layers. It can accelerate the training speed of the VGG16 network and improve its accuracy. The fixed weights have been pretrained on the ImageNet dataset, and the updated weights can be achieved from labeled images of vibration signals. The image size of the input layer is 224 × 224. The output layer is modified to twelve classes according to diesel engine valve leakage images. The dropout rate is set as 0.4, the number of neurons in the first two fully connected layers is set as 512, and the final softmax output classification is changed to twelve classes. Setting a low learning rate to ensure the network weight does not change too fast due to the difference in the target dataset during the fine-tuning process. In this way, the learning time and required image number decrease dramatically. Figure 3 shows the flow chart of transfer learning.

3. Dataset Preparation

In this research, a combination of time domain, frequency domain, and wavelet decomposition data are extracted from the vibration signal of the diesel engine cylinder head surface. The vibration signal is preprocessed through low-pass filtering (Chebyshev filter) and windowing (Hamming window), which can filter out noise and spurious components, improve the signal-to-noise ratio, and makes the analysis data smoother. The low-pass filter frequency is set as 12,800 Hz, and the windowing length is set as 32 points [33].

3.1. Time Domain Signal Processing

In a working cycle of the diesel engine, the impact of the intake valve closing down, the exhaust valve closing down, the throttling effect of the exhaust valve opening, and the combustion in the cylinder are the main sources of excitation. The injection advance angle is determined by the oil supply system, which determines the moment when the combustion burst pressure is generated. The opening and closing angles of the intake and exhaust valves are determined by the gas distribution mechanism [34]. Therefore, the measured time-domain waveform of the cylinder head surface vibration signal is closely related to the valve timing. The 4120SD diesel engine valve timing is shown in Figure 4.

Since the vibration signals sampled on the cylinder head surface generated by excitation sources offset greatly in timing, the time domain signal of each excitation can be identified. A typical sampled signal waveform from a single cycle of a 4120SD diesel engine is shown in Figure 5. Based on the exhaust valve operating conditions simulated in the experiment, the vibration signal of the combustion section can be located, specifically 30° crankshaft rotation angle before top dead center (TDC) to 30° crankshaft rotation angle after TDC, i.e., 330°CA to 390°CA.

3.2. Frequency Domain Signal Processing

The valve leakage fault can be diagnosed when the vibration signal has a high characteristic frequency response in the frequency domain. Since the combustion causes low characteristic frequency, and as the degree of gas leakage increases, its characteristic frequency is shifted to the higher frequency region, valve leakage faults in working condition can be diagnosed [12].

Through the fast Fourier transform (FFT), the frequency domain of the signal can be obtained. The sampled vibration signals are subject to interfere from various excitation sources and environmental noise, so most of the collected vibration signals are not ideal data. Figure 6 shows the PSD of the ideal and noise disturbed vibration signals based on the normal exhaust valve.

(a)

(b)

(c)

(d)

3.3. Wavelet Decomposition Signal Processing

Wavelet decomposition was proposed by Wang et al. in performing the orthogonal wavelet basis construction. According to the theory of multiresolution analysis, the larger the scale of its decomposition, the smaller the length of the decomposition coefficients [35].

In this study, a three-layer wavelet is adopted, since the characteristic frequency of valve leakage is in the high-frequency domain. The three-layer wavelet decomposition data of the single-cycle and combustion section are obtained, which is shown in Figures 7 and 8, respectively.

(a)

(b)

(c)

(a)

(b)

(c)

4. Experimental Verification

4.1. Description of the Experimental Environment and Dataset

The diesel engine valve leakage test rig based on a 4120SD (four-cylinder four-stroke nonsupercharged diesel engine) is shown in Figure 9. The main parameters are listed in Table 1. The test rig consists of a diesel engine, a magnetoelectric sensor, three vibration sensors, and a cylinder pressure sensor. For the 4120SD engine, every two cylinders share a cylinder head. In the experiment, 1 # cylinder (the first cylinder from the flywheel end) is used to simulate the valve leakage fault. Figure 9 shows the location of sensors: the magnetoelectric sensor is installed at the flywheel end of the diesel engine to examine TDC, the vibration sensors are installed on the cylinder head, and the cylinder pressure sensor is mounted in a predrilled hole of the cylinder head. The engine load is 0% and the sampling frequency is 51.2 kHz. Four exhaust valves are used to simulate different degree of valve leakage, which is shown in Table 2.

The time domain image, frequency domain signal image, and wavelet decomposition image are obtained through processing the vibration signal. In this study, the dataset C1 includes the single-cycle time-domain waveform image P1, combustion section time-domain waveform image P2, single-cycle PSD image P3, combustion section PSD image P4, single-cycle wavelet decomposition three-layer signal image P5, and combustion section wavelet decomposition three-layer signal image P6. The types of working conditions are normal, mild gas leakage, moderate gas leakage, and severe gas leakage. In this study, three different sensors were used to measure cylinder head vibration data at the same time, so the three groups of data were separated, and each condition type was divided into three groups to create labels together. Table 3 shows the dataset C1 fault classification and labels. For comparison, the combustion section time-domain waveform image P2 is selected as dataset C2. The description of the diesel engine valve leakage fault dataset is shown in Table 4.

4.2. Preprocessing of Image Datasets

In order to improve the generalization ability of the model and avoid overfitting for few samples, the data augmentation method is adopted for the C1 and C2. Data augmentation means generating more information from limited data, increasing its diversity and improving robustness, the data augmentation methods are as follows:(a)Image scaling: the size of the input image of VGG16 is 224 × 224, scaling the image datasets from 875 × 656 pixels to 224 × 224 pixels can greatly speed up the training of the model.(b)Image flipping: random horizontal or vertical flipping of the images in the training and test sets to increase the image diversity.(c)Normalization process: the fault datasets C1 and C2 are normalized on 3 different channels of red, blue, and green. The mean values of the 3 channels are 0.987, 0.989, and 0.989 and the standard deviations are 0.064, 0.051, and 0.053. The calculation formula is as follows: where output denotes the normalized input, input denotes the input of image, mean denotes the mean value of each channel, and std denotes the standard deviation of each channel.

4.3. Diagnostic Result Analysis

In this study, the hardware used is Intel Core i5-10400F CPU processor, with 16 GB RAM, and the GPU configuration is NVIDIA GeForce RTX3060. The software operating system is Windows 10, and the deep learning framework is PyTorch.

Through the experiment, the datasets C1 and C2 are processed with a, b, and c mentioned in Section 4.2, respectively. The dataset processing method groups and their diagnostic accuracy are listed in Table 5. The modified VGG16 deep convolutional neural network model is employed to classify four kinds of valve leakage conditions of diesel engine. The best result is obtained when the dropout rate is set as 0.4, and the number of neurons in the fully connected layer is changed to 512 for the accuracy improvement. Table 6 shows the accuracy rates for different dropout rates and the number of neurons in the fully connected layer in the test set of the C1. The max training epoch is set as 100 with a batch size of 2. Cross entropy loss function is used, which can measure the subtle differences, convex optimization function, and the optimal solution, which is gained by gradient descent methods.

This Adam optimization algorithm is applied in this study, which is a learning rate adaptive optimization algorithm that combines the advantages of GDM and RMSprop [36]. The Adam optimization algorithm dynamically adjusts the learning rate for each parameter based on the estimated first-order moments and second-order moments of the gradient of the objective function for each parameter. It makes the gradient diagonal scaling invariant and robust to the choice of hyperparameters, achieving efficient iterations in the parameter space. The iterative equation is as follows:where is the distance that the th parameter of the neural network should fall along the gradient direction at the th iteration. and are the exponentially decaying average of the squared historical gradient and the exponentially decaying average of the historical gradient at the th iteration for the th parameter, respectively. and are the bias corrections of and , respectively, the purpose is to eliminate the effect of small gradient weights caused by small at the beginning of the iteration.

However, the Adam optimization algorithm is easy to overfit the features that appear in the early stage, and the features that appear later are difficult to correct for the fitting problem, which may affect the effective convergence and miss the global optimal solution. Thus, the cosine annealing learning rate setting method is adopted, which accelerates the learning rate at first, and then decreases the learning rate. The learning rate quickly rises back to the initial value after each decrease in the minimum value. The learning rate will suddenly jump up with periodic restarting, which can result in jumping out of the local optimal solution and approaching the global optimal solution. In this study, the initial learning rate of cosine annealing is 10⁻⁵, and the minimum value of the learning rate is set as 10⁻⁶. Equation (11) is the formula of learning rate with the number of iterations, where denotes the number of restarts, and are the range of learning rate, denotes the currently executed epochs, and denotes the total number of epochs in the th run

Freezing the weights of the pretrained model can greatly speed up the training of the network and improve the accuracy rate, as shown in Figure 10. The highest accuracy of weighting was 0.952 for the first 7 layers of freezing.

The training set accuracy of dataset C1 reached 0.987 after 100 epochs of training, while the test set accuracy reached 0.952. Figure 11 shows the loss and accuracy curves for the training and test sets of dataset C1. The accuracy of the training set of dataset C2 reached 0.903 after 100 epochs of training, while the accuracy of the test set reached 0.892. Figure 12 shows the loss and accuracy curves of the training and test sets of dataset C2.

4.4. Comparison with Other Methods

In order to verify the effectiveness of the proposed modified VGG16 transfer learning model, AlexNet, ResNet50, Inception-v3, and MACGAN are constructed for comparison. Figure 13 shows the accuracy comparison of each model on the test set, and Figure 14 shows the comparison of the loss of each model on the test set. The highest accuracy of the AlexNet is 0.937, the highest accuracy of the MACGAN is 0.915, the highest accuracy of the Inception-v3 is 0.893, and the highest accuracy of the ResNet50 is 0.587. On the contrary, the highest accuracy of the modified VGG16 transfer learning model is 0.952, which is 0.015 more than the AlexNet, 0.037 more than the MACGAN, 0.059 more than the Inception-v3, and 0.365more than the ResNet50. Compared with AlexNet, the modified VGG16 has a smaller convolution kernel, deeper network, and is more stable, which achieves higher recognition accuracy. Although ResNet50 and Inception-v3 have deeper layers, they are prone to overfit and lower the accuracy rate. MACGAN also performs better in small sample datasets, but its accuracy is less than that of modified VGG16. Abovementioned results demonstrate that the modified VGG16 transfer learning model achieves the highest accuracy and the least loss with the least amount of time. The model has a greater advantage in diesel engine valve leakage fault diagnosis.

5. Conclusion

In this study, a modified VGG16 transfer learning fault diagnosis model is proposed to distinguish the leakage degree of marine diesel engine valves under various interference sources and noise environments. This method can automatically extract features for fault diagnosis. The modified VGG16 deep convolutional neural network is pretrained using the ImageNet dataset. The well-trained model is adopted to train and test the valve leakage dataset. The model can be excellently used for the vibration and environmental interference encountered in the valve leakage fault diagnosis of marine diesel engines. The main contributions of this study are:(1)For the marine diesel engine valve leakage fault diagnosis, the proposed diagnosis method adopts multiple signal processing methods to describe the data from multiple dimensions in the case of a small sample dataset. The modified VGG16 transfer learning fault classification method can exclude the interference of noise environment and has higher accuracy than traditional signal analysis methods and other neural network models, when the dataset is small. The diagnosis accuracy reaches as high as 0.952.(2)Combined with transfer learning, the data enhancement method can prevent model overfitting and improve the robustness of the model by scaling image and normalization of the image datasets. Integrating the cosine annealing learning rate setting and the Adam optimization algorithm can find the optimal solution for the learning rate.

The proposed method can be used in diagnosis cases and extended to other data-driven tasks, including condition monitoring and anomaly detection. In the following studies, two essential points will be explored for fault diagnosis of diesel engine valve leakage by deep learning: (I) improving the accuracy of fault diagnosis for diesel engine in noise environment, and (II) building a new fault diagnosis model to realize early-stage fault diagnosis of diesel engine valve leakage.

Data Availability

The data that support the findings of this study are available from the first author upon reasonable request.

Conflicts of Interest

The authors declare that they have no conflicts of interest.

Acknowledgments

This research was sponsored by the Ministry of Industry and Information Technology of China (Grant No. [2019] 360), the National Natural Science Foundation of China (Grant No. 52005168), Key Research and Development Program of Hubei Province of China (Grant No. 2020BBB084), Natural Science Foundation of Hubei Province (Grant No. 2022CFB882), the High Level Talent Fund of Hubei University of Technology (Grant No. BSQD2020010), and the Open Fund of Hubei Key Laboratory of Modern Manufacturing Quality Engineering (Grant No. KFJJ2021012).

References

G. Xiong, W. Ma, N. Zhao, J. Zhang, Z. Jiang, and Z. Mao, “Multi-type diesel engines operating condition recognition method based on stacked auto-encoder and feature transfer learning,” IEEE Access, vol. 9, pp. 31043–31052, 2021.
View at: Publisher Site | Google Scholar
Y. Cai, G. Xu, A. Li, and X. Wang, “A novel improved local binary pattern and its application to the fault diagnosis of diesel engine,” Shock and Vibration, vol. 2020, Article ID 9830162, 15 pages, 2020.
View at: Publisher Site | Google Scholar
Y. Liu, J. Zhang, and L. Ma, “A fault diagnosis approach for diesel engines based on self-adaptive WVD, improved FCBF and PECOC-RVM,” Neurocomputing, vol. 177, pp. 600–611, 2016.
View at: Publisher Site | Google Scholar
Z. Gao, C. Cecati, and S. X. Ding, “A survey of fault diagnosis and fault-tolerant techniques—Part I: fault Diagnosis with model-based and signal-based a,” IEEE Transactions on Industrial Electronics, vol. 62, no. 6, pp. 3757–3767, 2015.
View at: Publisher Site | Google Scholar
L. Song, H. Wang, and P. Chen, “Vibration-based intelligent fault diagnosis for roller bearings in low-speed rotating machinery,” IEEE Transactions on Instrumentation and Measurement, vol. 67, no. 8, pp. 1887–1899, 2018.
View at: Publisher Site | Google Scholar
G. Chen, P. Wei, H. Jiang, and M. Liu, “Formal Language generation for fault diagnosis with spectral logic via adversarial training,” IEEE Transactions on Industrial Informatics, vol. 18, no. 1, pp. 119–129, 2022.
View at: Publisher Site | Google Scholar
R. Zeng, L. Zhang, Y. Xiao et al., “An approach on fault detection in diesel engine by using symmetrical polar coordinates and image recognition,” Advances in Mechanical Engineering, vol. 6, Article ID 273929, 2014.
View at: Publisher Site | Google Scholar
D. Zhang, X. Ding, W. Huang, and Q. He, “Transient signal analysis using Parallel time-frequency manifold filtering for bearing health diagnosis,” IEEE Access, vol. 7, pp. 175277–175289, 2019.
View at: Publisher Site | Google Scholar
S. M. Ramteke, H. Chelladurai, and M. Amarnath, “Diagnosis and classification of diesel engine components faults using time–frequency and machine learning approach,” Journal of Vibration Engineering & Technologies, vol. 10, no. 1, pp. 175–192, 2022.
View at: Publisher Site | Google Scholar
W. Liu, X. Yang, and S. Jinxing, “An integrated fault identification approach for rolling bearings based on dual-tree complex wavelet packet transform and generalized composite multiscale amplitude-aware permutation entropy,” Shock and Vibration, vol. 2020, Article ID 8851310, 18 pages, 2020.
View at: Publisher Site | Google Scholar
A. Naimi, J. Deng, P. Doney, A. Sheikh-Akbari, S. R. Shimjith, and A. J. Arul, “Machine learning-based fault diagnosis for a PWR nuclear power plant,” IEEE Access, vol. 10, pp. 126001–126010, 2022.
View at: Publisher Site | Google Scholar
X. Bi, S. Cao, and D. Zhang, “Diesel engine valve clearance fault diagnosis based on improved variational mode decomposition and bispectrum,” Energies, vol. 12, no. 4, p. 661, 2019.
View at: Publisher Site | Google Scholar
L. Yuan, D. Lian, X. Kang, Y. Chen, and K. Zhai, “Rolling bearing fault diagnosis based on convolutional neural network and support vector machine,” IEEE Access, vol. 8, pp. 137395–137406, 2020.
View at: Publisher Site | Google Scholar
L. Hou, J. Zou, C. Du, and J. Zhang, “A fault diagnosis model of marine diesel engine cylinder based on modified genetic algorithm and multilayer perceptron,” Soft Computing, vol. 24, no. 10, pp. 7603–7613, 2020.
View at: Publisher Site | Google Scholar
A. Khorram, M. Khalooei, and M. Rezghi, “End-to-end CNN + LSTM deep learning approach for bearing fault diagnosis,” Applied Intelligence, vol. 51, no. 2, pp. 736–751, 2021.
View at: Publisher Site | Google Scholar
Z. Jiang, Y. Lai, J. Zhang, H. Zhao, and Z. Mao, “Multi-factor operating condition recognition using 1D convolutional long short-term network,” Sensors, vol. 19, no. 24, p. 5488, 2019.
View at: Publisher Site | Google Scholar
T. Lu, F. Yu, B. Han, and J. Wang, “A generic intelligent bearing fault diagnosis system using convolutional neural networks with transfer learning,” IEEE Access, vol. 8, pp. 164807–164814, 2020.
View at: Publisher Site | Google Scholar
W. Li, X. Zhong, H. Shao, B. Cai, and X. Yang, “Multi-mode data augmentation and fault diagnosis of rotating machinery using modified ACGAN designed with new framework,” Advanced Engineering Informatics, vol. 52, Article ID 101552, 2022.
View at: Publisher Site | Google Scholar
J. Duan, C. Hu, X. Zhan, H. Zhou, G. Liao, and T. Shi, “MS-SSPCANet: a powerful deep learning framework for tool wear prediction,” Robotics and Computer-Integrated Manufacturing, vol. 78, Article ID 102391, 2022.
View at: Publisher Site | Google Scholar
J. Duan, X. Zhang, and T. Shi, “A Hybrid attention-based paralleled deep learning model for tool wear prediction,” Expert Systems with Applications, vol. 211, Article ID 118548, 2023.
View at: Publisher Site | Google Scholar
Z. Zhao, Q. Zhang, X. Yu et al., “Applications of unsupervised deep transfer learning to intelligent fault diagnosis: a survey and comparative study,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–28, 2021.
View at: Publisher Site | Google Scholar
Z. Chen, K. Gryllias, and W. Li, “Intelligent Fault diagnosis for rotary machinery using transferable convolutional neural network,” IEEE Transactions on Industrial Informatics, vol. 16, no. 1, pp. 339–349, 2020.
View at: Publisher Site | Google Scholar
P. Ma, H. Zhang, W. Fan, C. Wang, G. Wen, and X. Zhang, “A novel bearing fault diagnosis method based on 2D image representation and transfer learning-convolutional neural network,” Measurement Science and Technology, vol. 30, no. 5, Article ID 55402, 2019.
View at: Publisher Site | Google Scholar
W. Zhang, T. Zhang, G. Cui, and Y. Pan, “Intelligent machine fault diagnosis using convolutional neural networks and transfer learning,” IEEE Access, vol. 10, pp. 50959–50973, 2022.
View at: Publisher Site | Google Scholar
W. Zhang, P. Zhang, X. He, and D. Zhang, “Convolutional neural network based two-layer transfer learning for bearing fault diagnosis,” IEEE Access, vol. 10, pp. 109779–109794, 2022.
View at: Publisher Site | Google Scholar
B. Zhao, X. Zhang, Z. Zhan, and S. Pang, “Deep multi-scale convolutional transfer learning network: a novel method for intelligent fault diagnosis of rolling bearings under variable working conditions and domains,” Neurocomputing, vol. 407, pp. 24–38, 2020.
View at: Publisher Site | Google Scholar
S. Zhong, S. Fu, and L. Lin, “A novel gas turbine fault diagnosis method based on transfer learning with CNN,” Measurement, vol. 137, pp. 435–453, 2019.
View at: Publisher Site | Google Scholar
J. Yang, S. Li, and W. Xu, “Active learning for visual image classification method based on transfer learning,” IEEE Access, vol. 6, pp. 187–198, 2018.
View at: Publisher Site | Google Scholar
H. Huangfu, Y. Zhou, J. Zhang, S. Ma, Q. Fang, and Y. Wang, “Research on inter-turn short circuit fault diagnosis of electromechanical actuator based on transfer learning and VGG16,” Electronics, vol. 11, no. 8, p. 1232, 2022.
View at: Publisher Site | Google Scholar
Y. Fu, X. Chen, Y. Liu, C. Son, and Y. Yang, “Multi-source information fusion fault diagnosis for gearboxes based on SDP and VGG,” Applied Sciences, vol. 12, no. 13, p. 6323, 2022.
View at: Publisher Site | Google Scholar
A. Khan, J. S. Kim, and H. S. Kim, “Damage detection and isolation from limited experimental data using simple simulations and knowledge transfer,” Mathematics, vol. 10, no. 1, p. 80, 2021.
View at: Publisher Site | Google Scholar
B. Petrovska, E. Zdravevski, P. Lameski, R. Corizzo, I. Štajduhar, and J. Lerga, “Deep learning for feature extraction in remote sensing: a case-study of aerial scene classification,” Sensors, vol. 20, no. 14, p. 3906, 2020.
View at: Publisher Site | Google Scholar
X. An, H. Zeng, and C. Li, “Demodulation analysis based on adaptive local iterative filtering for bearing fault diagnosis,” Measurement, vol. 94, pp. 554–560, 2016.
View at: Publisher Site | Google Scholar
J. Flett and G. M. Bone, “Fault detection and diagnosis of diesel engine valve trains,” Mechanical Systems and Signal Processing, vol. 72-73, pp. 316–327, 2016.
View at: Publisher Site | Google Scholar
X. Wang, C. Liu, F. Bi, X. Bi, and K. Shao, “Fault diagnosis of diesel engine based on adaptive wavelet packets and EEMD-fractal dimension,” Mechanical Systems and Signal Processing, vol. 41, no. 1-2, pp. 581–597, 2013.
View at: Publisher Site | Google Scholar
G. Li, J. Wu, C. Deng, Z. Chen, and X. Shao, “Convolutional neural network-based bayesian Gaussian mixture for intelligent fault diagnosis of rotating machinery,” IEEE Transactions on Instrumentation and Measurement, vol. 70, pp. 1–10, 2021.
View at: Publisher Site | Google Scholar

Copyright

Copyright © 2023 Yijie Cai et al. This is an open access article distributed under the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

PDF Download Citation

Download other formats

Order printed copies