
==== Front
Heliyon
Heliyon
Heliyon
2405-8440
Elsevier

S2405-8440(24)12642-4
10.1016/j.heliyon.2024.e36611
e36611
Research Article
Compressor fault diagnosis and result visualization based on fusion of vision transformer and improved residual network
Duan Xianling a
Hu Shaolin hfkth@126.com
b⁎
Wang Sijing c
Duan Ru d
a School of Information and Control Engineering Jilin Institute of Chemical Technology, Jilin, Jilin, 132000, China
b Automation School Guangdong University of Petrochemical Technology, Maoming, Guangdong, 52500, China
c School of Optoelectronic Engineering, Xi'an Technological University, Xian, Shanxi, 710000, China
d Xi'an Peri Power Semiconductor Converting Technology Co., Ltd, Xian, Shanxi, 710000, China
⁎ Corresponding author. hfkth@126.com
23 8 2024
15 9 2024
23 8 2024
10 17 e366119 8 2024
19 8 2024
19 8 2024
© 2024 The Authors
2024
https://creativecommons.org/licenses/by-nc-nd/4.0/ This is an open access article under the CC BY-NC-ND license (http://creativecommons.org/licenses/by-nc-nd/4.0/).
Compressors are important production equipment in the petrochemical industry, and the accuracy of their fault diagnosis is critical. In order to detect and diagnose compressor equipment faults in a timely manner, this paper constructs a deep residual shrinkage visual network (DRS-ViT). The network comprises a modified residual network (ResNet) and a vision transformer (ViT). The obtained compressor vibration signals were transformed into gram angle sum field (GASF) plots using gram angle field (GAF). The resulting image is the passed through a modified ResNet network to extract initial features. The extracted feature images are subsequently input into the ViT model for fault classification. The experimental results demonstrate that the fault diagnosis accuracy achieved by the DRS-ViT model is 99.5 %. The visualization of the model indicates that it can effectively identify the fault points. The validity and robustness of the DRS-ViT model are confirmed through comparison and analysis with various models.

Keywords

Compressor
Residual network
Vision transformer
Gram angle field
Fault diagnosis
Deep learning
==== Body
pmc1 Introduction

In the petrochemical industry, large compressor units [[1], [2], [3]], which are crucial power equipment, are susceptible to failure due to harsh operating conditions and complex control systems. In the past, the acquisition and real-time analysis of mechanical operation status was not feasible due to technological limitations [4]. However, with the emergence of the industrial Internet, it has become feasible to obtain and evaluate the real-time operational condition of machinery [5]. Conventional methods for fault analysis such as the fourier transform (STFT) [6], wavelet transform [7], and wavelet packet analysis [8], heavily depend on the analysis of vibration signals to detect possible mechanical faults by examining the spectral, amplitude, and phase characteristics of these signals. These techniques encounter certain limitations when handling complex or nonlinear mechanical signals.

With the advancement of artificial intelligence technology, the limitations encountered by conventional fault diagnosis methods have been partially resolved. Numerous deep learning models have been implemented in the domain of fault diagnosis [9], enabling the automatic extraction and acquisition of valuable features from unprocessed vibration signals. Consequently, this facilitates more precise fault detection and classification. For instance, Bao et al. [10] have developed a novel deep learning model known as the group self-attention network (GSAN). This model employs intra-group self-attention and inter-group self-attention mechanisms to effectively capture both local and global features. Additionally, it integrates these features through the fusion gate technique. The experimental results indicate that GSAN attains a high level of prediction accuracy and exhibits exceptional noise robustness when compared to other models based on RNN [[11], [12], [13]] or CNN [[14], [15], [16]]. Cabrera et al. [17] utilized a simplified long short-term memory (LSTM) network [18,19] to analyze vibration signals. They employed Bayesian methods [20] to optimize hyperparameters and achieved a diagnostic accuracy of up to 93 % in identifying double-stage mechanical intake and exhaust valve faults. Wei et al. [21] team's Data augmentation composite multi-scale network (DACMSN) extends training samples through the use of data augmentation techniques. By incorporating a multi-scale network architecture, it enhances the model's capability for feature extraction and improves its robustness. Zhang et al. [22] innovatively utilized a method based on Convolutional deep belief networks (CDBN), for denoising and feature extraction of signals from various sensors. The accuracy of fault diagnosis in real-world applications achieved a rate of 91.89 %. Deng et al. [23] proposed a semi-supervised integrated fault diagnosis framework called adversarial decoupled auto-encoder and low-dimensional feature distance metric (ADAE-LFDM). This framework effectively separates normal and fault samples, allowing for the application of the adversarial decoupled auto-encoder (ADAE) to capture and decouple fault features, resulting in high accuracy fault diagnosis. Zhang et al. [24] introduced the convolutional block attention mechanism conditional regularized least squares generative adversarial network (CBAM-CRLSGAN) method, which achieves a diagnostic accuracy of 99.35 % in scenarios with small fault sample. Liu et al. [25], proposed a bearing fault diagnosis method that utilizes the gray level cooccurrence matrix (GLCM) and dual channel convolutional neural network (DCCNN) to diagnose bearing faults under variable speed conditions.

The aforementioned researchers have proposed various models for fault diagnosis. However, due to the nature of deep learning networks as black-box diagnostic algorithm, it is impossible to determine whether the algorithm accurately identified the fault point or not, and there is no reasonable interpretation of the diagnostic results. Therefore, this paper is dedicated to enhancing the accuracy of fault diagnosis and visualizing of diagnostic results through the design of a DRS-ViT model. The DRS-ViT model effectively enhances the feature representation and improves fault diagnosis accuracy. The study's contributions can be summarized as follows: (1) The integration of a network shrinkage layer into the traditional ResNet model is intended to address the limitations in extracting crucial feature information from complex signals. (2) A novel DRS-ViT neural network model has been developed, which combines an enhanced ResNet with ViT to improve the accuracy of compressor fault diagnosis. (3) The vibration signal is transformed into a two-dimensional image using a Gram angle field, and the coding method ensures a unique mapping of the image to the vibration signal while preserving losing important information.

2 Materials and methods

2.1 Materials

The rapid advancement in the field of fault diagnosis can be attributed to the continuous improvement of neural network models [[26], [27], [28]]. These models have the capability to effectively process and analyze vast amounts of data collected from different sensors. Notably, residual learning and Vision Transformer models have emerged as crucial factors in enhancing the performance of these networks.

2.1.1 Residual learning

Residual learning is a technique used to train deep neural networks by incorporating "residual blocks". It was first introduced by He et al. [29] in 2015. This approach is particularly suitable for constructing deep network architectures, such as the residual network (ResNet), which have demonstrated significant advancements in various vision tasks.

In conventional neural networks, each layer aims to directly learn a target mapping. However in residual networks, the differences between the input and output are learned to represent a residual mapping. This enables residual networks to effectively tackle the degradation commonly encountered in deeper networks during the training process. The residual block, as depicted in Fig. 1, is defined as y=H(x,W)+x [29], where x and y represent the input and output vectors, and H denotes the learned residual function.Fig. 1 Residual module.

Fig. 1

2.1.2 Vision transformer

ViT is a neural network architecture that was initially proposed by Dosovitskiy et al. [30] at google research for image recognition tasks. ViT breaks down an image into a series of small patches, which are then treated as elements in a sequence. It utilizes a standard Transformer architecture to process these image patches, effectively converting the image into sequence inputs. ViT employs self-attention mechanisms to handle the relationships between pixels [31], and ultimately, classification is performed using an MLP (multi-layer perceptron). ViT is well-suited for tasks related to image recognition and computer vision [32].

Due to the complex operating environment of the compressor, the vibration signals collected by sensors exhibit significant fluctuations. This necessitates the use of more convolution operations to accurately extract effective features for diagnosing fault types. As the depth of the network increases, the model may encounter issues such as gradient vanishing or exploding [33]. Taking inspiration from deep residual networks and ViT, this paper proposes a deep residual shrinkage visual network (DRS-ViT). This network not only leverages the ability of convolutional operations to extract features, but also addresses the problem of gradient vanishing in deep network training and improves the convergence speed of the network.

2.2 Methods

2.2.1 Preprocessing of experimental data and sample construction

In the experiment, a total of 6000 images were generated, with 1000 images generated for each type of signal. Out of these, 4800 images were utilized for training the network model, while the remaining 1200 images were employed to test the model's recognition and classification capabilities. To ensure the fairness and reliability of the experimental outcomes, each type of model was repeated five times, with all parameters being held constant.

Centrifugal compressors are highly intricate mechanical devices comprised of multiple subsystems. When conveying gases through centrifugal compressor units, it is essential to utilize various shaft-end sealing devices in order to prevent or minimize gas leakage into the atmosphere along the rotating shaft end of the compressor. This is necessary to maintain the normal operation of the main engine, reduce material and energy consumption, prevent environmental pollution, and ensure the safety of personnel and equipment. The four commonly used forms of centrifugal compressor seals are pumping or inflatable labyrinth seals, floating ring seals, mechanical seals, and dry air seals. According to statistics on the causes of centrifugal compressor failures, failures in the sealing oil and lubricant system account for as much as 55%–65 % of the total.

Dataset A: Data A was collected through the northern branch of a petroleum pipeline limited liability company, specifically from the Tangshan station. The data consists of real-time operational data on six types of failures of centrifugal compressors. This includes the supply gas temperature for the gas seal, inlet temperature, lubricating oil sump temperature, electronic stator temperature, auxiliary thrust bearing temperature, and primary thrust bearing temperature. The codes assigned to each dataset are as follows: GQWD, YSJJKWD, YXWD, ZDJDZRZW1, YSJZTZCWD, YSJZTZCFW.

Dataset B: Data B was collected through the northern branch of a petroleum pipeline limited liability company, specifically from the Gaoling station. The data consists of real-time operational data on six types of failures of centrifugal compressors. This includes the lubricating oil tank liquid level, lubricating oil tank temperature, lubricating oil cooler outlet temperature, humidity inside the variable frequency drive cabinet, high-pressure dry gas seal vent pressure A, and low-pressure dry gas seal vent pressure A. The codes assigned to each dataset are as follows: LIA310, TIA311, TICA331, Fcchum, PISA501A, PISA502A.

The time series data that was obtained was transformed into 2D images using the GAF method. The GAF method combines coordinate transformation and the Gramian matrix to encode the time series into images. The Gramian matrix is made up of inner products of pairwise vectors, which can capture the temporal dependencies of the time series but may not effectively distinguish valuable information from Gaussian noise. Therefore, before undergoing the Gramian matrix transformation, the time series needs to undergo a spatial transformation. A common approach is to convert Cartesian coordinates into polar coordinates (radius, angle) [34,35]. The main steps in constructing a GAF are as follows:1) The original time series signal X={xt,t=1,2,…,N} is scaled to the range of [-1,1] using equation (1), and subsequently normalized

(1) x˜i=(xi−maxX)+(xi−minX)maxX−minX

where x˜i maps to an inverse cosine function.2) Thus we can represent the rescaled time series X˜ in polar coordinates by encoding the value as the angular cosine and the time stamp as the radius with the equation below [35]:

(2) {φi=arccos(x˜i),−1≤x˜i≤1,x˜i∈X˜ri=tiN,ti∈N

where ti is the time stamp, N is a constant factor to regularize the span of the polar coordinate system. This polar coordinate based representation is a way to understand time series. As time increases, corresponding values warp among different angular points on the spanning circles, like water rippling. The encoding map of equation (2) has two important properties. First, it is bijective as cos(φ) is monotonic when φ∈[0,π]. Given a time series, the proposed map produces one and only one result in the polar coordinate system with a unique inverse map. Second, polar coordinates preserve absolute temporal relations.3) After transforming the rescaled time series into the polar coordinate system, we can easily exploit the angular perspective by considering the trigonometric sum between each point to identify the temporal correlation within different time intervals. The GASF defined as follows [35]:

(3) GS=(cos(φ1+φ1)⋯cos(φ1+φn)⋮⋱⋮cos(φn+φ1)⋯cos(φn+φn))

Through this transformation, the vibration signal is converted into a Gram matrix that is symmetrically distributed along the diagonal.4) Finally, each element of the matrix is scaled to a value between 0 and 255, which corresponds to the pixel size of each point in the image. This process converts the 2D matrix into GASF.

2.2.2 An intelligent fault diagnosis methods combining ViT and improved deep residual networks

The time series signals collected from compressors in industrial environments are frequently affected by changes in operating conditions, resulting in non-stationary characteristics where their statistical features change over time. The traditional ResNet model has limitations in extracting important features from complex signals, which hinders its ability to achieve the desired accuracy in fault diagnosis. In order to enhance the accuracy of compressor fault diagnosis, this study improves the ResNet model by refining the data processing pipeline and adjusting the network architecture to better handle complex industrial signals. The enhancements to the model primarily focus on the following aspects:(1) To maximize the model's ability to learn various features, the size of the convolutional kernel in the initial convolutional layer is modified from 7 × 7 to 3 × 3, with a stride of 1 and padding of 1. Once it goes through the Batch Normalization (BN) layer and ReLU activation function, it is then inputted into the residual module.

(2) In the basic residual block, a network contraction layer is incorporated, which uses a convolutional kernel size of 3 × 3 and a stride of 1. The network contraction layer calculates the absolute values of each element in the input tensor and then computes the mean along the first dimension to obtain a global feature representation. The output tensor maintains the same dimensions as the input tensor, allowing for the weighting of features and highlighting important characteristics.

(3) To mitigate the issue of overfitting caused by the increase in network depth, a Batch Normalization (BN) layer is incorporated after each convolutional layer to enhance the model's generalization ability.

As depicted in Fig. 2, the DRS-ViT model integrates the enhanced residual network with ViT to enhance the accuracy of fault diagnosis. The incorporation of a network shrinkage layer into conventional ResNet models overcomes the constraints associated with extracting vital feature information from intricate signals. The enhanced ResNet model is then combined with ViT to formulate a DRS-ViT network, enabling the network to acquire more profound feature representations and consequently enhancing the training efficiency and performance of the network.Fig. 2 DRS-ViT model network structure.

Fig. 2

As depicted in Fig. 3, the intelligent fault diagnosis process of the compressor using the DRS-ViT model can be divided into the following steps:(1) Time series data related to the centrifugal compressor from Tangshan Station and Gaoling Station were collected through the Northern Branch of a specific petroleum pipeline company.

(2) The collected time series data were converted into GASF 2D images using GAF. Subsequently, these images were divided into a test set and a training set.

(3) The deep residual model was enhanced by adjusting the network structure, hyperparameters, and adding a network contraction layer. The images transformed by GAF were then input into the network model.

(4) The images processed by DRS were fed into the ViT model for further feature extraction. The model parameters were updated through backpropagation until the accuracy met the expected requirements.

(5) Construction of the DRS-ViT model. DRS-ViT was compared with other models such as ResNet [29], multiscale convolutional neural network (MsCNN) [36], ConvNeXt [37], GoogleNet [38], ResNetVit and MobileNetV4 [39] to analyze the performance of each model. Additionally, the models were visualized using t-SNE.

Fig. 3 Fault diagnosis flowchart.

Fig. 3

3 Results and discussion

3.1 Performance analysis of model DRS-ViT and comparison with other models

To evaluate the performance and validate the DRS-ViT model, experiments were conducted on two distinct datasets. The samples were designed with a shape of 1024 × 1. These samples were then transformed into images of 480 × 640 using GAF and fed into the neural network model. The Adam optimization algorithm was utilized for gradient updates. The optimal hyperparameters were determined through a network search. The training process consisted of 50 epochs with a batch size of 32. The learning rate was set to 0.001. The training hardware environment included of an AMD EPYC 7543 processor with 80 GB of memory and an NVIDIA A40 graphics card with 48 GB of memory. PyTorch was employed as the deep learning framework for conducting the experiments.

As depicted in Fig. 4, Fig. 5, on dataset A and dataset B, the DRS-ViT model exhibits a substantial increase in both training and testing accuracy as the number of epochs increases, indicating its remarkable rapid learning capabilities. The subsequent convergence of accuracy towards a high value close to 1.0 suggests that the model's predictions are highly precise. After an initial rapid rise, the accuracy remains consistently high, demonstrating its sustained stability. Furthermore, the test accuracy closely aligns with the training accuracy, indicating that the model possesses good generalization ability for unseen data.Fig. 4 Accuracy and loss on dataset A

Fig. 4

Fig. 5 Accuracy and loss on dataset B.

Fig. 5

In terms of training and testing loss, the data reveals a substantial decrease in the model's loss during the initial few epochs, indicating its capacity for rapid improvement. Following the initial decline, the loss values stabilize and remain consistently low. The convergence and close proximity of the training and testing loss curves indicate that the model exhibits strong convergence without overfitting.

The DRS-ViT model demonstrates superior generalization ability and stability. The accuracy and loss maps of the training and test data indicate minimal overfitting, suggesting that the model performs similarly well on both the training and unseen test data. After an initial phase of rapid learning, the model reaches a stable state, maintaining its performance as the amount of data increases or over time. Furthermore, the model exhibits fast convergence to high accuracy and low loss, making it suitable for fault identification in engineering practice.

The validation of our proposed model, the deep residual shrinkage visual network (DRS-ViT), is compared with four other models: ResNet, MsCNN, ConvNeXt, GoogleNet, ResNetVit, and MobileNet. Cross-validation is conducted six times for each method, and the final result is the average of the five experiments. This approach helps eliminate the potential error of a single experiment that could be influenced by a local optimum.

The results are presented in Table 1, where accuracy assesses the validity of the method and standard deviation (STD) evaluates the stability of the model. Text in blue font indicates the method with the highest accuracy within the same dataset. The slanted font is used to indicate the experimental results of DRS-ViT.Table 1 Accuracy and STD of 5 models on different datasets.

Table 1Model	Dataset	
Train accuracy ± STD	Test Accuracy ± STD	
A	B	A	B	
ResNet	74.38 % ± 0.13	83.83 % ± 0.12	67.25 % ± 0.16	82.67 % ± 0.25	
MsCNN	99.90 % ± 0.30	99.21 % ± 0.21	96.83 % ± 0.29	98.92 % ± 0.22	
ConvNeXt	99.48 % ± 0.31	60.71 % ± 0.10	99.42 % ± 0.31	49.75 % ± 0.07	
GoogleNet	48.85 % ± 0.07	55.00 % ± 0.09	34.42 % ± 0.05	46.67 % ± 0.06	
ResNetVit	99.98 % ± 0.15	85.57 % ± 0.14	99.67 % ± 0.17	83.75 % ± 0.17	
MobileNetV4	100.0 % ± 0.15	81.65 % ± 0.27	99.42 % ± 0.16	82.00 % ± 0.28	
DRS-ViT	99.35 %±0.19	99.10 %±0.20	99.50 %±0.23	99.58 %±0.20	

On the A training dataset, the MobileNetV4 model achieved the highest accuracy of 100 % with a STD of 0.15. On the A test dataset, the DRS-ViT model achieved an accuracy of 99.5 %, which is 0.08 % higher than that of the MobileNetV4 model, with a STD of 0.23. Although GoogleNet has a minimum STD of 0.05, its accuracy is relatively low.

On the B training dataset, the MsCNN model achieved the highest accuracy of 99.21 % with a STD of 0.21. Despite being less accurate than MsCNN on the training dataset, the DRS-ViT model has a STD of 0.20. On the B test dataset, the DRS-ViT model outperforms both the ResNetVit and MobileNetV4 models with an accuracy of 99.58 %, which is approximately 15 % higher. The STD for the DRS-ViT model on the B test dataset is 0.2.

On the A test dataset, the accuracy of the DRS-ViT model, although not the highest, was only 0.17 % lower than the highest performing ResNetVit model. On the B test dataset, the DRS-ViT model achieves the highest accuracy of 99.58 %, which is significantly higher than the accuracy of the ResNetVit model.

The accuracy convergence curves for the seven methods are depicted in Fig. 6. It is evident from Fig. 6(a) and (b) that the proposed model exhibits faster convergence on both dataset A and dataset B. The loss convergence curves are presented in Fig. 7. It can be observed from Fig. 7(a) and (b) that although the proposed model experiences a sudden increase in dataset A, it quickly stabilizes. On dataset B, the proposed model achieves the fastest rate of stabilization. The effectiveness of the DRS-ViT models is maximized when they demonstrate high accuracy and low volatility.Fig. 6 Accuracy convergence curves on test sets A and B.

Fig. 6

Fig. 7 Loss convergence curves on test sets A and B.

Fig. 7

3.2 Comparison of confusion matrices between models

In order to further demonstrate the effectiveness of the proposed method for fault diagnosis, a comparison is made between the confusion matrices of the various models on data set A and data B. The confusion matrix for each model in dataset A is shown in Fig. 9, with rows representing predicted labels and columns representing true labels. The diagonal cells indicate the rate of correctly recognized samples and non-diagonal cells correspond to the rate of incorrectly recognized samples. The percentage of the total number of samples is displayed in each cell.

The confusion matrix in Fig. 8(a) illustrates the recognition accuracies of ConvNeXt for six compressor signals, namely GQWD, YSJJKWD, YSJZTZCFW, YSJZTZCWD, YXWD, and ZDJDZRZW1. The accuracies for these signals 99 %, 99 %, 99.5 %, 100 %, 99.5 %, and 99.5 % respectively.Fig. 8 Confusion matrix a(ConvNeXt), b(GoogLeNet), c(MsCNN), d(ResNet), e(DRS-VIT), f(ResNetVit), g(MobileNetV4).

Fig. 8

Fig. 9 Confusion matrix a(ConvNeXt), b(GoogLeNet), c(MsCNN), d(ResNet), e(DRS-VIT), f(ResNetVit), g(MobileNetV4).

Fig. 9

On the other hand, the confusion matrix in Fig. 8(b) demonstrates the recognition performance of GoogLeNet for the same six compressor signal types. The accuracies achieved by GoogLeNet are 93.5 % for GQWD, 12 % for YSJJKWD, 19 % for YSJZTZCFW, 26.5 % for YSJZTZCWD, 26.5 % for YXWD, and 29 % for ZDJDZRZW1. It is worth noting that GoogLeNet's recognition performance on dataset A is poor.

As depicted in the confusion matrix of MsCNN in Fig. 8(c), MsCNN achieves high accuracies in recognizing different types of compressor signals. Specifically, the accuracies for GQWD, YSJJKWD, YSJZTZCFW, YSJZTZCWD, YXWD, and ZDJDZRZW1 are 99 %, 87.5 %, 100 %, 95.5 %, 99 % and 100 % respectively. It is worth nothing that the model achieves a perfect accuracy of 100 % in recognizing the compressor lubricating oil tank temperature and the compressor main thrust bearing temperature. However, the accuracy in recognizing the compressor inlet temperature is relatively low.

In Fig. 8(d), ResNet incorrectly predicts a higher number of test samples between YSJZTZCFW and YSJZTZCWD. Specifically, the recognition accuracies of ResNet for the six types of data(GQWD, YSJJKWD, YSJZTZCFW, YSJZTZCWD, YXWD, and ZDJDZRZW1) are 99 %, 72 %, 24 %, 48.5 %, 60.5 %, and 99.5 % respectively. As shown in Fig. 8(e), our DRS-VIT method achieves intelligent recognition of the six signal types of the compressor, with a recognition accuracy of 99.5 % for each type.

From the confusion matrices presented in Fig. 8(f) and (g), it is evident that ResNetVit achieves accurate recognition of the six compressor signals (GQWD, YSJJKWD, YSJZTZCFW, YSJZTZCWD, YXWD, and ZDJDZRZW1) with accuracies of 100 %, 99 %, 99.5 %, 99.5 %, 100 %, and 100 % respectively. On the other hand, MobileNetV4 demonstrates a recognition accuracy of 99.5 % for all six compressor signals.

The confusion matrix for the five models in dataset B is depicted in Fig. 9. It is evident from the confusion matrices in Fig. 9(a) and (b) that both ConvNeXt and GoogLeNet made incorrect predictions for a higher numberi of test samples between TIA311 and Fcchum. The recognition accuracies of ConvNeXt for the six types of data (Fcchum, LIA310, PISA501A, PISA502A, TIA311, and TICA331) were 87.5 %, 56.5 %, 43 %, 67 %, 8 %, and 36.5 % respectively. Similarly GoogLeNet achieved recognition accuracies of 81 %, 65 %, 26.5 %, 72.5 %, 1.5 %, and 33.5 % for the six types of data. The confusion matrix of MsCNN in Fig. 9(c) demonstrates that MsCNN achieves accurate recognition of the six signal types of the compressor with accuracies of 100 % (Fcchum), 99.5 % (LIA310), 99 % (PISA501A), 97 % (PISA502A), 100 % (TIA311), and 98 % (TICA331). It is noteworthy that the model accurately identifies the humidity inside the compressor inverter cabinet as well as the lubricant mailbox temperature with 100 % accuracy. Furthermore the confusion matrix of ResNet in Fig. 9(d) reveals, it can be seen that the recognition accuracies of ResNet for the six compressor signals (Fcchum, LIA310, PISA501A, PISA502A, TIA311, and TICA331) are 96.5 %, 95 %, 77.5 %, 70.5 %, 67.5 %, and 89 % respectively.

From the confusion matrices presented in Fig. 9(f) and (g), it is evident that ResNetVit accurately identifies the six compressor signals (Fcchum, LIA310, PISA501A, PISA502A, TIA311, and TICA331) with accuracies of 99.5 %, 99.5 %, 82.5 %, 22.5 %, 99 %, and 99.5 %, respectively. On the other hand, MobileNetV4 were 99 %, 96 %, 63 %, 49.5 %, 89.5 %, and 99.5 % for the same signals.

Finally, as depicted in Fig. 9(e), our DRS-ViT method achieves intelligent recognition of the six signal types of the compressor with recognition accuracies of 99.5 % (Fcchum), 99.5 % (LIA310), 100 % (PISA501A), 99.5 % (PISA502A), 99.5 % (TIA311), and 99.5 % (TICA331). It is evident that our proposed model achieves 100 % accuracy in recognizing the dry gas seal venting pressure A at the high pressure end of the compressor, and 99.5 % accuracy in recognizing the other five signal types. Therefore, it can be concluded that the proposed model exhibits higher recognition accuracy on both dataset A and dataset B, surpassing the performance of the other four models.

The models were validated uisng dataset A and dataset B. The DRS-ViT model proposed in this paper achieves an average accuracy of 99.5 % for both datasets, which is significantly superior to other models. In dataset A, DRS-ViT accurately recognizes each fault type with a 99.5 % accuracy rate. In dataset B, the term PISA501A refers to a problem with the dry gas sealing system at the high-pressure end (typically the outlet end of the compressor), while PISA502A refers to a problem with the dry gas sealing system at the low-pressure end (typically the suction end). Despite occurring in different locations, having different characteristics, and requiring different approaches for resolution, both failures fall under the category of dry gas sealing system failure. The ResNetVit and MobileNetV4 models were unable to accurately identify both fault types, resulting in a decrease in their accuracy on dataset B. On the other hand, the DRS-ViT model proposed in this paper demonstrates excellent recognition of different fault types and achieves a 100 % accuracy rate for TICA331 recognition.

3.3 Visual analysis

In order to enhance the validation of the feature learning capability of each model, t-SNE was employed to reduce the dimensionality of the features in the fully connected layers of ConvNeXt, GoogLeNet, MsCNN, ResNet, DRS-ViT, ResNetVit, and MobileNetV4 for visualization and analysis. Fig. 10(a–g) depict the visual characteristics of the seven models in the fully connected layer on test set A. Each point in the figure represents a sample, and the various colors indicate different types of signals from the compressor.Fig. 10 A test set of T-SNE visualization of fully connected layer features a(ConvNeXt), b(GoogLeNet), c(MsCNN), d(ResNet), e(DRS-VIT), f(ResNetVit), g(MobileNetV4).

Fig. 10

As depicted in Fig. 10, Fig. 11 (a-g), the ConvNeXt and MsCNN models exhibited superior within-class clustering performance for the same on the A test set, while also demonstrating greater separation performance for different classes. However, on the B test set, the ConvNeXt model displayed inadequate feature separation, as evidenced by the feature overlap observed between Fcchum, LIA310, PISA501A, PISA502A, TIA311, and TICA331. The MsCNN models of PISA501A and PISA502A exhibited even more feature overlap.Fig. 11 B test set T-SNE visualization of fully connected layer features a(ConvNeXt), b(GoogLeNet), c(MsCNN), d(ResNet), e(DRS-VIT), f(ResNetVit), g(MobileNetV4).

Fig. 11

As depicted in Fig. 10, Fig. 11, both on test set A and test set B, the GoogLeNet model demonstrates inadequate ability to separate features, making it challenging to distinguish between different types of compressor signals. From Fig. 10, Fig. 11, it can be observedthat on test set A, the ResNet model exhibits superior intraclass clustering performance for GQWD and ZDJDZRZW1, as well as better separation performance from other signals. However, there is an overlap between YSJJKWD, YSJZTZCFW, YSJZTZCWD, and YXWD. On test set B, ResNet demonstrates poor feature separation, with overlap between PISA502A and PISA502A, LIA310 and TIA311, TIA311 and TICA331, and Fcchum and a small amount of LIA310.

As depicted in Fig. 10, Fig. 11 and 10–11(g), the ResNetVit and MobileNetV4 models demonstrate improved within-class clustering performance for the same type of signal on the A test set. Additionally, these models exhibit greater separation performance among different classes. However, in the B test set, there is a higher degree of feature overlap between the two models in both PISA501A and PISA502A.

Finally, as depicted in Fig. 10, Fig. 11, our proposed DRS-ViT model demonstrates strong intra-class clustering capability when handling similar signal types on both test set A, and test set B. It efficiently clusters similar features, indicating a high sensitivity to the inherent similarity of similar signals and the ability to accurately capture them. Additionally, the model exhibits significant separability in the feature distribution of different signal categories, enabling clear differentiation between them. This proves the effectiveness of the model in recognizing and distinguishing between different signal types. Overall, DRS-ViT accurately captures intrinsic features and effectively differentiates various types of signals during the learning and recognition process.

4 Conclusions

In this paper, we propose a new method for intelligent fault diagnosis of DRS-ViT compressors. By combining the improved ResNet network with the powerful feature extraction capability of ViT, we are able to improve the fault diagnosis accuracy. Finally, we conducted experiments on both the Tangshan and Gaoling station datasets, which resulted in better diagnostic results. Based on these findings, we draw the following conclusions:(1) The compressor vibration signals that have been obtained are transformed into 2D images using the GAF coding method. GAF utilizes a combination of coordinate transformation and Gram matrix to ensure that the time-dependent nature of the signals is preserved while capturing all the relevant information.

(2) In comparison to traditional CNN, the fault diagnosis model DRS-ViT proposed in this paper utilizes an enhanced ResNet with ViT as a feature extractor. The enhanced ResNet ensures superior feature extraction while retaining crucial information. Additionally, the ViT model employs self-attention to capture the interrelationships among various positions in the input sequence, thereby enabling the model to effectively handle long-range dependencies. Consequently, the fault diagnosis accuracy of the DRS-ViT model is enhanced.

(3) On both dataset A and dataset B, the DRS-ViT model demonstrates superior performance compared to other models such as ConvNeXt, GoogLeNet, MsCNN, ResNet, ResNetVit, and MobileNetV4, based on a comprehensive comparison. For dataset A, a recognition accuracy of 99.5 % was achieved for all six compressor signal types (GQWD, YSJJKWD, YSJZTZCFW, YSJZTZCWD, YXWD, and ZDJDZRZW1). For dataset B, a recognition accuracy of 100 % was achieved for the PISA501A compressor fault type. The accuracy for recognizing the other five fault types (Fcchum, LIA310, PISA502A, TIA311, and TICA331) was 99.5 %. Therefore, this method can be effectively applied in practical engineering for fault identification.

Data availability statement

Data associated with the study was not been deposited into a publicly available repository. The data that has been used is confidential.

CRediT authorship contribution statement

Xianling Duan: Writing – review & editing, Writing – original draft, Visualization, Validation, Methodology, Investigation, Formal analysis, Data curation, Conceptualization. Shaolin Hu: Writing – review & editing, Supervision, Resources, Funding acquisition. Sijing Wang: Writing – review & editing, Investigation, Data curation. Ru Duan: Writing – review & editing, Investigation, Data curation.

Declaration of competing interest

There are no conflicts of interest to declare. The publication of this work is tacitly and explicitly approved by all the authors and corresponding institutions.

Acknowledgements

This work was supported by the 10.13039/100014717 National Natural Science Foundation of China (62373115 ); Guangdong Basic and Applied Basic Research Foundation (2023A1515012341 , 2024A1515010870 ).
==== Refs
References

1 Chauhan R. Sartape R. Minocha N. Goyal I. Singh M.R. Advancements in environmentally sustainable technologies for ethylene production Energy Fuels 37 17 2023/09/07 2023 12589 12622
2 Ghanta M. Fahey D. Subramaniam B. Environmental impacts of ethylene production from diverse feedstocks and energy sources Applied Petrochemical Research 4 2 2014/06/01 2014 167 179
3 Stadtherr M.A. Rudd D.F. Systems study of the petrochemical industry Chem. Eng. Sci. 31 11 1976/01/01/1976 1019 1028
4 Nadeem I. A comprehensive survey on ‘circular polarized antennas’ for existing and emerging wireless communication technologies J. Phys. Appl. Phys. 55 3 2021/10/18 2022 033002
5 Alibakhshikenari M. A comprehensive survey on antennas on-chip based on metamaterial, metasurface, and substrate integrated waveguide principles for millimeter-waves and terahertz integrated circuits and systems IEEE Access 10 2022 3668 3692
6 Griffin D. Jae L. Signal estimation from modified short-time Fourier transform IEEE Trans. Acoust. Speech Signal Process. 32 2 1984 236 243
7 Torrence C. Compo G.P. A practical guide to wavelet analysis Bull. Am. Meteorol. Soc. 79 1 1998 61 78
8 Yen G.G. Lin K.-C. Wavelet packet feature extraction for vibration monitoring IEEE Trans. Ind. Electron. 47 3 2000 650 667
9 Kibrete F. Engida Woldemichael D. Shimels Gebremedhen H. Multi-Sensor data fusion in intelligent fault diagnosis of rotating machines: a comprehensive review Measurement 232 2024/06/15/2024 114658
10 Bao G. Zhang H. Wei Y. Gu D. Liu S. Fault diagnosis of reciprocating compressor based on group self-attention network Meas. Sci. Technol. 31 6 2020/04/02 2020 065014
11 Zhu J. Jiang Q. Shen Y. Qian C. Xu F. Zhu Q. Application of recurrent neural network to mechanical fault diagnosis: a review J. Mech. Sci. Technol. 36 2 2022/02/01 2022 527 542
12 Choi D.J. Han J.H. Park S.U. Hong S.K. Comparative study of CNN and RNN for motor fault diagnosis using deep learning 2020 IEEE 7th International Conference on Industrial Engineering and Applications (ICIEA), 16-21 April 2020 2020 693 696
13 Qiao M. Yan S. Tang X. Xu C. Deep convolutional and LSTM recurrent neural networks for rolling bearing fault diagnosis under strong noises and variable loads IEEE Access 8 2020 66257 66269
14 Ince T. Kiranyaz S. Eren L. Askar M. Gabbouj M. Real-time motor fault detection by 1-D convolutional neural networks IEEE Trans. Ind. Electron. 63 11 2016 7067 7075
15 Janssens O. Convolutional neural network based fault detection for rotating machinery J. Sound Vib. 377 2016/09/01/2016 331 345
16 Zhong S.-s. Fu S. Lin L. A novel gas turbine fault diagnosis method based on transfer learning with CNN Measurement 137 2019/04/01/2019 435 453
17 Cabrera D. Bayesian approach and time series dimensionality reduction to LSTM-based model-building for fault diagnosis of a reciprocating compressor Neurocomputing 380 2020/03/07/2020 51 66
18 Chen X. Zhang B. Gao D. Bearing fault diagnosis base on multi-scale CNN and LSTM model J. Intell. Manuf. 32 4 2021/04/01 2021 971 987
19 Jalayer M. Orsenigo C. Vercellis C. Fault detection and diagnosis for rotating machinery: a model based on convolutional LSTM, Fast Fourier and continuous wavelet transforms Comput. Ind. 125 2021/02/01/2021 103378
20 Amin M.T. Khan F. Ahmed S. Imtiaz S. A data-driven Bayesian network learning method for process fault diagnosis Process Saf. Environ. Protect. 150 2021/06/01/2021 110 122
21 Wei Y. A novel data augmentation and composite multiscale network for mechanical fault diagnosis IEEE Trans. Instrum. Meas. 72 2023 1 12 37323850
22 Zhang Y. Ji J. Ma B. Fault diagnosis of reciprocating compressor using a novel ensemble empirical mode decomposition-convolutional deep belief network Measurement 156 2020/05/01 2020 107619
23 Deng C. Deng Z. Miao J. Semi-supervised ensemble fault diagnosis method based on adversarial decoupled auto-encoder with extremely limited labels Reliab. Eng. Syst. Saf. 242 2024/02/01/2024 109740
24 Zhang J. CBAM-CRLSGAN: a novel fault diagnosis method for planetary transmission systems under small samples scenarios Measurement 234 2024/07/01/2024 114795
25 Liu F. Fault diagnosis of rolling bearings under varying speeds based on gray level co-occurrence matrix and DCCNN Measurement 235 2024/08/01 2024 114955
26 Jiao J. Zhao M. Lin J. Liang K. A comprehensive review on convolutional neural network in machine fault diagnosis Neurocomputing 417 2020/12/05/2020 36 63
27 Chen Z. Mauricio A. Li W. Gryllias K. A deep learning method for bearing fault diagnosis based on Cyclic Spectral Coherence and Convolutional Neural Networks Mech. Syst. Signal Process. 140 2020/06/01/2020 106683
28 Shao H. Xia M. Han G. Zhang Y. Wan J. Intelligent Fault diagnosis of rotor-bearing system under varying working conditions with modified transfer convolutional neural network and thermal images IEEE Trans. Ind. Inf. 17 5 2021 3488 3496
29 He K. Zhang X. Ren S. Sun J. Deep residual learning for image recognition 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) 2016 27 30 June 2016
30 Dosovitskiy A. An Image Is Worth 16x16 Words: Transformers for Image Recognition at Scale 2020 arXiv preprint arXiv:2010.11929
31 Lee C.P. Lim K.M. Song Y.X. Alqahtani A. Plant-CNN-ViT: plant classification with ensemble of convolutional neural networks and vision transformer Plants 12 14 2023 2642 37514256
32 Vaswani A. Attention is all you need Adv. Neural Inf. Process. Syst. 30 2017
33 Hochreiter S. The vanishing gradient problem during learning recurrent neural nets and problem solutions Int. J. Uncertain. Fuzziness Knowledge-Based Syst. 6 2 1998 107 116
34 Xiong L. He M. Hu C. Hou Y. Han S. Tang X. Image presentation and effective classification of odor intensity levels using multi-channel electronic nose technology combined with GASF and CNN Sensor. Actuator. B Chem. 395 2023/11/15/2023 134492
35 Wang Z. Oates T. Imaging Time-Series to Improve Classification and Imputation 2015 arXiv preprint arXiv:1506.00327
36 Cai Z. Fan Q. Feris R.S. Vasconcelos N. A unified multi-scale deep convolutional neural network for fast object detection Computer Vision–ECCV 2016: 14th European Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, Part IV 14 2016 Springer 354 370
37 Liu Z. Mao H. Wu C.-Y. Feichtenhofer C. Darrell T. Xie S. A convnet for the 2020s Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2022 11976 11986
38 Szegedy C. Going deeper with convolutions Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2015 1 9
39 Qin D. MobileNetV4-Universal models for the mobile ecosystem arXiv Preprint arXiv:2404.10518 2024
