An innovative approach to vibration signal denoising and fault diagnosis using attention-enriched joint learning

Abstract

Vibration signals play a crucial role in mechanical fault diagnosis. However, they are susceptible to various noise disturbances, presenting challenges for reliable fault detection. We propose an end-to-end Cross-task Attention Joint Learning (CTA-JL) model that concurrently denoises and diagnoses faults in noisy signals. This model utilizes a multi-task encoder, composed of task-shared and task-specific feature encoding units, along with a feature information exchange unit with a Cross-task Attention (CTA) mechanism, fostering information exchange across different tasks. By collectively executing diagnosis and denoising tasks and sharing valuable task information, the model enhances prediction accuracy and denoising performance. Under three noise conditions of SNR = −9 dB, −6 dB, and −3 dB, the prediction accuracy of CTA-JL on the rolling bearing datasets reached 91.38%, 97.95%, and 99.69%, respectively. Meanwhile, the result on elevator guide system datasets reached 87.31%, 95.58%, and 99.64%

Keywords

Vibration signal processing fault diagnosis attention mechanisms multi-task learning

1. Introduction

During prolonged operation, mechanical equipment experiences an increased risk of damage due to the continuous accumulation of degradation on various components. Abrupt breakdowns during operation often lead to catastrophic accidents. Therefore, timely and reliable fault diagnosis is essential to prevent potential losses (Dong et al., 2021). Depending on the distinct characteristics of equipment, diverse signals such as vibration, sound, temperature, and current can be utilized for fault diagnosis. Among them, vibration signal analysis is the most commonly employed technique for diagnosing mechanical faults (Vakharia et al., 2015) because the acquisition of vibration signals is non-intrusive, eliminating the need to disassemble or alter the mechanical system. It is relatively easy to implement. Besides, it can accurately reflect different components’ motion, amplitude, and frequency, demonstrating high reliability and accuracy. Therefore, it is significant to research and develop fault diagnosis methods based on vibration signal analysis.

In addition to the direct application of traditional deep learning techniques such as CNN (Mao et al., 2017), RNN (Liu et al., 2018), LSTM (Chen et al., 2017) for fault diagnosis, emerging deep learning technologies in recent years have increasingly focused on addressing potential real-world issues in vibration signal diagnosis. For instance, regarding high data dependency and the overlooked physical information in deep learning-based fault diagnosis, Ni et al. (2023) proposed a novel Physics-Informed Residual Network (PIResNet) that introduces physical-modal-attribute-guided generation layers, domain-conversion layer, and a parallel bi-channel residual learning architecture. It enables effective extraction of bearing fault features under variable speed and load conditions, providing a fault diagnosis solution consistent with physical laws under non-steady-state conditions. For multi-sensor signal diagnosis, Xu et al. (2023a) introduced a novel convolutional fusion framework called Collaborative Fusion Convolutional Neural Network (CFCNN), which realizes multi-level feature extraction and noise suppression through multiscale shrinkage denoising module (MSDM) and learns shared feature representations of multi-source sensor data through central fusion module (CFM). In response to the issue that mechanical equipment often operates in a fluctuating condition, where data distributions can easily change, Xu et al. (2023b) proposed an online knowledge distillation-based multiscale threshold denoising network (OKD-MTDN). The network captures discriminative features of mechanical signals through the multiscale convolutional module (MCM). Furthermore, an online knowledge distillation algorithm (OKD) enhances the model’s generalization ability. Additionally, Feng et al. (2023) proposed a novel approach that combines traditional signal analysis methods with deep learning techniques and conducts a gear health indicator (HI) through cyclic entropy and the Wasserstein distance and predicting the degradation trend of gear surface wear through an optimized gated recurrent unit (GRU).

However, in the process of vibration signal collection, noise interference is an inevitable situation that significantly impacts the accuracy of diagnosis. Particularly under high-noise backgrounds, the vibration analysis and diagnosis accuracy rate may decrease by 10%–30% (Wang et al., 2016). Therefore, it is essential to suppress such noise. The commonly employed methods can be categorized into traditional methods based on signal analysis and data-driven deep learning methods. The former includes filtering, wavelet analysis (WT) (Seryasat et al., 2010), and EMD, which require users to have rich experience and set appropriate artificial parameters to achieve excellent denoising effects. For instance, in the WT method, selecting suitable wavelet basis functions is decisive in denoising. On the other hand, deep learning techniques, such as Convolutional Neural Networks (CNN) and auto-encoders, can suppress redundant information in the input signal and highlight critical components by establishing a nonlinear mapping relationship between noisy and clean signals (Xiong et al., 2023). This approach avoids complex manual parameter settings and does not require extensive modifications to handle various tasks.

The basic process for fault diagnosis problems disturbed by noise is to eliminate background noise in vibration signals using pre-processing techniques and then perform fault diagnosis using intelligent algorithms. However, this serial structure design severely affects the model diagnosis effect by the denoising effect. If the denoising task is unreasonably designed, many key features may be misjudged as redundant information and filtered out, limiting subsequent intelligent algorithms’ feature extraction effect. Furthermore, fault diagnosis tasks mainly focus on the local characteristics of the signal, especially the specific patterns related to faults. These minute local changes may be regarded as noise and ignored in the denoising task.

This study aims to design a parallel architecture for fault diagnosis and signal denoising, ensuring direct feedback on diagnosis task performance while improving the model’s fault diagnosis anti-interference ability through learning the denoising task. Additionally, the denoising task learns latent features representing the global trend through signal reconstruction, introducing signal structure information into the diagnosis task learning. On the other hand, the fault diagnosis task uses classification label information to help the denoising network learn category difference information. Accurate fault diagnosis under a strong noise background is achieved by utilizing complementary information between tasks. The main contributions and noteworthy aspects of this research are summarized below:

1) A novel parallel framework is proposed for fault diagnosis and signal denoising, named the cross-task attention-enriched joint learning model (CTA-JL). This model simultaneously considers the signals’ category differences and inherent attributes during the feature extraction phase and enhances its noise-resistant diagnostic ability by integrating complementary information.

2) A cross-task encoder is utilized, which incorporates a task-shared encoder unit (TShEnc) and a task-specific encoder unit (TSpEnc), to extract common and exclusive features from vibration signals. Further, an information exchange unit (IEU) facilitates the interchange of feature information between different tasks, yielding fused features.

3) The efficacy of proposed joint framework is validated through two case studies. Experimental outcomes demonstrate that the CTA-JL model significantly surpasses the competitive algorithms, particularly in scenarios with high-intensity noise.

The remainder of this article is organized as follows: Section 2 describes the proposed CTA-JL model for noisy vibration signal diagnosis and denoising. Section 3 presents the experimental results and analysis of proposed method in comparison with state-of-the-art methods. Finally, Section 4 draws the conclusion of complete paper.

2. Methodology

2.1. Overall architecture of the CTA-JL model

The anti-noise fault diagnosis often employs traditional signal denoising techniques to eliminate noise interference in vibration-based fault diagnosis. However, these methods require prior knowledge of signal analysis and threshold selection. Improper selection of thresholds may lead to the filtering of compelling features, thereby affecting fault diagnosis accuracy.

To address this issue, a joint learning architecture-based approach called CTA-JL is proposed. It utilizes cross-task attention (CTA) to exchange information during the feature encoding stage and synchronously implements fault diagnosis tasks (F-task) and denoising tasks (D-task). As shown in Figure 1, the model consists of four main components: (i) feature embedding, (ii) Cross-task attention-based multi-task encoder module (CTAEnc), (iii) fault diagnosis decoder network (FdDec), and (iv) Signal denoising decoder network (SdDec). By using CTA, the CTAEnc achieves parameter soft sharing, and this approach can improve the overall performance of prediction by sharing information between different tasks and mitigate the risk of negative transfer to some extent (Wang et al., 2022). Finally, a multi-encoder composes the original signal to the latent space. Then, these features are sent to the FdDec and SdDec, respectively, to obtain the denoised vibration signal and its fault classification result under a strong noise background.

2.2. CTA-based joint-task encoder

As shown in Figure 2(a), the CTAEnc is composed of three components: a task-shared feature encoding unit (TShEnc), a task-specific feature encoding unit (TSpEnc), and a feature information-exchange unit (IEU). Firstly, the input vibration signal utilizes TShEnc to obtain its shared features, and for the F-task and the signal D-task, the deep features f_F and f_D of the corresponding task are obtained by different TSpEnc. Then the feature information of each task is exchanged by IEU, which provides the denoising-to-diagnosis feature f_D→F and the diagnosis-to-denoising feature f_F→D. Finally, the fused features ${\tilde{f}}_{F}$ and ${\tilde{f}}_{D}$ are obtained by adding the original task features to the information-exchanged feature.

2.2.1. Gated attention unit

In the joint-task encoder, each feature encoding unit comprises stacked gated attention units (GAU) and convolutional layers. The GAU extracts global feature dependencies of the vibration signal, while convolutional layers capture local feature relationships between time-domain signals and reduce computational costs. GAU was proposed by Hua et al. (2022), which combines the original Multi-Head Attention (MHA) and Feedforward Network (FFN) in the Transformer and introduces gating mechanisms. Its structure is shown in Figure 2(b). GAU improves computational performance and enhances the encoder model’s learning ability for fault and denoising features. The gating mechanism of GAU can be described as:

O = (U ⊙ AV) W_{o}

(1)

A = \frac{1}{n} {ReLU}^{2} (\frac{Q (X) K {(X)}^{T}}{\sqrt{d_{emd}}})

(2)

where A is the simplified attention matrix, simplifying the calculation of Q and K compared to the original MHA. It is obtained through affine transformation

Q (•)

and

K (•)

. In addition, the SoftMax function is replaced with the ReLU function to reduce the parameter quantity in the computation process. Then, the features are compressed through the convolutional layer while capturing the local features of the data.

2.2.2. Cross-task attention

Additionally, the CTA is used to exchange feature information between different tasks, enabling the interchange of complementary information across tasks while retaining their exclusive knowledge. CTA is a bidirectional IEU, allowing the F-task and D-task to each serve as the primary task and incorporate the auxiliary information from the other task, represented by CTA_D→F and CTA_F→D, respectively. Figure 3 mainly illustrates the internal structure of CTA_D→F, with the information flow for the diagnosis and denoising tasks feature represented by green and purple.

In the CTA_D→F unit, F-task is the primary task, while D-task is the secondary task. The purpose of CTA is to extract features from the D-task that can contribute to the F-task. This paper explores the beneficial features of the secondary task for the primary task from two aspects: task relevance and intrinsic features.

The cross-attention is used to capture the cross-task correlation, thereby guiding the contribution feature extraction of secondary tasks to the primary task. As shown in Figure 3, the query (Q) in cross-attention uses the input of the primary task, and the key (K) and value (V) are derived from the input of the secondary task. Then the cross-attention matrix is obtained by applying a SoftMax on the matrix multiplication Q × K^T and normalizing with d_emd, the correlation-guided attention features, denoted as C_{D→F is obtained} by multiplying the attention matrix with the projected secondary tasks features V.

C_{D \to F} = V \times softmax (\frac{Q \times K^{T}}{\sqrt{d_{e m d}}})

(3)

Furthermore, a self-feature extract consisting of a convolutional layer and a sigmoid function is defined, where self-feature is defined as a point-wise multiplication between the feature from Conv₁ and the dynamic mask provided by Conv₂

S_{D \to F} = {Conv}_{1} (f_{D}) ⊙ σ ({Conv}_{2} (f_{D}))

(4)

Finally, the two sets of features are connected by channels, and mapped to the same size as the input features through a 1 × 1 convolutional layer to obtain complementary features f_D→F, which are added to the original input features f_F to ${\tilde{f}}_{F}$ obtain fusion features

f_{D \to F} = p r o j e c t i o n ([C_{D \to F}, S_{D \to F}])

(5)

{\tilde{f}}_{F} = f_{D \to F} + f_{F}

(6)

2.3. Diagnosis and denoising decoder

The proposed CTA-JL model implements soft parameters sharing through CTA unit in the encoder network and compresses the original vibration signal into a latent space representation. In the decoder network, the denoised vibration signal is reconstructed from the latent space representation, achieving fault diagnosis. To describe the processes of the F-task and the D-task, the training datasets is defined as $D = {c_{i}, g_{i}, l_{i}}_{i = 1}^{N}$ , where c_i ∈ R^1×L represents the ith clean sample in datasets D, g_i represents the corresponding contaminated signal, and l_i is the corresponding fault type label.

The F-task mines fault features from the contaminated vibration signal g_i and predicts the current sample’s corresponding fault label l_i. The compressed features learned from the joint encoder need to be further compressed to one channel and mapped to the same dimension as the prediction category through fully connected layers. Finally, the Softmax function outputs the predicted probability q_i of fault type. The cross-entropy function is used as the loss function for the diagnosis task

L_{F D} = - \sum_{i = 1}^{N} l_{i} \log (q_{i})

(7)

The denoising task requires reconstructing a clean signal with a higher signal-to-noise ratio (SNR) from the contaminated vibration signal g_i. The compressed features learned from the joint encoder, which contain fault information, are passed through multiple convolutional layers and then mapped through the activation function tanh to obtain a denoising mask x_mask ∈ R^1×L with the same shape as the noise signal. The dot product of the denoising mask and the noise signal yields the reconstructed signal s_de, represented as s_de = x_mask ⊙ g. The denoising mask can suppress invalid signals in the noise signal and retain key feature signals. The mean squared error (MSE) between the reconstructed signal and the uncorrupted clean signal is used as the loss function for the denoising task in the diagnosis task

L_{D E} = \frac{1}{N} \sum_{i = 1}^{N} [\frac{1}{L} \sum_{n = 1}^{L} {(c_{i} [n] - s_{i} [n])}^{2}]

(8)

To better balance the denoising task and the diagnosis task, the joint architecture’s loss function is set with a loss weight γ, which coordinates the training of the joint task

L_{J L} = L_{F D} + γ L_{D E}

(9)

3. Experimental setup

3.1. Experimental setting

The proposed model is implemented with Pytorch 1.13 in Python3.9 and run on a GTX 3060 GPU. The Adam optimizer is used, with the learning rate is 0.005. The StepLR is employed as learning rate update scheduler, and the update step size is 20. The weight decay factor γ is set to 0.95, and the batch size is 128.

3.2. Comparison setting

Considering that the proposed CTA-JL can perform both denoising and fault diagnosis tasks, we will select comparison methods from two aspects to evaluate the comprehensive performance of the model in joint denoising and fault diagnosis.

Regarding the fault diagnosis task, we compared the performance of CTA-JL with seven other advanced methods:

(1) WDCNN: WDCNN (Zhang et al., 2017) is a five-layer-based convolutional model for fault detection in dynamic environments.

(2) AResNet (Zhong et al., 2022) introduces an Attention mechanism to the ResNet backbone to capture long-term features of the signal.

(3) QCNN: QCNN (Liao et al., 2023) replaces the inner product of traditional neurons with a simplified quadratic function called quadratic neurons and utilizes it to conduct a novel CNN network.

(4) DCA-BiGRU: DCA-BiGRU (Zhang et al., 2022) utilizes a 1D-signal attention mechanism and the GRU structure to capture the fault characteristics of bearings.

(5) BiLSTM (Zhao et al., 2020): BiLSTM is an enhanced LSTM model that has been successfully implemented in machine fault diagnosis.

(6) RNN-WDCNN (Zhang et al., 2017): RNN-WDCNN is a composite CNN model capable of extracting long-dependent feature information.

(7) Transformer (Jin et al., 2022): The Transformer acquires global dependent features of sequential characteristics through a multi-head self-attention mechanism.

For the denoising task, we selected classic signal analysis methods such as Wavelet Transform (WT) and Empirical Mode Decomposition (EMD) as comparative methods. Additionally, the Denoising Auto-encoder (DAE) model, built on multi-layer CNN, and the Fully Convolutional Network with Non-local Block (FL-CNN) (Xiong et al., 2023), constructed through a global NLB module and convolutional layers, were also employed as DL-based comparative methods.

3.3. Evaluation setting

Although noise with different probability distributions may occur in practical industrial environments, we assume that the noise sources are mutually independent for simplicity. According to the central limit theorem, as the number of noise sources increases, the probability distribution of the total noise becomes additive Gaussian white noise (AWGN) (Han et al., 2022). Therefore, we use AWGN as an example to simulate noise interference under actual operating conditions. Specifically, the collected signal is considered the original signal c, and a Gaussian noise n ∼ N (μ, σ²) is added to obtain a simulated noisy signal g = c + n. The goal of signal denoising is to construct a nonlinear mapping from the noisy signal g to the original signal c to eliminate the influence of the noise n.

For comparison purposes, we use two performance metrics: SNR and MSE. They are defined as follows

S N R = 10 \log_{10} \frac{\sum_{i = 0}^{N} x_{i}^{2}}{\sum_{i = 0}^{N} {(x_{i} - {\hat{x}}_{i})}^{2}}

(10)

M S E = \frac{1}{N} \sum_{i = 0}^{N} {(x_{i} - {\hat{x}}_{i})}^{2}

(11)

where

x_{i}, {\hat{x}}_{i}

represent the clean and denoised signals, respectively.

At the same time, we use accuracy as the evaluation metric for fault diagnosis, which is defined as

A c c u r a c y = \frac{T P + T N}{T P + F N + F P + T N}

(12)

where TP is true positive, TN is true negative, FP is false positive, and FN is false negative.

4. Experimental validation and result analysis

4.1. Case study 1: MFPT bearing datasets

4.1.1. Data description

The MFPT datasets (Dataset, 2020) is a publicly available dataset for bearing fault diagnosis provided by The Society for Machinery Failure Prevention Technology. This dataset consisted of three bearing sub-datasets. The baseline sub-datasets were sampled in each file at 97,000 Hz for 6 seconds. Additionally, seven outer ring fault (ORF) sub-datasets and seven inner ring fault (IRF) sub-datasets were sampled at 48,000 Hz for 3 s in each file. As the collected vibration data is a long time-series signal, we employed a sliding segmentation strategy to obtain more training samples. As shown, the stride was set to 256, and the window length was set to 2048, which ensured that each sample contained a complete cycle. Table 1 classifies one healthy state bearing and two fault bearings, including the inner ring fault and the ORF, into 15 categories (one healthy state and 14 fault states) based on different loads.

Table 1.

Description of the health conditions in the MFPT datasets.

Label	Fault type	Load (lbs.)
F1	N	270
F2-F8	ORF	25,50,100,150,200,250,300
F9-F15	IRF	25,50,100,150,200,250,300

4.1.2. Selection of the joint loss weight λ

Considering the significant impact of the choice of the joint loss weight λ on the results, we conducted experiments to investigate the influence of different λ values on the network’s performance. In our experiments, we set λ to 0.1, 0.5, 1.0, 2.0, 4.0, 5.0, 6.0, and 10.0, and conducted experiments on the MFPT datasets under the noise condition with SNR_dB = −6 dB. The diagnostic and denoising results of CTA-JL are presented in Table 2. It can be observed that the variation of λ has a minor impact on the diagnostic performance of the network. The fluctuation range of diagnostic accuracy remains below 0.92% as λ changes from 0.1 to 10. On the contrary, the denoising performance of the network improves to a certain extent with the increase of λ. This is because the increase in λ amplifies the importance of the denoising loss in the joint loss, enabling the network to focus more on the denoising task. Considering the diagnostic and denoising results of CTA-JL shown in Table 2, we select λ = 4 as the loss weight for CTA-JL.

Table 2.

Experimental results under different loss weights λ (SNR_dB = −6 dB).

	λ = 0.1	λ = 0.5	λ = 1	λ = 2	λ = 4	λ = 5	λ = 6	λ = 10
ACC	96.81% ± 0.76%	97.55% ± 0.65%	96.79% ± 0.61%	96.63% ± 0.87%	97.95% ± 0.75%	96.61% ± 0.77%	97.49% ± 0.61%	97.36% ± 0.81%
MSE	0.7891 ± 0.0992	0.7778 ± 0.1031	0.7892 ± 0.0814	0.7795 ± 0.0901	0.6971 ± 0.0893	0.6993 ± 0.0934	0.6988 ± 0.0911	0.6982 ± 0.0991
SNR	1.171 ± 0.699	1.201 ± 0.703	1.173 ± 0.603	1.191 ± 0.625	1.433 ± 0.683	1.463 ± 0.693	1.445 ± 0.663	1.343 ± 0.703

4.1.3. Performance on prediction accuracy

In this section, we simulated the noise interference in the vibration signal acquisition process by adding Gaussian noise. This allowed us to obtain contaminated noisy signals. To compare the CTA-JL model’s adaptability to different noise levels, we set three noise levels: SNR = −9 dB, SNR = −6 dB, and SNR = −3 dB. We performed diagnostic performance validation on both the CTA-JL and comparative models using these noise signals and raw signals.

Figure 4 shows the diagnostic performance results. For the fault diagnosis of the original signal, all algorithms achieved an accuracy of 98% and were able to identify the fault types effectively. However, in the presence of noise interference, the accuracy of other algorithms significantly declines, especially WDCNN and QCNN, which exhibit accuracies of approximately 66.79% and 63.21%, respectively, at SNR = −9 dB. Contrastingly, CTA-JL still maintains an accuracy of 91.38% under these conditions.

4.1.4. Performance on denoising metric

To evaluate the CTA-JL method’s signal denoising performance, we conducted validation experiments on different groups of samples with varying SNR obtained by adding Gaussian noise to samples in the MFPT dataset. The SNR of the denoised signals was calculated by comparing them with the pure signals without noise. We evaluated the proposed algorithm and compared algorithms’ signal denoising performance on each data group; the results are shown in Table 3. The results show significant improvements compared to traditional denoising methods. For example, when SNR = −9 dB, the original vibration signal was almost entirely masked by noise. After denoising, the MSE of the EEMD and WT methods were 1.981 and 0.893, respectively, and the SNR was around −3.18 dB and 0.621 dB, respectively. However, the denoised signal generated by the proposed CTA-JL method can achieve an SNR of 1.728 dB, with an MSE of only 0.703 compared to the pure signal. Compared to DL-based methods such as DAE and FL-CNN, the proposed CTA-JL also demonstrates certain advantages in terms of denoising performance.

Table 3.

Denoising performance of the CTA-JL model and comparative models under three noise environments on MFPT datasets.

	MSE			SNR _dB
	−9 dB	−6 dB	−3 dB	−9 dB	−6 dB	−3 dB
EEMD	1.981 ± 0.112	1.398 ± 0.201	1.01 ± 0.095	−3.18 ± 0.669	−1.06 ± 531	−0.291 ± 0.353
WT	0.893 ± 0.086	0.782 ± 0.115	0.586 ± 0.215	0.621 ± 0.511	1.323 ± 0.446	2.751 ± 0.321
DAE	0.956 ± 0.075	0.943 ± 0.319	0.871 ± 0.112	0.242 ± 0.419	0.3131 ± 0.342	0.7372 ± 0.542
FL-CNN	0.751 ± 0.011	0.637 ± 0.069	0.506 ± 0.012	1.431 ± 0.066	2.573 ± 0.721	3.151 ± 0.121
Mine	0.703 ± 0.009	0.602 ± 0.071	0.499 ± 0.035	1.728 ± 0.079	2.361 ± 0.045	3.174 ± 0.115

4.1.5. Effectiveness of joint learning

The CTA-JL method employs a joint learning architecture to exchange denoising and diagnostic feature-related information. To verify the effectiveness of this joint learning architecture, we compared the diagnostic accuracy and denoising performance of the proposed CTA-JL method, its denoising networks and diagnostic networks on the MFPT datasets under high noise environments (SNR = −9 dB, SNR = −6 dB, and SNR = −3 dB). The comparative results are presented in Table 4, where CTA-JL⁺ denotes the Denoising network part, and CTA-JL* denotes the Diagnosis network part in CTA-JL. It can be observed that the proposed joint learning method improves the predictive accuracy of diagnostic network by 5.52%, 3.64%, and 0.83%, respectively, under different noise level, compared to the methods without joint learning. Additionally, the denoising performance is also improved. These results demonstrate that the joint learning architecture enhances the model’s overall performance.

Table 4.

Comparisons of all metrics in CTA-JL, CTA-JL⁺ and CTA-JL* with strong noise on MFPT datasets. CTA-JL⁺ denotes the Denoising network part, and CTA-JL* denotes the Diagnosis network part in CTA-JL.

Noise level(dB)	Metric	CTA-JL	CTA-JL⁺	CTA-JL*
−9	ACC	91.38 ± 1.89	-	85.86 ± 3.25
	MSE	0.703 ± 0.009	0.887 ± 0.059	-
	SNR _dB	1.728 ± 0.079	0.587 ± 0.321	-
−6	ACC	97.95 ± 1.66	-	94.31 ± 2.14
	MSE	0.602 ± 0.071	0.871 ± 0.111	-
	SNR _dB	2.361 ± 0.045	0.737 ± 0.254	-
−3	ACC	99.69 ± 0.35	-	98.86 ± 0.98
	MSE	0.499 ± 0.035	0.563 ± 0.087	-
	SNR _dB	3.174 ± 0.115	3.051 ± 0.125	-

4.2. Case 2: Elevator guidance system datasets

4.2.1. Data description

The elevator guidance system (EGS) fault datasets was obtained from a simplified elevator test bench, as shown in Figure 5. The EGS datasets mainly includes two types of faults: wear of the elevator rolling guide shoes and deformation of the elevator guide rails. According to Table 5, we divided the diagnosis category by different fault locations and different fault types, resulting in a total of seven fault diagnosis categories (one healthy state and six types of faults). Data for each type of fault was obtained from three different load tests.

Table 5.

Description of the health conditions in the EGS datasets.

Label	Fault type	Fault location	Load (N)
Z1	Normal	Normal	0,10,20
Z2	Wear and tear of guide shoe	Upper single side guide shoe	0,10,20
Z3	Wear and tear of guide shoe	Upper double side guide shoe	0,10,20
Z4	Wear and tear of guide shoe	Lower single side guide shoe	0,10,20
Z5	Wear and tear of guide shoe	Lower double side guide shoe	0,10,20
Z6	Guide rail deformation	Single side guide rail	0,10,20
Z7	Guide rail deformation	Double side guide rail	0,10,20

In contrast to the MFPT datasets, the EGS datasets involves diagnosing faults in the entire mechanical system. A simulation experiment platform was built by scaling down a high-speed elevator with a scale ratio of 5:0.6, as shown in Figure 5. Vibration signals were collected through accelerators sensor placed on the elevator test bench, and different working conditions were simulated on the test bench. According to Table 5, the diagnosis category was divided by different fault locations and different fault types, resulting in a total of seven fault diagnosis categories (one healthy state and six types of faults). Data for each type of fault was obtained from three different load tests.

4.2.2. Performance on prediction accuracy

Similar to the process of MFPT datasets, we set three noise levels signals: SNR = −9 dB, SNR = −6 dB, and SNR = −3 dB with Gaussian noise. We performed diagnostic performance validation on both the CTA-JL model and the comparative models using these noise signals and raw signal. We evaluated the CTA-JL method’s noise adaptability by assessing the proposed algorithm’s diagnostic accuracy and comparing it with that of other algorithms on each data group. The results are presented in Figure 6.

As shown in Figure 6, the proposed method outperforms all other methods comprehensively. For the original signal, all algorithms can achieve good results with an accuracy rate of over 96%. This is because the fault characteristics in the original signal are very easy to distinguish. However, all algorithms’ diagnostic accuracy decreases in the face of noise interference, especially WDCNN, whose accuracy is lower than 70%. On the other hand, the proposed CTA-JL model outperformed all comparative models, maintaining a diagnostic accuracy of approximately 87.31% even under extreme noise interference with SNR = −9 dB.

4.2.3. Performance on denoising metric

To evaluate the CTA-JL method’s signal denoising performance, we conducted validation experiments on different groups of samples with varying SNR obtained by adding Gaussian noise to samples in the EGS datasets. The SNR of the denoised signals was calculated by comparing them with the pure signals without noise. We evaluated the proposed algorithm and compared algorithms’ signal denoising performance on each group of data, and the results are shown in Table 6. The results show significant improvements compared to traditional denoising methods. For example, when SNR = −9 dB, the original vibration signal was almost completely masked by noise. The MSE of the EEMD and WT methods after denoising were 1.631 and 0.732, respectively, and the SNR were around −5.575 dB and −2.278 dB, respectively. However, the denoised signal generated by the proposed CTA-JL method can achieved an SNR of 1.728 dB, with an MSE of only 0.703 compared to the pure signal.

Table 6.

Comparison of denoising metrics between CTA-JL and other methods on EGS datasets.

	MSE			SNR _dB
	SNR = −9 dB	SNR = −6 dB	SNR = −3 dB	SNR = −9 dB	SNR = −6 dB	SNR = −3 dB
EEMD	1.631 ± 0.117	1.021 ± 0.154	0.663 ± 0.079	−5.575 ± 0.566	−3.723 ± 0.732	−1.851 ± 0.251
WT	0.732 ± 0.211	0.601 ± 0.135	0.471 ± 0.097	−2.278 ± 0.755	−1.425 ± 0.553	−0.365 ± 0.351
DAE	0.956 ± 0.075	0.904 ± 0.095	0.871 ± 0.111	0.242 ± 0.419	0.454 ± 0.542	0.737 ± 0.638
FL-CNN	0.798 ± 0.011	0.742 ± 0.008	0.499 ± 0.012	0.976 ± 0.069	1.361 ± 0.072	3.174 ± 0.121
Mine	0.703 ± 0.009	0.602 ± 0.006	0.545 ± 0.013	1.728 ± 0.065	2.361 ± 0.077	2.689 ± 0.097

4.2.4. Effectiveness of joint learning

The CTA-JL method adopts a joint learning architecture that facilitates the exchange of denoising and diagnostic feature-related information. To validate the effectiveness of this joint learning architecture, we conducted a comparison of the diagnostic accuracy and denoising performance between the proposed CTA-JL method and its denoising and diagnostic networks using the EGS datasets under high noise environments (SNR = −9 dB,SNR = −6 dB and SNR = −3 dB). The comparative results are presented in Table 7, where CTA-JL⁺ refers to the denoise network part and CTA-JL* refers to the diagnosis network part within CTA-JL. It can be observed that the proposed joint learning method improves the predictive accuracy of the denoising and diagnostic algorithms by 9.37%, 2.94%, and 0.23%, respectively, under different noise levels compared to methods without joint learning. Additionally, the denoising performance is also enhanced, with an increase in SNR and a decrease in MSE. These results demonstrate that the joint learning architecture enhances the overall performance of the model.

Table 7.

Comparison of all metrics in CTA-JL, CTA-JL⁺ and CTA-JL* with strong noise on EGS datasets. CTA-JL⁺ denotes the denoise network part, and CTA-JL* denotes the diagnosis network part in CTA-JL.

Noise level(dB)	Metric	CTA-JL	CTA-JL⁺	CTA-JL*
−9	ACC	87.31 ± 1.75	-	73.96 ± 3.59
	MSE	0.703 ± 0.009	0.883 ± 0.017	-
	SNR _dB	1.728 ± 0.065	0.514 ± 0.121	-
−6	ACC	95.58 ± 1.11	-	91.62 ± 1.25
	MSE	0.602 ± 0.006	0.805 ± 0.023	-
	SNR _dB	2.361 ± 0.077	0.9628 ± 0.89	-
−3	ACC	99.29 ± 0.15	-	98.96 ± 0.13
	MSE	0.545 ± 0.013	0.763 ± 0.033	-
	SNR _dB	2.689 ± 0.097	1.057 ± 0.638	-

5. Conclusion

To enhance the anti-noise performance of fault diagnosis models, this paper proposes an end-to-end attention-enhanced joint learning (CTA-JL) model for cross-task feature information exchange. The CTA unit is used to exchange feature information between different tasks, obtaining complementary information from both task correlation and the intrinsic characteristics of auxiliary tasks. The F-task and D-task both can be taken as the primary tasks, and complementary features are obtained by combining auxiliary information from other tasks, which are then added to the original features to obtain fusion features. The decoder network corresponding to each task is used to get the results of fault diagnosis and denoised signal. Experimental results on the motor bearing and EGS datasets demonstrate that the CTA-JL model has better performance in both fault diagnosis and signal denoising than the compared methods. Furthermore, ablation experiments show that the joint architecture improves the performance of both diagnostic and denoising tasks.

The future work will focus on two aspects: (1) Under the impact of strong noise interference, the differences in data distribution among various fault categories are obscured by noise, resulting in significant prediction uncertainties. To address this, we will utilize Bayesian theory for confidence estimation in prediction, providing a new insight for maintenance decision-making. (2) The limited scope of collected data often excludes certain fault types from our training datasets. To mitigate this, we plan to use out-of-distribution detection to identify unknown fault categories, preventing false fault alarms.

Footnotes

Declaration of conflicting interests

The author(s) declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article.

Funding

The author(s) disclosed receipt of the following financial support for the research, authorship, and/or publication of this article: This work was supported by the National Natural Science Foundation of China (52275274), Pioneer and Leading Goose R&D of Zhejiang, China (2022C01051) and the Public Welfare Technology Application Projects of Zhejiang Province, China (LGG22E050008).

ORCID iDs

Zili Wang

Lemiao Qiu

References

Chen

Liu

(2017) Mechanical state prediction based on lstm neural netwok. In: 2017 36th Chinese Control Conference (CCC), Dalian, China, 26–28 July 2017.

Dataset MFPT (2020) Society for Machinery Failure Prevention Technology.

Dong

Jia

, et al. (2021) Non-iterative denoising algorithm for mechanical vibration signal using spectral graph wavelet transform and detrended fluctuation analysis. Mechanical Systems and Signal Processing 149: 107202.

Feng

, et al. (2023) A novel vibration-based prognostic scheme for gear health management in surface wear progression of the intelligent manufacturing system. Wear 522: 204697.

Han

Wang

Liu

, et al. (2022) Intelligent vibration signal denoising method based on non-local fully convolutional neural network for rolling bearings. ISA Transactions 122: 13–23. DOI: 10.1016/j.isatra.2021.04.022.

Hua

Dai

Liu

, et al. (2022) Transformer quality in linear time. In: International Conference on Machine Learning, Baltimore, 17–23 July 2022.

Jin

Hou

Chen

(2022) A time series transformer based method for the rotating machinery fault diagnosis. Neurocomputing 494: 379–395.

Liao

Dong

Sun

, et al. (2023) Attention-embedded quadratic network (qttention) for effective and interpretable bearing fault diagnosis. IEEE Transactions on Instrumentation and Measurement 72: 1–13.

Liu

Zhou

Zheng

, et al. (2018) Fault diagnosis of rolling bearings with recurrent neural network-based autoencoders. ISA Transactions 77: 167–178.

10.

Mao

, et al. (2017) Bearing fault diagnosis with auto-encoder extreme learning machine: a comparative study. Proceedings of the Institution of Mechanical Engineers - Part C: Journal of Mechanical Engineering Science 231(8): 1560–1578.

11.

Halkon

, et al. (2023) Physics-informed residual network (piresnet) for rolling element bearing fault diagnostics. Mechanical Systems and Signal Processing 200: 110544.

12.

Seryasat

Honarvar

Rahmani

, et al. (2010) Multi-fault diagnosis of ball bearing using fft, wavelet energy entropy mean and root mean square (rms). In: 2010 IEEE International Conference on Systems, Man and Cybernetics, Istanbul, 10–13 October 2010.

13.

Vakharia

Gupta

Kankar

(2015) A multiscale permutation entropy based approach to select wavelet for fault diagnosis of ball bearings. Journal of Vibration and Control 21(16): 3123–3131.

14.

Wang

Xiang

Markert

, et al. (2016) Spectral kurtosis for fault detection, diagnosis and prognostics of rotating machines: a review with applications. Mechanical Systems and Signal Processing 66-67: 679–698.

15.

Wang

Cao

Han

(2022) A combination of dilated self-attention capsule networks and bidirectional long-and short-term memory networks for vibration signal denoising. Machines 10(10): 840.

16.

Xiong

Yuan

, et al. (2023) A novel deep convolutional image-denoiser network for structural vibration signal denoising. Engineering Applications of Artificial Intelligence 117: 105507.

17.

Feng

Yan

, et al. (2023a) Cfcnn: a novel convolutional fusion framework for collaborative fault identification of rotating machinery. Information Fusion 95: 1–16.

18.

Zhang

Peng

, et al. (2017) A new deep learning model for fault diagnosis with good anti-noise and domain adaptation ability on raw vibration signals. Sensors 17(2): 425.

19.

Yan

Sun

, et al. (2023b) Online knowledge distillation based multiscale threshold denoising networks for fault diagnosis of transmission systems. IEEE Transactions on Transportation Electrification 1–1.

20.

Zhang

, et al. (2022) Fault diagnosis for small samples based on attention mechanism. Measurement 187: 110242.

21.

Zhao

, et al. (2020) Deep learning algorithms for rotating machinery intelligent diagnosis: an open source benchmark study. ISA Transactions 107: 224–255.

22.

Zhong

Wang

Ban

(2022) Development of a plug-and-play anti-noise module for fault diagnosis of rotating machines in nuclear power plants. Progress in Nuclear Energy 151: 104344.