Bearings are critical components in rotating machines used in various applications such as auto-manufacturing, chemical industries, railways, water pumping stations, etc. Their health has significant impacts on the operation of industrial systems and economy of industries, with 45%–55% of faults in industrial motors are caused by bearing faults [1,2]. Therefore, effective fault diagnosis of bearings can significantly increase the reliability and life span of machines in various industrial applications, which in turn reduces maintenance and operation costs [3,4]. Considering the importance of bearings in rotating machines, various methods have been applied by researchers to effectively monitor the bearing condition and to avoid downtime caused by the bearing failure.

Vibration is considered as an early and highly accurate indicator of various faults in mechanical equipment, and its measurement and analysis are widely employed with Artificial Intelligence (AI) based fault diagnosis and prognosis methods [5–7]. Typically, a fault diagnosis procedure comprises of three steps: data acquisition, feature extraction, and fault diagnosis. Firstly, the data acquisition step for machinery fault detection involves signals such as vibration, current, voltage, temperature, and acoustic emissions. Second step includes the extraction of time-domain features (root mean square, skewness, kurtosis, and gap factor), frequency domain (Fourier transform), and time-frequency domain features (wavelet transform, short-time Fourier transform, empirical mode decomposition, and Hilbert-Huang transform) [8–13]. Lastly, fault diagnosis step which has been widely explored by researchers using different model-based techniques and AI-based techniques. Machine Learning (ML) based methods are the foundation for effective fault diagnosis in the arena of AI. Various researches in literature have employed these techniques for fault diagnosis, for instance, Support Vector Machines (SVM), K-Nearest Neighbors (KNN), and Random Forest (RF) [14–16].

ML methods have yet proven to be an effective diagnosis method; however, these methods have some inherent limitations and their performance extremely relies on designed features. For combined bearing fault detection, it often requires advanced signal processing techniques [17]. Moreover, these methods are not effective in a practical industrial environment where background noise and interference of signal components are inevitable [18]. Thus, the progress in AI led to the emergence of DL which has been widely exploited by researchers in various fields of this era. DL as a subfield of AI and has provided an efficient way of automatic representative feature learning, even from raw and noisy input data without any human intervention [19,20]. Compared to ML methods, DL methods automatically learn rich features even from raw data using high-performance computing hardware [7,21–24]. Various DL architectures have been exploited for efficient bearing fault detection including Convolutional Neural Networks (CNN), Long Short-Term Memory (LSTM), Auto-Encoders (AE), Deep Belief Networks (DBN), and Deep Boltzmann Machines (DBM) [3,25–28]. Among these DL models, CNNs has been the most representative model and has demonstrated robust and remarkable performance as a supervised learning approach [29,30]. Furthermore, Full Convolutional Network (FCN) can perform pixel-level semantic segmentation of image data owing to their rich hierarchical architecture. It has also made substantial progress in achieving higher level of feature hierarchy by processing of lower level features. These models learn features through complex mapping which extracts spectral information from individual pixels [3,6]. Considering the merits of CNN model, it has been widely used to diagnose rolling bearing fault using 1D (one dimensional) and 2D (two dimensional) data with 1D-CNN and 2D-CNN configurations, respectively. Guo et al. [31] have used an adaptive CNN (ADCNN) to classify the 4 conditions of the rolling bearing with the Case Western Reserve University (CWRU) bearing data. The initial layers of the model comprise of LeNet-5 and then stacked hierarchical layers which determine two output components: fault type and fault size. The adaptive configuration of the learning rate allowed to maintain a trade-off between accuracy and training speed. This model yielded competitive performance to the earlier models.

In [32], a deep CNN model with multisensory vibration data for classifying the nine bearing conditions have been experimented. Their model was able to learn features from the raw data and avoided overfitting through dropout regularization. Li et al. [33] have employed deep CNN with data augmentation. The model was trained with 400 data samples of the CWRU bearing dataset with the ten bearing conditions. Raw vibration data was the input to the CNN layers then residual layers were stacked for better feature learning. The network was trained with batch-normalization and dropout layers for fast training and avoiding overfitting through data augmentation. Although these DL architectures have demonstrated effective performance in bearing fault diagnosis, but these models face multiclass window problem and loss of information owing to sliding window labelling technique. Thus, these models pose limitations in performing dense predictions.

Unlike the above-cited works, a novel bearing fault diagnosis method based on UNET and vibration images is presented in this paper. The proposed approach overcomes the multiclass window problem that occur in DL models and can predict label for each input data sample in the bearing data. The UNET model achieves pixel-level dense predictions through the down-sampling and up-sampling layers in the architecture. This model is successfully applied in various domains and has revealed excellent performance in terms of dense predictions [34–36]. The main contribution of this research article is twofold. The primary contribution is UNET based novel rolling bearing fault diagnosis method, which can perform dense predictions by overcoming the multiclass window problem, caused by sliding window-labelling technique. To the best of authors’ knowledge, UNET network is applied to diagnose bearing faults for the first time. The second is the dense labelled 2D vibration images without any data-augmentation given as input to the UNET model for sample-based bearing condition classification.

The remaining paper is organized as follows: Section 2 discusses dense labelling and fault classification using UNET, Section 3 explains the UNET architecture and training process, Section 4 discusses the dataset, experiment configuration, and experiment evaluation, Section 5 reports and discusses the results, and Section 6 concludes the investigation.

Conventionally, time-series or sequential signals are divided into fixed-length windows using sliding window techniques. Then, each window is assigned with the same label and this label assigning technique leads to a multiclass window problem. There are two sliding window techniques: the first technique assigns a label by selecting the most frequent sample class and the second technique assigns labels by selecting last sample class of time step. Using both the techniques, classifier generates incorrect output information and causes a decrease in accuracy owing to the multiclass window problem. This problem is solved by dense labelling technique which increases classifiers’ accuracy by providing correct label information. This method assigns label to each sample of dataset rather than labelling based on sliding windows. Hence, it upholds all the label information in dataset. Fig. 1 describes both the labelling techniques. Current sample and next sample are described as S1 and S2. In the case of sliding window labelling method, the most frequent appearing class is assigned with label S1. However, owing to sliding windows it contains information of both classes. Thus, it causes learning incorrect information which in turn brings down the recognition accuracy.

The traditional CNN restricts implementation of dense prediction, whereas UNET model reduces resolution of top-level output by using the max-pooling operation. For image recognition task, UNET reduces sensitivity of image shift. In the case of time-series data recognition, CNN causes a mismatch between input data length and output label length. To overcome this problem, UNET architecture has been successfully implemented by adding up-convolution layers or up-sampling layers in CNN architecture. It allows the use of dense labelling and preserves same length for input series and output label. Thus, UNET can be applied for dense predictions owing to addressing the multiclass window problem which is often posed by traditional CNN.

This section reports experiments performed on a rolling bearing dataset using UNET model. Furthermore, the investigation and evaluation process of performance of UNET is reported in this section. In addition, this section presents a comparison of UNET model performance with other DL architectures, which are employed in the recent research for bearing fault diagnosis. Firstly, the dataset employed in this research is explained, then the experimental configuration and evaluation metrics are discussed.

In this section, the performance of the four models is compared and summarized based on accuracy and F1-Score indexes. The highest results obtained with the UNET and comparative methods are shown in Tab. 3. The results clearly indicate effectiveness of the UNET method, which outperforms the comparative DL methods owing to its capability of dense predictions. The UNET has achieved the best accuracy of 98.91% and the highest F1-Score of 99% among all the employed DL models on the dataset in this research. FCN has achieved 97.61% of accuracy and 97.6% of F1-Score. While, LetNet-5 has demonstrated the second-best performance with accuracy and F1-Score of 96.74% and 96.8%, respectively. However, ResNet-50 has demonstrated the lowest performance in diagnosing the rolling bearing conditions. The comparative models have yielded lower accuracies than the UNET owing to loss of useful information caused by sliding window labelling technique. Contrarily, UNET model preserves useful information from the vibration images owing its hierarchical network architecture which in turn allowed it to perform better than the comparative models. The achieved results conclude the UNET model as an effective model among all the used DL models in this investigation.

Fig. 5 shows the F1-Score of the UNET model and comparative DL models in terms of the each bearing condition. Comparatively, UNET model has the best F1-Score for each bearing condition.

Fig. 6 shows the confusion matrices of the UNET and comparative DL methods in terms of the each bearing condition. As shown in Fig. 6a, UNET model accuracy for the individual classes varies within the range of 95.65% to 100%. Similarly, individual class accuracy of the LeNet-5 is shown in Fig. 6b, where the accuracy has the minimum value of 84.75% and the maximum of 100%. In Fig. 6c, individual class accuracy of the ResNet-50 model is shown which varies within the range of 82.61% and 100%. Lastly, confusion matrix of the FCN model shows the individual class accuracy changes between 82.61% and 100%. These confusion matrices confirm the UNET model as the best model for the individual bearing condition classification among the employed DL models.

Fig. 7 shows training accuracy and loss function of the employed DL models. These models were trained with early stopping callback which stops the training when a DL model achieves maximum accuracy without overfitting. Thus, number of the epochs varies for each DL model. Fig. 7a shows that the accuracy of the UNET model tends to be stable after the 30th epoch and achieves the accuracy of 98.91% and its loss function value improved after the 25th epoch. The LeNet-5 model achieves the accuracy of 96.74%, but the model response keeps fluctuating as shown in Fig. 7b. Similarly, the ResNet-50 model achieves the accuracy of 95.43% with the fluctuating response for both the accuracy and the loss function as depicted in Fig. 7c. Lastly, Fig. 7d shows accuracy of the FCN that is around 97.61% and its accuracy and the loss function response improve after 20 epochs. It can be noted from the figures that the UNET model depicts the most stable classification performance with the highest classification accuracy.

Although the UNET model was fed with the vibration images of resolution of 32 × 32 × 1, yet it yielded the excellent performance in terms of classifying the bearing conditions owing to the pixel-level learning capabilities. Moreover, with the adaptive learning rate the model learned features hierarchically thorough adjusting the learning rate, if the performance of the model did not improve for certain iterations. Another technique which is known as early stopping, helped to avoid the overfitting, if the model did not improve the performance. The UNET model has depicted robustness in the training process as shown in Fig. 7a.

The UNET model was remarkably able to predict sample-based rolling bearing conditions owing to its inherent property of dense predictions. Its comparison with the other popular DL models revealed that the UNET model has superior and robust performance in terms of rolling bearing fault diagnosis. It was observed from the results that the proposed algorithm possesses the excellent potential for fault diagnosis of rotating machines in various industrial applications.

In this paper, a novel DL method namely UNET is proposed for rolling bearing fault classification based on the vibration images. Compared to the existing DL methods, this model overcomes the multiclass window problem that is inherent in the sliding window labelling method. The UNET model used in this investigation has effectively performed dense predictions of the bearing conditions with an accuracy of 98.91% and F1-Score of 99%. The obtained results confirm that the UNET model outperforms the comparative DL models such as CNN, LeNet-5, and ResNet-50. The model yielded robust and better results than the comparative methods on short-term feature recognition. The robust classification results of the UNET model on this bearing dataset indicates its excellent potential for applications in other domains.