Classification of Bearing Fault Signals in Rotating Machinery Using Neural Networks

In modern life, Machinery serves as a utility integrated into domestic and industrial appliances, designed to operate reliably for extended periods. Rotating machines are the crucial component required for smooth and efficient machine operation. As the machines operate uninterrupted for extended periods, they are vulnerable to failures, and prediction of machinery failure is of paramount important. The risk of sudden breakdowns or failures of the Machinery can be minimized if regular maintenance is accomplished and failures are detected early [1]. Early fault detection of machinery optimizes operational efficiency, thereby averting costly downtime, production losses and potential safety hazards [2]. The predominant component contributing to these machine breakdowns is the rolling bearings component, and 90% of failures are related to bearing faults [3]. The root causes of bearing failures and their fault modes include defamation, fracture and wear [4], which are identified during predictive maintenance based on continuous machine monitoring. Early identification of faults is possible based on the analysis and inspection of available data collected from various sensors such as voltage, vibration, temperature, and pressure [5]. A specific predictive maintenance strategy utilizes vibration signal analysis, where data is collected with a microprocessor and distinct vibration frequency components, amplitude changes generated during machine operating conditions are analyzed. However, microprocessor-based data collected cannot capture accurate data and generate low-frequency vibrations [6]. Thus, signal analysis performed on low-frequency vibration components is not precise. These findings on signal have inspired researchers to propose various techniques to capture high-frequency data and diagnose the faults of bearings in rotating Machinery through vibration monitoring [7].

Most of the research hitherto has concentrated on identifying faulty Machinery by monitoring the signal amplitude and its frequency transitions, which include manual extraction of signal features: statistical time-domain, frequency-domain, spatial domain, and time-frequency features [8]. Statistical time features of the time domain include max, mean, root amplitude, standard deviation, variance, standard error, peak-peak, root mean square, skewness, entropy, kurtosis, margin, clearance, crest, shape and impulse factor. Frequency domain or spectral features include applying fast Fourier transform (FFT) and performing envelope analysis Time-frequency features include wavelet transforms, Fourier transforms like short-term Fourier transform (STFT); the derived spectrograms, wavelets, and decomposed wavelets are inspected, and wavelet analysis is performed [9]. According to the task requirements, a selective approach with only a subset of features is adopted to analyze machine faults. Many supervised and unsupervised machine learning models, namely, decision trees, k nearest neighbors, random forest, artificial neural networks, support vector machines, backpropagations and convolution neural networks (CNN), are trained on these selected features [10-13]. While many supervised learning models have been utilized in fault diagnosis, the incorporation of deep learning models remains crucial.

The application of neural networks across various domains has seen a surge, and recent advancements in deep learning have become popular in bearing fault diagnosis. The deep learning models applied on bearing datasets convert raw vibration signals to time-frequency images using various wavelet transforms. Wavelet transforms like continuous wavelet transform (CWT) [14-25], CWT of complex morlet [26], two-level discrete wavelet transform (DWT) [27], Non-uniform Fast Fourier Transform (NFFT) [28] and Short-time Fourier Transform (STFT) [29] were applied to convert one-dimensional vibration signals to 2-dimensional spectrograms. These time-frequency images are given as input to convolution neural networks for fault identification. While neural networks are integrated into the system, but their operations are not directly applied on the raw signal.

The following researchers employed deep learning approaches for fault diagnosis and achieved the following performance metrics. Xia et al. [30] proposed a CNN model incorporating multiple sensors. The model is trained using spatial and temporal information from raw data signals. Manual feature extraction and selection are avoided, and is tested, which results in 99.41% accuracy. Sonmez et al. [31] proposed real-time condition monitoring based on combining one—and two-dimensional deep convolution neural networks combined with wide first-layer kernels (WD-CNN). The model’s effectiveness is tested, and 96.45% accuracy is achieved on different operating loads. van den Hoogen et al. [32] introduced an adaptive wide kernel in one-dimensional CNN architecture to classify multivariate signals without data pre-processing. The model’s performance is trained and evaluated, and an average accuracy of 89.39% is achieved. Eren et al. [33] proposed an intelligent-based diagnosis using CNN without encapsulating distinct blocks for feature extraction, selection, and classification. One-dimensional CNN is applied directly to the raw time-series sensor data, learning optimal features automatically and achieving 93.22% accuracy. Jin et al. [34] proposed variational mode decomposition (VMD) to decompose the signal and CNN for bearing fault diagnosis. High-correlated components acquired are given as input to the classification model. The fault diagnosis method achieved 96.41% accuracy on CWRU dataset. Kulevome et al. [35] proposed rolling element-bearing classification using a deep neural network. The features from the input vibration signals are classified using a modified VGG16 architecture. The classification model achieved 99.23% accuracy and a healthy indicator framework for varying operating conditions of faults is proposed.

As evidenced by the previous studies, there remains a research potential to propose an efficient fault diagnosis model to diagnose bearing faults in rotating Machinery. While CNNs are primarily designed for two-dimensional image processing, their applicability can be extended to one-dimensional signals through appropriate modifications of parameters. The main objective of the proposed work focuses on,

1. Processing raw input vibration signals directly.

2. Signals are segmented to make them appropriate for neural networks.

3. A neural network and a two-level invariant Neural Network Model is proposed to classify fault signals and to improve the classification accuracy.

The following sections outline the paper’s organization, ‘Theoretical Background’ provides a functional overview of the theoretical underpinnings of the Neural Network Models. The ‘Proposed Methodology’ provides a systematic approach and description of the fault diagnosis model. ‘Experimental Results and Discussions’ demonstrates a comprehensive and detailed performance analysis of the proposed models’ capabilities. The section ‘Conclusion’ summarizes the essential findings and contributions of the paper, highlighting the potential of the proposed model in improving the efficiency and accuracy of model.

This study uses convolutional neural networks to automatically extract features and classify vibration signals for the diagnosis of bearing faults. The subsequent sections detail the data preparation for the proposed model and the findings obtained.

3.1 Data collection and processing

The dataset utilized in the research is the Case Western Reserve University (CWRU) bearing dataset [38], a publicly accessible benchmark dataset found in reference [39], which has been employed in the majority of prior publications. The bearing dataset is produced using a 2hp Reliance electric motor equipped with a dynamometer, encoder and transducer, as seen in Figure 3, sourced from publication [38].

3.png

Figure 3. CWRU bearing test rig

Sensors are placed at different locations. Data were collected at two different frequencies, 12 KHz and 48 KHz, from three different ends: bearing end, Drive-end (DE) and fan-end (FE).

A total of 161 data files were utilized, and defects were introduced to bearings with diameters varying from 0.007 to 0.040 inches. Data was collected at several speeds ranging from 1720 to 1797 rpm. Each data file comprises multiple properties, specifically: drive end signal (DE_Time) and fan end signal (FE_Time) – acceleration data measured in vertical direction, base signal (BA). Only DE_time data is utilized for analysis, as it is proximate to the bearing.

This research examines data sampled at a frequency of 48 kHz. Three categories of faults are examined: bearing ball fault, race inner and race outer fault, in conjunction with data from normal conditions. All files are in .mat format. Each file is renamed to indicate the defect kind, its diameter, and the motor run load. The file name I7_2 signifies that the fault type is Inner Fault, occurring while the motor operates at a 2hp load, with a fault diameter of “0.007”, the file name B21_1 indicates the fault type is Bearing Fault generated when the motor is run at 1hp load, and the fault diameter is “0.021” and the file name O14_3 indicates the fault type is Outer Fault generated when the motor is run at 3hp load, and the fault diameter is “0.0014”. The data is designated with distinct names for each class type and provided as input to the models.

Despite 48 KHz sampling rate for 10 seconds, certain data files exhibit a smaller number of data points indicating potential error during data acquisition, storage or hardware malfunctions. To standardize the varying sample rate of data files recorded across the dataset, 48 KHz samples are resampled to a consistent 420 samples of 1024 data points each to streamline computations and facilitate focused analysis of data processing. To enable accurate diagnostics and predictions, fault diameter under varying loads is considered which provides the impact of fault severity and its influence on the machine. Therefore, the vibration signals are divided into 420 segments, each containing 1024 data points for processing. The sampled signals are categorized into four types of dataset cases: A, B, C, and D. Dataset A comprises 9,728 data samples of 0.007 diameter faults, encompassing three distinct fault types and normal faults. Dataset B has 6,475 instances of three distinct fault types at a diameter of 0.014. Dataset C has 9,759 data samples with a fault diameter of 0.0021. Dataset D encompasses all varieties of faults under three distinct load circumstances and fault sizes. Table 1 presents the configuration of instances from the CWRU dataset under consideration.

Table 1. Dataset cases of CWRU

3.2 Classification models

A multiclass fault diagnosis system is proposed for various data-organized groups of bearing faults. The data samples of each dataset are partitioned into training and testing sets. The training samples include 80% of the data, whereas the testing samples constitute 20% of the data. The classification models are executed for 35 epochs with a batch size of 128.

3.2.1 CNN

In this study, Layer 1 of the Convolution has a kernel size of (3, 3) with 64 filters, and the input shape is (30, 30, 64), featuring 640 parameters which facilitates balance between diverse feature extraction capability and computational efficiency of the signals. The second layer of the convolutional layer comprises 128 filters with a kernel size of (3, 3), and the input shape is (15, 15, 64). The produced feature maps have a shape of (13, 13, 128) and contain 73,856 parameters. The employed activation function is a ReLU function, characterized by continuity and an unbounded signal. Layer 3 of the Convolution layer comprises 256 filters with a kernel size of (3, 3), and the input shape is (6, 6, 128) which facilitates the hierarchical extraction of complex features contributing to fault detection. The produced feature maps have a shape of (4, 4, 256) and contain 295,168 parameters. To standardize the activations of preceding layers, batch normalization is utilized. The feature maps from the first, second, and third Conv2D layers are downscaled by a factor of two by the selection of the maximum value. The pooling layer’s feature maps are converted into a one-dimensional vector appropriate for the following layers. The initial dense layer comprises 1024 neurons that process flattened feature vectors to generate a 128-dimensional vector. A dropout rate of 0.5 is established to mitigate overfitting. The model’s final output layer executes categorization by transforming a 128-dimensional vector into six output classes. The condensed architecture of the network model is presented in Table 2.

Table 2. Overview of CNN layer configurations

3.2.2 CINN

Feature extractor: The convolutional layers of the model manage the feature extraction process. The three convolutional layers acquire intricate aspects of the signal and utilize 64 filters of kernel size (3, 3) with a ReLU activation function. The pooling layer down-samples by selecting the maximum value with a stride, so mitigating overfitting and reducing computational expenses. This pertains to learning and feature extraction. The pooling layer’s feature maps are converted into a one-dimensional vector appropriate for the following layers.

The initial dense layer of the classifier has 1024 neurons utilizing ReLU activation, processing a flattened feature vector that captures the intricate patterns and characteristics identified in the preceding feature-extraction phase. The second layer, utilizing softmax activation, generates the predicted probabilities for each class within the dataset.

The combined model integrates the feature extractor and classifier into a singular sequential framework for comprehensive training and prediction.

Table 3 presents a summarized architecture of the Convolution Invariant Neural Network Model.

Table 3. Overview of CINN layer configurations