Data-driven fault diagnosis framework of taper roller bearings using statistically ranked feature sets and machine learning algorithms Authors: A. Anwarsha, T. Narendiranath Babu Published: 2026 ABSTRACT Fault diagnosis of taper roller bearings needs to be accurate and efficient to ensure industrial machinery reliability. This research has developed a vibration-based fault diagnosis method which combines statistical feature ranking and machine learning classification. Vibration signals for the five different health conditions (healthy, inner race fault, outer race fault, roller fault, and cage fault) were collected from an SKF 32,206 taper roller bearing with changing speeds and loads. Two types of features, time-domain and frequency-domain, were derived and the features with the highest discriminating power were determined by one-way ANOVA and Kruskal Wallis statistical tests. Different combinations of features were tested with six classifiers: support vector machines (SVM), neural networks, discriminant analysis, naive bayes, decision trees, and nearest neighbor. It was found that Kruskal-Wallis feature-ranking benefits not only the result accuracy but also the computational efficiency, and the best feature set had 18 features. Thus, the linear SVM classifier yielded a classification accuracy of 99% and an AUC of 1, while requiring the least training time, demonstrating that it is suitable for real-time purposes. This work proposes a novel, dependable, and computationally efficient method of identifying faults in taper roller bearings, thus leading to greater automation of condition monitoring in industrial plants. Introduction Condition monitoring of rotating machinery has gained a lot of importance for industrial systems nowadays, especially in the sectors of oil and gas, manufacturing, and energy. Among the important parts of rotating machines, the taper roller bearings are the key components that ensure the smooth running of the machines and the avoiding of mechanical failures. Finding faults in these bearings at the beginning stage is imperative to get rid of major breakdowns and unplanned downtime. Machine Learning (ML) has gradually been recognized as the most beneficial and powerful means bringing positive changes to fault detection and classification in the recent years by being capable of understanding very complicated nonlinear relationships in vibration signals1,2. However, the success of ML models is highly dependent on two fundamental steps. They are the feature extraction and the feature selection. Features that are well-extracted and properly selected will not only improve the performance of a model but also make it more understandable and increase its computational efficiency3. Feature extraction converts raw vibration signals into numerical expressions that provide useful information and basically depict the condition of the bearing. In an article, Zhang et al.4 mentioned that vibration signals from defective bearings are mostly nonlinear and nonstationary, which makes it necessary to apply sophisticated signal processing methods in the analysis. Simple features in the time domain like mean, standard deviation, kurtosis, skewness, crest factor, and RMS can serve as good indicators of fault symptoms. On the other hand, second domain analysis, acquired from the Fast Fourier Transform (FFT), enables the identification of fault-related characteristic frequencies brought about by inner or outer race defects. A number of recent works on this topic have mentioned time-frequency domain methods such as wavelet packet decomposition (WPD), empirical mode decomposition (EMD) and Hilbert-Huang transforms, as the ways to achieve better localization of transient fault events5. Such techniques make it possible to get features that would represent the time-varying and frequency-dependent behavior of bearing signals. Hence, the choice of a suitable feature extraction method is a very important step towards successful fault classification with the aid of machine learning algorithm Recent studies have demonstrated the effectiveness of signal processing and machine learning techniques for fault diagnosis in rotating machinery. For example, the use of wavelet-based methods together with machine learning techniques has proven to be effective for monitoring the state of tools and classifying faults. Likewise, diagnosing faults from vibration signals by means of machine learning models has yielded very good results on electric motor test benches and gearbox systems. Besides, classification through statistical feature extraction by means of different algorithms, e.g., support vector machines, k-nearest-neighbors, and artificial neural networks, has been extensively used for fault detection in planetary and helical gear systems. These studies highlight the importance of effective feature extraction and selection in improving classification performance across different mechanical systems6–9. After a rich set of features has been extracted; the following issue is to determine which features are most relevant for fault discrimination. If all the extracted features are included, this will typically result in high-dimensional data which may contain noise, redundancy, or irrelevant information—these are issues that lead to lower accuracy of the model and higher computational complexity. The goal of feature selection is to keep the most informative features that have the greatest impact on the classification performance while discarding the less important ones. Kumar et al.10 note that feature selection not only enhances the performance of the model but also its interpretability, as the engineers will be able to link the physical parameters to different fault types more clearly. Feature selection techniques are mostly divided into filter, wrapper, and embedded methods. Filter-based statistical methods such as One-Way ANOVA and Kruskal–Wallis tests have been particularly popular for their ease of use, speed, and independence from particular machine learning algorithms11. One-Way Analysis of Variance (ANOVA) is a parametric statistical method employed to evaluate whether the meaning of a given feature differ significantly in a set of fault categories. The test assumes that the data are normally distributed and that variances are equal across groups. In bearing fault detection, each group has a different fault condition, e.g., normal, inner race defect, outer race defect, or rolling element fault. A large F-statistic or a p-value less than 0.05 means that the average feature values of the different fault types are significantly different, thus, the feature is a good discriminator of various fault conditions. For example, Sun et al.11 utilized the One-Way ANOVA test to prioritize the vibration features derived from taper roller bearings. Features with higher F-statistics were used for classification with a Support Vector Machine (SVM), which resulted in better diagnostic accuracy and less computational effort. Among other things, ANOVA provides a quantitative ranking of the features. Based on their F-values, features can be ranked and thus a subset of the most discriminative features can be selected. However, if the data significantly violate the normality assumption or are heteroscedastic—typical of vibration datasets—ANOVA’s results become less trustworthy12. The Kruskal Wallis (KW) test is a non-parametric alternative to the ANOVA method. Hence, it is appropriate for cases when the assumptions of data normality or equal variances are not valid. Rather than comparing the means of different groups, the Kruskal Wallis test checks if the median ranks of features are different across various fault categories13. It first ranks all the feature values and then calculates an H-statistic which measures the difference among those groups. Usually, bearing vibration signals correspond to distributions that are not symmetric and have the presence of extreme values. Therefore, the KW approach might provide a more stable and reliable indication of feature importance as compared to traditional methods. According to Chen et al.12, fault diagnosis through non-Gaussian vibration features selected by the Kruskal Wallis test is more efficient than those obtained from ANOVA feature selection methods. Likewise, Liu and Xu13 proposed a combined feature selection technique based on the Kruskal Wallis test and correlation analysis to remove redundant features. The authors said that this feature set led to higher classification accuracy with the help of Random Forest classifier and Deep Neural Network. Because the Kruskal Wallis test is non-parametric, it is a very handy tool for real-life industrial datasets where experimental noise and signal anomalies are almost unavoidable. Secondly, the test does not require large computational resources which also make it suitable for real-time fault monitoring systems. One Way ANOVA and Kruskal Wallis are the two main statistical methods that can be used to select features significant in bearing fault diagnosis. When data satisfy the assumptions required, ANOVA can be a very powerful tool for feature ranking. However, under real-world condition monitoring applications, vibration signals are mostly non-Gaussian and heteroscedastic. Therefore, the Kruskal Wallis test will be more robust and accurate for such cases than ANOVA method14. Integrating these statistical techniques with machine learning not only guarantees higher diagnosis accuracy but also a faster computation. For example, Zhao et al.14 showed that an SVM classifier coupled with a Kruskal Wallis feature selection can boost the classification accuracy of taper roller bearing faults by 6% as compared to a model trained on all the extracted features. This result points out that statistical feature selection plays a vital role in building reliable and efficient machine learning-based fault diagnosis systems. Lately, many studies have been done on machine learning (ML) methods for fault detection and diagnosis in rotating machinery via vibration signal analysis. Most of these papers tested and optimized classification methods like Support Vector Machines (SVM), Random Forests (RF), and Artificial Neural Networks (ANN)15–24. Still, the major initial steps of feature extraction and feature selection, which are the main factors influencing the performance of the model, have hardly ever been touched. If features are not selected properly, models receive redundant, noisy, or irrelevant data that cause them to overfit and their diagnostic accuracy to go down. Besides, the lack of systematic feature ranking methods has been limiting not only the interpretability but also the effectiveness of ML-based fault detection systems up to now. Particularly in the area of taper roller bearing fault diagnosis, only a few papers have addressed the impact of statistical feature ranking techniques on the selection of the most influential features in classification accuracy. The gap in the literature outlined here suggests the need for a comprehensive investigation of feature importance and its effect on the performance of ML-based bearing fault detection systems. This study is motivated by the fact that the performance of a ML-based diagnostic system is not just a matter of using a complex model but also, and mostly, relying on good input features. Although machine learning techniques have been extensively applied for fault classification in rotating devices, one of the limitations of these methods is that their output may not almost always correspond to the fault condition, especially when they are trained on high-dimensional feature maps containing irrelevant and weakly discriminative features. Therefore, the existence of a proper step of feature extraction and ranking that would lead to a rise in both classification accuracy and computational efficiency is important. Furthermore, in industrial environments where high-precision instruments characterize production environments such as oil and gas, fault diagnostic techniques should be not only accurate but also simple to understand. Thus, statistical feature ranking approaches such as One-Way ANOVA and Kruskal Wallis tests can help to identify the key features that represent fault conditions. Once these methods are incorporated into machine learning algorithms, they serve to reduce data volume and computational time. At the same time, they assist in creating models that are more resistant to noise and signal variations. This work’s novelty rests on the thorough examination of feature-level optimization for bearing fault diagnosis. Feature-level optimization is relatively less explored compared to model-centric approaches. Traditional research mainly compares different machine learning classifiers. However, this paper focuses on the significance of feature ranking and choosing the best feature subset in improving diagnostic performance. Highlighted main points of the paper: A comprehensive feature extraction framework incorporating both time-domain and frequency-domain features for taper roller bearing fault signals.A dual statistical feature ranking approach using one-way ANOVA (parametric) and Kruskal–Wallis (non-parametric) methods to evaluate feature discriminative capability from complex raw signals.Identification of an optimal feature subset (18 features) that maximizes classification accuracy while minimizing computational cost.A comparative evaluation of multiple machine learning models, including SVM, CNN, neural networks, k-NN, naïve Bayes, and discriminant analysis, under varying feature set sizes.A combined analysis of classification performance and computational time, demonstrating the impact of feature dimensionality on model efficiency and scalability.Development of a feature-level optimization framework that enhances model interpretability, robustness, and suitability for real-time industrial applications. A comprehensive feature extraction framework incorporating both time-domain and frequency-domain features for taper roller bearing fault signals. A comprehensive feature extraction framework incorporating both time-domain and frequency-domain features for taper roller bearing fault signals. A dual statistical feature ranking approach using one-way ANOVA (parametric) and Kruskal–Wallis (non-parametric) methods to evaluate feature discriminative capability from complex raw signals. A dual statistical feature ranking approach using one-way ANOVA (parametric) and Kruskal–Wallis (non-parametric) methods to evaluate feature discriminative capability from complex raw signals. Identification of an optimal feature subset (18 features) that maximizes classification accuracy while minimizing computational cost. Identification of an optimal feature subset (18 features) that maximizes classification accuracy while minimizing computational cost. A comparative evaluation of multiple machine learning models, including SVM, CNN, neural networks, k-NN, naïve Bayes, and discriminant analysis, under varying feature set sizes. A comparative evaluation of multiple machine learning models, including SVM, CNN, neural networks, k-NN, naïve Bayes, and discriminant analysis, under varying feature set sizes. A combined analysis of classification performance and computational time, demonstrating the impact of feature dimensionality on model efficiency and scalability. A combined analysis of classification performance and computational time, demonstrating the impact of feature dimensionality on model efficiency and scalability. Development of a feature-level optimization framework that enhances model interpretability, robustness, and suitability for real-time industrial applications. Development of a feature-level optimization framework that enhances model interpretability, robustness, and suitability for real-time industrial applications. This study demonstrates that feature selection has a greater impact on classification performance than increasing model complexity. Methodology Within the framework developed for this research, feature extraction and feature selection are considered two pivotal components in the machine learning-based taper roller bearing fault detection operation. The method consists of five major steps, as shown in Fig. 1. They are data collection, signal pre-processing, feature extraction, feature ranking and feature selection using statistical tests, model training and performance evaluation. This practice ensures that the classification stage is carried out with the most relevant and distinctive features only. Thus, resulting in higher accuracy of fault diagnosis with minimum computation time. Fig. 1Various stages of the proposed methodology. Various stages of the proposed methodology. Data acquisition Different bearing conditions, including normal, inner race fault, outer race fault, roller fault, and cage fault were utilized to obtain the bearing vibration data from the experimental test rig. The sampling frequency and load conditions are adjusted so that sufficient temporal and spectral information can be acquired. Later, the data was randomly split into training and testing sets to evaluate the generalization ability of the proposed method. Signal preprocessing Raw vibration signals usually contain noise, baseline drift, and other types of unwanted signals that may mask the fault features easily. To lower the adverse effects of these problems, the signal undergoes a pre-processing step with the help of the Tunable Q-factor Wavelet Transform (TQWT). The advantage of using TQWT is that the users can make small alterations to its main parameters like the q-factor, level of decomposition, and redundancy factor according to the oscillatory nature of the signal to be extracted. This method can distinguish the fault-related components differently from the noise and other interfering components that are being suppressed25,26. In this case, the technique makes the defect signals extraction of taper roller bearings extremely localized, thus allowing the correct identification of bearing faults even if other vibrations of the machine and noise are present. Tunable Q-factor Wavelet Transform (TQWT) belongs to the category of signal decomposition that can be customized to the nature of the signal, allowing for the analysis of non-stationary and oscillatory signals. TQWT, unlike the usual wavelet transforms, offers the possibility of elegantly altering the oscillatory character of the basic functions through three main parameters, namely the quality factor (Q), redundancy (R), and the number of decomposition levels (J). These features make TQWT a natural choice for rolling element bearing vibration-based fault diagnosis. The quality factor (Q) specifies how oscillatory the wavelet is and can be calculated using Eq. 1.1\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:Q=\frac{{f}_{c}}{BW}$$\end{document} Here, fc stands for the center frequency and BW for the bandwidth. If the Q-factor is high, the oscillations will be narrowband; if it is low, it tends to capture transient and impulsive features, which are mainly associated with bearing faults. The TQWT breakdown is achieved through a filter bank described by scaling parameters α and β, and the redundancy can also be written as a function of α and β see Eq. 2.2\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:R=\frac{\beta\:}{1-\alpha\:}$$\end{document} Where R > 1 guarantees the presence of redundancy in the transform, which results in better handling of noises and more faithful reconstructions. The limit of decomposition levels J is mainly governed by the length of the signal N and the scaling parameters α and β, and the approximate expression is given in Eq. 3.3\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:{J}_{max}=\left[\frac{log(\beta\:N/8)}{\mathrm{l}\mathrm{o}\mathrm{g}(1/\alpha\:)}\right]$$\end{document} In essence, all these parameters control the time-frequency resolution aspect of the transform and make it possible to decompose the vibration signal into several subbands. Proper adjusting Q, R, and J gives the TQWT the ability to cleanly separate the components of the fault by removing noise and irrelevant frequency components. The key role of TQWT in this article is its capability to strengthen the fault signature through enhancing impulsiveness and that it can provide a neat expression of the signal, both of which lead to fault detection accuracy. Bearing faults like inner race and outer race ones tend to cause transient impulses. However, in many cases, these are drowned out in the noise present in the raw vibration signals. TQWT-based preprocessing is a technique that highlights these characteristics and achieves a better representation in both time and frequency domains. In turn, this leads to more reliable and accurate feature extraction and classification. A detailed mathematical derivation and comprehensive discussion of TQWT can be found in26. The present paper sees TQWT used as a potent preprocessing tool to raise the signal-to-noise ratio and to upgrade the diagnostic yield of the newly introduced framework. Feature extraction Feature extraction is a key activity in the entire bearing condition monitoring chain. As in this phase, the raw time-series vibration signals are converted into some numerical features that characterize the operational state of the bearings. In this case, the components have been extracted from both the time and frequency domains to get the complete picture of the vibration signals’ characteristics. Time-domain features include signal features, time series features, and non-linear features that provide information about the temporal behavior of the bearing vibrations as given in Fig. 2. The frequency-domain features include peak frequency, peak amplitude, and band power, are also extracted to describe the spectral properties of the signals as given in Fig. 3. The combination of all these features gives a detailed feature matrix which sufficiently describes the bearing’s behavior under different fault conditions and thus lays the foundation for the proper fault diagnosis and condition monitoring27–29. Fig. 2Time domain features. Time domain features. Fig. 3Frequency domain features. Frequency domain features. Feature ranking and selection Basically,‍‌‍‍‌ the success of the fault diagnosis of taper roller bearings depends to a great extent on the extracted features from vibration signals, as well as, on the decision of which features are the most informative and therefore which ones to use in differentiating between a healthy and a faulty condition. Here the feature ranking is done through statistical analysis methods, mainly One-way ANOVA and the Kruskal–Wallis test, which measure the importance of each extracted feature. One-way ANOVA (Analysis of Variance) is a parametric statistical technique that tests if the means of a certain feature are significantly different across several fault groups. By comparing the inter-group variance to the intra-group variance, ANOVA reveals the features that are changing in a consistent way with the fault type. Features that are marked with higher F-values have a stronger discriminating power and thus, they are the ones that lead the feature ranking process. On the other side, the Kruskal–Wallis test is a non-parametric method which is also a good choice when the data does not come from a normal distribution, hence it is appropriate for those vibration features whose distribution is either skewed or non-Gaussian. In this case, all the readings are ranked, and the test evaluates if the distributions of the features are significantly different across the fault conditions. The higher the Kruskal–Wallis H-statistic a feature has, the more classified and fault identified it is. After these statistical procedures, a feature ranking that depicts the most informative features is obtained, which not only reduces the dataset dimensions but also keeps the most informative ones. This greatly enhances the performance of the classifier and the fault diagnosis framework for taper roller bearings is not only more reliable but also computationally more efficient. Moreover, the use of both parametric and non-parametric methods in feature selection suggests a robust and flexible approach to adapting to different working conditions and signal characteristics. Model training and performance evaluation The features identified are learned by various machine learning models or algorithms namely Decision Trees, Discriminant Analysis, Naive Bayes, Support Vector Machine (SVM), Nearest Neighbor and Neural Network. Each model learns the distinguishing characteristics of different fault classes from the training dataset, and then their performance is checked on a separate test dataset. Decision trees (DT) Decision Trees represent supervised learning methods that can classify a dataset by segmenting the feature space multiple times according to the feature values. DTs in this paper are basically used to differentiate between noise, interferences, and other faults in the components of the roller bearing by finding out from the given features the simple decision rules that can be used to recognize the different fault conditions. Their explainability and low computational cost make them good candidates for initial fault classification30. Nearest neighbor (k-NN) Classifier The nearest neighbor classifier assigns a label to a test data point based on the majority label of the ‘k’ nearest training data points in the feature space. In this paper, k-NN is used for classification of roller bearing faults by comparing features extracted from vibration signals of unknown samples with those of the labeled training set. Its straightforwardness and high competence in recognizing patterns make it a very appropriate choice for fault diagnosis31,32. Naive bayes (NB) classifier Naive Bayes is a probabilistic classifier that makes the strong assumption of feature independence and hence calculates the posterior probability of each class. For fault detection of tap roller bearings, NB is utilized to infer the probabilities of the different fault states conditioned on the features extracted, thus, resulting in a classification method that is very fast and computationally efficient33,34. Discriminant analysis (DA) Discriminant Analysis is a statistical classification technique that tries to describe the difference between the classes by determining the linear or the quadratic combinations of the features that will give maximum class separability. In this work, DA is used to distinguish between the healthy and the faulty bearing conditions, thus providing the trade-off point between interpretability and the classification accuracy35. Support vector machine (SVM) Support Vector Machines are supervised learning algorithms that identify the best hyperplane that separates the data of two classes with the largest margin. By applying SVM to the vibration features of taper roller bearings, fault classification with high accuracy can be realized even if the features are in a high-dimensional space and the classes are not linearly separable36–38. Narrow neural network (NN) A narrow Neural Network is one that has a very limited number of neurons in its single hidden layer, yet it is still capable of learning the complex nonlinear relationships present in the data. NN in this study is used for the very purpose of learning the complex dependencies in the vibration features and classifying various fault conditions of taper roller bearings with high accuracy39–41. In order to determine the impact of feature ranking on the performance of the fault classification, models that have been trained with different sets of features are tested and their performance is quantified in various ways including classification accuracy and computational time. Further, Confusion Matrices and Receiver Operating Characteristic (ROC) plots are produced to visualize the performance of classifiers. In general, feature ranking that results in a higher Area Under the Curve (AUC) value is considered better because it indicates stronger feature discrimination and higher model reliability. The results are thus interpreted to decide whether the differences in the inferred feature rankings are real or just by chance. The comparative study is aimed at finding the best method among One-Way ANOVA and Kruskal-Wallis in terms of the quality of the results obtained under different signal distributions. The interpretability of selected features is also discussed in order to relate the statistical findings to the underlying physical mechanisms in taper roller bearings. This ensures that the outcomes are not merely data-driven but also aligned with the needs of engineering. The method proposed in this paper provides an integrated framework for the examination of feature extraction and statistical feature ranking impact on fault detection of taper roller bearings using machine learning classifiers. Integrating One-Way ANOVA method and Kruskal Wallis test in the diagnostic process brings together statistical analysis and intelligent fault classification. Experimental setup The experimental set-up for the current study consisted of a three-phase induction motor, a speed controller, two bearing blocks, a test bearing, a shaft, a loading mechanism, a tachometer, an accelerometer, a data acquisition system, and a personal computer. The layouts along with the photographs of the set-up are depicted in Figs. 4 and 5, respectively. The research utilizes a water-cooled three-phase induction motor as the drive motor. This motor is capable of producing a constant rotational speed over a wide range of operations. In the experiment, the test bearing is an SKF 32,206 taper roller bearing. Its details are given in Table 1. The bearing testing is done in block 2, and a 30 mm diameter, 1200 mm long shaft conveys the rotational motion of the motor. Also, to support and align properly, the other end of the shaft is fixed in bearing block 1. The loading on the test bearing is carried out by a simple and controlled mechanical system based on circular disc weights. One disc is 3 N weight, so surplus loading of the bearing can be achieved. During the test, the bearing is exposed to three different loading conditions: 0 N (no load), 3 N (one disc), and 6 N (two discs). Stepwise loading not only allows to run the operating conditions in a consistent and repeatable way but also helps in evaluating the bearing behavior under different load levels. To study the change in vibration characteristics of the bearing with speed, the bearings are tested at discrete rotational speeds from 1000 rpm to 2500 rpm. This change speed conditions will make sure that the vibration data that is collected is truly representative of real operating scenarios for bearing fault diagnosis. The bearing’s rotational speed is managed with a three-phase induction motor with a speed regulator. A tachometer is used to adjust and monitor the operating conditions accurately. The test runs are conducted at discrete speed levels of 1000, 1200, 1500, 1800, 2100, and 2500 rpm. During each individual test, the bearing is run at a constant speed. For each speed, vibration data is recorded under steady-state conditions only. This method not only results in a systematic study of the effects of rotational speed on vibration characteristics and fault symptoms but also gives assurance that the recorded signals represent stable and reliable operating conditions. Moreover, the combination of controlled load variation and discrete speed setting ensures that the collected data effectively portrays real operating scenarios for bearing fault diagnosis. Table 1Bearing dimension.DimensionBore diameterOutside diameterWidth, inner ringWidth, outer ringTotal widthContact angleValue30 mm62 mm20 mm17 mm21.25 mm14.0360 Table 1Bearing dimension.DimensionBore diameterOutside diameterWidth, inner ringWidth, outer ringTotal widthContact angleValue30 mm62 mm20 mm17 mm21.25 mm14.0360 Bearing dimension. A KISTLER 8702B50 accelerometer, which is a piezoelectric accelerometer, is used to measure the vibrations. The vibratory signals are picked up by this accelerometer fixed on the top surface of bearing block 2. Its measurement is made in the vertical (radial) direction, which is perpendicular to the shaft axis. One of the reasons why this direction is chosen is that it has a high sensitivity to variations in bearing stiffness and load distribution; in fact, it can be used as a means to capture the vibrations resulting from faults in the bearing. The consistent placement and orientation of the accelerometer ensures reliable and repeatable data acquisition over all experimental conditions. This accelerometer has a frequency response of 0.5–10, 000 Hz, ± 50 g measurement range, and a sensitivity of 100 mV/g. The vibration signals are captured by the data acquisition system at a uniform sampling frequency of 20 kHz. The sampling frequency is sufficiently high to capture the characteristic fault frequencies and their harmonics associated with taper roller bearing defects. In order to ensure uniformity and reliability in the acquired vibration data, the same sampling frequency has been used in all experimental conditions. Initially, the motor was operated at different speeds to check the alignment of the machinery and other variables visually. Ultimately, the speed is adjusted to the required value using the tachometer. Following the acquisition of the required working parameters, the vibration data from the test bearing is obtained and stored in the computer’s storage. In order to obtain more data, different rotational speed and load combinations were executed. This procedure was repeated back and forth five bearing samples to accommodate multiple stages of data collection. Five types of test bearings are used in this study: the healthy one, the one with an inner race fault, the one with an outer race fault, the one with a roller fault, and the one with a cage fault (Fig. 6). The bearing defects were deliberately made by the wire Electrical Discharge Machining (EDM) method so as to have controlled and repeatable fault conditions. Defects of 0.25 mm in size were created on the bearing components by a wire. The seeded defects are not natural ones, but they closely resemble the types of localized faults that happen in real-life applications, such as spalling, pitting, and surface degradation of bearing components. Generally, these defects develop in real operating conditions as a result of several factors. Material fatigue, improper lubrication, contamination, misalignment, or excessive loading over long periods of operation are some of them. These faults first show small and localized imperfections on the raceways or rolling elements. If left undetected, it gradually turns into more severe damage. Besides being a major tool for in-depth understanding and implementation of the bearing fault diagnosis research, the intentional defects are also a non-motive source of fault. The EDM-induced defects brought about the tested characteristics of the fault to be produced with the highest level of fact, likeness, and fault generation dynamics, which are important factors in reliable characterization and CAD-PDM model creation. As such, the dependencies emerging and dependencies within interactions are often quite associated with the hits of the technical tool of the Analytical Machinery Fault Diagnosis System which also serves the needs of the creation of new picture and simulation perfect diagnosis and has the intellectual recognition of the current fault diagnosis techniques. Fig. 4Schematic representation of the experimental setup26. Schematic representation of the experimental setup26. Once the data has been gathered, the vibration signals that were recorded are going to be analyzed for retrieving specific information about the fault, a stage that is typically known as signal processing or data analysis. In this work, the acquired vibration signals are decomposed into a collection of frequency-localized subbands through the Tunable Q-factor Wavelet Transform (TQWT). MATLAB software was chosen since the TQWT algorithm implementation could be facilitated there only from the data stored. Every subband is thoroughly examined, and the one containing the most useful information is chosen for the fault characteristic frequency extraction. The next part briefly explains how different bearing faults can be identified through fault frequencies. Fig. 5The experimental setup. The experimental setup. Fig. 6Test bearing defects (a) Outer ring (b) Inner ring (c) Roller (d) Cage. Test bearing defects (a) Outer ring (b) Inner ring (c) Roller (d) Cage. Results and discussions The MATLAB environment (version R2022a) was used to process the collected vibration data from the bearings for a fault classification purpose. Stage 1: Signal preprocessing. Raw vibration signals are usually a mixture of the actual signal and noise caused by environmental conditions, structural vibrations, and other parts of the machine. In this paper, noise elimination is done by a preprocessing method based on a Tunable Q-factor Wavelet Transform (TQWT). This method allows for the separation of the fault-related components from noise effectively. First, using TQWT, the raw vibration signal is broken down into subbands, where each subband corresponds to a separate frequency range. Thanks to the fact that its Q-factor can be changed, the TQWT can separate oscillatory components (which are related to fault signatures) from non-oscillatory or noise components. By choosing suitable values for the parameters Q, R, and J, one can make the transform highlight the impulsive features and at the same time reduce the level of the unwanted signal components. Once the signal has been split into subbands, those subbands with the most important fault-related information are found by looking at their energy distribution and frequency content. Subbands that mainly consist of noise are either thrown away or their level is reduced but subbands that contain fault components are saved. At this point, a “cleaned-up” version of the signal is attained by the amalgamation of the only-the-selected subbands. The end-product is an upgraded signal with an elevated signal-to-noise ratio (SNR). The function of this step is to clean the signal from noise effectively and increase in both the time and frequency domain the visualization of fault-induced impulses. The better quality of the signal enables more precise feature extraction and serves as a factor for the boosted performance of the next step, i.e. classification of faults based on machine learning. Representative vibration signals both in time and frequency domains are given in Figs. 7, 8, 9, 10 and 11 to demonstrate how well the preprocessing process worked. They display the frequency spectrum and how the signal changes over time under inner race fault (IRF) and outer race fault (ORF) conditions, before and after applying TQWT. The frequency spectrum and the time waveform of the healthy bearing are shown in Fig. 7 as a baseline to allow for a direct comparison with the raw and denoised IRF and ORF signals. The fault features are not easily visible in the raw signals because of noise and interference from other mechanical components. After TQWT-based preprocessing, the signals have better clarity. The time domain signal exhibits clearer, more prominent impulses and the frequency domain shows well-defined peaks at the characteristic fault frequencies. Fig. 7Frequency Spectrum and Time Waveform of the Healthy Bearing rotates at 2100 rpm under no-load condition. Frequency Spectrum and Time Waveform of the Healthy Bearing rotates at 2100 rpm under no-load condition. Fig. 8Frequency spectrum of the OR fault rotates at 2100 rpm under no-load condition. Frequency spectrum of the OR fault rotates at 2100 rpm under no-load condition. Fig. 9Time domain plot of the OR fault rotates at 2100 rpm under no-load condition. Time domain plot of the OR fault rotates at 2100 rpm under no-load condition. Fig. 10Frequency spectrum of the IR fault rotates at 2100 rpm under no-load condition. Frequency spectrum of the IR fault rotates at 2100 rpm under no-load condition. Fig. 11Time domain plot of the IR fault rotates at 2100 rpm under no-load condition. Time domain plot of the IR fault rotates at 2100 rpm under no-load condition. The fault frequencies and their harmonics which are responsible for reliable fault identification are clearly seen in the enhanced spectral representation of the vibration signal. These findings verify the ability of TQWT to enhance the signal-to-noise ratio effectively. Besides, TQWT enables the diagnosis from the feature extraction stage of the signal. The higher fidelity of the signal obtained through TQWT, therefore, explains why subsequent steps, e.g., feature extraction and fault classification using machine learning, perform better. A detailed study on the application of TQWT in signal processing of taper roller bearing faults under various test conditions has been reported and published in26. The present study includes representative results to demonstrate the effectiveness of TQWT preprocessing and at the same time to keep the paper short and concise. Table 2Bearing defect frequency equations.Defect typeAbbreviationFrequency equationOuter race faultBPFO \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{n}{2}{f}_{s}\left(1-\frac{d}{D}cos\propto\:\right)$$\end{document} Inner race faultBPFI \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{n}{2}{f}_{s}\left(1+\frac{d}{D}cos\propto\:\right)$$\end{document} Cage faultFTF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{f}_{s}}{2}\left(1-\frac{d}{D}cos\propto\:\right)$$\end{document} Roller faultBSF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{D.f}_{s}}{2d}\left[{1-\left(\frac{d}{D}cos\propto\:\right)}^{2}\right]$$\end{document} Table 2Bearing defect frequency equations.Defect typeAbbreviationFrequency equationOuter race faultBPFO \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{n}{2}{f}_{s}\left(1-\frac{d}{D}cos\propto\:\right)$$\end{document} Inner race faultBPFI \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{n}{2}{f}_{s}\left(1+\frac{d}{D}cos\propto\:\right)$$\end{document} Cage faultFTF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{f}_{s}}{2}\left(1-\frac{d}{D}cos\propto\:\right)$$\end{document} Roller faultBSF \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$\:\frac{{D.f}_{s}}{2d}\left[{1-\left(\frac{d}{D}cos\propto\:\right)}^{2}\right]$$\end{document} Bearing defect frequency equations. Based on the bearing parameters and the operating speed of the shaft, one can compute the typical fault frequencies of rolling element bearings. A few of the common fault frequencies are: the ball pass frequency of the outer race (BPFO), ball pass frequency of the inner race (BPFI), ball spin frequency (BSF), and fundamental train frequency (FTF). The representations of these frequencies are given in Table 2. Besides n which is the number of rolling elements, fs is the shaft rotational frequency, d is the rolling element diameter, D is the pitch diameter, and α is the contact angle. The frequency spectra obtained after TQWT-based preprocessing reveal very clear peaks at these feature frequencies and their harmonics. It verifies the presence and the exact location of the bearing faults. On the other hand, these frequency components are a lot less pronounced in the raw vibration signals due to the noise and interference. Numerical frequency peaks acquired from experiments also resemble fault frequencies computed theoretically very closely. For example, outer race fault frequency is determined at 249 Hz, which is almost identical with the theoretical figure of 248.5 Hz, obtained from the application of the formulas in Table 2. Similarly, inner race fault frequency is discovered at 346 Hz which matches perfectly with its theoretical figure of 346.15 Hz. The bearing fault frequencies at various rotational speeds under different load conditions is given in Table 3. This justifies the correctness of the proposed diagnostic method. In fact, it is seen that the preprocessing technique proposed here increases the ease of recognizing fault-related frequency components. This supports on a very solid physical level the feature extraction and classification stages later on. Table 3Experimental fault frequencies of the bearings at various rotational speeds and loads.Sl No.Shaft speed (RPM)Outer race fault frequency (Hz)Inner race fault frequency (Hz)0 N3 N6 N0 N3 N6 N11000119.10119.63119.63164.90165.10165.7421200141.90142.68142.05197.20198.03198.9931500177.90176.41177.52247.80247.06248.0241800214.20213.62212.46296.10297.20297.9452100249.00247.86247.31346.00347.83347.5762500295.10296.76295.75412.30412.74413.66 Table 3Experimental fault frequencies of the bearings at various rotational speeds and loads.Sl No.Shaft speed (RPM)Outer race fault frequency (Hz)Inner race fault frequency (Hz)0 N3 N6 N0 N3 N6 N11000119.10119.63119.63164.90165.10165.7421200141.90142.68142.05197.20198.03198.9931500177.90176.41177.52247.80247.06248.0241800214.20213.62212.46296.10297.20297.9452100249.00247.86247.31346.00347.83347.5762500295.10296.76295.75412.30412.74413.66 Experimental fault frequencies of the bearings at various rotational speeds and loads. Stage 2: Feature extraction. Feature extraction was performed via the Diagnostic Feature Designer application in MATLAB. This application allows the extraction of a wide variety of features in the time domain and the frequency domain from the vibration signals. Time-domain features basically represent the statistical properties of the signal amplitude, whereas frequency-domain features essentially illustrate the energy distribution in the spectra of fault-induced vibrations. The features that are referred could be viewed as a kind of description of the condition of the bearing, since the signals characteristic of the fault, if the bearing is faulty contrast with the signals of the normal bearing in both the statistical and spectral domains. However, not all features lead to fault discrimination equally since some features are more reactive to the defects of the bearing than others which are even completely irrelevant or just redundant. Therefore, it is a must to determine the features that will give the best classification performance and at the same time reduce the computational complexity that is another purpose of this paper. Stage 3: Feature ranking & feature selection. The solution to the problem above is feature ranking and feature selection to find the strongest features for the matter of bearing fault classification. Two statistical feature ranking methods that can be found in the MATLAB Diagnostic Feature Designer tool, one-way Analysis of Variance (ANOVA) and Kruskal-Wallis (KW) test, have been used in this work. These methods look at each feature as a discriminative power and, therefore, they measure which features are better in distinguishing among the fault classes by means of statistical differences. The methodology section contains the explanation of the one-way ANOVA and KW methods. Besides the importance of the features, the number of features chosen also plays a role in the accuracy of the classification, that is the machine learning models performance in general. Eight different feature sets are made for each of the above-mentioned ranking methods such that each set contains the top features in different amounts. Feature sets containing more or fewer features can be chosen also, but here the choice is limited to eight for a simple and more straightforward evaluation of the effect of feature dimension on the classification performance. Stage 4: Fault classification. Next, the chosen feature vectors are brought into the Classification Learner app of MATLAB for the purpose of fault classification. The classification task involves five distinct bearing conditions: healthy bearing, inner race fault, outer race fault, roller fault, and cage fault. Supervised machine learning models are trained and tested using the selected feature sets to evaluate their ability to accurately distinguish between these bearing states. The classification outcomes achieved by various feature ranking methods and different feature set sizes are the topics of the analysis and the discussion carried out in the subsequent subsections, focusing on classification accuracy, robustness, and the impact of feature selection on fault diagnosis performance. Classification accuracy of machine learning models using one-way ANOVA –ranked feature sets Figure 12 shows the results of feature ranking by the one-way ANOVA method. The bar chart displays the F-values of single features, sorted from highest to lowest importance. This indicates their statistical significance in distinguishing between different bearing fault conditions. The superimposed curve illustrates the cumulative contribution of features in percentage. This helps in identifying the minimum number of features required to achieve effective fault classification. It is observed that a limited number of top-ranked features contribute significantly to the overall discriminative capability. The remaining features have relatively lower importance. The cumulative curve indicates that the majority of the feature contribution is captured within the top-ranked features. This supports the selection of an optimal feature subset. Fig. 12Feature ranking using one way ANOVA method. Feature ranking using one way ANOVA method. Machine learning (ML) models were used to classify data with the features ranked by the one-way ANOVA method. To ensure that the trained models can predict accurately in the future and can generalize well with the unseen data, a suitable validation scheme has been employed. Here, five-fold cross-validation is used during the training of the models to get the most reliable results. Moreover, 20% of the entire dataset is kept aside for final testing of the models. This way the models are tested with unseen data. The use of such a combined validation and testing scheme guarantees that the given performance metrics are truly indicative of the expected diagnostic performance in the real world. As explained previously, different ML classifiers such as support vector machines (SVM), neural network (NN) classifiers, naive Bayes (NB) classifiers, decision trees (DT), discriminant analysis (DA), and nearest neighbor (k-NN) classifiers are considered for the bearing fault classification task. Each of the above-mentioned classifiers are trained and assessed on the feature sets obtained through the ANOVA-based ranking method in order to determine the most appropriate classifier as well as the exact number of features necessary for diagnosing faults. Eight different feature sets (FS1 to FS8) are derived from the ranked features. FS1 carries the top 5 ranked features, FS2 contains 8, FS3 has 10, FS4 includes 12, FS5 holds 15, FS6 features 18, FS7 has 21, and FS8 includes 24. The Classification Learner application is fed with these feature sets one after another, and the classification accuracy of each ML model is checked. The classification results shown in Table 4 reveal that the support vector machine remains robust compared to other classifiers for every feature set it is tested on, as it tops the list of average classification accuracies at 97.33%. Using FS6, the highest accuracy of 98.8% is recorded, which further confirms the SVM’s good discriminative potential when the number of features is right. Discriminant analysis is the second-best classifier after SVM with an average accuracy of 96.31%. When FS7 is used, the best accuracy of 98% is reached. DA produces similar classification results to SVM for FS1 and FS7 which means that it can successfully work with both small and large feature sets. Table 4Classification accuracy for different models in different feature sets ranked by one-way ANOVA.Sl. No.Model nameClassification accuracy (%)FS 1FS 2FS 3FS 4FS 5FS 6FS 7FS 8Average1Decision trees92.091.893.295.093.595.594.294.893.752Discriminant analysis95.29694.597.097.597.898.094.596.313Naive bayes classifier93.592.293.893.596.297.597.895.294.964Support vector machine95.296.897.597.597.898.898.097.097.335Nearest neighbor classifier93.293.895.095.597.897.094.894.595.206Neural network classifier92.095.296.896.596.297.597.295.295.83Average93.5294.3095.1395.8396.5097.3596.6795.20 Table 4Classification accuracy for different models in different feature sets ranked by one-way ANOVA.Sl. No.Model nameClassification accuracy (%)FS 1FS 2FS 3FS 4FS 5FS 6FS 7FS 8Average1Decision trees92.091.893.295.093.595.594.294.893.752Discriminant analysis95.29694.597.097.597.898.094.596.313Naive bayes classifier93.592.293.893.596.297.597.895.294.964Support vector machine95.296.897.597.597.898.898.097.097.335Nearest neighbor classifier93.293.895.095.597.897.094.894.595.206Neural network classifier92.095.296.896.596.297.597.295.295.83Average93.5294.3095.1395.8396.5097.3596.6795.20 Classification accuracy for different models in different feature sets ranked by one-way ANOVA. Out of the six classifiers, NN is the 3rd closest to SVM in terms of average accuracy as it averages 95.83%, and it reaches 97.5% at its maximum by employing FS6. This confirms that the NN classifier can take advantage of the set of features large enough but at the same time not overly diverse. The nearest neighbor classifier is able to yield average accuracy close to 95.20%, and its peak performance of 97.8% is reached by using FS5. The naive bayes classifier attains an average accuracy of 94.96% and its maximum accuracy of 97.8% is when FS7 is used. On the other hand, the decision tree is the least performing classifier overall as its average accuracy is 93.75% while its maximum accuracy of 95.5% is when FS6 is used. Looking at the average accuracies of all classifiers, it appears that FS6 (18 features) is the optimum set of features, and the best classifier is able to classify with a mean accuracy of 97.35%. The top performance of the SVM, NN, and DT classifiers is also attained when FS6 power feature set is used. FS7 places second with an average accuracy of 96.67%, while DA and NB classifiers reach their highest classification accuracies using this feature set. FS5 ranks third with an average accuracy of 96.50% and provides the nearest neighboring classifier with an optimal performance level. The relationship between the size of the feature set and the classification accuracy can be seen quite clearly through the trend. As the number of features increases from FS1 to FS6, the average classification accuracy increases, which implies that the presence of additional relevant features makes the faults more distinguishable. However, after FS6, the performance goes down from FS6 to FS8, which implies that some redundant or less informative features have been included. This kind of phenomenon points to the trade-off between having too many features and a model’s ability to generalize. Considering above evidence, 18 features would be the perfect number of features for bearing fault classification in the presented framework here. Fig. 13Confusion matrices of different ML models using one-way ANOVA–ranked features. Confusion matrices of different ML models using one-way ANOVA–ranked features. Classification accuracy of machine learning models using Kruskal–Wallis–ranked feature sets In this subsection the classification results obtained by firstly ranking the features with the Kruskal Wallis (KW) statistical test and then using them as inputs to the machine learning models are presented. Feature ranking achieved by the Kruskal-Wallis method is displayed in Fig. 14. The features are ranked based on their non-parametric statistical significance across multiple fault classes. Each bar plot represents the ranking statistics of each feature, which are sorted in descending order of ranks. The overlaid curve depicts the cumulative in percentage contribution of the features. Compared to the ANOVA-based ranking, the Kruskal-Wallis method spreads the importance of features more evenly among multiple features. This leads to improved representation of fault-related characteristics. The cumulative contribution curve demonstrates that a sufficient level of discriminative information is achieved within a specific number of top-ranked features. This supports the selection of an optimal feature set for classification. Fig. 14Feature ranking using Kruskal-Walli’s method. Feature ranking using Kruskal-Walli’s method. Similar to the ANOVA-based study, eight separate feature sets were formed and given to the MATLAB Classification Learner application. These feature sets are named FS9 to FS16, where FS9 has the top five ranked features, FS10 has eight features, FS11 has ten features, FS12 has twelve features, FS13 has fifteen features, FS14 has eighteen features, FS15 has twenty-one features, and FS16 has twenty-four features. Table 5 reports the classification performance of the six machine learning models for the different feature sets ranked by KW. Out of all the classifiers, the support vector machine (SVM) continuously records better performance than the other classifiers for all feature sets, obtaining the highest average classification accuracy of 97.81%. The highest accuracy of 99% is recorded when the SVM is trained using FS14, thus, it is illustrated that SVM has a very good generalization capability when an optimal number of discriminative features is selected. Discriminant analysis (DA) follows SVM very closely and attains an average accuracy of 97.06%, with its best performance of 98% also obtained with FS14. Table 5Classification accuracy for different models in different feature sets ranked by Kruskal Wallis.Sl. No.Model nameClassification accuracy (%)FS 9FS 10FS 11FS 12FS 13FS 14FS 15FS 16Average1Decision trees89.894.594.894.594.595.29595.094.162Discriminant analysis96.296.597.297.297.89897.895.897.063Naive bayes classifier959195.293.596.597.598.293.595.054Support vector machine96.59797.89898.59998.297.597.815Nearest neighbor classifier95.895.594.597.59698.29895.596.386Neural network classifier95.594.295.295.297.897.296.896.596.05Average94.8094.7895.7895.9896.8597.5297.3395.63 Table 5Classification accuracy for different models in different feature sets ranked by Kruskal Wallis.Sl. No.Model nameClassification accuracy (%)FS 9FS 10FS 11FS 12FS 13FS 14FS 15FS 16Average1Decision trees89.894.594.894.594.595.29595.094.162Discriminant analysis96.296.597.297.297.89897.895.897.063Naive bayes classifier959195.293.596.597.598.293.595.054Support vector machine96.59797.89898.59998.297.597.815Nearest neighbor classifier95.895.594.597.59698.29895.596.386Neural network classifier95.594.295.295.297.897.296.896.596.05Average94.8094.7895.7895.9896.8597.5297.3395.63 Classification accuracy for different models in different feature sets ranked by Kruskal Wallis. The nearest neighbor (k-NN) classifier is in third place, with an average classification accuracy of 96.38%, and it reaches its highest accuracy of 98.2% using FS14. The neural network (NN) classifier has an average accuracy of 96.05%, and its peak accuracy of 97.8% is obtained with FS13. The naive bayes (NB) classifier has an average accuracy of 95.05%, and when FS15 is used, it achieves its highest accuracy of 98.2%. Similarly to the ANOVA-based results, the decision tree (DT) classifier is seen to have the lowest overall performance as the average accuracy is 94.16%, and the highest accuracy of 95.2% is achieved with FS14. The evaluation of average classification accuracy for all classifiers shows that FS14 (18 features) is the best feature set when using the Kruskal Wallis method for ranking, thus, it yields the highest average accuracy of 97.52%. Decision trees, discriminant analysis, support vector machines, and nearest neighbor classifiers all produce their best results with FS14. The second-best feature set is FS15, which has an average accuracy of 97.33% and brings about the optimal performance for the naive bayes classifier. FS13 is ranked third as it has an average accuracy of 96.85% and it provides the best accuracy for the neural network classifier. A steady trend in performance is seen across the feature sets: the average classification accuracy keeps on rising gradually from FS9 to FS14, which shows that the ability to discriminate against faults is increased as more relevant features are included. However, after FS14, there is a decrease in accuracy as it goes from FS14 to FS16, thus, it is assumed that the addition of redundant or less informative features takes place. These findings support that 18 features correspond to the feature dimensionality for the proposed fault diagnosis framework. When Tables 3 and 4 are compared, it is evident that the Kruskal Wallis feature ranking procedure gives better results than the one-way ANOVA method. The features ranked by KW result in a higher average classification accuracy for all machine learning models. Among the classifiers, the linear SVM model attains the highest overall accuracy of 99%, thereby being the most successful classifier in this study. Besides that, FS14, which is obtained by the KW method comes out as the feature set with the highest average accuracy of all the feature sets investigated, thus exceeding the FS6 (ANOVA-based) which has an average accuracy of 97.35%. Interestingly, FS14 and FS6 both have the top 18 features; thus, the conclusion is further supported that the feature set size is an important factor in classification performance. Figures 13 and 15 show the confusion matrices for each machine learning model when features are ranked using one-way ANOVA and Kruskal Wallis methods, respectively. The confusion matrix is a good overall measure of the classifier’s performance as it shows the distribution of correctly and incorrectly classified samples for each fault category. Figures 13 and 15 clearly show that the linear SVM classifier gives better results than the other models as it has the highest classification accuracy of 99%. Figure 16 shows the ROC curve of the linear SVM classifier at its maximum accuracy. The better the classification performance, the closer the ROC curve is to the upper-left corner. The area under the curve (AUC) corresponding to it lies between 0 and 1, where an AUC of 1 is equal to perfect classification and an AUC of 0.5 corresponds to random guessing. As can be seen in Fig. 16, linear SVM has achieved an AUC of 1, which is the proof that it is able to perfectly discriminate between all bearing faults and the healthy condition. Fig. 15Confusion matrices of different ML models using Kruskal Wallis–ranked features. Confusion matrices of different ML models using Kruskal Wallis–ranked features. Fig. 16ROC curves of linear SVM for different fault conditions using Kruskal–Wallis–ranked features. ROC curves of linear SVM for different fault conditions using Kruskal–Wallis–ranked features. Computational time analysis Tables 5 and 6 show the training time of various machine learning models using features ranked by one-way ANOVA and Kruskal Wallis (KW) methods, respectively. A thorough comparison of these results reveals that the feature ranking and the selection of the optimal feature set have a far greater effect on the efficiency of computation than the choice of a classifier alone. Table 7 both statistical ranking techniques significantly reduce the training time when a small set of informative features is chosen, thus emphasizing the importance of feature selection in the efficiency of bearing fault diagnosis. From the perspective of feature-set, the same pattern is seen in both ranking methods. With ANOVA-ranked features, the mean training time goes down from FS1 to FS6, reaching its lowest point (3.6526 s) at FS6 and then goes up for FS7 and FS8. In a similar way, for KW-ranked features, the mean training time gradually falls from FS9 to FS14, reaching its minimum (3.3691 s) at FS14, and then it rises for FS15 and FS16. These patterns suggest that feature sets with a low number of features do not provide enough fault-related information and thus the computational effort is higher, on the other hand, very large feature sets bring duplication and unnecessary overhead. Hence, FS6 (ANOVA) and FS14 (KW) are the best feature sets as they strike the right balance between the amount of information and the computational cost. It is worth mentioning that both FS6 and FS14 have the same number of features 18 top-ranked features, which means that this feature dimension is enough to grasp the main fault characteristics and at the same time, prevent redundancy irrespective of whether parametric (ANOVA) or nonparametric (Kruskal Wallis) ranking is used. A per-classifier study adds further credibility to these findings. Each machine learning model has equally a steep drop followed by a rise in the training time pattern corresponding to the size of the feature set for both ranking methods. Support Vector Machine (SVM) and Neural Network are the classifiers that demonstrate the greatest shortening of the training time as the feature sets are almost at their optimal sizes, thus showing their reliance on the features being discriminative and compact. The statistical and distance-based classifiers, e.g., Naive Bayes, Discriminant Analysis, and Nearest Neighbor, reveal a relatively smooth variation; but even so, their lowest training times (minima) coincide with the location of the optimal feature sets. Decision Tree classifiers, generally characterized by increased training costs, are also among the winners of the optimal feature selection process, thus confirming their significantly lowered training time at FS6 and FS14. Optimal feature sets and best overall performance An immediate comparison of the optimal settings indicates that the Support Vector Machine is a winner in terms of training performance, being the best option most of the time under both feature ranking methods. SVM reaches its minimum training time of 2.9184 s at FS6 using ANOVA-ranked features, whereas with KW-ranked features, it decreases further to 2.7439 s at FS14. From these experiments, it is clear that SVM efficiently takes advantage of the discriminative power of both parametric and nonparametric feature ranking methods when used together with an optimal feature subset. In general, even though both ANOVA and Kruskal Wallis methods are equally capable of enhancing computational efficiency, the KW-ranked optimal feature set together with SVM results in the lowest training time, thus making it the most computationally efficient setup for on-line and resource-limited fault diagnosis scenarios(Figs. 17 and 18). Table 6Training time of different machine learning models using ANOVA-ranked features.ML modelsFS 1FS 2FS 3FS 4FS 5FS 6FS 7FS 8AvgSupport vector machine5.38605.64484.44044.57753.29182.91843.13504.11284.1883Neural network classifier5.50654.97954.59374.70883.71653.43003.86224.67484.4340Naïve bayes classifier5.90905.46305.21384.56844.13134.16104.49844.68904.8292Nearest neighbor classifier6.28425.87605.35154.57134.29533.76454.29505.05684.9368Discriminant analysis5.76705.65154.98965.26875.02613.80264.27645.47285.0318Decision tree6.18705.71305.48285.25225.42803.83884.56705.91455.2979Average5.84005.55465.01204.82454.31483.65264.10574.9868 Table 6Training time of different machine learning models using ANOVA-ranked features.ML modelsFS 1FS 2FS 3FS 4FS 5FS 6FS 7FS 8AvgSupport vector machine5.38605.64484.44044.57753.29182.91843.13504.11284.1883Neural network classifier5.50654.97954.59374.70883.71653.43003.86224.67484.4340Naïve bayes classifier5.90905.46305.21384.56844.13134.16104.49844.68904.8292Nearest neighbor classifier6.28425.87605.35154.57134.29533.76454.29505.05684.9368Discriminant analysis5.76705.65154.98965.26875.02613.80264.27645.47285.0318Decision tree6.18705.71305.48285.25225.42803.83884.56705.91455.2979Average5.84005.55465.01204.82454.31483.65264.10574.9868 Training time of different machine learning models using ANOVA-ranked features. Fig. 17Training time of different machine learning models using ANOVA-ranked features. Training time of different machine learning models using ANOVA-ranked features. Table 7Training time of different machine learning models using Kruskal Walli’s-ranked features.ML modelsFS 9FS 10FS 11FS 12FS 13FS 14FS 15FS 16AvgSupport vector machine4.65904.20943.97743.99033.95052.74393.86493.97143.9209Neural network classifier5.10784.61444.41204.12893.63422.88543.12774.16364.0093Naïve bayes classifier4.78504.72104.55404.49904.47153.25404.44864.42124.3943Discriminant analysis5.02584.98404.59414.41323.72623.88104.13505.08904.4810Nearest neighbor classifier5.64695.62105.14314.98184.29853.81253.86254.91054.7846Decision tree5.92035.75875.59845.10214.99513.63804.44204.74895.0254Average5.19084.98484.71324.51924.17933.36913.98014.5508 Table 7Training time of different machine learning models using Kruskal Walli’s-ranked features.ML modelsFS 9FS 10FS 11FS 12FS 13FS 14FS 15FS 16AvgSupport vector machine4.65904.20943.97743.99033.95052.74393.86493.97143.9209Neural network classifier5.10784.61444.41204.12893.63422.88543.12774.16364.0093Naïve bayes classifier4.78504.72104.55404.49904.47153.25404.44864.42124.3943Discriminant analysis5.02584.98404.59414.41323.72623.88104.13505.08904.4810Nearest neighbor classifier5.64695.62105.14314.98184.29853.81253.86254.91054.7846Decision tree5.92035.75875.59845.10214.99513.63804.44204.74895.0254Average5.19084.98484.71324.51924.17933.36913.98014.5508 Training time of different machine learning models using Kruskal Walli’s-ranked features. Fig. 18Training time of different machine learning models using Kruskal Walli’s-ranked features. Training time of different machine learning models using Kruskal Walli’s-ranked features. Conclusions The study presented a highly reliable fault diagnosis system for taper roller bearings based on vibration signal analysis. The system consists of an advanced feature extraction method and statistical feature selection integrated with machine learning classification algorithms. Vibration data for five different health states of the bearing, namely healthy, inner race fault, outer race fault, roller fault, and cage fault, were collected in the experimental setup under varying operating conditions. Wire-cut electrical discharge machine was used to create artificial defects which ensured repeatability and controlled the severity of faults. Plenty of time-domain and frequency-domain features were extracted from the signals with the aid of the Diagnostic Feature Designer in MATLAB. In order to select the best features that represent the data while at the same time removing unnecessary features, two statistical feature ranking methods, one-way ANOVA and Kruskal Wallis tests, were performed. Various feature sets differing in the number of features were formed so that the effects of feature selection and number of features on the classification performance could be thoroughly analyzed. From the classification studies it appeared that the support vector machine (SVM) exceeded the performances of the other ML models by a wide margin. The highest classification accuracy of 99% was reported when the SVM was trained on the features ranked using the Kruskal Wallis method. The performances of the discriminant analysis and nearest neighboring classifiers were close to that of SVM, whereas the decision trees yielded the least accurate results. The receiver operating characteristic evaluation also confirmed the excellent capabilities of the suggested method. A linear SVM achieved an area under the curve (AUC) of 1, which means that it perfectly separated all the classes of bearing faults. A major takeaway from this work is that the quantity of the selected features considerably determines the classification accuracy and computational speed as well. Sets of 18 best-ranked features from both feature ranking methods enabled the models to obtain the highest performance. Too few features in the feature sets did not hold enough distinguishing information, whereas the feature sets that were too large contained unnecessary features that not only worsened the classification accuracy, but also increased the computational cost. The feature set FS14 (Kruskal Wallis, ranked) stands out from the rest because it led to the highest mean classification accuracy (97.52%) and the shortest training time (3.3691 s), thereby representing the most effective and computationally efficient setup. The analysis of the time required for computation points to the fact that, among the models compared here, the Support Vector Machine achieves the best training efficiency when combined with optimal feature sets (FS6 for ANOVA and FS14 for Kruskal Wallis). Though Neural Network classifiers are slightly slower than SVM, they also greatly benefit from feature selection. Above all, feature ranking by Kruskal Wallis results in lesser training time for all classifiers as compared to ANOVA-ranked features. Thus, it not only demonstrates the ability of KW ranking in reducing computational cost but also highlights its suitability for real-time or online condition monitoring applications. In summary, the proposed feature selection and classification methods are capable of diagnosing faults in taper roller bearings with high accuracy and low computational cost. It is demonstrated that bearing condition monitoring can be reliably and scalably performed by coupling Kruskal Wallis feature ranking with the right-sized feature set and linear SVM classification. The further steps in research entail taking the suggested approach into the real-time industrial settings, introducing adaptive feature selection, and combining deep learning models for fully automated fault diagnosis.