CNC Machine Bearing Fault Detection using Hybrid Signal Processing

The e�ciency of the Flexible Manufacturing System (FMS) is highly in�uenced by computer numerical control (CNC) machine tools. The most common causes of CNC machine tool failure is the bearing faults that in�uences the performance of manufacturing system. The detection and diagnosis of bearing defects are crucial to the reliable functioning of revolving machinery. This paper suggests an intelligent vibration-based condition and fault diagnostic technique for the identi�cation of bearing faults in CNC machine tool. Investigational vibration data obtained for various bearings and operational requirements were analyzed in order to create a structure for the monitoring and classi�cation of bearing defects in order to determine the health of the machine. Fault diagnosis was made using hybrid signal decomposition (HSD) methodology for the decomposition of the vibration signal. Vibration features derived from the received decomposed raw signal were chosen using the key component review to eliminate redundant features. Subsequently, these related features were input into support vector machines (SVM) and arti�cial neural network (ANN) for classi�cation of various bearing faults in CNC machine tool. Experimental outcome suggest that the proposed approach has an immense capacity to prevent unplanned and unnecessary device shutdowns due to loss of bearings in CNC machine tool.


Introduction
In manufacturing industry, precision CNC machine tools are commonly used. CNC machine tools are frequently made up of several servo axes, and complex processing programmes which can be implemented according to the synchronized function of each servo axis [1]. The ndings by using simulation model of CNC system show that presented approach is responsive to incipient rotating part degradation and provides an complement solution for servo axis health monitoring [2]. Computer control based machine tools such as CNC lathe, milling, drilling and grinding machines are used by almost all the FMS and the heart of any exible manufacturing system (FMS). Modern manufacturing rms are concentrating on FMS to gain a competitive edge, meet demands, minimise direct labor costs, and boost performance through improved customer service and on-time delivery. The FMS is made up of a large number of computerized machine tools that are linked by an automated material handling device which can handle medium volumes of various parts at the same time [3]. In FMS, the in uence of versatility is one of the key factors in the study of the e ciency of the production system. Machine exibility found the maximum degree of exibility between the fteen exibility choices and the fteen as attributes based on the Preference Selection Index (PSI) methodology [4]. In addition, machine tools with the highest rating in FMS are at the core of every scalable production method.
Bearing is a vital and often experienced component of CNC machines, vulnerable to danger due to its heavy load and long operation [5][6]. It operates under extremely di cult operating conditions and produces nonlinear vibrations that are in uenced by a variety of factors including friction, clearance, stiffness, and high load [7][8]. Different techniques, such as variational mode decomposition (VMD), hybrid signal decomposition (HSD), arti cial intelligence models, and others, are used to identify the faults and are used to identify defects in bearings automatically. These methods are created using extracted indicators and then used to identify defects. Arti cial neural networks (ANN) [9][10][11] and support vector machines (SVM) [12][13] are the most commonly used models for defect detection. Analysis of time and frequency VMD (variational mode decomposition) is a popularly used adaptive tool for fault diagnosis and detection of rotating machinery [14]. Fault diagnosis of rotating machines using VMD is a recently used signal processing approach that implements effective solutions: the envelope kurtosis used as an detector, and then a frequency band entropy is used to know the optimized IMF(s) of VMD that contain important for fault detection [15]. Vibration-based research has been proposed as a key method for assessing mechanical health [16][17][18]. Impacts from incipient defects, like local defects of support bearings, gearboxes, or ball screws, can only cause minor vibrations of the structure for precision CNC machine tools' servo axis [19]. A fault detection, isolation, and prognostics scheme based on the Mahalanobis-Taguchi system (MTS) has been introduced, and experiments show that the proposed method produces adequate results [20]. Diagnosis practice based on the adaptive wavelet transform (AWT) and the adaptive neuro-fuzzy inference method (ANFIS) has been proposed for detecting bearing damages [21].
Therefore, detection of fault bearings is a critical activity to ensure the protection of the manufacturing system. Analysis of fault detection in a CNC machine tool gearbox with a planetary bearing con guration is still a di cult topic for researchers. Upwards of fty percent of CNC machine breakdowns were reported to be due to faults. In fact, rolling failure will lead to serious shaking of equipment, unplanned shutdown, a halt to output, and even health and nancial losses. Study on bearing fault detection is of great importance in the practical use, which can be due to the fact that the health state of the bearings is closely related to the reliability and durability of the operation of the machinery. The early fault function is usually very low and interferes with all the high ambient noise generated by other parts of the device.
Standard architecture methods receive defect location information from the vibration analysis signals in either the time-domain or the frequency-domain.Criteria roles are then programmed to consider the tness of the bearing. However, it is quite complicated to reliably assess the bearing condition by way of an analysis of both realms. Signi cant focuses has recently been paid to identifying and diagnosing defects in bearings. The vibration-based condition monitoring and defect diagnostic approach is traditional and useful methods of all form of bearing loss diagnostic techniques. In vibration-based bearing loss diagnostics, there are basically two distinct methods which are shown to be valuable in the detection of faults for signal processing and pattern recognition. Vibration analysis-based methods are studied as the key methods during a long period in fault diagnosis of rotating machines [22]. Popular signal processing approaches such as rapid Fourier transformation (FFT), wavelet transformation (WT) and empiric mode decomposition (EMD) have been used to diagnose bearing defects and have gained some e cacy [23]. Among these, EMD has been reported to be e cient in extracting vibration characteristics dependent on the local time scale characteristics of the signal. This technique will tailor a nuanced multi-component signal to a range of IMFs whose instantaneous vibrations are mechanically signi cant [24]. The raw vibration signal characteristics data can be retrieved in a much more reliable and effective manner by implementing the enveloping procedure to each part of the IMF. In addition, the frequency elements involved in each intrinsic mode feature are correlated to the sample frequency component. In addition to health indicators determined using time-signal based analysis, the vibration signal data on the oscilloscope or the graph recorder for observation of the periodic peak which is further used to detect local fault bearings [25][26]. The shock pulse system has been extensively studied for carrying fault detection with the aid of structural resonance detection in high frequency [27][28]. Different commonly willful approach to time-domain measurement is the time synchronous average (TSA). TSA can derive the periodic signal by the cumulative raw signal average over several revolutions thathave been successfully used for local fault diagnosis [29][30].
Frequency dependent signal is the most dominant approach for fault detection in bearings [31]. The frequency based analysis are able to detect quickly and classify those related frequency sub components, there are certain bene ts over time based analysis. Fast Fourier transform (FFT) method is a traditional frequency based analysis and can be effectively calculate the narrowband spectrum of faulty signal. In particular, the bearing defect introduces a short-term pulse, resulting in very high energy in that frequency band. TSA is used to obtain bearing faults indicators in frequency domain [32]. Another e cient complementary method for spectrum analysis includes spectrum visualization, envelope analysis, frequency lters, and sideband analysis [33]. Cepstrum based indicators like power cepstrum [34] and bispectrum [35] have couple of bene ts over conventional approaches.
Time and frequency based analysis is expanding the capability of frequency component to nonstationary vibration signals. A couple of methods based on time-frequency based like wavelet Transform (WT) [36], Wigner-Ville Distribution [37], and short-term Fourier Transform [38], can applied for nonstationary vibration signals to detect the bearing failure. WT have demonstrated their successfulness in the detection of bearing failure [39]. However, the Wigner-Ville distribution have energy leakage issue as discussed in [40] as an effective time and frequency based analysis tool used to diagnose rotating machinery faults, Centered on the top of study, this paper suggests a novel approach to hybrid signal decomposition for the diagnosis of fault bearings in the CNC machine. The hybrid signal decomposition device will be tested with vibration signals received from a bearing coupled to a CNC machine tool. This paper provides an HSD-based CNC machine tool diagnostic approach using SVM and ANN machine learning approaches.
Experiments were performed in varying bearing conditions under different operating conditions. Vibration data has been decomposed using HSD to minimize background noise. PCA was used to pick the most important features and, subsequently, these preferred features were feed as input vectors to SVM and ANN machine learning methods for classi cation and performance assessment.
The Article shall be formulated as follows. Section 2 includes a description of machine learning and the use of algorithms to detect faults. Section 3 includes a recommended methodology and a detailed description of the experimental setup and data processing protocols for HSD-dependent vibration-based bearing loss diagnostics in CNC machines. Section 4 outlines the results of numerous machines learning algorithms and, ultimately, the conclusions and possible prospects are summarized in Section 5.

Machine Learning
The basis of all ML processes is a large training data. To construct e cient ML and DL architectures for carrying out predictive maintenance. Since natural bearing decay is a sluggish process that will lager time to complete, most individuals conduct tests and gather data by using arti cially induced fault bearings or using rapid life-testing techniques [41]. For several years before the latest DL explosion, i.e. the arti cial neural network (ANN) [42]. It takes a great deal of experience in the eld and di cult function engineering to implement these algorithms. Usually, deep data processing is done rst on the acquired dataset by dimensional reduction methods for feature extraction, such as main component analysis (PCA), etc.
Lastly, along with the ML algorithm, the most representative functions have been moved. The information base of various disciplines can be very different, and with each region it often requires extensive advanced expertise, an e cient extraction function, or to sustain a fair degree of transfer-ability of ML models learned in one eld to be extended or adapted to other environments or contexts. One of the earliest studies discussing the use of arti cial intelligence (AI) methods in motor failure detection [43], which extensively outlines the signature fault frequencies for different forms of motor fault, and explores similar papers using ANN and fuzzy systems Consequently, in addition to being able to practice the ANN model in a more accurate way, most of the papers on the basis of ANN [44] all require a certain degree of human experience to guide their system of selection of features. PCA has proved itself to be an e cient and comprehensive feature selection scheme that offers instructions for classi cation purposes on the manual selection of the most representative features. One of the early adoptions of PCA can be found in [45] for fault-bearing diagnostics. Experimental studies have shown that the gain of using only the de ned PCA features of the initial 13 features is substantial as the precision of the diagnosis of fault is improved from 88% to 98%. The study shows that, relative to the use of all features, the suggested PCA strategy is successful in classifying load faults with higher precision and a less input indicator. Similarly, the remainder of the PCA-based papers [46] takes advantage of their produce a manual collection method for features. The early implementation of the k-NN classi er for fault bearing diagnosis [47], k-NN algorithm for the acoustic signal dependent ceramic data mining fault classi er. Similarly, k-NN [48][49] use k-NN to do a separation analysis of each new series of measurements and to determine a certain fault class. SVMs learning that evaluate data used for study of non-probabilistic classi cation and regression problems. In [50] results obtained by the SVM are satisfactory to distinguish bearing faults, [51][52][53] presented the e ciency and durability of the use of SVM as a classi er of faults. In addition to the frequently used ML to mentioned to identify fault bearings with different features and bene ts, including, extreme learning machines (ELM) [54], transformation learning [55]

Experimental Setup And Data Acquisition
The CNC machining centre is commonly used for large-scale precision machining of gear, shaft, and frame. However, conventional health appraisal approaches are very di cult to apply in the in-service environment, constrained by the dynamic servo-axis structure and poor signatures of incipient decay. The suggested approach is adopted for the incipient deterioration evaluation of the servo axis of these CNC machining centers in order to address this constraint. The trials are carried out in an in-service MCL-12, with a 1.5KW spindle style AC induction motor rating and a spindle speed of 2800 rpm, and eld tests are done during the entire servicing cycle on its X-axis. The experimental system is illustrated in the encoder signal acquisition system shown in Fig.1. Outer race defect The X-axis consists of an AC motor, ball screw, gearbox, and bearings for these machine tools. In the current study the two sets of type accelerometers are used. The outward accelerometer was mounted in the radial direction on the outer race and inner accelerometer is xed in the axial direction while the internal accelerometer rotated with the planet carrier. Testing at xed shaft speed of 14 Hz with various torque loads, including 30, 50 and 70 Nm, carried out during the CNC process. Checking was often done for bearings of various status (normal bearing, inner and outer race defected bearing). During each test, the acceleration signals with both the internal and external accelerometer were collected. Table 1 displays the acquired data from the test environments.

Proposed Methodology For Cnc Machine Tool Fault Diagnosis
The technique followed for controlling the state of rotating machinery in the current work as seen in Fig.  2. Signal and data processing have been grouped into four sections.
a. Training and testing vibration data of faulty and healthy bearing of CNC machine tool were collected in such a way that adequate numbers of data sets wereavailable to reliably diagnose faults.
b. The acquired vibration data was then processed using HSD, and the technique for the extraction of features was addressed.
c. PCA was used to eliminate the redundant features for enhancing the classi cation accuracy.
d. Finally, classi cation and evaluation of results were discussed using SVM and ANN classi ers.

Hybrid Signal Decomposition (HSD) Techniques
For bearing with a aw, couple of impacts on bearing passes across the defect site. Due to the basic geometry of bearing, balls in the bearing is 8, and the touch angle is 0 o , and 37.9mm is pitch diameter, 8.7mm is ball diameter, 25mm is bore diameter, outer diameter 52mm. The outer race and inner race defect frequencies of the ball pass can be determined using equations (1) and (2), n is number balls, (β) is the radial plane load angle, (d) is the roller diameter, (D) is the diameter of pitch, f o is outer race direction, and f i is the velocity of the inner race.
In this case, f o =11.52Hz and f i =5.13Hz can be seen with shaft speed of 14 Hz. BPFO=58.87Hz, and BPFI=93.93Hz are the resulting fault frequencies. The signature frequency of outer race defect is 58.87 Hz while the input shaft's rotating speed was 840 RPM. And there's 93.93 Hz for the inner race. In the 50 Nm torque load case, the impulsiveness becomes much easier with both the outer and inner race defects, as seen in Fig.3 and Fig.4 after the HSD ltering. The next step was to nd the right band for demodulation using the spectrum kurtosis (SK) technique [35]. The optimized SK is shown in Fig.5 (a), optimized by the maximal kurtosis value of the SK as the optimal demodulation core frequency for outer race defects. Fig.5 (b) Envelope processing was then carried out at fault information outer race defect of the planet bearing can be clearly speci ed at 58.87 Hz. It is straightforward to identify the envelope range with the peak value, the internal race defect frequency of 93.93 Hz of the planetary bearing. The optimized SK is shown in Fig.6 (a), optimized by the maximal kurtosis value of the SK as the optimal centre frequency for internal race defect demodulation. Fig.6 (b), Envelope processing was then carried out at inner race fault at 93.93 Hz of the planet bearing. An HSD technique was applied in this paper, which is a collection of simpli ed SK and envelope analysis suggested by compute the problem of demodulation band to remove the fault characteristics in the measured signal and further optimize the measurement diagnostic performance based on the accelerometer sensor for fault diagnosis of CNC machine.

Feature Extraction and Dimension Reduction
To obtain information about the fault that is masked in complex signals, extraction of features is necessary. A total of 11 indices of mathematical time-domain situations, viz. Mean, rms, standard deviation, kurtosis, form factor, skewness, impulse factor, crest factor, energy, entropy and margin factor were extracted using a moving window size of 600 samples from the enveloped vibration signal of 15k samples and 90% overlaps with its adjacent window. It is normalized by minimum and maximum difference within the range of [0,1] as expressed, To reduce the number of input variables is prudent to both reduce the modeling computational costs and, in some cases, improve the model's performance. Hence, selection of statistical features from the coe cients of the HSD as an input for principal component analysis (PCA) was used to train the model.
PCA is the most well approach used to decrease the dimensionality of the predictor variables and helps to avoid over tting of the classi cation algorithm. To pick the most important attributes for knowledgebased algorithms to make decisions, the distances between all the extracted features have been calculated. A total of ve strongest features, i.e. mean, standard deviation, skewness, energy, entropy, and kurtosis, were selected from among all the extracted features. Furthermore, the non-relevant features were ignored and modi ed to the degree of relevance limit. Finally, for further grouping, a function matrix of the chosen features integrating the data at various loads and speeds was used as the inputvector.

Result And Discussion
This section explains the results obtained using machine learning models SVM and ANN for the various bearing states of CNC machine tool. Training and testing are carried out for classifying the bearing faults of CNC machine tool using different classi ers. For assessing the performance of HSD based bearing fault diagnosis methodology for rotary machinery application, 5-fold cross validation is used to train, test and validate the model. Classi cation accuracy and error rate was recorded as performance measures.
Multi-state bearing prediction has been implemented using the 2D confusion matrix, comprising of a column matrix representing the predicted states, and a row matrix representing true bearing states. The selection of fault as a class attribute initiates the process of categorization and the performance of the classi er involves comprehensive class precision, confusion matrix and evaluation of favorable numerical prediction. The confusion matrices obtained using machine learning models SVM and ANN for identifying the different bearing conditions for HSD based preprocessed signal are given in Table 2. It was observed that both SVM and ANN achieved the acceptable classi cation accuracies of different conditions. In SVM, healthy outer race and inner race bearing condition are classi ed with 99%, 95% and 99 % respectively. In ANN, it outer perform SVM and achieved higher accuracies in healthy bearing is 100% classi ed, 99% for outer race defect and 100% for inner race defect. Although, the misclassi cation rate was very less for all the bearing states, the confusion matrices for different bearing conditions of CNC machine tool using SVM and ANN are given in Table 2. In order to decide a hyperplane with the maximum margin in a high-dimensional feature space, SVM is basically generated as a quadratic optimization method for both classi cation and regression issues, and the training data based on vibration signals are categorized by the hyperplane into ve classes. It has been observed that the e ciency of the classi er for classifying bearing failures is superior which may be due to the fact that in practice a prior learning and regularization concept is often inserted into the model to prevent over tting and enhance price of success. For both the classi er, individual-class metrics were calculated such as Precision, Recall, F-score. In SVM, Precision, Recall, F-score are 98.65%, 98.21% and 98.45% respectively. In ANN, Precision, Recall, F-score are 99.95%, 99.11% and 99.78% respectively. These per-class or individual-class metrics were averaged to fetch a single value called Recall, Precision, and Fscore for these metrics for various Machine Learning models applied. Precision is the ratio of correct predictions for a particular class. Recall is de ned as the ratio of the class instances which are predicted correctly. F-Score is the weighted average of the performance parameters Precision and Recall. Through means of an experiment, the e ciency of the proposed approach is investigated and this technique is extended to the diagnosis of bearing fault and deterioration assessment of an in-service CNC vertical machine tool. The ndings have shown that the proposed approach is vulnerable to incipient deterioration of the rotating components and provides an alternative option for CNC machine health assessment.

Conclusion
An intelligent vibration-based condition monitoring and fault diagnosis technique for the detection of bearing faults in CNC machine tool is proposed in this article. In order to create a system for tracking and classi cation of bearing faults in order to assess the machine's health, investigational vibration data collected for different bearings and operating speci cations were analysed.A hybrid signal decomposition (HSD) technique for the decomposition of the vibration signal was used to diagnose the fault. In order to remove redundant features, vibration features extracted from the obtained decomposed raw signal were chosen using the main component review. Subsequently, these similar features were implemented into classi cation algorithms, including support for the detection and classi cation of different bearing faults by vector machines (SVM) and arti cial neural network (ANN). Experimental ndings show that the suggested solution has an enormous potential to avoid unplanned and unwanted shutdowns of devices due to rotary machine loss of bearings.
In order to effectively incorporate an automated fault detection strategy, the HSD solution has proved to be useful in ltering non-stationary signals.
To achieve the optimum function vector to enhance the accuracy of the classi er, the collection of PCA-based features has been found to be useful.
The success rate achieved using ANN is outperformed SVM, the performance rate obtained for de ning various bearing states dependent on vibration signals. Competing interests: The writers note that they do not have any con icts of interest.      Spectrum of outer race defect (a) Optimized SK (b) Envelope spectrum