VIBRATION SIGNATURE ANALYSIS BY HYBRID MULTI-LAYER NEURO- FUZZY SYSTEM

A proficient fault detection model has to be sketched for detecting slight variations of the vibrating signal of rotating machine whereas the diagnosis process prominently stuck with the inefficient extraction of effectual features of a signal in reduced time. Existence of above stated issue results in the confinement of inventive Module 1 of the Vibration Signature Analysis by Hybrid Multi-Layer Neuro – Fuzzy System (V-HMNFS), which could collect the RKC (RMS, Kurtosis, Crest factor) signal features for every instantaneous signal unit while eliminates noise thereby reducing pre-processing task. This in turn lessens time consumption and at the end yields learnt extracted faulty features. Accurate classification of faulty features can be accomplished by casting inimitable Module 2 classifier which provokes a good path to provide accurate classification based on learnt features. This responsible classifier collectively organises the RKC features of respective signal units and does accurate classification of faulty occurrences based on the features in less time.


INTRODUCTION
One of the most significant parts of rotating machines are Rolling bearings, thus conditioned monitoring of rotating machine prevent the failure of machine [1]. When faults are generated in the rotating bearing, it causes vibration and noise, which will make the machine breakdown/shutdown [2].In industrial perspective production of product vacant, sometime severe vibrations of bearings can even lead the whole system to operate incorrectly [3]. To overcome these issues there are several monitoring ways accessible, that need costly sensors and specialised tools.
This resulted the necessity for a less expensive and precise technique for identifying and preventing the machine faults rather than curing issues. Effective and economical fault diagnosing to prevent machine fault is usually a difficult task due to large process involved like data acquisition, feature extraction, and fault detection and identification structure [4]. In order to prolong life of bearings, it becomes pertinent to develop a consistent system to predict faults in bearings which may help in easing the three problems mentioned above.
There are several different bearing fault prediction techniques, one of the foremost popularly used algorithm is 'Vibration Signal Analysis'. It is long since understood that vibration signals and sound/acoustic signal from rotating machines are connected with their structural dynamics [5].
Dominant diagnostic information can be attained from non-healthy vibration signals from rotating parts by using appropriate process technique. In recent years, varied fault detection ways are 637 VIBRATION SIGNATURE ANALYSIS been implemented for extracting the useful features from the vibration signals taken as input. Also for the fault diagnosis, BPNN is implemented [14].
This neural network concept lead the grouping of SOMNN with PCA for the purpose of fault diagnosing [15]. The stuck in inefficient feature extraction by PCA brought wavelet entropy along with SVM [16]. The time complexity in SVM build Deep learning [17] which is one of the strongest representational learning algorithms and DBNs [18,19] are experts in constructing deepbelief structures.
Deep Belief Nets are thought of as an awfully intricate non-linear feature extraction algorithms, in this all hidden layers learn to characterise required features which are useful and that attain higher order associations in the originally collected input data set. Recently developed applications of DBNs are: hand-written character recognition [20], speech recognition [21], 3Dobject recognition [22], and extraction of road maps from clustered-aerial images [23] as well as knowledge retrieval [24]. The maintenance department is to inherit the best from above requisites and stay rotating machinery and plant equipment in smart operative condition that prevents failure and production loss.
Monitoring the rotating machine has been a difficult task for engineers particularly in industries because the vibration signals emitted by rotating parts of machinery have a nonstationary characteristic but the actual signal are weak and possess low energy generated by the faulty components, but with sturdy noise. This necessitate the effective extraction of features for diagnosis which represents main fault information of vibration signals in least time. Though effective extractions of features are made success for each and every instantaneous frequency, there arises a need to predict accurately the fault by precise collection of all the features in less time.
Henceforth a proficient algorithm is presented in the proposed methodology section.

PROPOSED FEATURE EXTRACTION AND CLASSIFICATION ALGORITHMS: V-HMNFS
Fault detection is a vital process in analysis of vibrational signal yielded from rotating machinery. Vibration signals have to be analysed in depth for that there is an increased need to 638 SUMIT KUMAR SAR, RAMESH KUMAR extract the features such as RMS, Kurtosis and Crest factor from every signal unit whereas obtaining features from each individual unit is more tedious. To analyse all such aspects, here we propose a system consisting of two modules: first, which has the stochastic nature to accurately extract the required features based on the characteristics for feature extraction which considers every signal unit and the second, for an efficient fault diagnosis process where there is an increased need to exactly classify faulty signals in reduced time. In the first module, the effectual features are accurately obtained by the aid of Hilbert Huang Transform which utilizes EMD (Empirical Mode Decomposition) for effective IMF (Intrinsic Mode Function) extraction. The features are learned individually based on the faulty features using DBN which at end yields the learned faulty features. In the second module of the proposed system, the classifier utilizes the random forest algorithm which makes a list of available outputs. The problem of overfitting in RF is tackled by the ANFIS which utilizes the knowledge based rule for prediction.
The techniques used are explained and the overall proposed architecture described below. This proposed architecture show V-HMNF System for feature extraction and Fault Diagnosis which aims at improving automatic identification of faults accurately. One of the primary requirements for this system to work is the availability of vibration signals from a healthy machine of the same species. We are interested in accurate fault prediction as part of Condition Based Monitoring (CBM), hence we have considered Vibration Signature (VS). In our proposal, the 639 VIBRATION SIGNATURE ANALYSIS system starts with collection of raw data emitted by machines from appropriate rotating machine parts by various data acquisition methods. This raw vibration data is mostly very noisy and is also exposed to numerous environmental impurities. In order to detect features effectively identifying and eradicating noise, the first module was developed. This feature extraction technique automatically pre-processed the raw data with the help of high speed training and utilizes the advantage of DBN thereby concentrating on the extraction of only required RKC features. The first module determines the features of the original signal for every instantaneous frequency irrespective of signal type by the greedy layer wise learning enabling fast and active sorts without redundancy and noise. Every signal's RKC features are determined and learnt which holds the faulty signal information. After this process in order to detect the fault diagnosis automatically, an intelligent pattern classification method is introduced as the second module. Finally faulty and healthy or VS feature are compared which is based on the pattern matching concept and the predicted output is finally retrieved.

DATA ACQUISITION
Fault Diagnosis starts with data acquisition, where machine individualities are captured for further analyses and our work is focused on vibration signature based fault prediction as discussed in eqn (1) Let Vi (t) is the set of data of original vibration signal from various mechanical components of the machine. Any required feature of a rotating machinery can easily be calculated from the raw vibration signal data, however the data needs to be pre-processed. But in this work the raw data is automatically pre-processed with the help of our proposed feature extraction technique. The Module 1 pre-training procedure treats each consecutive pair of layers in the learning process, whose joint probability is defined as,

MODULE 1 FOR FEATURE SELECTION
.
The above eqn (1) The RBM parameters can be efficiently trained in an unsupervised fashion by maximizing the likelihood of the joint probability in equation (2) , ( , ( )) This L is over training samples of vibrated signal v(t).
For extracting the IMFs of a complex (vibration plus acoustic) signal in feature extraction, the training phase of Module 1 algorithm adopts the EMD technique to determine the monocomponent of the original vibration raw signal.
EMD can efficiently decompose any given signal into its individual mono-component 641 VIBRATION SIGNATURE ANALYSIS signals called Intrinsic-Mode Function (IMF). An IMF must satisfy two conditions given below: 1. In the considered data set, the difference between the number of extrema and the number of zero-crossings must be 0 or at max 1, and 2. At any instance of time, the mean value of the envelope defined by local maxima and the local minima is zero.
The instantaneous mono-component signal, IMF is obtained in order to analyse the signal individually for better confinement of effectual features from each unit which is obtained by EMD methodology [25]. Post EMD, a signal V (t) for eqn (2) is expressed as eqn (4), Where i x is the i th decomposed IMF of the signal () Vt , n y is the noisy signal. Here the EMD determines the mono-component of the original signal individually but there is an extended need to analyse every mono-component individually based on its instantaneous frequency, so that the effectual features can be obtained from each individual vibration signal and also in addition this EMD process results in some limited amount of noise, which must also be handled. Above stated issues are handled by obtaining the Hilbert Haung transform which results in a selective preprocessing of the effectual signal features obtained. Here HHT is implemented on both sides of Eqn(5), the Hilbert-Haung spectrum of V(t), P (ω, t), is obtained from the following equation (6): Where 'Re' is the operator of real part, () i v t is the function of the amplitude and ( ) i t  denotes the function of the instantaneous frequency. It is notable that the residual value n y in Eqn (5), which uses very small amount of energy from the signal, is ignored; thus noise is avoided, which is equivalent to pre-processing of the signal by utilizing eqn (6) ,which in turn reduces the time factor for extraction of faulty features.
The marginal spectrum of HHT which extracts RKC features is expressed by an integrated 642 SUMIT KUMAR SAR, RAMESH KUMAR spectrum w.r.t. time as in eqn (7) dt t P p where T is number of the features of ()  (9) such that v(t) is the raw time series at point n, µ is the mean of the trained data,σ 2 is the variance of the data, N is the total number of data points.
(iii) Crest Factor : Crest Factor is a better and more useful feature and is defined as the ratio of the peak value of the input signal to the RMS value. Thus, peaks in the time series signal give a proportionate increase in the crest factor value.

FAULT DIAGNOSIS USING MODULE 2
The elimination classification is carried out in Module 2. In our work this classifier offers The above eqn (12) is used to construct a tree with different bootstrap sample from original data using a tree classification algorithm. Where 'm' is the number of features which are extracted from Module 1. After the forest is formed, object that needs to be trained is put down under each of the trees in the forest for training. The training features are described in the eqn (13) ) ( max Next, the accurate prediction process of the Module 2 classifier post deep learning is done by utilizing the Neuro-fuzzy inference system which utilizes knowledge based pattern identification for the prediction of vibration fault signal.
It is well observed that ANFIS modelling is better organized plus less dependent on expert knowledge. This makes it more objective. For the purpose of generality as well as simplicity, it may be considered that the ANFIS in this work has 2 inputs, x and y and 1 output f.

VIBRATION SIGNATURE ANALYSIS
Here we assume that the considered rule base consists of only 2 'if-then' rules of 1 st order Sugeno type, then the given ANFIS structure is discussed using the following example:  (17) The above algorithm is an amalgamation of the Gradient Descent(GD) method as well as Least Squares Estimate(LSE) which consists of 2 steps:-First step is that, the premise parameters are presumed to be constant while the optimal consequent parameters are subsequently identified by LSE.
Second step is that, the consequent parameters are presumed to be constant while the premise parameters are improvised by Back-Propagation GD method based on the error values, thus yielding the optimized classification by pointing the correct prediction. Thus, by the usage of efficient feature extraction by utilizing Module 1 and the intelligent pattern recognition by means of Module 2 classifier, the V-HMNF System correctly classifies fault with the assistance of learning based feature extraction. This is the well-lit process, where due to the expelling of pre-processing stage and accurate prediction process, there is optimal use of time. The result validation in the below section will be an added proof for the efficiency of the work.

RESULT AND DISCUSSION
Our proposed system has been described in detail in the preceding section. In the results section, we discuss in detail about its performance and the actual analysis. Our proposed system is implemented in MATLAB 2015a, which requires Windows 10 and minimum 8 GB RAM.

DATA COLLECTION
We collected the Data set from 'a single-stage reciprocating-type air compressor' installed     In order to predict the accurate vibrated signal, intelligent pattern classification is required, so in this research paper focuses on Module 2 for classification and prediction of the accurate faulty signal automatically. The Figure 13 given below describes the decision tree but it doesn't predict the accurate signal because it has apriori knowledge less nature

COMPARISON ANALYSIS
Comparison was made by proper analysis of computation time, accuracy, Diagnosis 651 VIBRATION SIGNATURE ANALYSIS Accuracy, Testing Prediction Time and Diagnosis Noise are described below section.

COMPUTATION TIME
Computation time during feature extraction is defined as the time required for extracting the necessary features from raw vibration data. The computation formulas are described below.

ACCURACY
The ability to differentiate between the faulty and healthy condition is the Accuracy of any algorithm test. It is actually the proportion of true positive vis-a-vis true negative in all considered cases.

TESTING PREDICTION TIME
The testing prediction time is defined as the time taken to predict the precise fault. to all other existing algorithm and also precisely diagnose the fault with 0.068 secs which is quite higher. This time is acceptable since we do training twice and we are receiving this less time.

COMPARISON OF DIAGNOSIS ACCURACY
The Diagnosis Accuracy is defined as the overall probability that a fault will be correctly classified based on the learning sample data set. The Diagnosis Accuracy formula is described below a+d Diagnosis Accuracy= a+b+d+c (19)   In the above figure 20, that describes the comparison of different existing feature selection method and our proposed for fault diagnosis of rotating machine. In existing work large data set are used to extracting the selected features for diagnosing the fault. Similarity,in this work we are focusing to learn the large number of raw vibration signal for necessary features selection. Table 5 describes the training and testing samples for accurate feature selection using different classes for features extraction. Overall perspective of our proposed feature selection technique attains higher accuracy for selecting the features because the training data set is huge compared to existing works on techniques such as LCD-SVD, TDF-FDF and TDF which is easily attainable to extract the feature. But our proposed feature such RMS, Crest factor and Kurtosis, which are not easily achievable to extract the feature but in this work our proposed algorithm to extract the accurate needed feature easily with the help of stochastic nature. Finally our proposed diagnosing accuracy value is being increased 98.28 when compared to all other existing classifiers like CRO, PNN, etc,

COMPARISON OF DIAGNOSIS NOISE
Diagnosis Noise is most easily defined from the Mean Square Error(MSE). If we consider a noise-free m×n mono-component feature I and its noisy approximation K, MSE can be calculated as: 11    In the above Figure 21 it is clear that proposed Module 2 Classifier has statistically substantial advantage over the various methods of comparison while processing the test samples at different degrees or levels of noise. The given values in Table 4, for Extreme Learning Machine, Probabilistic Neural Network and Support Vector Machine, a noticeable decline appears when the SNR is lowered below 22dB. However, our proposed Module 2 notably performs well given a longer range of SNR, due to mainly the denoising ability of our proposed module.

TIME CONSUMPTION OF PREDICTION IN DIFFERENT CLASSIFIERS
Time consumption of prediction is the ratio of total time taken for prediction to the time taken for completion.   In the later course of this work, we deliberate to make changes in this proposed V-HMNFS to improve quality of the final results. Thus, the proposed system comprising of various techniques, has been effectively implemented on the rotating parts of an air compressor and our system is able to predict faults with comparatively better precision and speed. Fault prediction was done in a run time of 0.021 secs, 98.2% accuracy and prediction time of 0.049ms using our structured feature extraction and classification framework V-HMN system.

ACKNOWLEDGEMENT
This work is supported by Department of Computer Science and Engineering, Bhilai Institute of Technology, Durg and Dr. Ramesh Kumar. We also thank Electrical & Electronics department for providing air compressor machines available for research and study.