Rahul K. Sevakula

Intelligent Condition Based Monitoring Using Acoustic Signals for Air Compressors

Intelligent fault diagnosis of machines for early recognition of faults saves industry from heavy losses occurring due to machine breakdowns. This paper proposes a process with a generic data mining model that can be used for developing acoustic signal-based fault diagnosis systems for reciprocating air compressors. The process includes details of data acquisition, sensitive position analysis for deciding suitable sensor locations, signal pre-processing, feature extraction, feature selection, and a classification approach. This process was validated by developing a real time fault diagnosis system on a reciprocating type air compressor having 8 designated states, including one healthy state, and 7 faulty states. The system was able to accurately detect all the faults by analyzing acoustic recordings taken from just a single position. Additionally, thorough analysis has been presented where performance of the system is compared while varying feature selection techniques, the number of selected features, and multiclass decomposition algorithms meant for binary classifiers.

Intelligent condition based monitoring of rotating machines using sparse auto-encoders

Support Vector Machine (SVM) has been very popular for use in machine fault diagnosis as classifier. In most of the complex machine learning problems, the main challenge lies in finding good features. Sparse autoencoders have the ability to learn good features from the input data in an unsuperivised fashion. Sparse auto-encoders and other deep architectures are already showing very good results in text classification, speaker and speech recognition and face recognition as well. In this paper, we compare the performance of sparse autoencoders with soft max regression, fast classifier based on Mahalanobis distance and SVM in fault diagnosis of air compressors.

Object Matching Using Speeded Up Robust Features

Autonomous object counting system can help industries to keep track of their inventory in real time and adjust their production rate suitably. In this paper we have proposed a robust algorithm which is capable of detecting all the instances of a particular object in a scene image and report their count. The algorithm starts by intelligently selecting Speeded Up Robust Feature (SURF) points on the basis stability and proximity in the prototype image, i.e. the image of the object to be counted. SURF points on the scene image are detected and matched to the ones on the prototype image. The notion of Feature Grid Vector (FGV) and Feature Grid Cluster (FGC) is introduced to group SURF points lying on a particular instance of the prototype. A learning model based on Support Vector Machine has been developed to separate out the true instances of the prototype from the false alarms. Both the training and inference occur almost in real time for all practical purposes. The algorithm is robust to illumination variations in the scene image and is capable of detecting instances of the prototype having different distance and orientation w.r.t. the camera. The complete algorithm has been embodied into a desktop application, which uses a camera feed to report the real time count of the prototype in the scene image.

Generating feature sets for fault diagnosis using denoising stacked auto-encoder

Recent advancements in sensor technologies and data driven model based techniques have made intelligent diagnostic systems prominent in machine maintenance frameworks of industries. The performance of such systems immensely relies upon the quality of features extracted and the classifier model learned. Traditionally features were handcrafted, where engineers would manually design them with statistical parameters and signal transforms based energy distribution analysis. Recently, deep learning techniques have shown new ways of obtaining useful feature representation that provide state of the art results in image and speech processing applications. This paper first presents a brief survey of traditional handcrafted features and later presents a short analysis of handcrafted features v/s features learned by deep neural networks (DNN), for doing fault diagnosis. The DNN based features in this paper were generated in 3 phases: 1) extracted handcrafted features using traditional techniques 2) initialized the weights of DNN by learning de-noising sparse auto-encoders with the handcrafted features in unsupervised fashion and 3) applied two generic fine tuning heuristics that tailor DNNs weights to give good classification performance. The experimentation and analysis were performed on 5 datasets: one each on Air compressor monitoring, Drill bit monitoring and Steel plate monitoring, and two on bearing fault monitoring data. The results clearly show the prospects of DNN obtaining good feature representations and good classification performance. Further, it also finds that Fast Fourier Transform based features with DNN are more suited for Support Vector Machine as classifier than Random Forest.

Template matching for inventory management using fuzzy color histogram and spatial filters

Automated counting of objects is useful for firms to keep a track of the number of objects present in the inventory. This in turn helps them to adjust their production rate accordingly and thus efficiently cater to the market demand of goods. In this paper, we have proposed a methodology for object counting using color histogram based segmentation and spatial filters. Given a prototype image, which refers to the objects image one wishes to count, the scene image is first segmented using color histogram to extract out the most likely location of the prototype. This is followed by the calculation of sum of squared distances and use of spatial filters to reject false alarms. To improve the systems robustness towards uneven lighting conditions, a fuzzy based color histogram has also been introduced. The paper then further compares the two histogram methods for their performance and computational complexity. The complete algorithm has been developed into a desktop application which uses a remotely connected camera to give real time object count. The methodology and application were tested by performing real time experiments. Both of them have shown good results under normal illumination conditions.

Unsupervised approach for object matching using Speeded Up Robust Features

Autonomous object counting system is of great use in retail stores, industries and also in research processes. In this paper, a Speeded Up Robust Feature (SURF) based robust algorithm for identifying, counting and locating all instances of a defined object in any image, has been proposed. The defined object is referred to as prototype and the image in which one wishes to count the prototype is referred to as scene image. The algorithm starts by detecting the interest points for SURF in both, prototype and scene images. The SURF points on prototype are first clustered using density based clustering; then SURF points in each cluster are matched with those in scene image. The SURF points in scene image that have been matched w.r.t. a single cluster, are clustered using the same clustering algorithm. Each cluster formed in scene image represents an instance of prototype object in the image. Homography transforms are further used to give exact location and span of each prototype object in the scene image. Once the span of each prototype is defined, SURF points within this span are matched with the prototype image and then Homography transform is once again applied while considering the newly matched SURF points; thus eliminating noisy detection/s of prototype. While the same process is repeated with each cluster, a novel centroid based algorithm for merging repeated detections of same prototype instance is used. Carrying the benefits of SURF and Homography transforms, the algorithm is capable of detecting all prototype instances present in scene image, irrespective of their scale and orientation. The complete algorithm has also been integrated into a desktop application, which uses camera feed to report the real time count of the prototype in the scene image.

Support vector machine for large databases as classifier

Support Vector Machine (SVM) has been successful in multiple areas and is widely accepted as the best off the shelf algorithm for classification. A standard SVM has O(n3) time and O(n3) space complexities, hence making it limited in its usability for large database. We know that in real world scenario, most of the databases where Data Mining is used are large. This paper reviews various algorithms and techniques that have been brought forth since 1995 by researchers for implementing SVMs in a practical manner for large databases.

Ranking of sensitive positions using empirical mode decomposition and Hilbert Transform

Condition Monitoring is the process of recognizing machine health status, by analyzing the various parameters of machine. For Acoustic Emission based condition monitoring, generally acoustic data needs to be taken from several positions and analyzed, which can be cumbersome and many times economically not viable. Thus there arises a need to define sensitive positions. Sensitive positions are positions which demonstrate relatively better features for fault recognition. Previously, the sensitive positions were found and ranked by analyzing certain statistical parameters of acoustic data. In this paper, the same has been done after extracting envelope of the relevant signal, using Empirical mode decomposition followed by Hilbert Transform. A case study was done on a reciprocating type air compressor for comparing the old and proposed technique for finding sensitive positions. Though similar results were found by both methods in normal conditions, when noise was introduced in some positions, the proposed method was found to be more robust w.r.t. noise.

Compounding General Purpose Membership Functions for Fuzzy Support Vector Machine Under Noisy Environment

Fuzzy support vector machine (FSVM) is accepted as a significant addition over soft margin SVM like

Vision Based Object Counting Using Speeded Up Robust Features for Inventory Control

C