Browse Prior Art Database

Machine Learning Algorithms for Smart Meter Diagnostics

IP.com Disclosure Number: IPCOM000242462D
Publication Date: 2015-Jul-16
Document File: 52 page(s) / 1M

Publishing Venue

The IP.com Prior Art Database

Abstract

Machine learning algorithms (MLAs) are statistical models that make generalizations about training data in order to properly classify future, unknown data. MLAs can be supervised (given labelled training data), semi-supervised (given some labelled data and a lot of unlabelled data), or unsupervised (given only unlabelled data). There are a huge number of practical applications for these algorithms including Internet search engines, voice recognition, robot vision, stock market analysis, and cancer screening, among many others.

One application of particular interest is machine health monitoring. It aims to detect abnormal operating conditions so that machinery can be repaired or replaced before a more serious problem develops. There has been a significant amount of research conducted in this field, and MLAs have proved to be some of the more effective solutions.

Smart meters produce a large amount of data that can be difficult for a human to analyze. Currently, automated health monitoring systems use empirically determined limits on readings (e.g. profile factors in ultrasonic meters) to detect faults. The problem with this system is that it is difficult to know every possible fault condition, and what specific readings or combination of readings are indicative of a fault. Unsupervised MLAs, on the other hand, can detect abnormal conditions without any prior knowledge of fault conditions.

Research in machine learning has produced a huge number of algorithms that can be used independently or in concert. The quality of the predictions is largely dependent on how well the algorithms are selected and tuned to suit the input data. Highly-optimized unsupervised algorithms have been shown to detect machine faults in some cases with 95–100% accuracy.

Every time an algorithm is deployed in a new metering station, it will likely need to be trained on data collected from that specific station. As long as the algorithm itself does not need to be changed, the training process may be easily automated.

This text was extracted from a Microsoft Word document.
At least one non-text object (such as an image or picture) has been suppressed.
This is the abbreviated version, containing approximately 3% of the total text.

          Machine Learning Algorithms for Smart Meter Diagnostics

May 26, 2015

Contents

1        Abstract                                                                                                                      4

2        Introduction                                                                                                                5

3        Background                                                                                                                6

4        Unsupervised Machine Learning Algorithms                                                        8

4.1         Preprocessing..................................................................................................... 8

4.2         Feature Selection................................................................................................ 9

4.3         Density-Based Spatial Clustering of Applications with Noise (DB- SCAN)      12

4.4         Ordering Points To Identify the Clustering Structure with Outlier Factors (OPTICS-OF)......................................................................................................................... 15

4.5         K-Means Clustering..................................................................................... 17

4.6         Gaussian Mixture Models........................................................................... 17

4.7         Kernel Density Estimation.......................................................................... 20

4.8         Robust Covariance Estimation........................................................................... 22

4.9         Support Vector Machines............................................................................ 23

4.10     Hidden Markov Models................................................................................ 25

4.11     Artificial Neural Networks............................................................................ 26

4.12     Model Evaluation.......................................................................................... 30

5        Case Studies                                                                                                            32

5.1         Hidden Markov Model Monitoring of a Drilling Operation..................... 32

5.2         Induction Motor Fault Detection and Diagnosis Using Artificial Neural Networks       34

5.3         Fault Diagnosis of Low-Speed Bearing Using Support Vector Ma-  chine          35

6        Applications to Previous Experiments                             ...