Browse Prior Art Database

Method and system for combining the information from heterogeneous asset systems for failure prediction

IP.com Disclosure Number: IPCOM000233285D
Publication Date: 2013-Dec-05
Document File: 5 page(s) / 86K

Publishing Venue

The IP.com Prior Art Database

Abstract

Pipe network in big city typically consists of hundreds of thousands pipes, whose age ranges from 1 year to 100+ years old, and pipe has rich information like material, length, diameter, depth at the individual level. A system and method for using the heterogeneous pipe system information to predict the failure behavior of the other pipe system which is usually not well recorded about the sample label, including: 1) Minimize the failure prediction classifier model risk w.r.t. the classifier model weight W on the sample set with labels; 2) Minimize the distance between two sets associated with profile attributes w.r.t the sample wise weights S; 3) Alternatively repeat the above optimization procedure until converge or the iteration time exceeds a certain threshold; 4) Use the obtained parameter W to build the failure prediction model for the heterogeneous system failure prediction (usually with fewer and no-labeled samples)

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 43% of the total text.

Page 01 of 5

Method and system for combining the information from heterogeneous asset systems for failure prediction

Pipe network in big city typically consists of hundreds of thousands pipes, whose age ranges from 1 year to 100+ years old, and pipe has rich information like material, length, diameter, depth at the individual level.The database and information management for the pipe network is still in its infancy, no industrial standard regarding how to categorize/digitalize the pipe networks. On the other hand, the pipe network is usually very sensitive to the city and agency, they are reluctant to publish the data details. This poses the difficult to directly share/leverage the detailed information associated with the pipe networks across cities, thus the agency need to manage their network by their own in a isolated fashion. Existing solutions naively train a model from one water system, and directly apply it to another system. This will ignore the difference between two systems, resulting in worse failure prediction accuracy. We are aimed to improve the accuracy by means of designing a complex optimization procedure to make full use of the hisotrical data from heterogenous system and improve the accuracy.

A system and method for using the heterogeneous pipe system information to predict the failure behavior of the other pipe system which is usually not well recorded about the sample label, including:


1) Minimize the failure prediction classifier model risk w.r.t. the classifier model weight W on the sample set with labels


2) Minimize the distance between two sets associated with profile attributes w.r.t the sample wise weights S


3) Alternatively repeat the above optimization procedure until converge or the iteration time exceeds a certain threshold.


4) Use the obtained parameter W to build the failure prediction model for the heterogeneous system failure prediction (usually with fewer and no-labeled samples)

Known solutions do not consider simultaneous optimizing w.r.t. both the prediction model parameters and the instance weights. They usually do it separately, which makes the optimization procedure stop early in one iteration, losing the opportunity to continuously improve the model performance.

Our solution design and enable the alternative and simultaneous optimization thus leading to improved performance.

Embodiment

Scenario 1: use fresh pipe data to build the model for salt pipe network


Fact: fresh water pipes are well recorded in terms of both pipe attributes and associated historical failure events, while salt pipes

1


Page 02 of 5

are recently constructed such the event information is not enough to build a reliable predictive model

Observation:


Some of the fresh pipes are having similar working conditions like the salt pipes, e.g. similar water turbidity, similar soil conditions, they are most informative training samples
as more failures come for the salt pipe, it allows to find the fresh pipes having such similar event pattern...