Request a Demo

Io-T based predictive maintenance for fleet management

By Horacio de La Fuente, Editor www.sharedmobility.news

During the 2nd International Conference on Emerging Data and Industry 4.0 (EDI40) April 29 – May 2, 2019, Leuven, Belgium, researchers from the University of Ottawa and Harbin University of Science and Technology, China, Patrick Killeen, Bo Ding, Iluju Kiringa and Tet Yeap presented an article regarding fleet management, then published in the scientific journal Procedia Computer Science, 151.

In the article researchers provide an IoT architecture designed to support Fleet Management. Here are some of the key issues.

1. Introduction

Internet of Things (IoT) is a new paradigm that is growing quickly. By 2020 many believe billions of devices will be connected to the internet. Some of the applications of IoT are smart farming, smart transport, smart health, smart cities, smart homes, and smart grids. Predictive maintenance, an example of smart transportation, attempts to predict the health of equipment using machine learning.

2. Background

IoT connects multiple devices, and the devices can sense and interact with the environment around them. IoT can be split into five layers: sensing, network, storage, learning, and application. The sensing layer gathers data from the environment and interacts with it using sensors and actuators. The network layer connects lower level nodes to the cloud/fog. A company named Libelium is working on providing global wireless sensors network coverage. The storage layer stores sensor data, aggregations, and other types of data. The learning layer performs data analytics on stored sensor data for knowledge discovery; for example, anomaly detection, or deviation detection, which attempts to detect when an instance deviates from its norm, can be performed in the learning layer. The application layer provides the interface to the IoT system by providing lower layer information access and control.

Predictive maintenance lowers costs by preventing failures, unscheduled maintenance, and downtime, and by ensuring the replacement of failing parts is done only when needed.

3. Predictive maintenance system architecture

The authors stress that the present work proposes an IoT architecture designed to support fleet management. It is divided into three layers (the perception layer, the middleware layer, and the application layer). The perception layer abstracts the fog and embedded systems. It performs sensing, lightweight storage, networking, and machine learning. It provides the interface to low-level nodes. The middleware layer abstracts the fog and the cloud and generally performs more heavy-duty storage, networking, and machine learning compared to the perception layer. It provides the interface to perception-layer nodes. The application layer is similar to the IoT application layer mentioned in section 2.

4. Semi-supervised sensor feature selection — ICOSMO

Researchers highlight that this work attempts to improve the sensor feature selection performed in COSMO, by using a semi- supervised machine learning approach, which is currently under development. A few definitions are necessary.

  1. a) Sensor Class: a J1939 sensor definition, defined as a PGN-SPN pair
  2. b) Sensor Instance: a physical J1939 sensor, which may or may not be installed on a vehicle
  3. c) COSMO Sensor: a sensor class that has been chosen as a selected feature in the unsupervised deviation detection model of the COSMO approach.

The algorithm proposed by the authors is named Improved Consensus self-organized models (ICOSMO) and makes the following assumptions: a) VSRDB, with repair records that describe details about faults that were repaired, is accessible; b) the repair records in the VSRDB are associated with true faults (not preventive repairs but instead reactive repairs); c) a document retrieval algorithm exists, which takes a mechanic’s repair record as a query, searches its indexed J1939 specification document, and estimates sensor classes involved in the failing/faulty components that the repair fixed (a black box document retrieval algorithm (BBDRA) is used in this work to simulate this assumption); and d) all buses in a fleet are of the same model, and each bus shares similar daily travel routes.

ICOSMO is designed in a data-driven fashion for conducting predictive maintenance, since mostly current and historical sensor data are accessible.

To verify the performance of ICOSMO, simulations and modeling will be conducted by generating data using the STO J1939 data dumps acquired from the MVP. The goal would be to design a fully-autonomous predictive maintenance system using fleet-wide and on-board data analytics.

5. Predictive maintenance system prototype

A minimally viable prototype (MVP) of the architecture mentioned in section has been implemented and is running in a live environment. The MVP does not have ICOSMO implemented yet, since ICOSMO is still currently under development.

The MVP is targeted for creating an IoT predictive maintenance fleet management system for the public transport buses of the Société de Transport de l’Outaouais (STO), Gatineau, Canada. Each bus will have a gateway installed, which reads sensor data and performs lightweight analytics. The goal is to discover novelties and to provide this information to the fleet managers to help them make better maintenance decisions. With each bus equipped with a gateway, fleet-wide data analysis will be possible. This enables the possibility of discovering some novelties that would not be obtainable when only monitoring individual vehicles.

Since deployment, the MVP allowed the authors to acquire J1939 data dumps daily. Approximately 1 GB of uncompressed J1939 data (200 MB when compressed) was acquired daily. Noteworthy sensor readings found in the data dump are listed below:

  • wheel speed information: axle angle and relative speed
  • vehicle distance information: trip distance, total distance, and remaining distance before running out of fuel
  • driver pedal positions
  • engine information: speed, torque, and temperature
  • oil information: temperature, pressure, and level
  • coolant information: temperature, pressure, and level
  • transmission fluid information: oil temperature and pressure

6. Conclusion

The researchers express that this work provides a novel IoT architecture for predictive maintenance. A semi-supervised machine learning approach is proposed for improving the sensor feature selection of the COSMO approach, which the authors name ICOSMO. A prototype of the architecture has been installed at a garage of the STO, Gatineau, Canada. Currently only a single bus is equipped with a gateway, but the plan is to expand the MVP and equip many more buses with gateways. This MVP is the foundation of a predictive maintenance machine learning and data analytic research. By using the J1939 data acquired from the buses, the researchers will train machine learning algorithms, and once complete these algorithms will be deployed to the MVP.

Future work will include completing the implementation of ICOSMO and running experiments.

Contacts Want to learn how Octo can transform your business?

We are happy to hear from you.
Discover our tailor-made solutions.

Get in touch
Become
a Contributor!
We’re always looking for interesting ideas and content to share within our community.
Get in touch and send your proposal to: press@octotelematics.com