Predictive Maintenance in the Internet of Things: Survival Analysis and Deep Learning on Time Series Data from a Large Scale Wireless Sensor Network

Abstract

A major challenge in deploying battery powered sensors for the Internet of Things (IoT) involves making accurate predictions for when the batteries in the sensors will reach the end of their life. Maintenance of large numbers of deployed sensors includes replacing the batteries, or in the case of inexpensive sensors, replacing the sensors themselves. In either case, the goal is to minimize downtime so that there are as few gaps in data transmission as possible. For smaller, consumer oriented deployments, the end user often is responsible for maintenance. However for larger, enterprise scale deployments a central organization is responsible for maintaining the sensors, along with analyzing the data the sensors generate on the client’s behalf. This organization may have many such clients, and therefore many deployments to maintain. Managing the maintenance of batteries in sensor networks can be costly due to a number of factors. In large enterprise networks particularly, the maintainer of the system will need to place bulk orders to the manufacturers ahead of time, well in advance of device failure, rather than waiting for individual sensors to stop reporting. In general, the number of individual sensors in the overall network will increase the expense relative to the per unit cost of replacing a sensor. However, sensors are often very cheap, and may even be disposable. The more important driver is the cost in personnel needing to travel to the locations of sensors in order to replace them. Therefore the number and spread of locations have a much higher impact on overall maintenance cost. As a result, it is of primary concern to the maintainer to be able to accurately gauge how many sensors will reach their end of life, and when. Knowing this information well in advance enables the maintainer to reduce maintenance costs by proactively addressing potential failures and reducing or removing the need for additional site visits. Additionally, accurate predictions allow bulk orders for new sensors to be placed with enough time for the manufacturer to build and ship the sensors. Importantly, if the maintainers of the network act preemptively to replace sensors before failure occurs, they are able to minimize downtime and therefore have greater continuity of sensor data. This project utilizes Deep Learning models in conjunction with Survival Analysis techniques in order to predict the time at which battery death of sensors will occur, based on their readings. Two Deep Learning models are developed and compared against a baseline Cox Proportional Hazards model. The first is a Deep variant of the baseline model, able to predict the probability of survival of a given sensor, at each time point, based on its features. The sequential, time-dependent nature of the sensor data motivates the development of a second, novel Deep Learning model that is able to perform iterative predictions of a sensor's probability of survival as it receives new readings from that sensor.