MoonRaker Product Description

Purpose of this Document

This document is a summary of Project MoonRaker Technology.

MoonRaker has been independently developed by Lankester Engineering LLP for aggregation, analysis and data visualization of telemetry from industrial devices. MoonRaker is 100% owned by Lankester Engineering LLP.

Introduction

MoonRaker is an Artificial Intelligence (AI) based system for detecting, predicting and reporting performance within networks of Industrial Internet of Things (IIoT) devices.

MoonRaker was originally developed for BacNet-based systems that manage networks at universities, corporate campuses, network operations centers and airports. We have adapted MoonRaker to use Internet Protocol (IP) based networks and we have tested it with datasets from online industrial equipment supporting Telemetry.

MoonRaker consists of AI algorithms and a user-friendly web-based application for presenting results. We intend to develop an Application Programming Interface (API) for aggregating with other systems and a set of WebHooks notifications and alerts to enable rapid response to detected and predicted problems.

MoonRaker complements dashboard-based systems that only present instantaneous snapshots of current performance.

The MoonRaker generates historical and real-time performance indicators.

• Real-time indicators are actionable pieces of information that provide engineers with immediate and potentially urgent updates about the equipment that they maintain.
• Historical indicators focus on long-term preventative maintenance, such as predicting Remaining Useful Life.

MoonRaker includes three Artificial Intelligence based systems.

• Anomaly Detector
• Remaining Useful Life (RUL) Predictor
• State of Charge (SOC) Predictor

Each of these are described in the sections that follow.

Anomaly Detector

Automatically detects and predicts problems, proven on vast networks.

The Anomaly Detector is trained on massive amounts of data from a system on the assumption that the system works most of the time. The model recognises normal function, and flags abnormal function. This allows us to discriminate between the two. We have trained the Anomaly Detector on a proprietary dataset of IoT devices. It has been trained on 6 million Failure Rates and Average Latencies, our Autoencoder produces an intuitive classification of operation.

A scatter plot that discriminates between functioning and nonfunctioning (anomalies) devices.

Insights from the Anomaly detector can be used to inform crews on-the-ground to prevent damage and down-time.

Comparison with CAT Connect

A comparable system to the MoonRaker Anomaly Detector is the CAT Connect with CAT Remote Asset Monitoring. A case study illustrated here highlights its functionality. In this case study, an asset operating in the North West Territories of Canada was taken offline for a routine maintenance. When the device was brought back online, the telematics registered a 10% drop in Oil Pressure. A “watch” message was subsequently sent to a Fleet Advisor, after which it was determined to be a risk and taken offline. The fault was identified as a piece of debris in the lubrication system that damaged a pressure check valve. The damage warranted a replacement of the part. While CAT presents this story as a “job well done,” they do not address the massive 6-hour delay between the abnormal telematic and any action being taken.

A line chart showing a gap in functioning from CAT Connect Remote Monitoring.

The timestamp is grainy due to the image resolution, but we can clearly see 4:01 AM. At the bottom we can see the readings were abnormal 6 hours prior.

The chart below illustrates the Anomaly Detector working in a simulation of the event. The Anomaly Detector establishes the presence of an anomaly instantly, and is ready to report immediately. Clearly, if the MoonRaker Anomaly Detector had been operating in this system, loss and damage could have been prevented.

This is the Anomaly Detector running the simulation. Problem is discovered immediately.

Remaining Useful Life (RUL) Detector

Optimizes and extends device lifespan based on performance.

Our next algorithm is the RUL Predictor, which includes a capacity approximator. The RUL Predictor was trained on a dataset produced by NASA, comprising four lithium ion batteries. These batteries were charged, discharged, then tested for impedance in a temperature controlled environment. Trained on a set of telematics from a NASA experiment on lithium ion batteries, the RUL Predictor accurately predicts the End of Life of the test batteries within 10% up to 20 cycles away.

Remaining Useful Life Detector

The RUL Predictor predicts the RUL accurately, within just one cycle. This chart was produced in MoonRaker’s web-based application.

The RUL Predictor can decrease risk and improve utilization. In general, when a capacity drops to 80% of the initial value, the battery reaches the end of its service life and is subsequently taken offline. [Jin et al.] The RUL Predictor allows functional batteries to operate safer for longer.

In training the RUL Predictor and the SOC Predictor, we limited our information to exclusively vectors returned by a PQube controller: voltage, current, temperature and time. Since these measures are commonly available, our algorithms do not rely on esoteric information which might be difficult to obtain.

State of Charge (SOC) Predictor

Aggregates network-wide information on the state of all devices.

Our final model is the SOC Predictor. Currently, there is no good way to monitor the state of charge of large lithium ion batteries. We tried to answer a simple question: How long does this battery have left? The SOC Predictor seeks to solve that by estimating the trajectory of a battery, given its initial conditions. By observing the usage of the first 10% of a battery cycle, the model is able to predict the State of Charge until the battery is flat.

The SOC Predictor is able to predict the end of the cycle within 2% of the total length across all tests.

The SOC Predictor

With this, we can create prudent emergency insights for engineers on the ground, so they can take action sooner.

The MoonRaker System

MoonRaker is a system for detecting and predicting performance, plus presenting and communicating with responsible people and other systems. MoonRakers includes:

• Algorithms for analysing and aggregating content to find problems
• A web-based application for interactive viewing

We intend to develop:

• An API for developers that allows integration with other systems
• A set of Web Hooks that are used to form alerts and notifications

The API, or Application Program Interface, is a system for pulling from MoonRaker data, MoonRaker insights and other information sources. A Webhook, by contrast, is a way to push information to a subscriber.

An API like opening your postbox, and a Webhook like a knock at the door. For typical insights, MoonRaker will generate information, and consumers will ask the API to access them. Urgent, time-sensitive information, such as insights generated from the Anomaly Detector and emergencies predicted by the SOC Predictor, will come in by a Webhook and generate notifications to users.

There will be two types of information available from the API: detailed data, and at-a-glance insights. The former describes the full information learned by our models, while the at-a-glance information provides the immediate, most important information. For example, the RUL Predictor will return a complete, detailed capacity trajectory, or a single period of time indicating the remaining useful life of a battery.

MoonRaker Roadmap

MoonRaker’s algorithms are fully developed and tested. The web-based application provides user-friendly data visualization. MoonRaker has been tested with several existing large datasets.

The MoonRaker API and WebHooks for notifications are yet to be developed.

In it’s current form, MoonRaker adds value to telemetry systems that need to provide accurate, actionable data from time series collections of device updates.

MoonRaker also fills a gap by providing historical graphs, tables, and reports on high value equipment. These complement instantaneous dashboards that are limited to current data snapshots.

Going forward, we expect to invest in the API and WebHooks for linking MoonRaker to outside systems.

MoonRaker is an aggregator that produces insights. It is also a publisher of information that can be further aggregated by other systems. This makes MoonRaker especially valuable in systems where many types of heterogeneous components are integrated, and many stakeholders have different risks, opportunities and information requirements. Such systems are inevitable as more devices come online and make their performance updates available.

Footnote

[1] Jin, Siyu, et al. “Overview of Machine Learning Methods for Lithium-Ion Battery Remaining Useful Lifetime Prediction.” Electronics, vol. 10, no. 24, Dec. 2021, p. 3126, https://doi.org/10.3390/electronics10243126. Accessed 15 Mar. 2022.