AWS Machine Learning Blog

Improve operational efficiency with integrated equipment monitoring with TensorIoT powered by AWS

Machine downtime has a dramatic impact on your operational efficiency. Unexpected machine downtime is even worse. Detecting industrial equipment issues at an early stage and using that data to inform proper maintenance can give your company a significant increase in operational efficiency.

Customers see value in detecting abnormal behavior in industrial equipment to improve maintenance lifecycles. However, implementing advanced maintenance approaches has multiple challenges. One major challenge is the plethora of data recorded from sensors and log information, as well as managing equipment and site metadata. These different forms of data may either be inaccessible or spread across disparate systems that can impede access and processing. After this data is consolidated, the next step is gaining insights to prioritize the most operationally efficient maintenance strategy.

A range of data processing tools exist today, but most require significant manual effort to implement or maintain, which acts as a barrier to use. Furthermore, managing advanced analytics such as machine learning (ML) requires either in-house or external data scientists to manage models for each type of equipment. This can lead to a high cost of implementation and can be daunting for operators that manage hundreds or thousands of sensors in a refinery or hundreds of turbines on a wind farm.

Real-time data capture and monitoring of your IoT assets with TensorIoT

TensorIoT, an AWS Advanced Consulting Partner, is no stranger to the difficulties companies face when looking to harness their data to improve their business practices. TensorIoT creates products and solutions to help companies benefit from the power of ML and IoT.

“Regardless of size or industry, companies are seeking to achieve greater situational awareness, gain actionable insight, and make more confident decisions,” says John Traynor, TensorIoT VP of Products.

For industrial customers, TensorIoT is adept at integrating sensors and machine data with AWS tools into a holistic system that keeps operators informed about the status of their equipment at all times. TensorIoT uses AWS IoT Greengrass with AWS IoT SiteWise and other AWS Cloud services to help clients collect data from both direct equipment measurements and add-on sensors through connected devices to measure factors such as humidity, temperature, pressure, power, and vibration, giving a holistic view of machine operation. To help businesses gain increased understanding of their data and processes, TensorIoT created SmartInsights, a product that incorporates data from multiple sources for analysis and visualization. Clear visualization tools combined with advanced analytics means that the assembled data is easy to understand and actionable for users. This is seen in the following screenshot, which shows the specific site where an anomaly occurred and a ranking based on production or process efficiency.

TensorIoT built the connectivity to get the data ingestion into Amazon Lookout for Equipment (an industrial equipment monitoring service that detects abnormal equipment behavior) for analysis, and then used SmartInsights as the visualization tool for users to act on the outcome. Whether an operational manager wants to visualize the health of the asset or provide an automated push notification sent to maintenance teams such as an alarm or Amazon Simple Notification Service (Amazon SNS) message, SmartInsights keeps industrial sites and factory floors operating at peak performance for even the most complex device hierarchies. Powered by AWS, TensorIoT helps companies rapidly and precisely detect equipment abnormalities, diagnose issues, and take immediate action to reduce expensive downtime.

Simplify machine learning with Amazon Lookout for Equipment

ML offers industrial companies the ability to automatically discover new insights from data that is being collected across systems and equipment types. In the past, however, industrial ML-enabled solutions such as equipment condition monitoring have been reserved for the most critical or expensive assets, due to the high cost of developing and managing the required models. Traditionally, a data scientist needed to go through dozens of steps to build an initial model for industrial equipment monitoring that can detect abnormal behavior. Amazon Lookout for Equipment automates these traditional data science steps to open up more opportunities for a broader set of equipment than ever before. Amazon Lookout for Equipment reduces the heavy lifting to create ML algorithms so you can take advantage of industrial equipment monitoring to identify anomalies, and gain new actionable insights that help you improve your operations and avoid downtime.

Historically, ML models can also be complex to manage due to changing or new operations. Amazon Lookout for Equipment is making it easier and faster to get feedback from the engineers closest to the equipment by enabling direct feedback and iteration of these models. That means that a maintenance engineer can prioritize which insights are the most important to detect based on current operations, such as process, signal, or equipment issues. Amazon Lookout for Equipment enables the engineer to label these events to continue to refine and prioritize so the insights stay relevant over the life of the asset.

Combining TensorIoT and Amazon Lookout for Equipment has never been easier

To delve deeper into how to visualize near real-time insights gained from Amazon Lookout for Equipment, let’s explore the process. It’s important to have historic and failure data so we can train the model to learn what patterns occur before failure. When trained, the model can create inferences about pending events from new, live data from that equipment. This, historically, is a time-consuming barrier to adoption because each piece of equipment requires separate training due to its unique operation and is solved through Amazon Lookout for Equipment and visualized by SmartInsights.

For our example, we start by identifying a suitable dataset where we have sensor and other operational data from a piece of equipment, as well as historic data about when the equipment has been operating outside of specifications or has failed, if available.

To demonstrate how to use Amazon Lookout for Equipment and visualize results in near real time in SmartInsights, we used a publicly available set of wind turbine data. Our dataset from the La Haute Borne wind farm spanned several hundred thousand rows and over 100 columns of data from a variety of sensors on the equipment. Data included the rotor speed, pitch angle, generator bearing temperatures, gearbox bearing temperatures, oil temperature, multiple power measurements, wind speed and direction, outdoor temperature, and more. The maximum, average, and other statistical characteristics were also stored for each data point.

The following table is a subset of the columns used in our analysis.

Variable_name Variable_long_name Unit_long_name
Turbine Wind_turbine_name
Time Date_time
Ba Pitch_angle deg
Cm Converter_torque Nm
Cosphi Power_factor
Db1t Generator_bearing_1_temperature deg_C
Db2t Generator_bearing_2_temperature deg_C
DCs Generator_converter_speed rpm
Ds Generator_speed rpm
Dst Generator_stator_temperature deg_C
Gb1t Gearbox_bearing_1_temperature deg_C
Gb2t Gearbox_bearing_2_temperature deg_C
Git Gearbox_inlet_temperature deg_C
Gost Gearbox_oil_sump_temperature deg_C
Na_c Nacelle_angle_corrected deg
Nf Grid_frequency Hz
Nu Grid_voltage V
Ot Outdoor_temperature deg_C
P Active_power kW
Pas Pitch_angle_setpoint
Q Reactive_power kVAr
Rbt Rotor_bearing_temperature deg_C
Rm Torque Nm
Rs Rotor_speed rpm
Rt Hub_temperature deg_C
S Apparent_power kVA
Va Vane_position deg
Va1 Vane_position_1 deg
Va2 Vane_position_2 deg
Wa Absolute_wind_direction deg
Wa_c Absolute_wind_direction_corrected deg
Ws Wind_speed m/s
Ws1 Wind_speed_1 m/s
Ws2 Wind_speed_2 m/s
Ya Nacelle_angle deg
Yt Nacelle_temperature deg_C

Using Amazon Lookout for Equipment consists of three stages: ingestion, training, and inference (or detection). After the model is trained with available historical data, inference can happen automatically on a selected time interval, such as every 5 minutes or 1 hour.

First, let’s look at the Amazon Lookout for Equipment side of the process. In this example, we trained using historic data and evaluated the model against 1 year of historic data. Based on these results, 148 of the 150 events were detected with an average forewarning time of 18 hours.

For each of the events, a diagnostic of the key contributing sensors is given to support evaluation of the root cause, as shown in the following screenshot.

SmartInsights provides visualization of data from each asset and incorporates the events from Amazon Lookout for Equipment. SmartInsights can then pair the original measurements with the anomalies identified by Amazon Lookout for Equipment using the common timestamp. This allows SmartInsights to show measurements and anomalies on a common timescale and gives the operator context to these events. In the following graphical representation, a green bar is overlaid on top of the anomalies. You can deep dive by evaluating the diagnostics against the asset to determine when and how to respond to the event.

With the wind turbine data that was used in our example, SmartInsights provided visual evidence of the events with forewarning based on results for Amazon Lookout for Equipment. In a production environment, the prediction could create a notification or alert to operating personnel or trigger a work order to be created in another application to dispatch personnel to take corrective action before failure.

SmartInsights supports triggering alerts in response to certain conditions. For example, you can configure SmartInsights to send a message to a Slack channel or send a text message. Because SmartInsights is built on AWS, the notification endpoint can be any destination supported by Amazon SNS. For example, the following view of SmartInsights on a mobile device contains a list of alerts that have been triggered within a certain time window, to which a SmartInsights user can subscribe.

The following architecture diagram shows how Amazon Lookout for Equipment is used with SmartInsights. For many applications, Amazon Lookout for Equipment provides an accelerated path to anomaly detection without the need to hire a data scientist and meet business return on investment.

Maximize uptime, increase safety, and improve machine efficiency

Condition-based maintenance is beneficial for your business on a multitude of levels:

  • Maximized uptime – When maintenance events are predicted, you decide the optimal scheduling to minimize the impact on your operational efficiency.
  • Increased safety – Condition-based maintenance ensures that your equipment remains in safe operating conditions, which protects your operators and your machinery by catching issues before they become problems.
  • Improved machine efficiency – As your machines undergo normal wear and tear, their efficiency decreases. Condition-based maintenance keeps your machines in optimal conditions and extends the lifespan of your equipment.

Conclusion

Even before the release of Amazon Lookout for Equipment, TensorIoT helped industrial manufacturers innovate their machinery through the implementation of modern architectures, sensors for legacy augmentation, and ML to make the newly acquired data intelligible and actionable. With Amazon Lookout for Equipment and TensorIoT solutions, TensorIoT helps make your assets even smarter.

To explore how you can use Amazon Lookout for Equipment with SmartInsights to more rapidly gain insight into pending equipment failures and reduce downtime, get in touch with TensorIoT via contact@tensoriot.com.

Details on how to start using Amazon Lookout for Equipment are available on the webpage.


About the Authors

Alicia Trent is a Worldwide Business Development Manager at Amazon Web Services. She has 15 years of experience in Technology across industrial sectors and is a graduate of the Georgia Institute of Technology, where she earned a BS degree in chemical and biomolecular engineering, and an MS degree in mechanical engineering.

Dastan Aitzhanov is a Solutions Architect in Applied AI with Amazon Web Services. He specializes in architecting and building scalable cloud-based platforms with an emphasis on Machine Learning, Internet of Things, and Big Data driven applications. When not working, he enjoys going camping, skiing, and just spending time in the great outdoors with his family.

Nicholas Burden is a Senior Technical Evangelist at TensorIoT, where he focuses on translating complex technical jargon into digestible information. He has over a decade of technical writing experience and a Master’s in Professional Writing from USC. Outside of work, he enjoys tending to an ever-growing collection of houseplants and spending time with pets and family.