Selecting the best automatic machine learning to meet your manufacturing needs

Introduction

Machine learning (ML) has become a core technology in manufacturing, but it can be difficult to know which ML services and tools are best for your industrial operations. We will define and explain the use cases of when to use different Amazon Web Services (AWS) ML services.

In an age of rapid innovation, manufacturing industries are looking at different methods to reduce production disruption, reduce environmental impact, improve quality control, and reduce per-unit production cost to be more competitive.

One of the core technologies being adopted to achieve this goal is automated machine learning (ML). Automatic ML, known as AutoML, removes the tedious, iterative, and time-consuming work across the ML workflow, from data acquisition to model operationalization, so businesses can spend less time on low-level details and more time using ML to improve business outcomes. AutoML tools handle sourcing and preparing data, engineering features, training and tuning models, deploying models, and ongoing model monitoring and updating.

Though these AutoML services ease the model development, one significant challenge, which is often unnoticed and might lead to a longer time to develop a good ML model, is finding and using the right service. The challenge that IT / operational technology (OT) poses when they start the journey of using these AutoML capabilities is to answer these questions:

1. What are the primary purposes of these services?
2. Can these services be used for any use case, or are they designed for a specific scenario?
3. Which service should be used, and when should it be used?

AWS Machine Learning Services

AWS provides several automated ML services for industrial machine use cases. As a result, the manufacturing industry can develop ML models faster without deep ML expertise. Some AutoML services from AWS include:

Amazon Monitron

• Purpose: a complete system that uses ML to detect abnormal conditions in industrial equipment and facilitating predictive maintenance by capturing vibration and temperature data from equipment through wireless sensors.
• Good use cases: variety of rotating machinery in production lines and warehouses—for example, gearboxes, motors, pumps, compressors, and fans.
• Not good use cases: non-factory use cases, such as machinery used in outdoor power generation plants, equipment in offshore oil-and-gas stations, and consumer appliance.

Amazon Lookout for Equipment

• Purpose: ML industrial-equipment-monitoring service that detects abnormal equipment behavior so you can act and avoid unplanned downtime.
• Good use cases: industrial-process equipment that operates continuously and with low variability in operating conditions—for example, compressors, pumps, motors, turbines, boilers, heat exchangers, and inverters.
• Not good use cases: Lookout for Equipment might not be effective on highly variable equipment, such as construction equipment (cranes or trucks), vehicles, robots, or CNC machines.

Amazon Lookout for Metrics

• Purpose: ML used to detect and diagnose anomalies within business and operational data.
• Good use cases: business metrics related to customer engagement, operations, sales, and marketing in industries like retail, gaming, ad tech, and telecom.
• Not good use cases: multivariate data; faster processes that require detecting anomalies on higher frequency sensor data—for example, less than 5 minutes (Amazon Lookout for Equipment can only detect anomalies in a period of higher than 5 minutes); or anomalies that build over time, like metrics that degrade over weeks or months rather than a few hours of the day.

Amazon Lookout for Vision

• Purpose: an ML service that helps increase industrial production quality and reduce operational costs by identifying visual defects in objects.
• Good use cases: most common use case includes detecting damages to parts, identifying missing components.
• Not good use cases: a few limitations to keep in mind—Amazon Lookout for Vision supports image file sizes up to 5 MB and supports JPEG and PNG image formats only.

Amazon SageMaker

• Purpose: It’s used to build, train, and deploy ML models for nearly any use case with fully managed infrastructure, tools, and workflows.
• Good use cases: building custom models when no managed ML service fits the use case well. Amazon SageMaker provides ML engineers complete control when building custom models. They can pick the best algorithm, perform hyperparameter tuning, and perform retraining.

AWS Panorama

• Purpose: a collection of ML devices and a software development kit (SDK) that brings Computer Vision (CV) to on-premises internet protocol (IP) cameras to make automated predictions with high accuracy and low latency.

ML Services by Data Type

It’s important to evaluate the data variables and data types when selecting the most appropriate ML service.

Data variables include:

• Multivariate data: these data types involve two or more dependent variables resulting in a single outcome—for example, measurements taken by the temperature, pressure, and vibration sensors installed on a compressor to predict equipment failure, considering all these variables
• Univariate data: these are a single series of time-dependent variables—for example, energy consumption in kWh by the manufacturing equipment

There are three types of data to evaluate when selecting a ML service:

• Sensor data: time-series data that is collected as output from the sensor device receiving input from the physical environment—for example, weather sensors, manufacturing equipment, car-speed sensors, and more
• Business metrics: univariate data from different domains—for example, operational metrics, sales data, marketing data, and customer-engagement metrics
• Images: a picture composed of an array of pixels

The chart below will help you navigate the different criteria and select an appropriate service for a particular use case.

Summary

There is no one-size-fits-all when selecting ML services for your use case. However, there are simple ways to evaluate criteria to select the best ML service for your operation. A few guidelines to remember include:

• To carry out anomaly detection or prediction on the multivariate dataset, check out Amazon Lookout for Equipment.
• To carry out anomaly detection on a univariate dataset that does not involve anomalies built over time, check out Amazon Lookout for Metrics.
• For use cases requiring CV, check out Amazon Lookout for Vision and AWS Panorama.
• For other use cases that require custom model development and more control of the model—for example, hyperparameter tuning—check out Amazon SageMaker.