Guidelines for Automating ML Model Deployment to Achieve Quality Production on the Manufacturing Floor

Introduction 

Customers often need to take immediate action based on data from plant equipment. Machine Learning (ML) models are utilized to gain insights into predictive maintenance, remaining useful life, and optimal recipe selection in process industries like tire manufacturing, pharmaceuticals, distilleries, etc.

Machine Learning at the Edge (ML@Edge) refers to running ML models on edge devices such as mobile or IoT devices. This enables the model to be triggered by the edge application. ML@Edge is crucial in situations where raw data requires action as close to its source as possible, especially when the data is collected from sources distant from the cloud.

In this blog, we will explain how to manage end-to-end ML model deployment on edge devices.

Amazon SageMaker for MLOps is an Amazon Web Services (AWS) solution designed to assist companies in automating and standardizing processes throughout the machine learning lifecycle.

Business problem 

To minimize waste and improve profitability while efficiently producing high-quality goods, manufacturers aim to detect anomalies early.

Delays in detecting issues on the manufacturing assembly line can lead to a waste of time and resources. Consequently, manufacturers often require real-time data insights from plant equipment to act swiftly. They are seeking an approach for streamlined deployment of machine learning models on edge devices.

Solution: Automate ML deployment

Before diving into the discussion about ML deployment, we'll start by exploring how to create a model easily, followed by the advantages of ML@Edge, Efficient Data Management in Edge AI Deployments, and use cases of ML@Edge.

Amazon Lookout for Equipment simplifies the creation of an optimized ML model using historical data from industrial equipment, requiring minimal ML knowledge or experience. Alternatively, you can utilize Amazon SageMaker to build and train ML models with fully managed infrastructure.

After successfully creating an optimized ML model, there are numerous advantages to deploying and running it locally on an edge device.

Advantages of ML@Edge

• Facilitates real-time decision-making for early issue detection and mitigation.
• Enables edge applications to operate offline or in areas with unreliable internet connectivity, as the ML model runs locally.
• Addresses data privacy concerns by keeping critical data within the local network, avoiding the need for cloud transfers.
• Offers faster response times since the ML model operates locally.

Data Management Strategies for Efficient Edge AI Deployments

Efficient management of data in edge AI deployments is crucial for ensuring processing efficiency, minimizing bandwidth usage, and upholding data security and privacy. Let's explore the significance of data management in edge deployments and how edge devices handle tasks such as data storage, synchronization, and security.

1. Data Pre-processing: Sensors and IoT devices frequently transmit noisy data to edge devices. Implementing methods like noise reduction, data cleansing, and normalization elevate data accuracy, leading to optimized bandwidth utilization and enhanced analysis efficiency thereafter.
2. Data Filtering: Edge devices performs preliminary data filtering to extract required information/events. This practice guarantees that only essential data is sent, thereby lowering network congestion and decreasing latency.
3. Data Summarization: At the edge, summarization techniques compact data sets into concise formats. These condensed versions can either be sent to the cloud for analysis or stored locally, resulting in decreased bandwidth needs.
4. Data Storage: Effective management of data storage is crucial for edge devices, considering their limited storage capacities compared to cloud servers. They need to manage storage for temporary or offline operation efficiently.
5. Data Synchronization: Edge devices synchronize data with either the cloud or local servers when they establish a connection. This is vital, especially when devices have limited network connectivity or operate offline.
6. Data Security: Security measures, including encryption, access controls, and secure protocols, are essential for safeguarding data during transmission and storage in edge deployments.
7. Data Privacy: Ensuring data confidentiality is paramount, particularly when dealing with sensitive or personal data. Edge devices adhere to privacy guidelines and apply techniques such as data anonymization and differential privacy to safeguard individual identities and uphold data privacy.

ML@Edge Use Cases  

Numerous use cases can benefit from ML@Edge, and the following are just a few examples.

1. IoT and Wearable Devices: The growth of IoT-connected devices is exponential, with Statista predicting 29.42 billion such devices globally by 2030. These devices, including smart watches, programmable lights, and security systems, can aid healthcare providers in remote patient monitoring. Edge processing ensures swift alerts for practitioners, allowing them more time for critical decisions.
2. Augmented Reality / Virtual Reality: Seamless experiences in virtual and augmented reality rely on minimal latency and robust data processing capabilities. Edge computing in AR/VR enables remote collaboration, immersive gaming, and realistic virtual environments.
3. Autonomous Vehicles: For self-driving cars, even a fraction of a second in latency can be catastrophic. Edge computing coupled with 5G connectivity is crucial for real-time sensor data analysis and rapid response to traffic changes.
4. Optimized Traffic Flow and Smart Cities: Edge computing optimizes traffic flow by analysing sensor data to adjust traffic lights accordingly. Smart city solutions can react to events in real time and allocate resources based on congestion patterns.
5. Predictive Maintenance: Edge computing and AI collaborate in predictive maintenance, where machinery with installed sensors transmits data to algorithms predicting maintenance needs before breakdowns occur.
6. Personalized Retail Experiences: AI and edge computing enhance personalized retail experiences by analysing shopper behaviour at the edge. This data drives in-store promotions and tailored product recommendations.

These examples illustrate the diverse use cases of ML@Edge, highlighting the importance of continuous training of ML and automating ML deployment to edge devices.

ML Deployment
Amazon SageMaker for MLOps simplifies the management of processes throughout the ML lifecycle, including the automation of ML deployment.

Here are the high-level steps to automate ML deployment:

1. Begin with a trained and optimized model, either created using Amazon Lookout for Equipment or developed with Amazon SageMaker.
2. Utilize SageMaker Model Registry to register the model with a new version pending approval. An engineer, responsible for managing the process flow, validates the model version and provides approval.
3. Upon approval, a new deployment will be triggered exporting the model to the ONNX format.
4. Publish this model to Edge devices for real time inference.
5. Inference logs can be utilized for retraining the model.

System Architecture:

In this blog, we've discussed:

1. The concept of ML@Edge, along with its advantages and applications.
2. How to use MLOps for automating processes throughout the ML lifecycle.

Article co-authors:

 Krishnasamy, Gurumoorthy