AWS Partner Network (APN) Blog

Automating Quality Machine Inspection Infused with Edge AI and Digital Twins for Device Monitoring

By Karunakaran Samuel, Director, Software Engineering, Architecture – Kyndryl
By Raju Karuppiah, Director, Chief Architect – Kyndryl
By Chandra Mohan Ravanan, Associate Director, Data Architecture – Kyndryl
By Mac Mahalingam, Sr. Partner Solutions Architect – AWS

Connect with Kyndryl-1

In the manufacturing domain, industry 4.0 machineries provide data points which help organizations improve operational effectiveness. Automating the quality process as and when the material is manufactured helps companies reduce waste in the manufacturing process while improving worker efficiency and optimizing shift coverage.

Combining the quality process with machine monitoring helps to reduce machine downtime, worker safety, and maintenance cost, and avoid rework when a defective part moves further in the production line.

With market analysis and research, Kyndryl discovered that 20-30% of defective parts are not detected at the point of occurrence where rework cost is very low. This cost goes higher as the defective part moves forward in the production line, and if it’s not identified in the factory then recalls and rework costs are 10x higher with a negative impact on customer satisfaction.

Detecting and correcting where and when defects happen has tremendous impact on product quality and reduces waste in the manufacturing process. Artificial intelligence (AI)-based quality inspection delivers 10x accuracy improvements in detecting defects with a reduction in time to inspect.

In this post, we will discuss an AI-based solution Kyndryl has built on Amazon Web Services (AWS) to detect pores on the welding process using acoustic data and a custom-built algorithm leveraging voltage data. We’ll describe how Kyndryl collaborated with AWS to design an end-to-end solution for detecting welding pores in a manufacturing plant using AWS analytics services and by enabling digital twins to monitor welding machines effectively.

Kyndryl is an AWS Premier Tier Services Partner and Managed Service Partner (MSP) that’s a global IT infrastructure services provider that’s relentlessly innovating to help customers with cloud-native transformation and make the journey seamless.

Porosity Detection

Industry 4.0 welding machines can provide the voltage, current, and audio details as the welding progresses for a given type of material. Kyndryl’s data and AI team created a custom-built algorithm using voltage streaming data to detect pores based on the voltage fluctuation, an unsupervised classification machine learning (ML) model using audio streaming data to detect pores.

Combining the results of both methods, Kyndryl identified the location and diameter of pores with accuracy rates of 90% and above. The detected pores are then compared against welding standard guidelines to decide whether the welding process met quality guidelines, requires re-welding for a specific length, or disqualifies the welding itself.

As the quality of welding is inspected during the welding process, the operator can make decisions immediately and thus help to avoid waste, rework at a later point in time, and even avoid product recalls.

Automating Quality Inspection of Welding Operations

In the manufacturing industry, a typical production line will have equipment and machineries used to transform raw materials into finished goods. This can range from essential hand tools to highly automated and computer-controlled robotic machines, which are connected to AWS as well as the edge node in a manufacturing plant.

Kyndryl’s solution flow consists of collecting acoustic data with voltage and current from welding machines, processing and inferencing data at the edge to detect welding pores while providing actionable insights to welding operators.

Additionally, data is streamed to the cloud to perform historical analysis and improve operational efficiency and product quality over time. A digital twin is enabled to monitor the welding operation in real-time with warnings created to proactively manage the asset when predefined thresholds are met.


Figure 1 – Logical flow of data pipeline.

Key components of the solution on AWS for edge analytics are:

  • MECAI: Managed Edge Computing and Analytics Infrastructure (MECAI) is Kyndryl intellectual property (IP) to manage devices at the edge.
    • Mapper: Script which connects to the edge device and manages data collection.
    • Edge store: Message Queuing Telemetry Transport (MQTT) queue stores the collected data and creates a pub-sub interface to support multiple consumers.
  • AWS IoT Greengrass edge manager agent: Deploys the AI components and manage it at the edge node.
  • Inference script: Wrapper script which calls the ML model to get the inference based on acoustic data and custom algorithm based on voltage data.
  • SQLite: Ephemeral data store keeps the detected information for a given duration and supports visualization of actionable insights.
  • Amazon Kinesis producer script: Pushes the raw data and inferred porosity details into Amazon Kinesis Data Firehose.
  • Grafana: Dashboard at the edge for welding operators with actionable insights of recent welding.

Historical analytics components are:

  • Amazon Kinesis Data Firehose: Sinks data into Amazon S3 buckets.
  • Amazon S3: Stores the raw and porosity details in a folder structured organized for user consumption through an Amazon Athena query engine.
  • AWS Glue Data Catalog: Crawler builds the Glue catalog of raw data and porosity details.
  • Amazon Athena: Query engine supports ad-hoc query access to raw and porosity details and enable an Amazon QuickSight dashboard.
  • Amazon QuickSight: Dashboard with various dimensions to analyze historical data and identify trends and process improvement opportunities.

Model lifecycle management components are:

  • Amazon SageMaker Notebook: Used for data preparation, building, training, and testing the acoustic ML classification model.
  • Amazon SageMaker edge manager: Compiles and packages the AI model for the edge node using Amazon SageMaker Neo. Alternatively, Open Neural Network Exchange (ONNX) runtime can be used to compile and package AI models for edge node deployment. The model is deployed as a component in AWS IoT Greengrass.
  • AWS IoT Greengrass: Deploys and manages the ML model component at the edge node with transport layer security (TLS) enabled.
  • Amazon CloudWatch: Monitors the metrics of all services including deployed models at the edge.

Real-time monitoring components are:

  • AWS IoT SiteWise and AWS IoT TwinMaker: The IoT sensor data from a welding machine is securely published to AWS IoT TwinMaker via AWS IoT SiteWise using certificates from AWS Certificate Manager and a configured threshold for each dataset to raise alerts when breached.
  • AWS Managed Grafana: Dashboard for AWS IoT TwinMaker built on AWS Managed Grafana monitors the welding machine in real-time.


Figure 2 – Porosity detection flow using AWS components.

Porosity Detection Process Flow

The porosity detection process starts with data collection from welding machine along with welding operator inputs like welding material type and length.

The MECAI wrapper helps connect the welding machine for data collection and manages the connectivity between the machine and edge node. It also handles the availability of the edge node and restoration of its last known good configuration in the event of failure. This is achieved by having a high available cloud instance synchronized continuously with the edge node.


Figure 3 – Porosity detection process flow.

Data collected from the machine is placed in the edge store for further consumption, and MECAI gives a flexibility to pre-process the collected data (if required) before publishing it for further consumption.

The collected raw data is streamed into the cloud using a Kinesis Producer Library (KPL) script with server-side encryption via Amazon Kinesis Data Firehose, and the data then sink into Amazon Simple Storage Service (Amazon S3). Data is organized device-wise, which allows organizations to scale up supported welding machines.

Inference scripts at the edge detect the porosity in the current welding by using a custom-built algorithm and AI model. The results are stored in SQLite and streamed to Amazon S3 via Kinesis Data Firehose for historical analysis.

A sample of negligible pores and large pores of cold metal transfer (CMT) welding process is shown below.


Figure 4 – Pores on actual welding.

Approximately 15,000 messages per minute will be generated from a welding machine with a message size range from 256-512 bytes. This high velocity data is collected on Amazon S3 and cataloged using AWS Glue Crawlers. The cataloged data is used for model building, re-training when there’s a data/model drift, and analytics report on various dimensions for business consumptions and operational improvement insights.


Figure 5 – Analytics dashboard for welding operation.

Amazon SageMaker Neo is used to optimize the trained machine learning model towards the edge node configurations and packaged for deployment. An acoustic ML model is then deployed in the edge node for inference using AWS IoT Greengrass.

A digital twin is enabled to monitor the welding machine in real-time with a threshold configured to raise alarms based on the data streamed from device.


Figure 5 – AWS IoT SiteWise and AWS IoT TwinMaker dashboard in Grafana.


Kyndryl’s solution increases accuracy by up to 10x in detecting the defects at the place of occurrence, as per industry standards. The solution also improves efficiency by up to 2x, reduces the impact on supply chain by avoiding further production line rework and recall, reduces carbon footprint and customer churn, and improves customer satisfaction. The metrics will of course vary based on customer use cases and scenarios.

With this custom-built algorithm and AI model, Kyndryl can detect the welding pores effectively with 90% and above accuracy, and suggest the next best action based on welding standards as and when welding is completed.

The solution can be deployed with minimal customization and is built for high scalability using serverless AWS components. It can also support any number of devices in production.


Kyndryl – AWS Partner Spotlight

Kyndryl is an AWS Premier Tier Services Partner and MSP that’s a global IT infrastructure services provider that serves 1000+ clients and relentlessly innovating to help customers manage SAP applications on AWS.

Contact Partner | Partner Overview | AWS Marketplace