Seagate factories amplify IIoT benefits through AI.

Seagate’s Global Factory Information Technology (GFIT) team set out to assess AI-based solutions to apply within the company’s factories with an emphasis on realizing value. Within the company’s manufacturing facilities, a diverse assortment of complex semiconductor wafer processing equipment contains a significant deployment of industrial internet of things (IIoT) devices. These generate an immense amount of data, creating a rich and complex dataset that is ideal for AI analysis and processing. This abundance of information provides fertile ground for AI systems to identify patterns, draw insights, and make predictions that would be nearly impossible for humans to achieve manually.

Semiconductor manufacturing is complex and multi-faceted. Reliance on stable global supply chains requires maintaining high-quality product delivered through best-in-class manufacturing equipment and process conditions. To ensure that equipment and process conditions deliver the stability required to meet customer specifications, control must be maintained through the monitoring of key process input variables (KPIVs) and IIoT sensors. KPIVs from time series sensor arrays record the conditions throughout the setup, processing, and post-processing of wafers within the semiconductor process equipment set.   

Many original equipment manufacturers (OEMs) provide comprehensive sensor monitoring on more modern equipment sets. In comparison, legacy platforms often require a retrofit of modern sensing capabilities to improve the data collection capability of the process conditions. Within the semiconductor IIoT space, most equipment-related faults occur within signals for any impending failure and are so diverse that labels to capture the entire domain are not practical. Within the Industry 4.0 and smart manufacturing space, the “three Vs of data” (velocity, volume, and variety) accurately describe the challenges that confront stakeholders. Modern hardware and software development workflows allow organizations to adequately overcome volume and velocity challenges. However, variety must be tackled using AI deployments of advanced algorithms.   
 
Sensors within, or connected to, manufacturing tools are essential for data acquisition, as they measure the physical conditions of machines. They gather data to inform operators and supervisors about the machines' status and performance during operations. These sensors are crucial for enhancing efficiency through automation in industrial production facilities. The sensors are connected to monitoring systems to provide real-time feedback and data during the manufacturing process.  
 
The team started this project because they knew their tools generated vast amounts of data that could inform process and tool health. However, engineers were using a small fraction of it because they had to manually create numerous charts with static control limits. Traditional statistical process control (SPC) has demonstrated limited success within the factory control space. However, with recent exponential growth in the volume and variety of IIoT data, and the complexity of the processes that demand increasingly stringent controls, these approaches have reached their limitations.   
 
The systems team knew that continuing with SPC solutions was an unsustainable way to control factory tools. They soon developed the concept of an autonomous monitoring system, able to monitor sensor-connected tools in the factory automatically, with integrated workflows to provide alerts on deviations. At its core, a Reconstruction Error-based Deep Learning 1-Dimensional Convolutional AutoEncoder effectively learns a golden tool fingerprint that represents the optimal process setup. Such learning allows future unseen samples to be accurately discriminated, with explainability, from the desired historical entitlement. It informs engineers to areas of interest through localized and globalized reconstruction errors that direct root cause analysis investigations that ultimately result in impactful process intervention. The team pays close attention to ensuring explainability in the derived results, as detecting an anomaly is only the initial step in the process. Identifying areas of potential optimization or correct action to the equipment and process engineering organizations is where realized value occurs.

The mission was to utilize sensor data to drive operational efficiencies and product quality. They aimed to install an unsupervised detection and containment system leveraging AI models to monitor process and tool health. This would allow the engineering organization to identify deviations from historical populations and enact a control plan to contain or correct excursions. The team deployed the system across multiple Seagate factory commodities, while continuing to identify upgrade opportunities in the framework. The AI framework was intended to assist process optimization by integrating key process output variables (KPOVs).

The GFIT team faced challenges in their initial steps to create their new autonomous monitoring system, including issues related to data accuracy and reliability, complexity and scalability, complex decision-making and real-time response, and situational awareness and context. 

• Data accuracy and reliability. Often, data collected at the required rate and frequency can have some issues. This can result from normal sensor degradation but, more frequently, it can be caused by network congestion and signal interference. Additionally, factories often have diverse machines and systems from different manufacturers that might not use consistent data formats. This can lead to complexity and possible issues when data is integrated. 
• Complexity and scalability. Expanding an autonomous monitoring system to cover a diverse set of sensors connected to a wide range of toolsets in multiple factories can present challenges, such as handling inconsistent data formats, managing integration complexities, and ensuring scalability.
• Intricate decision-making and real-time response. Autonomous monitoring systems often need to identify and act on abnormal events or behaviors. Determining what constitutes an anomaly and accurately detecting it in real time requires a comprehensive solution.
• Situational awareness and context. Advanced autonomous monitoring requires a contextual understanding of events combined with the ability to make decisions. This can be difficult for AI systems to achieve without substantial adaptation.
• Validation. Validating system reliability and that they meet business and operational goals requires significant time and resources. The subject matter experts (SMEs) often juggle validation work with other responsibilities, limiting the time they can dedicate to the process. 
• Consumption. In a large-scale manufacturing plant, the autonomous monitoring system generated thousands of alerts for potential failures that overwhelmed maintenance teams. The significant amount of outputs (predictions, alerts, metrics, logs, etc.) was also a challenge for the systems team. 

Solving these challenges required taking advantage of ongoing advancements in sensor technology, AI processing, and robust system and process frameworks to make the intended autonomous monitoring system as effective and scalable as possible.

The team carefully reviewed their previous efforts, strengths, and challenges, which led to the successful development and deployment of a new autonomous IIoT data monitoring platform. This AI-based system monitors Seagate’s factory equipment and processes, scanning the IIoT domain to identify areas needing “human-in-the-loop” validation. AI-generated inferences are contextualized and delivered to relevant SMEs for decision-making, effectively closing the feedback loop with actionable insights or additional data collection.   
 
“We’re developing and driving the technology, but the engineering partnership and continuous engagement provides valuable feedback that pushes us to continually improve the system,” said Mark Gorman, senior staff data scientist on Seagate’s GFIT team.    

The current AI platform consists of multiple integrated solutions. For the machine learning operations (MLOps) platform, it utilizes a combination of databases for system input/output and model training data. The main dashboard, which engineers use to interact with the system, is built with a JavaScript front end. Messaging is facilitated through a queueing mechanism, and data collection is accomplished via a local historian and a tool data layer (TDL) that implements the SEMI Equipment Communications Standard/Generic Equipment Model (SECS/GEM) protocol, a semiconductor equipment interface standard for equipment-to-host data communications.

Seagate’s use of unsupervised anomaly detection through deep learning automates the initial stages of grouping analysis and data dispositioning. By the time a human enters the decision-making loop, the analytics are already complete. This democratization of analytics empowers equipment and process engineering teams to leverage AI. Previously, only a select few with specialized skills could utilize these technologies, but now, real-time insights are delivered directly to SMEs with sufficient context to inform sound decisions.  

The new autonomous monitoring system has delivered numerous transformative benefits including lower tool cost of ownership (TCO), democratized analytics, and enhanced decision-making capabilities.  
 
In an environment with potentially millions of configuration combinations, managing such complexity would be overwhelming without AI. The autonomous monitoring system requires minimal upfront setup with a continuous online learning module that updates the factory deployment portfolio to the given factory process and product mix. Processing billions of data points per day, the system effectively consumes, processes, and drives effective actions within the engineering organization.   
 
Built on top of the real-time pipeline service is an AI enabled visualization application. Contextualizing AI anomalies with relevant factory context and historical process parameters, the application enables engineers to discriminate 100 times more data without the requirement for coding or data science skills. The application is based upon the premise of democratization in AI and analytics capabilities where users are empowered to effectively navigate the data space with confidence.   
 
By integrating unsupervised anomaly detection capabilities with human-in-the-loop engineering feedback, factory context, and historical processing parameters, engineers have unparalleled insights into their equipment operation.   
 
The Seagate factory and IT teams have realized several additional value-based benefits in the application of their new autonomous monitoring system. These include enhanced maintenance actions, improved tool performance, real-time insights, and reduced rework and scrap. They also include greater scalability and flexibility, adaptability to new data, and cost efficiency and resource optimization.  
 
Enhanced maintenance actions have reduced downtime and saved on manufacturing-related costs. Through sensor data pattern analysis, the ML models can predict equipment failures that can cause large yield events, allowing maintenance teams to perform repairs proactively. This reduces unplanned downtime and extends equipment lifespans. Consistent monitoring of tool sensor data helps minimize repair costs by catching issues early, avoiding costly unscheduled or emergency repairs or replacements.