Some engineers are understandably skeptical about the promises surrounding artificial intelligence (AI), especially when it comes to deploying AI at the edge. Lucrative concepts and buzzwords like “real-time analytics” often fail to account for the complexities of integrating AI into an existing environment. 

While AI at the edge can deliver real value, many engineers know the path to adoption is riddled with challenges.

This article addresses those challenges and provides practical advice for engineers who want to take concrete steps toward AI deployment. 

The appeal of AI at the edge lies in its ability to process data locally, close to the point of generation. Decisions can be made in real time without the need to send data back to centralized cloud systems. The reduced latency enables immediate action and supports data privacy by keeping sensitive information on-site.

AI at the edge also offers an opportunity to improve operations through real-time analytics, predictive maintenance, and quality inspection. AI algorithms can detect anomalies or variations in production processes far earlier than traditional methods, allowing manufacturers to optimize performance and reduce waste.

In today’s competitive environment, organizations must look for opportunities to leverage AI to boost quality and productivity. And, the edge presents a critical intersection of data generation and processing, allowing faster, more secure, and cost-effective AI solutions.

Deploying AI at the edge has clear advantages, but in practice, there are also real challenges to overcome. The following graphic from Google reveals the vast array of infrastructure and data requirements. These complexities can slow down adoption and undermine the expected results.

Data is the backbone of any AI system. Incomplete data sets and poor data quality can lead to incorrect AI predictions and unreliable outcomes, negating AI’s benefits.

In a manufacturing environment, data comes from various sources—sensors, machines, IoT devices, and manual inputs. It’s a challenge to collect raw data from diverse sources and collate the information consistently to create a real-time, single source of truth.

Data verification is also necessary to ensure the data is clean and reliable. These processes involve removing duplicates, dealing with missing or inconsistent data, aligning data from different systems, and verifying data against multiple sources.

Implementing AI at the edge requires a robust infrastructure capable of handling complex computations. Hence, it is challenging to balance AI’s performance requirements with existing infrastructure and resource limitations.

Many manufacturing environments operate with a patchwork of older technologies, and integrating the latest AI tools into these systems can be tricky, to say the least. Engineers are tasked with ensuring that AI solutions are compatible with outdated hardware and software while also ensuring that the infrastructure is scalable to support future upgrades.

Managing CPU and GPU resources, memory, storage, and network bandwidth are also critical to ensure optimal AI performance without overloading existing systems. Many organizations struggle to scale these resources to meet increasing workloads and ensure availability while controlling costs.

The decentralized nature of many manufacturing operations—spread across multiple production lines, facilities, and even geographic regions—adds another layer of complexity, making it difficult for AI to deliver solutions across the enterprise.

Raw data is often complex and noisy. For AI models to work effectively, raw data must be transformed into meaningful and useful features that a machine learning model can use to make predictions or decisions. This process, known as feature extraction, is crucial for improving model accuracy.

Feature extraction often requires deep domain expertise to select the most significant variables and generate new ones through transformations. Contextualizing wafer geography is an example of feature extraction. Raw data provides X and Y coordinates, but based on domain expertise, the distance from the center is a more valuable feature for analysis. By transforming the raw X and Y coordinates into a distance measurement, critical relationships within the data are captured, improving the model’s performance and relevance.

In addition to specific measurements, it might also calculate cycle time and yields. The features should capture relevant information representative of the underlying processes without biases or oversimplifications that could distort the model’s predictions.

Handling high-dimensional data can make feature extraction difficult in complex manufacturing environments where multiple factors interact in ways that aren’t always obvious. Extracting the right features will make the difference between a high-performing AI model and one that fails to capture the nuances of reality—triggering false alarms or, worse, allowing quality escapes.

Despite its complexities, AI is a reality to embrace. For engineers looking to start their AI deployment journey, here are several best practices to simplify the process and facilitate successful adoption.

Engineers need to assess whether their current systems can support AI deployment on the edge, particularly in terms of data collection, storage, and computational resources. By evaluating infrastructure early on, potential barriers and bottlenecks can be addressed proactively.

Data collection and preprocessing are often the most time-consuming aspects of AI deployment. Engineers should focus on developing automated systems for data cleaning, validation, and transformation to reduce manual processes and improve data quality. This will ensure that data is ready for AI models to process without significant delays.

Once AI models are trained, integrating them into existing workflows is critical. Engineers need to ensure that AI models are deployed at the edge without disrupting current processes. APIs or microservices can be used to embed models into production systems. Seamless integration will lead to better accuracy and user adoption, as the AI system can continuously learn from new data and improve over time.

To mitigate risks, engineers might begin with a pilot project to validate the value of AI deployment. Once the system proves successful, it can be scaled across other parts of the operation. This incremental approach ensures that engineers can make adjustments before full-scale implementation, smoothing adoption at every stage.

AI models are not static—they require ongoing monitoring to detect performance drifts or deviations in accuracy. The models should also be adapted periodically to reflect new data and trends as production processes evolve.

Engineers should implement monitoring systems that track model performance and highlight anomalies. A human-in-the-loop (HITL) approach is also beneficial, allowing engineers to validate AI decisions and adjust models as necessary.

There’s no need to start from scratch. Leverage software platforms designed for AI deployment at the edge. Pre-configured solutions simplify the deployment process with a streamlined platform, offering data storage, rules engines, analytics, machine learning capabilities, and automated monitoring.

The most effective software solutions will integrate seamlessly into the organization. Because of this, it’s critical to evaluate how a new software solution will complement existing manufacturing and IT systems and processes to boost quality and productivity.

Deploying AI at the edge offers tremendous potential to enhance manufacturing processes, but the journey requires careful planning and execution. NI is a trusted partner in your AI journey, with a deep understanding of the challenges engineers face and the cutting-edge solutions needed to overcome them.