How to Harness Applied AI in Industrial Manufacturing

Summary

Industry 4.0 has continued to evolve and grow its presence within the industrial automation community. As a result, the community faces considerable challenges due to the growth of Industrial Internet of Things (IIoT) devices and technologies. One of these technologies is artificial intelligence (AI). Applied AI has been around as a concept for many years in various fields, but industrial automation has been cautious to adopt the technology. This article will explore a brief history of applied AI and its usage within industrial automation.
 

History of applied AI

The industrial automation and manufacturing sectors are facing unprecedented levels of pressure. These pressures include increasing demand for products, large backlogs due to supply shortages from the COVID-19 pandemic and significant labor shortages due to lack of skillset ready to enter these markets.

Due to the difficulty of hiring and retaining talent and the skills gap shortage, businesses need to turn to alternatives to help ease the pressures they are facing. Many manufacturers turned to robotics to help solve the labor challenges with various levels of success.

A second problem that manufacturers are attempting to solve is the data problem. With computing and processing costs at the lowest levels ever, data from devices and equipment is more prevalent. Manufacturers are struggling on their digital journeys with how to consume the growing volumes of data they’re producing. Extracting insights to drive useful outcomes is not easy.

The solution for many manufacturers is applied AI. The impact of applied AI is promising in industrial automation and manufacturing. A major consulting firm has estimated that more than 30 hours of a typical work week would become automated by AI by the year 2030. However, properly using AI in automation and manufacturing environments requires first understanding some key concepts.
 

How NLP solves the data problem

Natural language processing (NLP), a field at the intersection of computer science and linguistics, has evolved significantly from its preliminary concepts in 17^th-century philosophy to its formal establishment with the dawn of computing in the 1940s. These early ideas laid the groundwork for machine translation and the first computational models of language.

The field has seen steady progression, with early rule-based (symbolic) methods being supplemented by statistical models and eventually overtaken by today’s advanced neural network approaches. Among the most transformative neural network architectures for NLP is the “transformer,” introduced in the seminal paper “Attention is All You Need” under the umbrella of Google’s research initiatives in 2017.

Transformers have laid the foundation for the diverse array of NLP-powered AI now being developed across the technology sector. From compact models designed for budget devices to enormous architectures operating on cutting-edge cloud computing resources, the scope and application of NLP models have never been broader. Rather than only being present in research fields, this is offered in forms that are relevant to industry adoption or hyper-specific domain-bound tasks.
 

Nondeterminism neural networks

Understanding the nondeterministic nature of transformer-based neural networks is fundamental to appreciating the challenges highlighted in this article. These models, which underpin contemporary advances in natural language AI, exhibit inherent stochastic behavior often seen within manufacturing. Not only does nondeterminism enable the generation of diverse and fluent responses across various domains, it also presents unique challenges in standardization and quality assurance when deploying these models into production environments.

Contrary to traditional software, where deterministic input-output relationships allow for standardized testing methods such as unit, integration, and system testing, transformer-based models defy these conventional practices due to their probabilistic outputs. The variability in responses means that results from traditional testing can no longer guarantee consistent model performance.

This inherent nondeterminism necessitates the development of novel testing and validation frameworks attuned to the probabilistic nature of these AI models. Organizations must adapt by implementing strategies like A/B testing, continuous monitoring and dynamic error analysis, which accommodate the variability of responses.

Moreover, product teams must be educated in stochastic model behavior, establishing realistic expectations and a deeper understanding of the tradeoffs associated with leveraging these powerful but unpredictable models in commercial applications.
 

Applied AI example in industrial automation

Building on the discussion of prompt injection techniques, such as retrieval-augmented generation (RAG), an experiment was conducted to evaluate their efficacy in improving model accuracy and reducing instances of generated hallucinations. The authors used the GPT-3.5-turbo model developed by OpenAI for this investigation, structuring our prompts to solicit specific information. The initial prompt was constructed as follows:

“What is the name of alarm 10041 on the PowerFlex 755T Variable Frequency Drive? Provide only the alarm name, no other text.”

To guide the model’s responses, we incorporated an exemplar user/assistant exchange:

User: “What is the name of alarm 10012 on the PowerFlex 755T Variable Frequency Drive? Provide only the alarm name, no other text.”

Assistant: “Brake Slipped - Drive Stopped”

Upon presenting this structured prompt to the model 10,000 times, we observed 6,934 unique responses, none of which were the correct answer “Precharge Open Alarm.” This result suggests an absence of the necessary data within the model’s training corpus. The most frequent responses are listed in Table 1.

Table 1: The initial GPT-3.5-turbo prompt yielded these responses.

To address this variation in response, the authors performed a second trial, introducing a JSON object into the system context, derived directly from the relevant source material:
{
 “Condition Type”: “Alarm 2”,
 “Condition Code”: “10041\n11041”,
 “Display Text”: “PrechargeOpenAlm”,
 “Full Text”: “Precharge Open Alarm”,
 “Fault”: “The internal precharge-circuity-bypass relay (for drives) or main contactor (for CBIs) was commanded to open while the drive was stopped (PWM was not active) due to low DC bus voltage.”,
 “Action”: “Investigate low DC bus voltage or the reason the drive entered precharge.”,
 “Fault Action”: “—”,
 “Configuration Parameter”: “0:37 [Prchrg Control]\n0:190 [DI Precharge]\n0:191 [DI Prchrg Seal]\n”,
 “Configurable Action”: “—”
}
 
With this additional context, a subsequent set of 10,000 queries yielded only 19 unique responses. The majority appropriately identified the alarm as shown in Table 2.

Table 2: Responses from the contextual prompt.
 
The introduction of contextual data significantly increased the frequency of the correct response. However, despite the narrowed range of responses, the nondeterminism of the model continued to produce slight variations in the output.
 

Final thoughts

Harnessing applied AI in industrial manufacturing requires properly understanding key concepts and using necessary contextual data in AI model training. There is much opportunity to use applied AI in manufacturing, but along with that opportunity come many challenges. We are in an exciting time to see how this evolves.

_{This feature originally appeared in the October 2024 issue of InTech digital magazine.}