“Move fast and break things,” has been an anthem of Silicon Valley since the early 2000’s. Part philosophy, part mantra, it’s been used to motivate and accelerate software development, even at the expense of mistakes.
The mindset has led to great advances in enterprise apps, social media, gaming, and artificial intelligence (AI) – especially generative AI (GenAI), which has advanced at a breakneck pace since ChatGPT launched in late 2022. But as with anything moving too quickly, reliability can be among the first qualities to fail. In its short history, ChatGPT has been plagued by hallucinations (completely fabricated results), challenges with biases, apprehension around privacy, copyright concerns, and more.
Industrial AI is different
The booming realm of industrial AI demands more serious consideration. When AI malfunctions in industrial settings, from energy to transportation, the impact is real. Costly damage to industrial equipment, physical harm, detrimental exposure to cyber vulnerabilities, and loss of revenue and reputation, are just a few of the things that can come from “AI-gone-bad” in industry. In other words, the health and wellbeing of business and customers, as well as society, are at stake with industrial AI.
Such is the view of Hitachi. With deep roots in industrial equipment and operational technology, its history extends from the early days of electric motor development in Japan all the way to the advanced transformers and high-speed trains of today. The company is also steeped in digital innovation, with decades of research and development in data, infrastructure, and AI – with solutions deployed across every industry.
In fact, Hitachi is among a select few companies that can fully support the challenges and identify the opportunities for AI across every industry – from utilities reimagining the grid or navigating the energy transition, to manufacturers hungry for more automation and autonomous processes, or transportation companies yearning for more uptime and greater efficiencies.
Industrial customers count on Hitachi solutions and expertise to help them visualize possibilities, architect scalable solutions, and swiftly move from experimentation to production with a clear understanding of the desired outcomes, expected ROI, and challenges of the real world.
The pace of AI is undeniable and must be met with vigor to stay competitive. But the need for speed at any cost is not an option in industrial AI. Too much is at stake. For Hitachi, industrial AI is about moving fast and breaking nothing.
Pouring AI into industry
The criticality of industrial AI is clear. However, there’s more to integrating AI into industrial settings than what meets the eye.
In addition to a heightened focus on safety and reliability, successful industrial AI implementations demand things like support for small and heterogenous data, the incorporation of deep domain knowledge, explainability, alignment with physical and engineering constraints, and the ability to model system of systems, among others.
Complicating the environment further, data types found within industries are routinely heterogenous. They can include combinations of sensor data, event data, maintenance records, and operational data, not to mention acoustic, vision, and multimodal documents. Understanding these variables – per industry – and how they may or may not interconnect is crucial to successful industrial AI deployments.
A quick breakdown:
Small and heterogenous data. Industrial settings often involve data that is sparse, noisy, and highly heterogeneous, like unstructured text, sensor logs, video feeds, control signals, and time-series data. This information is typically collected through domain-specific protocols or legacy systems and is rarely curated for AI applications. Even when data is abundant, it may be inconsistently labeled or embedded in formats like operator notes, maintenance records, or audio transcripts.
Deep domain knowledge. Only experience and expertise in a particular field can generate deep domain knowledge. Distinct from understanding horizontal applications, learning the nuances and complexities of vertical industries takes hands-on experience gained over years. Without such industrial expertise, the AI that’s implemented will not live up to expectations; time will be wasted and revenue lost.
Explainability. Growing in importance in enterprise AI, explainability is critical in industrial AI. This boils down to the software clearly describing the function as well as the expected outcomes of machine learning models. It also provides digital breadcrumbs to explain how the model arrived at the final outcomes. A growing requirement from regulators, explainability drives trust in the models and the companies that use them.
Alignment with physical and engineering constraints. Safety and reliability is critical to industrial operations. To ensure both are addressed, we build models that respect the laws of physics and “understand” the engineering constraints that go into designing a system. Humans intuitively understand the former and are trained on the latter. But teaching models these constraints is still an open problem.
Modelling system of systems. Although no official definition exists for system of systems, the concept is widely known and leveraged. Essentially it describes the engineering of multiple decentralized management systems to monitor a variety of disjointed systems and predict behaviors in order to optimize performance and generate positive outcomes. In industrial environments, like rail transport, for example, one can imagine the value of intertwining predictive maintenance with ticketing and scheduling systems to maximize uptime and reliability.
Guiding your principles
After the surge of large language models (LLMs) in early 2023, researchers began picking up on a descending trend: more and more AI projects were stalling in the pilot phase. Peter Bendor-Samuel, CEO and founder of Everest Group, estimated at the time that 90% of GenAI pilots would not move into the production phase “any time soon,” and that some may languish there indefinitely.
Why? The reasons were myriad and spanned from pilot “fatigue” – too many pilots to manage – to an inability to identify a clear return on investment. In other words, rather than losing more time and money on their projects, organizations were abandoning them all together.
The trend was becoming almost as disruptive as the rush to LLMs in the first place. It compelled Hitachi to craft a short list of top-line guiding principles to help customers visualize their AI journey. The pragmatic approach encourages organizations to:
- Be outcome oriented, with stated goals and functions from the outset;
- Leverage purpose-built tools that are designed for your industry and your types of data;
- Ensure that your AI is responsible and reliable; in fact, the more responsible the AI, the more reliable the outcomes.
Solving the use case
The guiding principles serve the industrial AI space well, which, unlike enterprise AI, demands solutions designed for specific industries. Areas like analytics, maintenance and repair, operations optimization, quality assurance, and supply chain management sound horizontal in nature. But even a cursory comparison between, for example, an energy utility and a car manufacturer reveals a critical need for expertise and specialization.
An energy utility can take advantage of AI accelerators for things like critical event command center optimization, energy forecasting and consumption, and substation image analytics. An auto manufacturer can leverage accelerators for model-based yield prediction, dynamic scheduling, and inventory optimization. Industrial AI – it’s not for the faint of heart.
The impact
Although distinct, enterprise AI and industrial AI share a common denominator: improving business performance and customer experience. Likewise, the advent of new technologies like GenAI and agents are only accelerating experimentation and adoption, as companies find new ways to leverage new capabilities.
For example, GenAI is augmenting industrial AI in significant ways and transitioning us from Industrial AI 1.0 to Industrial AI 2.0. Consider a typical industrial value chain that comprises design and engineering, procurement and supply chain management, manufacturing and production, installation and commissioning, maintenance, customer support, and circularity.
In Industrial AI 1.0, the primary data sources were time-series, and event data and included some image data and manuals. During this era, we were focused on the problems of supply chain management, manufacturing quality and process optimization, and maintenance.
With GenAI, we saw the beginning of Industrial AI 2.0. For the first time, it enabled organizations to begin going upstream to aid in things like product and process design, and procurement; and go downstream with installation and commissioning as well as customer support. It’s also helping with OT code generation, simulated data generation for automation and robotics, and make metaverse applications much more meaningful. In fact, we believe that GenAI has far more potential to impact industrial operations.
Furthering advancing the Industrial AI 2.0 era will be the increased use of AI agents and agentic systems in industry. Although it will be a steeper climb than GenAI, due to the fact that the domain knowledge in the industrial field is not as easily accessible and the cost of making mistakes are much higher. But there’s no question, the development and deployment of agents across industry will only add to the ongoing transformation of industry.
Progress and Perfection
The benefits of industrial AI are clear. Whether integrating GenAI for industrial process transformation, building an agentic architecture to aid frontline workers, or leveraging accelerators to catapult the deployment of autonomous capabilities, the market is poised to skyrocket. In fact, manufacturing is one of the fastest-growing segments of the industrial AI space. The World Economic Forum estimates that the global AI manufacturing market will balloon from its 2023 value of $3.2 billion, to $20.8 billion by 2028.
The real challenge for industrial companies becomes determining with whom to partner for this AI journey. For Hitachi, nothing trumps roots in industrial equipment and OT and decades of AI and digital development. Ideal partners will have experience and expertise in both. Industrial AI is too critical for anything less.
For more information, visit AI Resource Center – Hitachi Digital.
Read More from This Article: Industrial AI: Move fast, break nothing
Source: News