Similarly, for large rotating equipment, Huawei uses Pangu predictive models in conjunction with vibration and temperature data to move from a break-fix approach to predictive maintenance.
In cooperation with Changqing Oilfield, this method has improved the accuracy of identifying hidden risks in special operations to more than 94%.
“AI allows us to move from reactive response to proactive prevention, which is revolutionary for chemical safety,” said Charles.
Break down data silos
Industrial data is often fragmented across legacy systems, plants, and departments.
Huawei’s strategy is focused on piecing this together for AI.
For Charles, the approach is, and should be, “network-first, integration-based.”
“We are using advanced networking technology to connect production, office, and R&D data that was previously scattered,” he declares. “On the other hand, we offer an architecture based on cloud-edge collaboration. At the edge, we enable legacy devices to talk and collect real-time data. In the cloud, we manage this data through an integrated data lake and platform.”
“Only when data flows can the underlying models of AI have the fuel to truly deliver their value.”
Control AI securely and transparently
The entry of AI into core control loops clearly raises concerns about misjudgments and hallucinations in critical systems.
But Charles is keen to stress that there is no room for error.
“Our technological path is mechanics + AI,” Charles explains. “We will not completely hand over control to black box AI. Instead, we will integrate models of physical and chemical mechanisms accumulated over decades by industry with the fundamental models of AI.”
Charles details that this mechanistic model acts as a defined safety boundary within which the AI will work to identify the best solution without exceeding that limit.
Make the scale realistic and human
Many in the industry are skeptical about scaling AI beyond pilots.
In this regard, Charles argues that this barrier is finally being broken.
“This is a common challenge across industries, but we are already seeing breakthroughs in the chemical sector,” he says.
“The key to large-scale replication lies in the generalizability of the model and standardized architecture.”
He added that customizing models for each project used to be expensive, but now, thanks to Huawei’s basic model that uses a pre-training and fine-tuning paradigm, only minor fine-tuning using a small amount of data is required.
He cited refining and rubber testing as evidence.
