Reinforcement Learning Greatly Improves Production Efficiency in Large Manufacturing Plants


Lenovo is a global leader in the consumer electronics industry and production of smart devices, and is the largest PC maker in the world. Lenovo consistently strives to provide the highest quality and best experiences to all of its customers, which include end consumer, commercial, and small to medium-sized businesses.

Each year, Lenovo provides global users with hundreds of millions of smart devices, including PCs, tablets, and smartphones. Lenovo accomplishes this intelligent transformation through technological innovations, not only in designing and building smart devices and infrastructures, but also in the daily operation and management of its massive supply chain and manufacturing system. 

Hefei Electronics Technology (LCFC), a subsidiary of Lenovo, produced more than 34 million PCs in 2020, accounting for nearly half of Lenovo’s PC production. Due to the large scale and complex production system at LCFC, production scheduling is a critical process to optimize manufacturing efficiency and throughput. The primary task for the planning team in LCFC is taking batches of orders twice a day and making corresponding production schedules for the manufacturing plants through long hours of intensive communication and coordination. Job orders are assigned to each production line with the intent to fully utilize the capacity of each machine.

Below is an overview of the challenging demands associated with the manufacturing scheduling process: 

• LCFC has four plants, containing 43 assembly lines in total.

• For each run of production scheduling, more than 6,000 production orders need to be scheduled, involving more than 200,000 laptops waiting for production.

• The products are categorized into more than 550 models with over 250,000 stock keeping

units (SKUs).

• A task corresponding to a specific model can only be executed on a portion of the total production line set.

• Orders with an urgent due date need to be scheduled preferentially to reduce the chance of a violation, which directly impacts the customer fulfillment rate and cost of manufacturing.

In order to improve performance of production scheduling, reduce the workload of the planning team, and respond to the scheduling demand more quickly, the Lenovo Research team and LCFC jointly developed an advanced production scheduling solution. Utilizing machine learning, artificial intelligence, and reinforcement learning, production scheduling in LCFC was modeled as a very large-scale combinatorial optimization problem with massive operational constraints. 

A deep reinforcement machine learning model based on an encoder-decoder architecture was used with improved representation ability added by using a multilayer forward convolution into the encoder and a masking mechanism that enforces the operational constraints to the output of the model. Through powerful self-learning and computing capabilities, the solution improved multiple production indicators and obtained complete scheduling results within minutes instead of the hours it previously took.

The new solution has significantly increased the production efficiency of LCFC, reformed the management process to reduce the backlog of production orders, and improved fulfillment rate, which has led to higher profitability and a better customer experience. Since 2019, this solution helped Lenovo increase revenue by more than 4.6 billion USD in total. 

Moreover, the analytics and operations research techniques in this project are not limited to just the production scheduling problem. This work is highly portable to other industrial planning and decision scenarios inside or outside of Lenovo, such as delivery routing and packing optimization, material requirements planning, and supply chain optimization.