Reinforcement Learning for Manufacturing Due Date and Scheduling

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current manufacturing systems lack an effective method to simultaneously determine optimal due date quotations and production scheduling policies, which are mutually dependent, leading to inefficiencies and increased risks of past due deliveries.

Innovation Solution

A virtual environment is created to simulate factory production lines and order arrivals, where policies for due date quotations and scheduling are learned through reinforcement learning, updating both policies simultaneously based on feedback to improve target metrics such as tardiness, and then applied to the actual operation system for real-time decision-making.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If factory emulation is used to simulate production lines for due date quotation, then the accuracy of due date estimation is improved, but the computation time increases and it becomes difficult to generate a due date reply within a desired period

Engineering Contradiction:
Improvedue date estimation accuracyVSAvoidcomputation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary simulation runs during an offline learning phase to train a machine learning model. The simulation environment pre-computes various production scenarios and stores the results in a trained model, so that during online operation, due date quotations can be generated instantly without performing full simulations in real-time.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates a virtual copy of the factory production lines as a simulation environment. This digital twin allows the system to learn from simulated scenarios without affecting actual production, and the learned patterns are then applied to real due date quotations, avoiding the need to run full simulations for each new order.

Inventive Principle:
Principle #26Copying

2Loss of time

If production scheduling rules are fixed to estimate lead time using machine learning, then the computation time is reduced, but the ability to adapt to changing production conditions is limited

Engineering Contradiction:
Improvecomputation timeVSAvoidadaptability to production changes
Core Design Contradiction:
Loss of timeVSAdaptability or versatility

Solution Approach 1:

The system transitions from fixed scheduling rules to dynamic, learned policies. The machine learning model is trained offline to capture complex production patterns and relationships, enabling it to adapt to changing conditions without requiring real-time computation or reconfiguration of fixed rules. The model dynamically adjusts predictions based on the current state inputs.

Inventive Principle:
Principle #15Dynamics

3Productivity

If optimal production scheduling is searched given fixed due dates, then the scheduling efficiency is improved, but the due date quotation and scheduling become mutually dependent requiring simultaneous consideration

Engineering Contradiction:
Improvescheduling efficiencyVSAvoidpolicy coordination complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system performs preliminary joint optimization offline by training the machine learning model to simultaneously learn due date quotation and scheduling policies. The simulation environment pre-explores the joint search space of due dates and schedules, capturing their mutual dependencies in the trained model. During online operation, both due date quotations and scheduling decisions are generated together based on the pre-learned policies, avoiding the need for iterative coordination.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP3783547B1System and methods for reply date response and due date management in manufacturing
Publication Date: 2023.02.22 HITACHI LTD
  • EP3783547B1 patent drawingFigure 1
  • EP3783547B1 patent drawingFigure 2
  • EP3783547B1 patent drawingFigure 3

AI summary

Example implementations described herein involve methods and systems with one or more machines on a factory floor. Example implementations involve, in response to received orders, determining an initial scheduling policy for internal processes to meet the order and a due date policy for the order; a) executing a simulation involving scheduling decisions and due date quotations based on the initial scheduling policy and the due date policy; b) executing a machine learning process on the simulation results to update the scheduling policy 261 and the due date policy by evaluating the scheduling decisions and the due date quotations according to a scoring function which is common for evaluating the scheduling decisions and evaluating the due date quotations; iteratively executing a) and b) until a finalized scheduling policy and the due date policy is determined; and output the finalized scheduling policy and the due date policy in response to the order.