Elevator Control Using Neural Network Probability Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing elevator control algorithms become complex and inefficient in managing multiple elevators in very tall buildings, especially when conventional rule-based methods struggle to optimize trip scheduling.
Innovation Solution
A computer-implemented method using an artificial neural network to assign new elevator calls to eligible cars, where eligibility is determined by sensor data and predefined rules, and the neural network provides probability values for each car to improve assignment accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If conventional rule-based control algorithms are used to manage multiple elevators in very tall buildings, then the control logic becomes very complex and difficult to optimize, but the system can still operate basic elevator functions
Solution Approach 1:
The patent replaces conventional rule-based control algorithms with an artificial neural network model. The neural network learns optimal trip scheduling strategies through reinforcement learning, substituting complex manual rule design with an adaptive machine learning system that automatically optimizes elevator dispatching without requiring explicit programming of control logic
Solution Approach 2:
The patent transforms the control approach by changing from fixed rule-based parameters to dynamic learned parameters. The neural network continuously adjusts its internal parameters (weights and biases) based on reinforcement learning feedback, enabling adaptive optimization of trip scheduling rather than relying on static pre-programmed rules
2Reliability
If an artificial neural network is used to make control decisions, then better control decisions can be made, but the system requires training using reinforcement learning which adds system complexity
Solution Approach 1:
The patent implements a training phase before actual operation where the neural network learns optimal control strategies through reinforcement learning in a simulated environment. This preliminary training action allows the system to pre-learn complex decision-making patterns, so that during actual operation the neural network can make accurate control decisions without requiring real-time complex computations or adjustments
3Adaptability or versatility
If the neural network outputs probability values for all cars, then comprehensive assignment options are available, but the accuracy is reduced because ineligible cars are not properly filtered
Solution Approach 1:
The patent segments the car selection process into two distinct stages: first, the neural network generates probability values for all cars based on current state inputs; second, a separate eligibility filtering mechanism divides cars into eligible and ineligible groups based on operational constraints. This segmentation allows the system to maintain comprehensive probability assessments while ensuring only qualified cars receive call assignments
Solution Approach 2:
The patent introduces an intermediary eligibility filtering mechanism that sits between the neural network output and the final call assignment. This intermediary layer evaluates operational constraints (car capacity, direction, floor position) and filters out ineligible cars before final assignment, ensuring that the neural network's comprehensive probability assessments are translated into accurate, constraint-compliant decisions
Data Source
AI summary
An elevator has a plurality of cars movable along different vertical axes between building floors and a sensor system providing sensor data indicative of the elevator current state. An elevator control method includes: receiving the sensor data including current position of each car relative to the floors, list of assigned current calls and a new call for assignment each indicating a destination floor; generating a list of eligible cars using the sensor data and at least one rule with which each car should comply when fulfilling the new call; inputting the sensor data as input data into an artificial neural network trained to convert the input data into one output value for each car indicating a probability and/or tendency for assignment of the car to the new call; determining one of the eligible cars as a selected car using the output values; assigning the new call to the selected car.