US 12,140,917 B2
Training server and method for generating a predictive model for controlling an appliance
Francois Gervais, Lachine (CA)
Assigned to DISTECH CONTROLS INC., Brossard (CA)
Filed by Distech Controls Inc., Brossard (CA)
Filed on Mar. 7, 2018, as Appl. No. 15/914,610.
Prior Publication US 2019/0278242 A1, Sep. 12, 2019
Int. Cl. G05B 19/042 (2006.01); F24F 11/63 (2018.01); G06N 3/08 (2023.01); G06N 5/022 (2023.01); G06N 5/04 (2023.01)
CPC G05B 19/042 (2013.01) [F24F 11/63 (2018.01); G06N 3/08 (2013.01); G06N 5/022 (2013.01); G06N 5/04 (2013.01); G05B 2219/2614 (2013.01)] 18 Claims
OG exemplary drawing
 
1. A method for generating a predictive model of a neural network used for controlling an appliance, the method comprising:
storing in a memory of a training server a predictive model of a neural network;
storing in the memory a set of rules;
receiving by a processing unit of the training server at least one room characteristic, the at least one room characteristic comprising a type of activity performed by humans occupying the room;
receiving by the processing unit at least one current environmental characteristic value and at least one set point from an environment controller via a communication interface of the training server;
determining by the processing unit one or more commands for controlling an appliance based on the at least one current environmental characteristic value, the at least one set point and the at least one room characteristic;
transmitting by the processing unit the one or more commands for controlling the appliance to the environment controller via the communication interface;
receiving by the processing unit at least one updated environmental characteristic value from the environment controller via the communication interface;
determining by the processing unit a value of a reinforcement signal by applying the set of rules to inputs comprising the at least one current environmental characteristic value received from the environment controller before generation and transmission to the environment controller of the one or more commands, the at least one set point, the at least one updated environmental characteristic value received from the environment controller after transmission to the environment controller of the one or more commands and the at least one room characteristic, the application of the set of rules generating the value of the reinforcement signal, the value of the reinforcement signal being one of positive reinforcement or negative reinforcement; and
executing by the processing unit a neural network training engine implementing reinforcement training to update the predictive model of the neural network based on:
inputs comprising the at least one current environmental characteristic value, the at least one set point, and the at least one room characteristic;
one or more outputs consisting of the one or more commands; and
the value of the reinforcement signal.