EP-4254114-B1 - PREDICTIVE MODELING AND CONTROL SYSTEM FOR BUILDING EQUIPMENT WITH GENERATIVE ADVERSARIAL NETWORK

Inventors

• JIANG, ZHANHONG
• RISBECK, Michael J.
• LEE, YOUNG M.
• KULANDAI SAMY, Santle Camilus
• ZHANG, Chenlu

Dates

Publication Date

20260506

Application Date

20230330

Claims (15)

1. A method for predicting faults in building equipment and initiating responsive actions, the method comprising: - training a conditional generator (904) by operating a generative adversarial network (900) comprising the conditional generator; - generating, by the conditional generator, synthetic timeseries data corresponding to a plurality of fault types, wherein labels for the plurality of fault types are used as inputs to the conditional generator and correspond to the plurality of fault types which are desired to be represented in the synthetic timeseries data, wherein the conditional generator uses the labels to generate the synthetic timeseries data corresponding to the plurality of fault types indicated by the labels; - training a fault prediction model (808) using the synthetic timeseries data; - predicting a fault for building equipment by applying the fault prediction model to real timeseries data relating to the building equipment; and - initiating an automated action in response to predicting the fault for the building equipment.
2. The method of Claim 1, wherein operating the generative adversarial network further comprises creating, by an embedder, a representation of preprocessed training data having reduced dimensionality relative to the preprocessed training data, attempting, by a recovery, to reconstruct the preprocessed training data from the representation, and attempting, by a discriminator, to discriminate to determine whether the synthetic timeseries data is synthetic.
3. The method of Claim 2, comprising: - receiving, by the embedder and the generator, the preprocessed training data; - providing, by the generator, generated data to the discriminator; - providing, by the embedder, a first output to the discriminator; and - providing, by the embedder, a second output to the recovery.
4. The method of any of Claims 2-3, comprising enabling, by the embedder, learning of temporal dynamics from a latent space.
5. The method of any of Claims 1-4, comprising using, by the generative adversarial network, a reconstructed loss, a weakly supervised loss, and an unsupervised loss.
6. The method of any of Claims 1-5, wherein training the conditional generator is based on first actual timeseries data for a first unit of building equipment, the method further comprising updating the conditional generator for a second unit of building equipment by operating the generative adversarial network based on actual timeseries data for the second unit of building equipment.
7. The method of any of Claims 1-6, wherein training the fault prediction model using the synthetic timeseries data comprises ranking a plurality of sets of the synthetic timeseries data and selecting a highest ranked of the plurality of sets of the synthetic timeseries data for use in training the fault prediction model.
8. The method of Claim 7, wherein ranking the plurality of sets of the synthetic timeseries data comprises comparing the plurality of sets of the synthetic timeseries data to historical training data.
9. The method of any of Claims 1-8, wherein the automated action comprises altering an internal operation of the building equipment to correct, mitigate, or prevent the fault, and/or altering a load on the building equipment to mitigate or prevent the fault, and/or performing maintenance on the building equipment to mitigate or prevent the fault.
10. One or more non-transitory computer-readable media storing program instructions that, when executed by one or more processors, cause the one or more processors to execute operations comprising: - training a conditional generator (904) by operating a generative adversarial network (900) comprising the conditional generator; - generating, by the conditional generator, synthetic timeseries data corresponding to a plurality of fault types, wherein labels for the plurality of fault types are used as inputs to the conditional generator and correspond to the plurality of fault types which are desired to be represented in the synthetic timeseries data, wherein the conditional generator uses the labels to generate the synthetic timeseries data corresponding to the plurality of fault types indicated by the labels; - training a fault prediction model (808) using the synthetic timeseries data; and - predicting a fault for building equipment by applying the fault prediction model to real timeseries data relating to the building equipment, and - initiating an automated action in response to predicting the fault for the building equipment.
11. The one or more non-transitory computer-readable media of Claim 10, wherein the operations further comprise creating, by an embedder, a representation of preprocessed training data having reduced dimensionality relative to the preprocessed training data, attempting, by a recovery, to reconstruct the preprocessed training data from the representation, and attempting, by a discriminator, to discriminate to determine whether the synthetic timeseries data is synthetic.
12. The one or more non-transitory computer-readable media of Claim 11, the operations further comprising: - receiving, by the embedder and the generator, the preprocessed training data; - providing, by the generator, generated data to the discriminator; - providing, by the embedder, a first output to the discriminator; and - providing, by the embedder, a second output to the recovery.
13. The one or more non-transitory computer-readable media of any of Claims 10-12, wherein the operations further comprise updating the conditional generator for a new unit of building equipment by operating the generative adversarial network based on actual timeseries data for the new unit of building equipment.
14. The one or more non-transitory computer-readable media of any of Claims 10-13, wherein training the fault prediction model using the synthetic timeseries data comprises ranking a plurality of sets of the synthetic timeseries data and selecting a highest ranked of the plurality of sets of the synthetic timeseries data for use in training the fault prediction model; wherein ranking the plurality of sets of the synthetic timeseries data preferably comprises comparing the plurality of sets of the synthetic timeseries data to historical training data.
15. The one or more non-transitory computer-readable media of any of Claims 10-14, wherein the operations further comprise mitigating or preventing the fault by altering an internal operation of the building equipment, altering a load on the building equipment, or causing maintenance to be performed on the building equipment.

Description

The present disclosure relates generally to building management systems. The present disclosure relates more particularly to fault detection for connected equipment in a building management system. A building management system (BMS) is, in general, a system of devices configured to control, monitor, and manage equipment in or around a building or building area. A BMS can include, for example, a HVAC system, a security system, a lighting system, a fire alerting system, any other system that is capable of managing building functions or devices, or any combination thereof. Systems and devices in a BMS often generate temporal or time-series data that can be analyzed to determine the performance of the BMS and the various components thereof and/or predict future events such as faults, errors, malfunctions, etc. of the building equipment. For example, data can be examined and alert a user to repair the fault before it becomes more severe when the monitored system or process begins to degrade in performance, or to provide other advantageous technical benefits. However, many fault detection or prediction approaches are dependent on pre-existence of a robust set of historical data with multiple instances of different types of fault events. Such robust data is often not available in practice. SUMMARY One implementation of the present disclosure is a method according to claim 1 for predicting faults in building equipment and initiating responsive actions. The method includes training a conditional generator by operating a generative adversarial network including the conditional generator, generating, by the conditional generator, synthetic timeseries data corresponding to a plurality of fault types. Labels for the plurality of fault types are used as inputs to the conditional generator. The method also includes training a fault prediction model using the synthetic timeseries data, predicting a fault for building equipment by applying the fault prediction model to real timeseries data relating to the building equipment; and initiating an automated action in response to predicting the fault for the building equipment. In some embodiments, operating the generative adversarial network further includes creating, by an embedder, a representation of preprocessed training data having reduced dimensionality relative to the preprocessed training data, attempting, by a recovery, to reconstruct the preprocessed training data from the representation, and attempting, by a discriminator, to discriminate to determine whether the synthetic timeseries data is synthetic. In some embodiments , the method also includes receiving, by the embedder and the generator, the preprocessed training data, providing, by the generator, generated data to the discriminator, providing, by the embedder, a first output to the discriminator, providing, by the embedder, an a second output to the recovery. The method may include enabling, by the embedder, learning of temporal dynamics from a latent space. The method may include using, by the generative adversarial network, a reconstructed loss, a weakly supervised loss, and an unsupervised loss. In some embodiments, training the conditional generator is based on first actual timeseries data for a first unit of building equipment. The method may also include updating the conditional generator for a second unit of building equipment by operating the generative adversarial network based on actual timeseries data for the second unit of building equipment. In some embodiments, training the fault prediction model using the synthetic timeseries data includes ranking a plurality of sets of the synthetic timeseries data and selecting a highest ranked of the plurality of sets of the synthetic timeseries data for use in training the fault prediction model. In some embodiments, ranking the plurality of sets of the synthetic timeseries data includes comparing the plurality of sets of the synthetic timeseries data to historical training data. In some embodiments, the automated action includes altering an internal operation of the building equipment to correct, mitigate, or prevent the fault. In some embodiments , the automated action includes altering a load on the building equipment to mitigate or prevent the fault. In some embodiments, the automated action includes performing maintenance on the building equipment to mitigate or prevent the fault. Another implementation of the present disclosure is one or more non-transitory computer-readable media according to claim 10 which stores program instructions that, when executed by one or more processors, cause the one or more processors to execute operations. The operations include training a conditional generator by operating a generative adversarial network including the conditional generator and generating, by the conditional generator, synthetic timeseries data corresponding to a plurality of fault types. Labels for the plurality of fault types are used as inputs to the c