Fusion of physics and AI based models for end-to-end data synthesization and validation

A system, comprising:

• a memory storing instructions;
• one or more communication interfaces;
• and
• one or more hardware processors coupled to the memory via the one or more communication interfaces, wherein the one or more hardware processors are configured by the instructions to:
• obtain a plurality of sensor data corresponding to one or more entities, wherein an entity of the one or more entities is one or more gears of a machine, and the plurality of sensor data corresponds to a type of fault of the entity, a nature of fault of the entity, a location of fault of the entity, and geometrical data of the fault of the entity;
• simulate, via a Finite Element Analysis (FEA) technique, the plurality of sensor data corresponding to the one or more entities to obtain a set of simulated data, wherein simulating the plurality of sensor data corresponding to the one or more entities is based on a real sensor dataset, wherein the real sensor dataset is a combination of input data and the plurality of sensor data, wherein simulation is performed using the FEA technique for healthy sensor data and faulty sensor data of a machinery;
• synthesize, via a variational autoencoder based generative model, the set of simulated data to obtain a set of synthesized data, wherein synthesizing the set of simulated data comprises synthesizing the healthy sensor data and the faulty sensor data comprised in the set of simulated data, and wherein the set of simulated data is synthesized by:
• generating a latent representation of the set of simulated data;
• and
• embedding one or more physical characteristics corresponding to the one or more entities into the generated latent representation and an associated loss function, wherein synthetically generated data from the variational autoencoder based generative model comprises the healthy sensor data and the faulty sensor data;
• and
• validate the set of synthesized data to obtain a set of validated synthesized data, wherein the set of synthesized data is validated by at least one of:
• performing a first comparison of (i) one or more statistical features comprised in the set of synthesized data and (ii) one or more corresponding statistical features obtained from experimental data to obtain a validation of the set of synthesized data;
• computing realistic bounds of the one or more statistical features for the synthesized data and discarding the one or more statistical features that are outside the realistic bounds, wherein the realistic bounds are based on the real sensor data and the FEA generated simulated data;
• and
• performing a second comparison of (i) performance of a classifier trained using the experimental data and (ii) performance of the classifier trained using the set of synthesized data to obtain a validation of the set of synthesized data, wherein the generated synthesized data is added in increasing amounts stepwise to a portion of the real sensor dataset;
• feed a portion of the simulated sensor data to the variational autoencoder based generative model that includes the set of synthesized data to generate both the healthy sensor data and the faulty sensor data;
• train a plurality of deep learning models using the healthy sensor data and the faulty sensor data, for synthetic machine signal analysis while keeping a quality of data high as the data generated by the FEA is very close to real scenario;
• and
• use the trained deep learning model for self-correction of the real sensor data which aids in better fault diagnosis of the machinery.

One or more non-transitory machine-readable information storage mediums comprising one or more instructions which when executed by one or more hardware processors cause:

• obtaining, a plurality of sensor data corresponding to one or more entities, wherein an entity of the one or more entities is one or more gears of a machine, and the plurality of sensor data corresponds to a type of fault of the entity, a nature of fault of the entity, a location of fault of the entity, and geometrical data of the fault of the entity;
• simulating, via a Finite Element Analysis (FEA) technique executed by the one or more hardware processors, the plurality of sensor data corresponding to the one or more entities to obtain a set of simulated data, wherein simulating the plurality of sensor data corresponding to the one or more entities is based on a real sensor dataset, wherein the real sensor dataset is a combination of input data and the plurality of sensor data, wherein simulation is performed using the FEA technique for healthy sensor data and faulty sensor data of a machinery;
• synthesizing, via a variational autoencoder based generative model executed by the one or more hardware processors, the set of simulated data to obtain a set of synthesized data, wherein synthesizing the set of simulated data comprises synthesizing the healthy data and the faulty sensor data comprised in the set of simulated data, wherein synthesizing the set of simulated data comprises synthesizing healthy sensor data and faulty sensor data comprised in the set of simulated data, and wherein the step of synthesizing by the variational autoencoder based generative model comprises:
• generating a latent representation of the set of simulated data;
• and
• embedding one or more physical characteristics corresponding to the one or more entities into the generated latent representation and an associated loss function, wherein synthetically generated data from the variational autoencoder based generative model comprises the healthy sensor data and the faulty sensor data;
• and
• validating the set of synthesized data to obtain a set of validated synthesized data, wherein the step of validating comprises at least one of:
• performing a first comparison of (i) one or more statistical features comprised in the set of synthesized data and (ii) one or more corresponding statistical features obtained from experimental data to obtain a validation of the set of synthesized data;
• computing realistic bounds of the one or more statistical features for the synthesized data and discarding the one or more statistical features that are outside the realistic bounds, wherein the realistic bounds are based on the real sensor data and the FEA generated simulated data;
• and
• performing a second comparison of (i) performance of a classifier trained using the experimental data and (ii) performance of the classifier trained using the set of synthesized data to obtain a validation of the set of synthesized data;
• feeding, a portion of the simulated sensor data to the variational autoencoder based generative model that includes the set of synthesized data to generate both the healthy sensor data and the faulty sensor data;
• training, a plurality of deep learning models using the healthy sensor data and the faulty sensor data, for synthetic machine signal analysis while keeping a quality of data high as the data generated by the FEA is very close to real scenario;
• and
• using the trained deep learning model for self-correction of the real sensor data which aids in better fault diagnosis of the machinery.