Hybrid neural network and autoencoder

One or more non-transitory machine-readable media comprising program code for denoising sensor data, the program code to:

• based on receiving sensor data, input at least a first subset of the sensor data into a deep neural network (DNN),
• generate a first set of reduced-noise sensor data based, at least in part, on the first subset of the sensor data and a physics-based model;
• evaluate the first set of reduced-noise sensor data based on a calculation of a value of a cost function, wherein the cost function incorporates the physics-based model;
• and
• generate a trained DNN from the DNN based on evaluation of the first set of reduced-noise sensor data.

An apparatus comprising:

• a processor;
• and
• a machine-readable medium having program code executable by the processor to cause the apparatus to,
• train a deep neural network (DNN) based on sensor data and a cost function, wherein the cost function incorporates a physics-based model;
• determine reduced-noise sensor data with the DNN during training of the DNN;
• pass the reduced-noise sensor data to an autoencoder;
• and
• train the autoencoder based on the reduced-noise sensor data to generate a trained autoencoder, wherein the autoencoder outputs a reconstructed representation of the reduced-noise sensor data.