Autoencoders are widely used to detect faults in manufacturing equipment. This article examines a MakinaRocks patent concerning autoencoder training.
[Patent]
Click the image to view the granted patent.
According to this patent, the sub model is an autoencoder.
Each of the multiple autoencoders is taught using training data obtained in different environments, such as time to create, manufacturing processes, and manufacturing recipes.
The submodels are trained sequentially rather than independently in parallel. After the training for sub model 1 is completed, the training for sub model 2 is carried out. In this case, at least some of the weights of sub model 1 are shared (or reused) when learning of sub model 2.
This learning leads to the next model without forgetting the knowledge learned from the previous model. Thus, the performance of the next model can be improved and the overall training time can be reduced. In other words, The key to this patent is to improve the performance of the model through transfer learning.
This article reflects the information available when it was published. Contact us to discuss your circumstances.