System and method for unsupervised anomaly prediction
Some embodiments are associated with a system and method for deep learning unsupervised anomaly prediction in Internet of Things (IoT) sensor networks or manufacturing execution systems. The system and method use an unsupervised predictive GAN model with multi-layer perceptrons (MLP) as generator and discriminator.
1. A computer-implemented unsupervised anomaly prediction system, comprising:
a data service configured to receive time series data from a plurality Internet of Things (IoT) sensors or manufacturing execution system and pre-process the time series data;
a model training and deployment service configured to receives and stores the pre-processed data from the data service in a distributed database and generate a risk score or a probability for next failure occurring within a time period, wherein the time period ranges from one day to thirty days; and
a user interface service configured to present the risk score or the probability for the next failure occurring within the time period in the IoT sensor networks or manufacturing execution system on a dashboard along with supporting evidence;
wherein the data service comprises a data pre-processing service configured to characterize the sensors' time series data by a state, wherein the states include at least one of the following: a failed state which characterizes situations that are problematic or inoperational; a warning state which characterizes situations that lead up to a failure; and a normal state which characterizes all other situations; and
wherein the data pre-processing service is configured to remove the time series data characterized as the failed state as part of a data preparation step.
2. The system of claim 1 , wherein the data pre-processing service uses a feature engineering script to prepare the sensors' time series data.
3. The system of claim 1 , wherein the model training and deployment service comprises a remaining useful life (RUL) calculator configured to calculate how much time is left before a next failure.
4. The system of claim 3 , wherein the RUL calculator employs an unsupervised predictive GAN model with multi-layer perceptrons (MLP) as generator and discriminator (MLP-MLP GAN).
5. The system of claim 4 , wherein the unsupervised predictive GAN model is trained with the pre-processed data in the normal or the warning state to generate the risk score or the probability of next failure occurring within the time period.
6. A computer-implemented method for unsupervised anomaly prediction, comprising:
receiving and pre-processing, via a data service, time series data from a plurality Internet of Things (IoT) sensors or manufacturing execution system, wherein the data service comprises a data pre-processing service;
characterizing, via the data pre-processing service, the sensors' time series data by a state, wherein the states include at least one of the following: a failed state which characterizes situations that are problematic or inoperational; a warning state which characterizes situations that lead up to a failure; and a normal state which characterizes all other situations; and
removing, via the data pre-processing service, the time series data characterized as the failed state as part of a data preparation step;
receiving and storing, via a model training and deployment service, the pre-processed data from the data service in a distributed database;
generating a risk score or a probability for next failure occurring within a time period, wherein the time period ranges from one day to thirty days; and
presenting, via a user interface service, the risk score or the probability for the next failure occurring within the time period in the IoT sensor networks or manufacturing execution system on a dashboard along with supporting evidence.
7. The method of claim 6 , wherein the data pre-processing service uses a feature engineering script to prepare the sensors' time series data.
8. The method of claim 6 , wherein the model training and deployment service comprises a remaining useful life (RUL) calculator configured to calculate how much time is left before a next failure.
9. The method of claim 8 , wherein the RUL calculator employs an unsupervised predictive GAN model with multi-layer perceptrons (MLP) as generator and discriminator (MLP-MLP GAN).
10. The method of claim 9 , wherein the unsupervised predictive GAN model is trained with the pre-processed data in the normal or the warning state to generate the risk score or the probability of next failure occurring within the time period.