Topology-inspired neural network autoencoding for electronic system fault detection

1. A method for fault detection in a sensor network, the method comprising:

receiving sensor data from sensors in the sensor network with a communication device;

analyzing the sensor data to determine if the sensor data is indicative of a fault with a fault detection model, the fault detection model including:

predicting the sensor data with an autoencoder by encoding the sensor data and decoding encoded the sensor data;

autoregressively modelling the sensor data with an autoregressor;

combining the modeled sensor data and the predicted sensor data with a combiner to produce reconstructed sensor data;

augmenting the reconstructed sensor data using the autoregressor to account for a lack of synchronicity between sensor data from two or more time-series, the two or more time series being modeled by:

Y t =c+Σ i=1 p ρ i D t-i + ϵ t ,

where Y is reconstructed sensor data according to the autoregression, c is a constant, i is a data point with a time series of the sensor data, p is a total number of data points, ρ is a learned parameter, D is a vector of the sensor data, ϵ is a white noise vector, and t is a time;

comparing the reconstructed sensor data to the sensor data with an anomaly evaluator to determine anomalies;

producing an anomaly classification by comparing the anomalies to historical anomalies with an anomaly classifier; and

automatically mitigating faults in the sensor network with a processing device by at least one of shutting down or resetting a faulty portion of the sensor network based on the anomaly classification.

2. The method as recited in claim 1 , further including ranking the anomalies according to a difference between the reconstructed sensor data and the sensor data.

3. The method as recited in claim 1 , further including logging sensor data in an event log to form time-series of sensor data.

4. The method as recited in claim 1 , wherein encoding the sensor data and decoding the encoded sensor data includes generating hidden states with long short-term memory units in a recurrent neural network in the autoencoder.

5. The method as recited in claim 1 , wherein comparing the reconstructed sensor data to the sensor data includes determining a reconstruction error.

6. The method as recited in claim 5 , further including training the autoencoder and the autoregressor according to the reconstruction error corresponding to sensor data associated with normal operating behavior.

7. The method as recited in claim 1 , wherein producing the anomaly classification includes:

extracting a fingerprint of the anomalies; and

comparing the fingerprint to historical anomaly fingerprints.

8. The method as recited in claim 7 , further including extracting features from the fingerprint according to importance.

9. The method as recited in claim 1 , further including adding the anomalies and the mitigation to the historical anomalies.

10. A method for fault detection in a sensor network, the method comprising:

receiving sensor data from sensors in the sensor network with a communication device;

logging sensor data in an event log to form time-series of sensor data;

analyzing the sensor data to determine if the sensor data is indicative of a fault with a fault detection model, the fault detection model including:

predicting the sensor data with an autoencoder by encoding the sensor data and decoding encoded the sensor data;

autoregressively modelling the sensor data with an autoregressor;

combining the modeled sensor data and the predicted sensor data with a combiner to produce reconstructed sensor data;

augmentin the reconstructed sensor data using the autoregressor to account for a lack of synchronicity between sensor data from two or more time-series, the two or more time series being modeled by:

Y t =c+Σ i=1 p ρ i D t-i + ϵ t ,

where Y is reconstructed sensor data according to the autoregression, c is a constant, i is a data point with a time series of the sensor data, p is a total number of data points, ρ is a learned parameter, D is a vector of the sensor data, ϵ is a white noise vector and t is a time;

comparing the reconstructed sensor data to the sensor data with an anomaly evaluator to determine anomalies;

ranking the anomalies according to a difference between the reconstructed sensor data and the sensor data;

producing an anomaly classification by comparing the anomalies to historical anomalies with an anomaly classifier; and

automatically mitigating faults in the sensor network with a processing device by shutting down systems and equipment determined to be malfunctioning based on the anomaly classification.

11. The method as recited in claim 10 , wherein encoding the sensor data and decoding the encoded sensor data includes generating hidden states with long short-term memory units in a recurrent neural network in the autoencoder.

12. The method as recited in claim 10 , wherein comparing the reconstructed sensor data to the sensor data includes determining a reconstruction error.

13. The method as recited in claim 12 , further including training the autoencoder and the autoregressor according to the reconstruction error corresponding to sensor data associated with normal operating behavior.

14. The method as recited in claim 10 , wherein producing the anomaly classification includes:

extracting a fingerprint of the anomalies; and

comparing the fingerprint to historical anomaly fingerprints.

15. The method as recited in claim 14 , further including extracting features from the fingerprint according to importance.

16. The method as recited in claim 10 , further including adding the anomalies and the mitigation to the historical anomalies.

17. A sensor network with a fault detection system to detect faults, the system comprising:

a communication device that receives sensor data from sensors in the sensor network;

a fault detection model that analyzes the sensor data to determine if the sensor data is indicative of a fault, the fault detection model including:

an autoencoder that encodes the sensor data and decodes the encoded sensor data to predict the sensor data;

an autoregressor that autoregressively models the sensor data;

a combiner that combines the modeled sensor data and the predicted sensor data to produced reconstructed sensor data, wherein the reconstructed sensor data is augmented using the autoregressor to account for a lack of synchronicity between the sensor data from a time-series the time series being modeled by:

Y t =c+Σ i=1 p ρ i D t-i + ϵ t ,

where Y is reconstructed sensor data according to the autogregression, c is a constant, i is a data point with the time series of the sensor data, p is a total number of data points, ρ is a learned parameter, D is a vector of the sensor data, ϵ is a white noise vector, and t is a time;

an anomaly evaluator that compares the reconstructed sensor data to the sensor data to determine anomalies;

an anomaly classifier that compares the anomalies to historical anomalies and produces an anomaly classification; and

a processing device that automatically mitigates faults in the sensor network by at least one of shutting down or resetting devices determined to be faulty based on the anomaly classification.

18. The method as recited in claim 10 , further including an anomaly ranker that ranks the anomalies according to a difference between the reconstructed sensor data and the sensor data.

19. The method as recited in claim 10 , further including an event log that logs sensor data to form time-series of sensor data.

20. The method as recited in claim 10 , wherein the autoencoder includes long short-term memory units in a recurrent neural network that encodes the sensor data and decodes the encoded sensor data by generating hidden states.