17541453. LOCALIZING FAULTS IN MULTI-VARIATE TIME SERIES DATA simplified abstract (INTERNATIONAL BUSINESS MACHINES CORPORATION)

LOCALIZING FAULTS IN MULTI-VARIATE TIME SERIES DATA

Organization Name

INTERNATIONAL BUSINESS MACHINES CORPORATION

Inventor(s)

Joshua M. Rosenkranz of White Plains NY (US)

Pranita Sharad Dewan of New York NY (US)

Mudhakar Srivatsa of White Plains NY (US)

Praveen Jayachandran of Bangalore (IN)

Chander Govindarajan of Chennai (IN)

Priyanka Prakash Naik of Mumbai (IN)

Kavya Govindarajan of Chennai (IN)

LOCALIZING FAULTS IN MULTI-VARIATE TIME SERIES DATA - A simplified explanation of the abstract

This abstract first appeared for US patent application 17541453 titled 'LOCALIZING FAULTS IN MULTI-VARIATE TIME SERIES DATA

Simplified Explanation

An ensemble of autoencoder models is trained using different seeds.

• Autoencoder models are a type of neural network that can learn to reconstruct input data.
• Training the ensemble with different seeds helps capture different patterns in the data.

The trained ensemble is used to generate predictions for new time series data.

• Time series data refers to data that is collected over a period of time, such as stock prices or weather data.
• The ensemble of autoencoder models can analyze the new data and make predictions about future values.

Reconstruction errors are calculated for the predictions.

• Reconstruction errors measure how well the autoencoder models are able to reconstruct the input data.
• Higher reconstruction errors indicate that the models are struggling to accurately predict the data.

Dimensions with the highest reconstruction errors are selected based on a threshold.

• Dimensions refer to different variables or features in the time series data.
• By selecting dimensions with high reconstruction errors, the models can focus on areas where they are less accurate.

The predictions are segmented based on bursts of reconstruction errors over time.

• Segmentation involves dividing the predictions into smaller sections based on certain criteria.
• Burst of reconstruction errors may indicate periods of time where the models are struggling to make accurate predictions.

A common pattern among the selected dimensions across the temporal segments is identified as a failure fingerprint.

• A failure fingerprint refers to a consistent pattern of errors that can indicate a potential failure or anomaly in the data.
• By identifying these patterns, the models can help detect and predict failures in the time series data.

Potential applications of this technology:

• Predictive maintenance: The models can be used to identify potential failures in machinery or equipment based on patterns in the time series data.
• Financial forecasting: The ensemble of autoencoder models can help predict stock prices or market trends based on historical data.
• Anomaly detection: By analyzing reconstruction errors, the models can identify unusual or abnormal patterns in the time series data.

Problems solved by this technology:

• Accurate prediction: The ensemble of autoencoder models helps improve the accuracy of predictions for time series data.
• Early detection of failures: By identifying failure fingerprints, the models can help detect potential failures before they occur.
• Handling multiple dimensions: The models can analyze time series data with multiple variables or features, providing a more comprehensive analysis.

Benefits of this technology:

• Improved efficiency: The ensemble of autoencoder models can process large amounts of time series data quickly and accurately.
• Cost savings: By detecting failures early, companies can avoid costly repairs or downtime.
• Enhanced decision-making: The predictions and failure fingerprints provided by the models can help businesses make informed decisions and take proactive measures.

Original Abstract Submitted

An ensemble of autoencoder models can be trained using different seeds. The trained ensemble of autoencoder models can be run on new time series data to generate a prediction associated with the new time series data. The new time series data can include multiple dimensions per time step. Reconstruction errors can be determined for the prediction. Dimensions having highest reconstruction errors can be selected among the multiple dimensions based on a threshold. The prediction can be segmented based on bursts of the reconstruction errors over time, where temporal segments can be obtained. At least one common pattern including a set of dimensions among the selected dimensions across the temporal segments can be obtained to represent a failure fingerprint.