General Electric Company

Ge Infrastructure Technology, LLC

Daniel Francis Holzhauer

Lalit Keshav Mestha

Cody Joe Bushey

Justin Varkey John

Masoud Abbaszadeh

Fernando Javier D'Amato

Christopher Eugene Long

Weizhong Yan

John William Carbone

Jonathan Carl Thatcher

Benjamin Edward Beckmann

Ramu Sharat Chandra

Dean Alexander Baker

Susan Peterson

Darrin Glen Kirchhof

Walter Charles Yund

Rohan Saraswat

Susan Ruth Peterson

Ratna Manedhar Punjala

Luis Blasini

Anilkumar Vadali

Cody Bushy

Fernando D'amato

Heterogeneous monitoring nodes may each generate a series of monitoring node values over time associated with operation of an industrial asset. An offline abnormal state detection model creation computer may receive the series of monitoring node values and perform a feature extraction process using a multi-modal, multi-disciplinary framework to generate an initial set of feature vectors. Then feature dimensionality reduction is performed to generate a selected feature vector subset. The model creation computer may derive digital models through a data-driven machine learning modeling method, based on input/output variables identified by domain experts or by learning from the data. The system may then automatically generate domain level features based on a difference between sensor measurements and digital model output. A decision boundary may then be automatically calculated and output for an abnormal state detection model based on the selected feature vector subset and the plurality of derived generated domain level features.

Feature extractions to model large-scale complex control systems

A verification platform may include a data connection to receive a stream of industrial asset cyber-attack detection algorithm data, including a subset of the industrial asset cyber-attack detection algorithm data. The verification platform may store the subset into a data store (the subset of industrial asset cyber-attack detection algorithm data being marked as invalid) and record a hash value associated with a compressed representation of the subset of industrial asset cyber-attack detection algorithm data combined with metadata in a secure, distributed ledger. The verification platform may then receive a transaction identifier from the secure, distributed ledger and mark the subset of industrial asset cyber-attack detection algorithm data in the data store as being valid after using the transaction identifier to verify that the recorded hash value matches a hash value of an independently created version of the compressed representation of the subset of industrial asset cyber-attack detection algorithm data combined with metadata.

Industrial asset cyber-attack detection algorithm verification using secure, distributed ledger

A verification platform may include a data connection to receive a stream of industrial asset data, including a subset of the industrial asset data, from industrial asset sensors. The verification platform may store the subset of industrial asset data into a data store, the subset of industrial asset data being marked as invalid, and record a hash value associated with a compressed representation of the subset of industrial asset data combined with metadata in a secure, distributed ledger (e.g., associated with blockchain technology). The verification platform may then receive a transaction identifier from the secure, distributed ledger and mark the subset of industrial asset data in the data store as being valid after using the transaction identifier to verify that the recorded hash value matches a hash value of an independently created version of the compressed representation of the subset of industrial asset data combined with metadata.

Industrial data verification using secure, distributed ledger

A cyber-physical system may have monitoring nodes that generate a series of current monitoring node values over time that represent current operation of the system. A hierarchical abnormality localization computer platform accesses a multi-level hierarchy of elements, and elements in a first level of the hierarchy are associated with elements in at least one lower level of the hierarchy and at least some elements may be associated with monitoring nodes. The computer platform may then determine, based on feature vectors and a decision boundary, an abnormality status for a first element in the highest level of the hierarchy. If the abnormality status indicates an abnormality, the computer platform may determine an abnormality status for elements, associated with the first element, in at least one level of the hierarchy lower than the level of the first element. These determinations may be repeated until an abnormality is localized to a monitoring node.

Scalable hierarchical abnormality localization in cyber-physical systems

According to some embodiments, a validation platform computer may interpret at least one received data packet to identify a control command for a controller of an industrial asset control system. The at least data packet being might be received, for example, from a network associated with a current operation of the industrial asset control system. The control command may then be introduced into an industrial asset simulation executing in parallel with the industrial asset control system. A simulated result of the control command from the industrial asset simulation may be validated, and, upon validation of the simulated result, it may be arranged for the control command to be provided to the controller of the industrial asset control system. Additionally, in some embodiments failed validation of a simulated result will prompt a threat-alert signal as well as prevent the command (e.g., data packet) from continuing to the controller.

Validation of control command in substantially real time for industrial asset control system threat detection

According to some embodiments, streams of monitoring node signal values may be received over time that represent a current operation of an industrial asset control system. A current operating mode of the industrial asset control system may be received and used to determine a current operating mode group from a set of potential operating mode groups. For each stream of monitoring node signal values, a current monitoring node feature vector may be determined. Based on the current operating mode group, an appropriate decision boundary may be selected for each monitoring node, the appropriate decision boundary separating a normal state from an abnormal state for that monitoring node in the current operating mode. Each generated current monitoring node feature vector may be compared with the selected corresponding appropriate decision boundary, and a threat alert signal may be automatically transmitted based on results of said comparisons.

Multi-mode boundary selection for threat detection in industrial asset control system

In some embodiments, a plurality of real-time monitoring node signal inputs receive streams of monitoring node signal values over time that represent a current operation of the industrial asset control system. A threat detection computer platform, coupled to the plurality of real-time monitoring node signal inputs, may receive the streams of monitoring node signal values and, for each stream of monitoring node signal values, generate a current monitoring node feature vector. The threat detection computer platform may then compare each generated current monitoring node feature vector with a corresponding decision boundary for that monitoring node, the decision boundary separating a normal state from an abnormal state for that monitoring node, and localize an origin of a threat to a particular monitoring node. The threat detection computer platform may then automatically transmit a threat alert signal based on results of said comparisons along with an indication of the particular monitoring node.

Threat detection and localization for monitoring nodes of an industrial asset control system

A threat detection model creation computer may receive a series of monitoring node values (representing normal and/or threatened operation of the industrial asset control system) and generate a set of normal feature vectors. The threat detection model creation computer may identify a first cluster and a second cluster in the set of feature vectors. The threat detection model creation computer may then automatically determine a plurality of cluster-based decision boundaries for a threat detection model. A first potential cluster-based decision boundary for the threat detection model may be automatically calculated based on the first cluster in the set of feature vectors. Similarly, the threat detection model creation computer may also automatically calculate a second potential cluster-based decision boundary for the threat detection model based on the second cluster in the set of feature vectors.

Cluster-based decision boundaries for threat detection in industrial asset control system

Operation of an industrial asset control system may be simulated or monitored under various operating conditions to generate a set of operating results. Subsets of the operating results may be used to calculate a normalization function for each of a plurality of operating conditions. Streams of monitoring node signal values over time may be received that represent a current operation of the industrial asset control system. A threat detection platform may then dynamically calculate normalized monitoring node signal values based at least in part on a normalization function in an operating mode database. For each stream of normalized monitoring node signal values, a current monitoring node feature vector may be generated and compared with a corresponding decision boundary for that monitoring node, the decision boundary separating normal and abnormal states for that monitoring node. A threat alert signal may then be automatically transmitted based on results of those comparisons.

Dynamic normalization of monitoring node data for threat detection in industrial asset control system

A system to protect a fleet of industrial assets may include a communication port to exchange information with a plurality of remote industrial assets. An industrial fleet protection system may receive information from the plurality of remote industrial assets or a cloud-based security platform and calculate, based on information received from multiple industrial assets, a current fleet-wide operation feature vector. The industrial fleet protection system may then compare the current fleet-wide operation feature vector with a fleet-wide decision boundary (e.g., separating normal from abnormal operation of the industrial fleet). The system may then automatically transmit a response (e.g., a cyber-attack threat alert or an adjustment to a decision boundary of an industrial asset) when a result of the comparison indicates abnormal operation of the industrial fleet.

Threat detection for a fleet of industrial assets

According to some embodiments, a threat detection computer platform may receive a plurality of real-time monitoring node signal values over time that represent a current operation of the industrial asset. For each stream of monitoring node signal values, the platform may generate a current monitoring node feature vector. The feature vector may also be estimated using a dynamic model output with that monitoring node signal values. The platform may then compare the feature vector with a corresponding decision boundary for that monitoring node, the decision boundary separating a normal state from an abnormal state for that monitoring node. The platform may detect that a particular monitoring node has passed the corresponding decision boundary and classify that particular monitoring node as being under attack. The platform may then automatically determine if the attack on that particular monitoring node is an independent attack or a dependent attack.

Automated attack localization and detection

According to some embodiments, a threat detection model creation computer may receive a series of normal monitoring node values (representing normal operation of the industrial asset control system) and generate a set of normal feature vectors. The threat detection model creation computer may also receive a series of threatened monitoring node values (representing a threatened operation of the industrial asset control system) and generate a set of threatened feature vectors. At least one potential decision boundary for a threat detection model may be calculated based on the set of normal feature vectors, the set of threatened feature vectors, and an initial algorithm parameter. A performance of the at least one potential decision boundary may be evaluated based on a performance metric. The initial algorithm parameter may then be tuned based on a result of the evaluation, and the at least one potential decision boundary may be re-calculated.

Feature and boundary tuning for threat detection in industrial asset control system

A normal space data source stores, for each of a plurality of threat nodes, a series of normal values that represent normal operation of an industrial asset control system, and a threatened space data source stores a series of threatened values. A model creation computer may generate sets of normal and threatened feature vectors. The computer may also calculate and output at least one decision boundary for a threat detection model based on the normal and threatened feature vectors. The plurality of threat nodes may then generate a series of current values from threat nodes that represent a current operation of the asset control system. A threat detection computer may receive the series of current values from threat nodes, generate a set of current feature vectors, execute the threat detection model, and transmit a threat alert signal based on the current feature vectors and at the least one decision boundary.

Growing community of inventors

Burnt Hills, NY, United States of America

Daniel Francis Holzhauer