Using stochastic models to diagnose and predict complex system problems

• US 7,788,205 B2
• Filed: 05/12/2006
• Issued: 08/31/2010
• Est. Priority Date: 05/12/2006
• Status: Expired due to Fees

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Rahman, Javed, Chalasani, Nanchariah Raghuveera, Subramanian, Balan, Wesley, Ajamu A.
• Primary Examiner(s): HOLMES, MICHAEL B
• Application Number:
  US11/433,822
• Publication Number:
  US 20070265811A1
• Time in Patent Office: 1,572 Days
• Field of Search: 706/24
• US Class Current: 706/48
• CPC Class Codes: G06Q 10/04 Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem" market predictions or forecasting for commercial activities G06Q30/0202

Abstract

A plurality of stochastic models is built that predict the probabilities of state transitions for components in a complex system. The models are trained using output observations from the system at runtime. The overall state and health of the system can be determined at runtime by analyzing the distribution of current component states among the possible states. Subsequent to a low level component failure, the state transition probability stochastic model for the failed component can be analyzed by uncovering the previous states at N time intervals prior to the failure. The resulting state transition path for the component can be analyzed for the causes of the failure. Additionally, component failures resulting from the failure, or worsening state transition, in other components can be diagnosed by uncovering the previous states at the N times prior to the failure for multiple components in the system and then analyzing the state transition paths for correlations to the failed component. Additionally, transitions to worsening states can be predicted using an action matrix. The action matrix is created beforehand using state information and transition probabilities derived from a component'"'"'s stochastic model. The action matrix is populated probabilities of state transitions at a current state for given actions. At runtime, when an action is requested of a component, the probability of the component transitioning to a worsening state by performing the action can be assessed from the action matrix by using the current state of the component (available from the stochastic model).

Claims

1. A computer-implemented method for using stochastic models to assess component state information comprising:

building at least one stochastic model for determining a probability of a state transition between possible states for each of a plurality of components in a system;

receiving output data from the system, at least some of the output data being relevant to state transitions for at least some of the plurality of components in the system;

training each stochastic model for the plurality of components in a system with output data that is relevant for a respective component;

determining states of the plurality of components in the system from the at least one stochastic model; and

determining a state of the system from a distribution of the states for the plurality of components among the possible states.

4. The method recited in claim 1, wherein building at least one stochastic model for determining a probability of a state transition between possible states for each of a plurality of components in a system further comprises:

determining a plurality of internal states for each of the plurality of components in the system;

determining relevant output data for each of the plurality of components, said relevant output data to be generated by the system at runtime;

creating at least one initial stochastic model for each of the plurality of components in the system;

obtaining training data for the at least one initial stochastic model for each of the plurality of components, said training data comprising relevant output data for each of the plurality of components; and

training each of the at least one initial stochastic model for each of the plurality of components with the training data.

5. The method recited in claim 1, further comprising:

receiving a plurality of actions for a component;

receiving action response output data from the system in response to the component processing the plurality of actions;

training a stochastic model for the component with the action response output data;

determining probabilities of a state transition between possible states of the component from the stochastic model for each of the plurality of actions from the action response output data;

correlating the probabilities of a state transition between possible states of the component with the current state of the component and an action, for each of the plurality of actions; and

storing the probability correlations.

6. The method recited in claim 1, further comprising:

receiving a plurality of actions for the plurality of components;

receiving action response output data from the system in response to the plurality of components processing the plurality of actions;

training each stochastic model for the plurality of components in the system with action response output data;

determining probabilities of state transitions between possible states of the component from each stochastic model for the plurality of components for each of the plurality of actions from the action response output data;

correlating the probabilities of state transitions between the possible states of the plurality of components with the current state of the plurality of components and an action, for each of the plurality of actions; and

storing the probability correlations.

7. The method recited in claim 6, further comprising:

receiving a request for an action by a component running in the system;

assessing a current state of the component from a stochastic model for determining a probability of a state transition between possible states for the component; and

determining a probability of a state transition between the possible states of the component for the current state of the component and the requested action.

8. The method recited in claim 7, further comprising:

assessing a high probability of a transition to a worsening state for the component; and

invoking a preventative measure.

9. The method recited in claim 8, wherein the preventative measure comprises canceling the requested action.

10. The method recited in claim 8, wherein the preventative measure comprises notifying an operator.

11. The method recited in claim 1, further comprising:

receiving a component failure indication;

uncovering a path of state transitions from a stochastic model, said state transitions occurring at discrete times prior to the component failure; and

analyzing the path of state transitions for a cause to the component failure.

12. The method recited in claim 1, further comprising:

receiving a component failure indication;

uncovering a plurality of paths of state transitions from the at least one stochastic model for each of the plurality of components in the system, said state transitions occurring at discrete times prior to the component; and

analyzing the plurality of paths of state transitions for a correlations between state transitions of the plurality of components the component failure.

2. A computer usable tangible medium having computer readable instructions embodied therewith for using stochastic models to assess component state information, the computer readable instructions, when executed on a computer system, causing the computer system to perform the operations comprising:

building at least one stochastic model for determining a probability of a state transition between possible states for each of a plurality of components in a system;

receiving output data from the system, at least some of the output data being relevant to state transitions for at least some of the plurality of components in the system;

training each stochastic model for the plurality of components in a system with output data that is relevant for a respective component;

determining a states of the plurality of components in the system from the at least one stochastic model; and

determining a state of the system from a distribution of the states for the plurality of components among the possible states.

13. The computer usable tangible medium recited in claim 2, wherein building at least one stochastic model for determining a probability of a state transition between possible states for each of a plurality of components in a system further comprises:

determining a plurality of internal states for each of the plurality of components in the system;

determining relevant output data for each of the plurality of components, said relevant output data to be generated by the system at runtime;

building at least one initial stochastic model for each of the plurality of components in the system;

obtaining training data for the at least one initial stochastic model for each of the plurality of components, said training data comprising relevant output data for each of the plurality of components; and

training each of the at least one initial stochastic model for each of the plurality of components with the training data.

14. The computer usable tangible medium recited in claim 2, comprising:

receiving a plurality of actions for a component;

receiving action response output data from the system in response to the component processing the plurality of actions;

training a stochastic model for the component with the action response output data;

determining probabilities of a state transition between possible states of the component from the stochastic model for each of the plurality of actions from the action response output data;

correlating the probabilities of a state transition between possible states of the component with the current state of the component and an action, for each of the plurality of actions; and

storing the probability correlations.

15. The computer usable tangible medium recited in claim 2, further comprising:

receiving a plurality of actions for the plurality of components;

receiving action response output data from the system in response to the plurality of components processing the plurality of actions;

training each stochastic model for the plurality of components in the system with action response output data;

determining probabilities of state transitions between possible states of the component from each stochastic model for the plurality of components for each of the plurality of actions from the action response output data;

correlating the probabilities of state transitions between the possible states of the plurality of components with the current state of the plurality of components and an action, for each of the plurality of actions; and

storing the probability correlations.

16. The computer usable tangible medium recited in claim 15, further comprising:

receiving a request for an action by a component running in the system;

assessing a current state of the component from a stochastic model for determining a probability of a state transition between possible states for the component; and

determining a probability of a state transition between the possible states of the component for the current state of the component and the requested action.

17. The computer usable tangible medium recited in claim 16, further comprising:

assessing a high probability of a transition to a worsening state for the component; and

invoking a preventative measure.

18. The computer usable tangible medium recited in claim 17, wherein the preventative measure comprises canceling the requested action.

19. The computer usable tangible medium recited in claim 17, wherein the preventative measure comprises notifying an operator.

20. The computer usable tangible medium recited in claim 2, further comprising:

receiving a component failure indication;

uncovering a plurality of paths of state transitions from the at least one stochastic model for each of the plurality of components in the system, said state transitions occurring at discrete times prior to the component; and

analyzing the plurality of paths of state transitions for a correlations between state transitions of the plurality of components the component failure.

3. A computer hardware system for using stochastic models to assess component state information comprising:

a plurality of components;

a plurality of servers to run the plurality of components and to generate output data;

a monitoring engine to receive the output data, at least some of the output data being relevant to state transitions for at least some of the plurality of components;

a stochastic engine to build at least one stochastic model for determining a probability of a state transition between possible states for each of the plurality of components, to train each of the at least one stochastic model stochastic model for the plurality of components with output data that is relevant for a respective component and with the action response output data, and to determine states of the plurality of components from the at least one stochastic model; and

a matrix engine to receive a plurality of actions for a component and to receive action response output data from the system in response to the component processing the plurality of actions and pass the action response output data to the stochastic engine, to determine probabilities of state transitions between possible states of the component from each stochastic model for the plurality of components for each of the plurality of actions from the action response output data, and to correlate the probabilities of state transitions between the possible states of the plurality of components with the current state of the plurality of components and a particular action, for each of the plurality of actions.