Simulation-Trained Machine Learning for System Fault Localization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing debugging methods for systems, such as electronic devices and integrated circuits, are inefficient in identifying and correcting system-level faults due to the complexity and scale of error syndromes generated during workload execution.
Innovation Solution
A computer-implemented method using a machine learning model trained with simulation-based fault injection data sets to predict fault locations within a system, facilitating accurate fault identification and correction by analyzing error syndromes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional debugging methods are used to identify system-level faults, then debugging can be performed, but the process is inefficient due to the complexity and scale of error syndromes
Solution Approach 1:
The patent introduces a machine learning model as an intermediary between the error syndrome and the fault location. The model takes the complex error syndrome as input and outputs the predicted fault location, effectively mediating the debugging process and eliminating the need for manual analysis of complex error patterns
Solution Approach 2:
The patent creates a simulation model that copies the behavior of the actual system under test. By injecting faults into the simulation and recording the resulting error syndromes, the system builds a training dataset that replicates real-world fault scenarios without requiring actual system failures
2Measurement precision
If machine learning model is trained with comprehensive fault injection data, then fault prediction accuracy is improved, but data generation time and computational resources increase
Solution Approach 1:
The patent performs preliminary fault injection experiments to generate a comprehensive training dataset before the actual debugging process. By pre-collecting error syndromes from simulated faults across multiple locations and workloads, the system prepares the machine learning model in advance, so that during actual debugging, predictions can be made quickly without real-time data collection
Solution Approach 2:
The patent employs excessive fault injection by injecting faults at many different locations and with various workloads to ensure comprehensive coverage. This excessive sampling during the training phase creates a robust model that can accurately predict faults in production, trading off initial data generation time for improved prediction reliability
Data Source
AI summary
A process to train a machine learning model to predict fault location in a system includes generating a data set for training the machine learning model. The generating includes injecting, at a selected location of the system, a test fault into a simulation of the system using a workload, and recording a respective error syndrome generated by the simulation. Further, the generating includes repeating the injecting, at other selected location(s) of the system, of other test fault(s) into the simulation of the system, and the recording of respective, generated error syndromes. In addition, the process includes training, using the data set, the machine learning model, and providing the trained machine learning model for use in debugging the system, where the debugging includes predicting, using the trained machine learning model, a fault location within the system based on an error syndrome generated by the system due to the fault.