Survey
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project
Extracting Software Failure Knowledge from Anomaly Databases at NASA Lucas Layman, Davide Falessi Fraunhofer Center for Experimental Software Engineering NASA OSMA Software Assurance Research Program © 2014 National Aeronautics and Space Administration 1 Overview We want to learn more about how software fails in NASA missions in order to improve software assurance This is a work in progress – Will provide some early analyses – The planned path forward with data mining and machine learning – Comments, concerns, criticisms are welcome How could this research help you? © 2014 National Aeronautics and Space Administration 2 Outline • Background and motivation • Current uses of anomaly information • Early analyses and data exploration • Machine learning © 2014 National Aeronautics and Space Administration 3 BACKGROUND AND MOTIVATION © 2014 National Aeronautics and Space Administration 4 Background: Project Lifecycle Program approval Pre- Phase A Concept Studies Phase A Concept Development Spacecraft launch Phase B Preliminary Design Phase C Final Design & Build Phase D Assembly, Integration, and Test © 2014 National Aeronautics and Space Administration 5 Phase E Deployment, Operations Background: Anomaly Definition !!! Anomaly: A deviation from expected, nominal behavior in the Operations phase Anomaly report: Artifact documenting the anomaly occurrence and investigation © 2014 National Aeronautics and Space Administration 6 Background: Anomaly Reports • Date/time of event • Criticality • Description • Investigation • Cause • Corrective Action Anomaly reports are a rich source of information on how software fails in operation © 2014 National Aeronautics and Space Administration 7 Anomalies are Institutional Memory Goal: To learn across missions and centers to improve software development http://phonedirectory.ksc.nasa.gov/assets/docs/NASA_centers1_700.jpg © 2014 National Aeronautics and Space Administration 8 HOW ARE ANOMALIES REPORTS USED TODAY? © 2014 National Aeronautics and Space Administration 9 Current uses – tracking and reporting Current Status Submitted Project 0 Age of open anomalies # Mission Impact Open anomalies < 30 days 3 Critical 0 20 Major 2 Substantial 5 Minor 3 Open 23 30 days – 90 days Rejected 20 > 90 days Closed 0 562 Mishaps No Effect 13 Index Anomaly Date Days Mission Impact Mishap Open 0595 08/13/2011 Major 13 Title Y 338 Reaction Wheel 3 Dragged down to zero 0686 03/22/2012 Substantial Y DOY 2012-082 Processor Reset / Sun Safe 116 Incident 0651 11/22/2011 Substantial N 237 Overlapping LROC File Handle Assignments 0095 08/06/2009 Substantial 0674 02/25/2012 Major Plethora of Open Files Prevents New File 1075 from Opening, Causes LROC Handle Errors N © 2014 National Aeronautics and Space Administration N 10 142 OPSMPS1 Database Corruption Current uses - Investigating “How many times has this star tracker problem caused an anomaly?” “What problems can we expect from the Microwave Sounding Unit based on previous missions?” “How many of these erroneous tracks are due to an iridium flash vs. an error in the software? © 2014 National Aeronautics and Space Administration 11 Problems in the current usage Anomaly data is rarely shared across centers or missions Anomaly reports tend to be “write once, read never” Meta-data is unreliable © 2014 National Aeronautics and Space Administration 12 EXPLORING SOFTWARE ANOMALY DATA © 2014 National Aeronautics and Space Administration 13 Analysis approach Finding software-related anomalies is not trivial 1. Search for search for “software”ish text strings) 2. Manually root out false positives – False negatives (estimated %) 3. Label the anomaly with some tag we wish to count Anomalies Subset analyzed Software anom. Missions Launch dates Software related* JPL 13696 4621 1031 9 ‘89-’05 257 GSFC 25320 9921 2328 29 ‘97-’10 634+ Location © 2014 National Aeronautics and Space Administration 14 How severe are the software anomalies? Mission Impact of Software Anomalies 70% 60% 50% 40% Project 30% Astrophysics All projects 20% 10% 0% Critical Major Substantial Minor No Effect © 2014 National Aeronautics and Space Administration 15 How are we fixing software anomalies? Software Anomaly Repair Actions 60% 50% 40% 30% Project 20% Astrophysics 10% All projects 0% © 2014 National Aeronautics and Space Administration 16 What is the software anomaly growth rate? 50% © 2014 National Aeronautics and Space Administration 17 Data drawn from anomaly samples of Goddard spaceflight projects WHAT CAN WE LEARN ACROSS MISSIONS? © 2014 National Aeronautics and Space Administration 18 “Why do I need Software Assurance?” JPL anomaly risk rating distributions All % caused by anomalies software Risk rating 1 - Unacceptable Risk 5.7% 26.8% 2 - Accepted Risk 34.4% 25.2% 3 - No Significant Risk 53.7% 34.7% 4 - No Risk 6.2% 25.1% Total 100% 30.4% Taken from a sample of JPL missions between 1989 and 2007. Sample size > 1000. © 2014 National Aeronautics and Space Administration 19 Where are the software anomalies? Flight vs. Ground Software Anomalies 100% 90% 80% 70% 60% 50% 40% 30% 20% 10% 0% Astrophysics Heliophysics Earth Planetary Flight software Ground software © 2014 National Aeronautics and Space Administration 20 ONGOING RESEARCH © 2014 National Aeronautics and Space Administration 21 Our goal: Lower the price of knowledge To do this analysis, you must know which anomalies below to a category, e.g. – Software cause – Attitude control system – Cruise phase Currently, determining these categories is an intense manual process Research Objective: Automatically label anomalies with high accuracy to enable anomaly “analytics” © 2014 National Aeronautics and Space Administration 22 The possibilities: Predicting causes and fixes !!! System Power subsystem Mission impact Low Likely cause Software logic error Likely corrective actions Procedure restart Responsible party L. Layman © 2014 National Aeronautics and Space Administration 23 Our automatic labeling approach Compute similarity using Natural Language Processing techniques 0.45 Compute NLP similarity score 0.72 0.93 Root cause: Operations error Root cause: Operations error Corrective action: Software fix Corrective action: Software fix © 2014 National Aeronautics and Space Administration 24 Finding the best NLP technique Input Variables Title Class Variables Description Root Cause Corrective Action 480 anomalies Class(X) ?= Class(most similar) NLP technique % correct Kappa LSA BINARY COSINE Porter stemmer 76.3 0.49 LSA RAW COSINE Porter stemmer 76.0 0.49 LSA TF_IDF COSINE Stanford (noun and verb)) 75.0 0.46 © 2014 National Aeronautics and Space Administration 25 Using NLP output for prediction Title Description Title Description Title Description Both NLP NLP #1 NLP #1 NLP NLP NLPNLP NLP NLP NLP NLPNLP NLP NLP NLP NLPNLP NLP NLP Data miner Data miner Data miner Data miner Anomaly class Anomaly class Anomaly class Anomaly class Anomaly class Anomaly class © 2014 National Aeronautics and Space Administration 26 Evaluation effort NLP techniques: 41 million comparisons – Over 1 month of computer runtime Data miners: 40,000 models evaluated © 2014 National Aeronautics and Space Administration 27 Predicting “Root Cause” types Title + Description Anomaly Title and/or Description NLP NLPNLP NLP NLP to predict Data miner Root Cause classes Software Hardware Operations Environmental Anomaly class Average Correctness Kappa 75.8% 0.21 Min 66.6% 85.4% -0.09 Promising, but need to test on a more diverse sample © 2014 National Aeronautics and Space Administration 28 Max 0.50 Labeling “Corrective Action” types Description Anomaly Title and/or Description NLP NLPNLP NLP NLP to predict Data miner Corrective Action classes Software Hardware Operations Ignore / none / unspecified Anomaly class Average Correctness Kappa 77.3% 0.53 Min 60.0% 93.3% 0.14 Very promising. Significant performance gains. © 2014 National Aeronautics and Space Administration 29 Max 0.86 Looking ahead Need to categorize anomalies in useful ways – how do engineers think about spacecraft? © 2014 National Aeronautics and Space Administration 30 Acknowledgements This work sponsored by NASA OSMA Software Assurance Research Program grants NNX11AQ40G and NNX08A260G Goddard Space Flight Center Lucas Layman Jet Propulsion Lab Allen Nikora NASA IV&V Facility Koorosh Mirfakhraie Dan Painter Keenen Bowens Wes Deadrick Ricky Forquer NASA Headquarters Martha Wetherholt Wallops Flight Facility Donna Smith Marshall Space Flight Center John McPherson http://commons.wikimedia.org/wiki/File:Nasa-logo.gif © 2014 National Aeronautics and Space Administration 31 Thank you Contact Lucas Layman [email protected] or [email protected] © 2014 National Aeronautics and Space Administration 32 BACKUP SLIDES © 2014 National Aeronautics and Space Administration 33 Does mission type affect SW anomaly discovery rate? % of anomalies discovered Normalized Software Anomaly Growth 100.0 90.0 80.0 70.0 60.0 50.0 40.0 30.0 20.0 10.0 0.0 0 Earth 20 40 60 80 Normalized time scale Heliophysics Planetary Astrophysics © 2014 National Aeronautics and Space Administration 34 100 120 all sw anomalies Who patches their software? Software anomaly repair category 80% 70% 60% 50% Astrophysics 40% Heliophysics 30% Earth 20% Planetary 10% 0% S/W patch or reconfigure Ops procedure or workaround None/Ignore © 2014 National Aeronautics and Space Administration 35 60 NLP techniques Each technique is a combination of: 1. Term extraction: {simple tokenization, Part of Speech tagging, stemming} 2. Term weighting: {TF, IDF, TF-IDF, Simple, Binary} 3. Model type: {Vector Space Model, WordNet} 4. Similarity computation: {Cosine, Jaccard, Dice, Resnik, Lin, Jiang, Pirro and Seco} © 2014 National Aeronautics and Space Administration 36 Acronyms • IV&V – Independent Verification and Validation • JPL – Jet Propulsion Laboratory • LSA – Latent Semantic Analysis • NASA – National Aeronautics and Space Administration • SW, S/W – software • TF – term frequency • IDF – inverse document frequency © 2014 National Aeronautics and Space Administration 37