You are on page 1of 25


(12) Unlted States Patent

(10) Patent No.:
(45) Date of Patent:

US 8,478,514 B2
*Jul. 2, 2013

Field of Classi?cation Search
USPC ............... .. 701/1, 2, 3, 29.4, 31.5, 32.9, 33.4,

(75) Inventor: Hillol Kargupta, Ellicott City, MD (US)

(73) Assignee: AGNIK, LLC, Columbia, MD (US)
(*) Notice: Subject to any disclaimer, the term of this

701/50, 300, 24, 25, 32.5, 33.6, 33.9, 114, 701/117, 118; 340/331, 531, 439, 905, 961, 340/989, 995.1, 991; 707/999.01, 999.104; 177/136, 139, 141; 709/203; 903/903; 180/65.21, 167, 168, 169; 303/152; 307/10.1;
_ _ 318/587; 901/1

patent is extended or adjusted under 35

See appl1cat1on ?le for complete search h1story.

U.S.C. 154(b) by 0 days.

4,839,835 A * 6,768,994 Bl *

References Cited

Tlhis patent is subject to a terminal disC almer'

6/1989 7/2004

Hagenbuch ................. .. 702/174 Howard et a1. ......... .. 1/1


APP1~ N91 13/173,408
Filed: Jun. 30, 2011

7,715,961 B1*
2002/0059075 A1*

5/2010 Kargupta ......... ..
Schick et a1. ..

..... .. 705/1

2007/0094268 A1*
. .

4/2007 Tabe ............................. .. 707/10


Prior Publication Data
US 2011/0258044 A1 Oct. 20, 2011

cued by exammer
Primary Examiner * McDieunel Marc

(74) Attorney] Agent] or Firm i AGNIK’ LLC

Related US. Application Data Continuation of application No. 11/ 116,889, ?led on


An improvement of methods and systems using mobile and (llsmbuted data Stream mlmng algomhms for mlmng Con
tmuously generated data from dlfferent components of a
vehicle. The system is designed for both on-board and remote

28 2005
’ ’ now



7 715 961
’ ’ '

(60) Provisional application No. 61/360,577, ?led on Jul. 1,
2010, provisional applicationNo. 60/565,840, ?led on

mining and management Ofthe data in Order I0 (1) detect ‘the
effect of Vanous ?nglne Paramaers on fuel Consumptlon

Apr_ 28’ 2004
(51) In‘; C]_ G06]? 1 7/10
Ulséc I

behavior, (2) predictive classi?cation of driving patterns and associative indexing of driver performance matrix, (3)
resource-constrained anomaly detection for onboard health monitoring, (4) vehicle-to-vehicle social networking and dis

(52) U 5 Cl
701/300_ 701/1_ 701/2_ 701/3_ 701/29 4,

tributed data mining, (5) adaptive placement of advertise
ments based on vehicle performance pro?le and (6) onboard

701/315; 340/331; 340/531; 340/439; 340/905; 340/961; 707/999.01; 707/999104; 180/6521; 180/167; 180/168; 180/169; 318/587; 901/1

emissions analytics computation for Wireless emissions
monitoring and smog test.

20 Claims, 14 Drawing Sheets

The MineFleet Onboard Data Mining Platform MF -DMP1 01 and MF-DMPZO'I
Serial Port 2 Serial Port 1
USB Port

Ethemet Port

Connects to the vehicle

diagnostic data port (eg. OBDJI port)

Connects to external Wireless modems through one of these


US. Patent

Jul. 2, 2013

Sheet 1 or 14

US 8,478,514 B2

Figure 1. The MineFleet Onboard Data Mining Platform MF-DMPlOl

Serial Port 2 Serial Port 1


USB Port

X Ethernet Port

Connects t0 m6 which

Connects to external wireless

diagnostic data port (e‘g' OED-ll port)

modems through one of these

Third Party Network Gateway Devices hosting the MineFleet Onboard software.g. 2.US. Wireless Modern Antenna Serial Port GPS Antenna / Connects to the vehicle diagnostic data port (e.514 B2 Figure 2. 2013 Sheet 2 or 14 US 8. Patent Jul. OBD-ll port) .478.

514 B2 Figure 3. Driver Points: 95 Number of Heaith Alerts: 5 Numberof Driver Alerts: 10 I‘ Fuel Economy: 22 MPG Keyboard .478. Handheld / Display " _. 2013 Sheet 3 0f 14 US 8. Patent Jul. Handheld cell-phones hosting the network gateway program and/or the MineFleet Onbaard software.US. 2.

The Bluetooth-enabled ()BD-II adapter. Patent Jul.478.US. (Bottom) .514 B2 Figure 4. 2013 Sheet 4 or 14 US 8. 2.

MINEFLEET Architecture Diagram E . 2. 2013 Sheet 5 0f 14 US 8.478.514 B2 Figure 5.US. Patent Jul.

Select All Deselect All Flagged Vehicles Name BWl Transport Vehicles 6 Drivers 10 Long Term Fuel Related Combustion Efficiency Long Term Fuel Related Combustion Efficiency Air intake Volume Inconsistency .US.478. 2.. 2013 Sheet 6 0f 14 US 8. A summary panel showing the different vehicle health tests performed. Patent Jul. Analysis F . Combustion Temperature Inequality Monitor Diagnostic Trouble Codes Fuel Economy Vehicle Health Tests Combustion Temperature Control Decay Monitor Oxygen Sensor Operating Condition Monitor Vehicle Diagnostic Test 02 Lean Vehicles 02 Rich Vehicles Driver Clustering Driving Performance Driver Avg.514 B2 Figure 6.

ignition system components. Long Term Fuel Trim Out of Range Compare with Benchmark Fuel Economy Vehicle Health Tests Test Failed Summary Thermal event detection Air intake volume Recommendation: inconsistency Long term fuel related combustion MineFleet recommends checking fuel pressure (too high). often significant collateral breakdown has occurred before the actual code will set. IVlineFleet recommends checking for clogged injector(s). we can often preempt the collateral damage through early detection of the deterioration. fuel delivery is also the most adaptive process the vehicle has to long term wear during normal engine operation. Patent Jul.478. oxygen sensor contamination and clogged air filter as most likely causes when fuel trim falls figh. fuel pressure (low). A Diagnostic Trouble Codes Fuel Economy Vehicle Health Tests Test Description: As part of the combustion formula. leaking fuel pressure regulator. or water intrusion on oxygen sensors as possible causes.514 B2 Figure 7. efficiency .US. clogged evaporative emissions system. Detailed information about the failure of a particular health test. injectors for leakage. While there is system Summary Vehicle Health Tests failure codes associated with reaching the limitations of its adaptability. This test History is designed to monitor changes well before they reach the breakdown stage. 2. 2013 Sheet 7 0f 14 US 8. By monitoring these changes within the fuel delivery portion as they occur over time.

Calculate Potential Savings Compare wit Benchmark elect Shift . It appears that if this vehicle were operating within the specified benchmark behavior. which is effective fuel economy.7 MPH.US. 2013 Sheet 8 0f 14 US 8. Patent Jul. Fuel consumption analysis summary panel. 2. The fuel economy would rise from 15. it would have a positive effect on the fuel economy.514 B2 Figure 8. The best fuel economy for this vehicle was obtained at Diagnostic Trouble Codes Fuel Economy speeds between 55 and 65 Miles per Hour (MPH) Summary Historical Fuel Economy Fuel Economy Prediction Feature Histograms Fuel Map Vehicle Health Tests According to the Benchmark Vehicle.478. the Oxygen Sensor — Bank 1 -—Sensor 1 appears to be operating outside of the normal parameters.7 to 16.

3% 9.US.514 B2 Figure 9.478.9% 6.8% 8.3 ?g“ Compare with Revert to Veh|cle Observed Distribution Modified Distribution Benchmark wee/we SelectShift Vehicle Speed Engine Speed O2 Sensor Engine Load Short Term Fuel Trim . Fuel system performance optimization using predictive models.2% 20% 3. 2. 20-40 40-50 50-60 60-70 70-80 80 90 90-100 1 0-120 Diagnostic Trouble Codes Fuel Economy Summary Historical Fuel Economy Fuel Economy Feature Histograms Fuel Map Vehicle Health Tests Pred'°“°_" 36.6% 7.2% Predicted Fuel Economy = 15. 2013 Sheet 9 of 14 US 8. Patent Jul.

478. 2. Patent Jul. 2013 Sheet 10 0f 14 US 8. Emissions analytics interface. forFleet AgnikFleet @ Based on TotaiCOZ Emissions @ Start Date 05/31/2009 End Date 06/30/2009 Trend Analysis T O I a I c Q 2 E m i S S i 0 n S Time ---> .514 B2 Figure 10. Agnik Home This graph disiplaysthe yehitieeniissii: trends‘ forithejéelected fie within e ‘specified time‘s'pan.US.

Agnik El Trend Analysis Fleet Performance Emissions Analytics Pending Requests Create Reseller Submit Problem "(DI Alerts Good Bad E5} Agnik . 2013 Sheet 11 0114 US 8. Patent Jul. 2.514 B2 Figure 11.478. The vehicle health score visualization interface in MineFleet.US.

40 Electrical System Air Conditioning Pending Requests Create Reseller Submit Prcblem @ Agnik . 2013 Sheet 12 0f 14 US 8.16 $835955 $8. 2.478.US.653. Predictive vehicle maintenance data analysis module linking maintenance and diagnostic data fOrFleet AgnikFleet @ lll c'eatepmdmke" Purchase Client Keys Start Date 05/31/2009 End Date 06/30/2009 Trend Analysis Total Cost Maintenance Tyne Fleet Performance $10.514 B2 Figure 12.259.16 Scheduled Service Engine Transmission Brakes Emissions Analytics $753609 $6. Patent Jul.330.

Vehicle and Accelerometer Frames of Reference. Patent Jul. 2013 Sheet 13 0f 14 US 8.514 B2 Figure 13.478.US. AW . 2.

2013 Sheet 14 0f 14 US 8. 2.514 B2 Figure 14. Multiple Linear Regression F-V Plane.478.US. Patent Jul. Observed accelerometer data Linear regression plane .

FIELD OF INVENTION the vehicle operation data generated thereby. 28. Pat. b.330. A playback mounting unit at a remote computer connects the remote The present invention relates generally to onboard and remote data stream mining of data collected from the vehicle computer to the host vehicle’s microprocessor module in order to establish digital communication Whereby the vehicle operation data and the analysis results processed therein are retrieved and displayed for a user. The on-board diagnostic systems process sensor readings and diagnostic information of the vehicle Embedded Control Sys tem in order to detect defaults. throWing aWay the collected data. SOCIAL 2 business models (note that onboard analysis dramatically reduces the data communication cost). 6. Pat. The system uses OEM proprietary or standardized interfaces. and c.g. coupled to the client computer device. process. Any evaluations to be made in light of the raW data are left for the user to make by themselves. d. shop test equipment.499 directs itself to a vehicle diagnostic and health monitoring system that includes a client computer device Within the vehicle. The overall methodology relies upon onboard analysis of the data. analyZing the data at the server for determining the 45 health condition of the vehicle. No. parts and lubricants (e. BACKGROUND OF INVENTION 35 computer system for use in evaluating an operator’s braking technique that employs a plurality of vehicle-mounted sen sors. presenting the results of the analysis to the user via a Web link to the server. Pat.514 B2 1 ONBOARD VEHICLE DATA MINING. The onboard computer in that system periodically receives and stores the parametric values associated With vehicle braking sensed by the sensors. Note that the main difference betWeen these tWo approaches is Where the data analysis is performed. 5. maintenance and diagnostics. c. For instance.207. for providing remote communication of data including data derived from internal monitoring sys tems of the vehicle. With short-range Wireless communica tion systems eventually replacing cable connections. No. No. No. monitoring is performed only during those discrete time intervals related to an instructional session. battery quality. sending the results generated by the onboard analysis to the remote server over the Wireless netWork. as the vehicle sensor monitoring system there is intended speci?cally as an instructional tool. outlier detection. 2. Physical connections link the vehicle and the Work d. US. brake pad Wear. sending the data to a server computer over the Wireless netWork.609.889.577.051 is directed to a vehicle condition 40 1. b. accelerometer.2005 Which are hereby incorporated by reference in its ule detachably coupled to the vehicle mounting unit a?ixed to and uniquely designated for a given host vehicle poles each vehicle sensor of that host vehicle to read. be divided into tWo groups: US. the vehicle needs to be mounted by a multiple number of sensors as opposed to using the standard OBDII data bus for getting the vehicle data data stream mining capabilities deal With advancement of the in the subject vehicle driver performance system. 6. That system does not per form any lightweight and sophisticated onboard data mining techniques on the data. and other related external sensor systems. J1708/J1939. coupled to the vehicle’ s monitoring systems. The 20 values of the parameters and is done remotely after the data is played back on the remote computer.. remote session management and user interaction. for data management. a server computer. and a remote service center including a vehicle data store. Off-board diagnostic systems acquire vehicle diagnos tics and sensor data or control on-board diagnostics and test 55 ing functions. vehicle to-vehicle social networking. The data thus gener ated by that computer is then available to be read later by an instructor Who compares the recorded parametric values to formulate further instructive steps. ?led Apr. and oil qual ity). Methodology 1: a. for example. GPS. predictive classi?cation of driving patterns.499. monitoring emissions and driver behavior. The maintenance systems on-board the vehicle continuously process sensor readings to vehicle. Methodology 2: a. 11/116. a diagnostic engine. and further aggregation and analysis of the analytics at the server or at the other onboard devices. Frequent acquisition of vehicle sensor and diagnostic data enables comparison With the created models to provide continuing analysis of the vehicle for repair. OBD to connect to the vehicle. ?led Jul. computing fuel economy. J1708/J1939. No. In addition. Several examples of knoWn art based on Methodology 1 exist. Non provisional application Ser. 50 determine the condition of the vehicle systems. 61/360. Provisional Application No. 5. monitoring emissions and driver behavior. 60 e. and store entirety. Collecting the vehicle diagnostic data onboard the vehicle from the onboard diagnostics port and accel erometer. associative indexing of driver performance matrix. Monitoring vehicle diagnostic data available from ports like OBD-II/CAN. and ment requires completely different sets of algorithms and . accelerometer. The loca tion of the data analysis has a huge impact on the technology used since real-time onboard monitoring of vehicle perfor mance data using resource constrained computing environ 65 US. presenting the results of the analysis to the user via a Web link to the server. 2010 and US. US.182 is directed to multiple vehicle component sensors mounted to a host vehicle measure vehicle component parameters indicative of a host vehicle’s driver performance. Current art can Furthermore.095 is directed to an onboard vehicle 25 technology for generating predictive data mining models. ADVERTISEMENT RELATED APPLICATIONS This application claims the bene?t of US. Pat. Collecting the vehicle diagnostic data onboard the monitoring system that employs machine learning and data mining technologies on data acquired from a plurality of vehicles in order to create models. computing fuel economy. A microprocessor mod NETWORKING. Also. the driver integrity-checking module is based on some pre-determined data bus like OBD-II/ CAN. a communication system. and GPS data over the Wireless netWork is a knoWn art. adaptive placement of adver tisement based on vehicle performance and vehicle emission data analysis. Analyzing the data onboard the vehicle for determin ing the health condition of the vehicle.US 8. This is an improvement of the currently available technology for vehicle onboard data mining and transmission of the resulting analytics to the server over Wired or Wireless netWorks. sending the results of the analysis to the remote server or to other onboard devices over Wired or 30 Wireless netWorks. 1.478.

. playing computer games. Published in the Special Issue on Learning from Data Streams. outlier detection. R. 4. Sarkar. Onboard Data Stream Mining for Detecting the Effect of Various Engine Parameters on fuel Consumption. Pat.US 8. Handy. driver behav ior scoring. Usually these data ports generate high throughput data streams and data stream mining algo rithms can be effectively used for various types predictive modeling. M. US. a ?eet comprised of multiple vehicles creates a large. M. “On Board Vehicle Data Stream Monitoring Using MineFleet and Fast Resource Constrained Monitoring of Correlation ing products and services that are triggered by the patterns identi?ed by the vehicle data analysis. 3. no provision is made for the assessment of driver performance based on any sensed parameters. these techniques are not capable of detecting linear and nonlinear complex driving patterns and they require an expensive process of transferring data to a remote monitoring station at a regular basis over the wireless net work. Gilligan. J1708/J1939. US. 1. D. 2004. Liu. Application of the onboard vehicle performance data mining technology in a distributed environment com prised of multiple vehicles connected over wireless net 50 works for insurance premium computation. P. Washington DC. Sarkar. Pat. H. Blair. predictive modeling. a second techniques that belong to Methodology 1 allow relatively simple tasks such as detection of a feature value crossing a limit set a priori. 7. Advanced data stream mining algorithms such as prin cipal component analysis. H. Resource-constrained Anomaly Detection for Onboard Vehicle Health Monitoring.961. vehicle-to vehicle social networking. SUMMARY OF THE INVENTION vehicle computer system for monitoring vehicle perfor mance. Web-based Presentation of the Results of Onboard Data Analysis and User-Interaction. prior art. K.715. 6. Onboard Predictive Classi?cation of Driving Patterns and Associative Indexing of Driver Performance Matrix. Vol 25 No. Similar data ports are also available in heavy-duty vehicles that follow different stan 25 Unfortunately. 2007. 2. No. Pro ceedings of SIAM Data Mining Conference. No. Various transducers for continually monitoring vari ous vehicle parameters are employed in that system. Pat. emissions monitoring and smog test. Sarkar. these techniques may check whether the driver crossed a speci?ed speed limit. It requires that various vehicle system parameters are made available through a standard data port. distributed data advanced onboard vehicle data stream mining techniques and their applications in different business processes.478. Accepted for Industrial full presentation in 2010 ACM SIGKDD Conference on 60 ment. and distributed envi ronment where each vehicle. V.034. dards like 11708 and 11939. Onboard and remote-server-based techniques for playing single and multi-party games based on the vehicle data. Kargupta. anomaly detec tion. The current work reports more limited communication-bandwidth and scalability of the overall system. fuel consumption and emissions data onboard the vehicle and sending the resulting analytics to the server over the wireless network. K. mobile. M. 3. OTHER REFERENCES Instead the different computing devices residing inside the vehicles perform on-board data stream mining and sends the data mining models and analytics to the remote control center. This approach advances the current technology by a large factor since onboard data stream mining using distrib uted data mining technology obviates the need for sending raw data to the remote server for analysis. “MineFleet: An Over view of a Widely Adopted Distributed Vehicle Perfor mance Data Mining System”. driver behavior. and other similar ports and external sensors. clustering. For example. Pat. LLC. Similarly. Kargupta. Agnik. Therefore. US. decision trees for analysis of the vehicle performance data onboard the vehicle. “Vehicle Data Stream Mining (VEDAS): A Mobile and Distributed Data Stream any aberrant operating conditions. Powers. The main differences are as follows: 40 stream mining system for onboard mining of vehicle data observed through the OBD-II. W. The current 35 is connected in the network using limited bandwidth wireless network. a moving source of data streams.961.894 directs itself to a self-diagnosis computer system onboard a motor vehicle wherein a plurality of detectors are mounted on that vehicle’s engine to detect 5 4 Matrices”. generation OBD system. Kargupta.715. Prior state-of-the-art is based on linear threshold-based The Environmental Protection Agency (EPA) requires vehicle manufacturers to install on-board diagnostics (OBD) system for emission control in light-duty automobiles and trucks. 3. clustering. vehicle health scoring. Bhargava. these techniques are applied after the data is uploaded to a remote desktop 20 computer from the vehicle. classi?cation using support vector machines. Automated adaptive algorithm generation and manage 1. Most mod ern automobiles come with the OBD-II system. This invention speci?cally deals with the following items: 1. 2.074. and K. Onboard and remote-server-based techniques for advertis 65 Knowledge Discovery and Data Mining. 5. is an exception that belongs to Methodology 2 and performs advanced sta the vehicle data patterns. University of Maryland Baltimore County and K. The patterns and statistical models generated by the onboard module are sent over the wireless or wired devices to a remote computing node for additional mining of the models and storage in databases or data ware 1. Onboard Vehicle Emissions Data Analysis. The invention features a multi-agent. Klein.144 is directed to an onboard Mining System for Real-Time Vehicle Monitoring”. A large number of both EPA-required and additional manufacturer-intro duced parameters can be observed using this port. and adaptive placement of advertisement based on vehicle 55 performance pro?le. mining technology for advanced fuel consumption mod eling. No. sending the vehicle data from all the vehicles over the wireless network is dif?cult because of the patent application is an improvement of the art reported in US. 5. which offers access to a large col lection of vehicle parameters. Klein. 8. however. 7. per se. H. 5. Moreover as noted earlier.514 B2 3 a communicator for communicating the results of analysis of vehicle information to the client computer device via the communication system. Puttagunta. Application of the onboard vehicle performance data 45 houses. ti stical data analysis and modeling for detecting patterns from vehicle performance. 7. New Generation Computing. and classi?cation of 30 However. Moreover. Although the computer system there performs continual monitoring while the vehicle is in operation. M. are not provided. comprehensive means for analyZing the measured vehicle parameters to characteriZe or assess driver performance. No. 2.

3. 5. .1 Architecture of MineFleet The architecture of the MineFleet system is designed to FIG. Ensemble-based approach for combining the output of any subset of these algorithms in a sound decision theoretic manner. 4. Onboard Data Stream Management System (DSMS) for e?icient continuous access to data. Fuel consumption analysis summary panel.514 B2 5 The on-board module is equipped With data stream man agement and mining algorithms that run on an embedded devices. A Web-server module for accessing the results of the 1. 4. 3) is based on the following main components: 1. 2. alerts. Analytics are generated on a continuous basis and sent over Wireless netWorks to the server either immediately or based on a schedule. Vehicle and Accelerometer Frames of Reference. 5. Onboard Module: a. Onboard predictive and monitoring algorithms for emission data analysis and correlating the emission characteristics With vehicle and driver behavior parameters. c. In addition. FIG. 1. 3. The main components of the MineFleet system are as fol loWs: 4. driving. the application alloWs the control sta tion to mine. FIG. This onboard system includes a data stream management system and modules for real-time 6 2. The MineFleet Onboard Data Mining Platform MF-DMP101 and MF-DMP201. monitor. b. Onboard analysis of the data stream and subsequent transmission of the diagnostic reports to a remote server over Wired or Wireless netWork. PDA or similar light-Weight computing devices con nected to the vehicle data bus. The system uses resource-constrained lightWeight data stream management. management and mining of doWnloaded data from the onboard system. FIG. The system is designed for both onboard or remote mining and management of the data in order to characterize a. The control sta tion and the Web server offers targeted advertisements Analysis and User-Interaction. FIG. FIG. 30 FIG. FIG. and anomaly detection across 60 the typical and unusual vehicle health. BRIEF DESCRIPTION OF DRAWINGS 25 tising products and services that are triggered by the patterns identi?ed by the vehicle data analysis. 14. predictive modeling. FIG. FIG. The MineFleet system Will minimiZe the Wireless data com module linking maintenance data With diagnostic data. The Web server offers a broWser d. and ?eet reports 65 tributed data mining techniques. Multiple Linear Regression F-V Plane. onboard stream mining. 8. b. b. Remote Web-Server Module: a. Onboard and remote-server-based techniques for adver triggered by the patterns detected by analyZing the vehicle related data analyZed onboard the vehicle and at the server. 3. Third Party Network GateWay Devices hosting the MineFleet Onboard softWare. 40 c. and dis the entire ?eet or subset of vehicles in a ?eet. The vehicle health score visualiZation interface in MineFleet. A summary panel shoWing the different vehicle health tests performed. 3. FIG. The Bluetooth-enabled OBD-II adapter. Report generation module for generating vehicle. Alert management module. Detailed information about the failure of a particu lar health test. Meta-level algorithms for activating and tuning of algo rithms based on vehicle and ?eet performance. e. The control center module runs on a desktop com algorithms for monitoring the multi-variate operating regimes of the vehicle system being monitored.478. iii. data analysis. Web-based Presentation of the Results of Onboard Data 20 based application for vieWing the analytics.US 8. Predictive vehicle maintenance data analysis 45 ii. Fuel system performance optimiZation using pre dictive models. 11. An onboard embedded system that supports the fol loWing tasks: i. MINEFLEET Architecture. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS 50 munication by performing the data analytics onboard the vehicle and schedule and/or priority/demand-based 55 The present invention is a method and system using mobile transmission of the analytics to the server. The sever and the Web server also alloWs single and multi party games based on the vehicle performance patterns. c. Rao-BlackWellised Particle Filtering (RBPF) based Dynamic Bayesian NetWorks for streams. 9. 7. Onboard module for adaptive targeted advertising based on the results of the analysis of the vehicle data. Communication Management Module: This module Will manage the transfer of data/ analytics from the onboard platform to the remote desktop machine. The control station also communi cates With a Web server. Meta-level dynamic selection and management of the algorithms for monitoring the vehicle. Support for long-term aggregate statistical analysis. d. 12. Various algorithms to analyZe the performance of the vehicle based on the data observed through the vehicle data bus and other sensors: a. 6. FIG. and diagnostic reports remotely over secure netWorks. Remote Control-Center Module: and distributed data stream mining algorithms for mining continuously generated data from different components of a vehicle. 2. Sensors for collecting data. Incremental support vector machines and kernel puter located at a remote control station and it Will support methods. 10. The onboard module is able to support tWo Way communication With the coordinator agent running at the control station at the remote server managing the vehicle(s) remotely. and query the performance of a single or collection of vehicles. Handheld cell-phones hosting the netWork gateWay program and/ or the MineFleet Onboard softWare. driver. and ?eet behavior. Long-term scalable data analytic stream management system. Principal component analysis-based fast incremental advanced analysis of the continuous data streams. 13. Emissions analytics interface. FIG. Onboard and remote-server-based techniques for play ing single and multi-party games based on the vehicle data. FIG. support predictive data stream mining for detecting the health of onboard systems and meta-level management of algo rithms for optimal performance of the system itself. The architecture of MineFleet (FIG.

Short Term Fuel Trim Bank l(%). Fuel System Status Bank 1 [Categ. Examples from the fuel sub-system include Air Fuel Ratio. 5 50 The symbol t denotes timestamp Whereas [t-i] represents the observation at time t-i. the ECM uses this signal to add the correct proportionate amount of fuel to the chamber for the reaction to occur. Events like this are triggered by the MineFleet. Start Up Engine Coolant Temp. the MineFleet server provides statistics about the different test 40 45 tive. The observed operating-condition-variables Which vary in an independent manner and the dependent Fleet makes use of the vehicle-performance characteristics and information from the entire ?eet to select Which set of algorithms should be used at a given time. MineFleet deploys a collection of data analysis techniques.). These algorithms are described in the folloWing sec tions. 1) Air Intake Volume Inconsistency Detection Test: Air and Fuel are the primary chemical components of the combustion process. Moreover. 3. Mass Air FloW Sensor 1(MAF) (lbs/min). 35 b. We discuss tWo such health tests in the folloWing for illustrating the approach. 4 shoWs the plot of the true data against the model output. MineFleet is equipped With a mobile data stream manage 8 potential mechanical breakdoWn. and Odometer (Miles). Start Up Intake Air Temperature (0 F. Fuel Level Sensor (%).514 B2 7 b. The algorithms are selected from a library of various domain knoWledge-based tests and 25 Engine Coolant Temperature (0 F. engine damage.5 Benchmarking a Vehicle Sub-System MineFleet supports benchmarking a vehicle or any of its particular sub-system With respect to another vehicle Which 60 formula and thereby causing combustion (and fuel) ine?i ciencies. Oxygen Sensor Bank 2 Sensor 1 (mV).). 3 . an automobile system Will also generate 2) Compare/benchmark the behavior of the vehicle With respect to other similar vehicles in the ?eet. driver behavior. The tests are performed onboard the vehicle. Instead of relying upon one single technique. Attrib. and the ?eet. Module for playing single and multi-party games based on the vehicle data. continued operation under these conditions can create violent mechanical failure due to a skeWed air/fuel ratio. We can use this technique to compare the behavior of a fuel subsystem of a bunch of 2000 International Truck With respect the behavior of the fuel sub system of 2007 International Truck. MineFleet acti vates a set of onboard and server-side data analysis algorithms features that change behavior in response to the changes in the 20 operating-condition-variables. FIG. If a particular test fails frequently in the ?eet then We may need to perform several related tests in order to further investigate the prob lem. For example. Throttle Position Sensor (degree). Once the data is segmented into different regimes.). Consider the linear model for Battery Volt age (BV[t]) generated as a function of Injector Pulse Width (IPW): failures in different vehicles of the entire ?eet. We can ascertain if there is negative impact to the the general technical approach for such data modeling and anomaly detection in MineFleet. The Mine Typically. On the other hand. HoWever. The data analysis tech niques are managed by the MineFleet meta-level algorithm management system Which selects and manages algorithms based on the current status of the vehicle. by identifying those Which are appropriate for the current status of the vehicle and the ?eet. Engine Torque. 3. 3) Generate a prescription and a schedule for the appropri ate algorithms based on the available monitoring algo rithms Which uses the folloWing kind of tests: a. A decline in engine vacuum may be a result of a minor leak or 65 in good shape. Calculated Engine Load (%). Data driven predictive modeling/anomaly detection techniques. Oxygen Sensor Bank 1 Sensor 2 (mV). approach that uses various algorithms to detect health-prob lems and combine their outputs to produce the overall out come of the system in a sound decision theoretic manner. mod els for each one of the regimes should be developed for the different regimes.).]. In order to maintain proper burn e?iciency. In order to do that We can compare the performance metric of the sub systems from dif ferent vehicles and also compare the distribution properties based on the estimations constructed by predictive modeling techniques. These tests are used for evaluating the status of a vehicle. 30 Vehicle Speed (Miles/Hour). (0 F.3 Vehicle Health Tests and Fleet-Level Trend Detection MineFleet is equipped With a collection of science-based vehicle health tests. also data mining techniques for predictive modeling and anomaly detection. or Worse. driver. This may result in high num ber of false positives.478. Injector Pulse Width #1 (msec). Oxygen Sensor Bank 1 Sensor 1 [mV]. Diagnostic Trouble Codes. and subse ment system (DSMS) for supporting all the data analysis and monitoring tasks. MineFleet uses several vehicle health. 3.US 8. Usually no single technique Works reliably in all scenarios. Science-based domain knoWledge-intensive tests. Intake Air Temperature (IAT) (0 F. 2) Engine Intake Vacuum Ine?iciency Detection: Vacuum is created through normal engine operation. FIG. By monitoring this signal. Oxygen Sensor Bank 2 Sensor 2 (mV). Rest of this section describes 55 shoWs MineFleet panels for presenting the vehicle diagnostic test results. other features that depend upon the operating conditions. Engine Speed (RPM).2 Automated Algorithm Management MineFleet Algorithm Generation and Management Engine is in charge of selecting and managing the algorithms that runs both onboard the vehicle and at the server. A Web-server module for advertising products and services that are triggered by the patterns identi?ed by the vehicle data analysis. Throttle Position Sensor (%). The methodology can be summarized as follows: 1) Evaluate the performance and the health-condition of the vehicle. Long Term Fuel Trim Bank l(%). Since operating conditions for a complex vehicle can be diverse. The ECM’s response to changes in vacuum is simply to alter fuel delivery. segmenting the distribution of values can be effec fuel-consumption and emission related data analysis algo rithms. Examples of operating-condi tion-variables in conventional automobiles include the fol loWing: Barometric Pressure.4 Onboard Statistical Data Modeling MineFleet is equipped With advanced multi-variate statis tical and data stream mining algorithms for automated moni toring and anomaly detection. . These data analysis techniques run on top of the data stream management system for e?icient execu quent fuel detonation. c. sensors in the vehicle subsystems generate tWo types of data. MineFleet makes use of an ensemble-based tion of the underlying operations. If a test fails then an alert is generated. and Manifold Absolute Pressure (Hg). the intake stroke of each piston draWs in the fuel/air mixture to be consumed. Idle Air Control Motor Position.

This clearly shoWs that the covariance matrix can be incrementally maintained. Use the operating regimes for detecting the anomalous behavior. Yl. If the change in the covariance matrix is small then We may not need to update the eigenvectors and the eigenvalues. . All the principal components are statistically uncorre lated and individual principal components are ordered With respect to the statistical variance of that component. The previous sections proposed incremental PCA technique to address that.6 Stream Algorithms for Onboard Principal Component 10 [Hall. We are interested in representing X by means of a small set of principal components (dimensionality 30 matrix E is de?ned as ||E||F:(Zi Zj (EU-)2)”. Therefore. 2. We Would like to extend a grid based clustering approach so that it can operate in a data Where Z'i:(Xi Xi)/ml. .478. 2002] on stream data clustering. .is uncorrelated With the previous principal components (Y 1. 3. This section explains the online cluster ing technique. the sample covariance Covt based on data blocks X1. Incremental clustering techniques can be used to dimensional matrix observed at time t (i. XS). Let 7012702. There exists little Work [Guha. . At any given moment one can update the covariance matrix and 55 stream environment by identifying those regions in the data compute the eigenvectors. . 2000] [O’Callaghan et al. X”)1 With mean E[X]:0 20 matrix perturbation theory [SteWart. The proposed research Will use 1) incremental cluster reduction).e. . . 1998] [Han. 1998] offers several algorithms to perform eigenanalysis in an incremental manner. . Where Xt is an mt><n Clustering is an important unsupervised data mining tech nique frequently used for identifying similar objects. Yk] be the ?rst k principal 35 stream in order to detect the normal characteristics and out components of X. . 3 . Apart from the polygon-based anomaly detection algorithm described in the previous section. We need incremental PCA algorithms that Will be able to update the eigenvectors and the eigenvalues of the covariance matrix With little computational cost.Where 6 is the difference betWeen the K1 and A2 of the matrix Covt.US 8. This section describes one of those based on the local outlier factor (LOF) approach [Breunig et . is a unit eigenvector of o. Consider the random vector XIQQ.Yl-_l) and has maximum variance. sometimes called its eigengap. The MineFleet is equipped With online incremental PCA algorithms for light-Weight data analysis on continuous data streams. Where k<<n. (1998). . In general. . 1 We denote our vectors as roW vectors. We can use the folloWing among the data points. . In that case. . it Will also deploy several other anomaly detection techniques. MINEFLEET Will make use of sev bottle-neck hindering scalable performance of the stream mining system. et al.. trans actions. They proposed an incremental k-Median clustering algorithm for stream data. X2. ||\(l — y'l||§(4||E||F)/(6—21/2||E||F)and|7tl—7t'l|§21/2||E||F. 1933].7 Onboard Clustering Techniques for Identifying Distribu tion Regimes A loW dimensional similarity-preserving representation of the potentially high dimensional data stream is very impor tant for scalable performance of a data stream mining system. Without explicitly performing the entire eigenanalysis for every neW observation from the stream. This technique projects the data to a loW dimensional feature space that preserves similarity space by an appropriate amount. Once the representation is constructed. 2001] that operate by dividing the data space into rectangular grids and then identifying those grids that contain dense regions of points.8. let y'liy'zi . 1990] We can Write. Y2. It involves linear transformation of a collection of related variables into a set of principal compo the sample covariance matrix com puted from only the data block X. These techniques Analysis Representation construction techniques are frequently used in data mining for de?ning a set of features that can Work by computing the changes in the basis introduced by the neW observation from the stream and then rotating the sub easily identify the underlying patterns in the data. The ith principal component of X is a linear combination YiIX aiT. Xt can be 45 identify the different regimes of the underlying distribution of the stream data. This is very likely to hap pen When the observed data stream is not producing any novel characteristics. ing and 2) techniques from computational geometry to address these issues. These algorithms Will be used for inequalities for bounding the changes in the eigenvectors and eigenvalues due to the changes in the covariance matrix. The PCA technique is used for modeling and benchmark ing of the vehicle subsystems and driver behavior. The proposed research Will implement and evaluate this algorithm. . . . Detecting Anomalies As mentioned earlier.514 B2 9 3. Where al. . . Using data and characteriZe the typical operating behavior. 270k be the domi the folloWing speci?c purpose: 1. Let E:Covt—Covt_l. . PCA is a statistical technique for analyZing multivariate data [Hotelling. . These principal components can be used to obtain a reasonable approximation of the original data as folloWs: X'IY' A'T Where the columns of A' consist of the ?rst k eigenvectors of Ox. These regimes Will capture the distribution of the sensor nant eigenvalues of the perturbed covariance matrix Covt. This research Will also evaluate and implement the incremental PCA technique proposed by Hall et al. These grid based clustering algo rithms can operate by performing only a small number of passes over the original data. X2. and behaviors of observed items from a given domain. X2. . The computer vision literature 65 eral anomaly detection techniques and combine their results for computing the likelihood of health-problem. mt observations are made at time t). Construct a loW-dimensional embedding of the multi variate operating regimes of the monitored system. There also exist gridbased clustering algorithms [AgraWal. This usu ally involves projecting high-dimensional data to a loW-di mensional space.. in many applications this is an extremely expensive proposition and it may create a 60 space that represent typical operating regimes. Principal Component Analysis (PCA) is a popular technique to do so. However. If the data has Zero-mean. Let Y':[Yl. Zy'k be the corresponding eigenvectors. They shoWed that it out-performs k-Means clus computed in a recursive fashion as folloWs: rel Em] Cov.C corresponding to the ith largest eigen value M. : 1 CovH + 7": 50 tering algorithm. We need to identify the different operational regimes of the liers. . Matrix perturbation theory also offers tools to deal With the problem of computing prin cipal components from stream data. 40 Consider a data stream mining problem that observes a series of data blocks Gil. In this case. The Frobenius norm of a 25 and covariance matrix Cov[X]:E[XTX]:ox. . .