As the largest mobile telecom carrier in the world, China Mobile has the world's largest wireless mobile network, based on the existing vehicle networking equipment (CAN-bus, OBD, ADAS, equipment fatigue warning system, GPS, driving recorder, etc.), which can provide vehicle networking service, based on vehicle networking data analysis and provide users risk assessment, vehicle real-time risk monitoring, and comprehensive financial institutions for the vehicle and provide data support for differentiated financial services.
The main contents include the following:
1. Vehicle and drivers data collection: Collecting information of vehicle's mechanical status, driving behavior, and surrounding environment through OBD, ADAS, fatigue warning system, GPS, and other equipment.
2. AI technology application: mainly include the identification of the driver's body state, the wine driving, the fatigue degree, and so on.
3. To improve the accuracy and applicability of the risk assessment model through machine learning.
Speaker
Duan Yunfeng, Chief Designer of China Mobile's big data system, China Mobile Communications Corporation
Gen AI in Business - Global Trends Report 2024.pdf
The case of vehicle networking financial services accomplished by China Mobile
1. Speaker : DuanYunfeng,LiShaonian
China Mobile Communications Corporation
13910211021@139.com
June 2018
Practice of vehicle networking
in financial industry
2. Founded in April 20, 2000, it is a mobile communication operator based on GSM, TD-SCDMA and
TD-LTE standard network
China Mobile is the largest telecommunications operator in the world by customer base and
network size
The company's operating income was 115 billion 700 million dollar, an increase of 4.5% over the
same period last year.
On the user side, the number of mobile phone customers reached 887 million, a net increase of 38
million 300 thousand households in the end of 2017
The total number of 4G base stations reached 1 million 870 thousand, covering 99% of the
population in China
The number of smart connections in the Internet of things(IOT) has increased by 126 million,
reaching 229 million.
[Introduction]The China Mobile Communication
Corporation
3. Name : Li Shaonian
Academic degree:Ph.D
Work experience :
2003.7-,China Mobile
Project manager of big data in Hunan
province
Expert in China Mobile
Name : DuanYunfeng
Academic degree:Post Ph.D in Peking University
Work experience :
General designer and founder of big data in China Mobile
17 years experience in data warehouse , big data
After 2000, he transferred to the research of data warehouse
and data mining, and obtained the post doctoral degree of
information and communication engineering of Peking University.
Presided over the design and construction of a data
warehouse system in China Mobile, and now becomes the largest
big data system in the world telecommunication industry (500PB
capacity).
In 2006, he was named China's first outstanding database
engineer by “Chinese computer newspaper” .
In 2006, he won the first prize in science and technology progress
of “the communications society in China”.
Training and forming China Mobile's big data technology team ,
more than 400 persons
[Introduction ] Speaker
4. [Appendix] Books written by DuanYunfeng:
‘Big Analysis’ concept
raised first time
Introduce internet
thought to Big Data
as methodology
《The foundation
of data
warehouse》,2005
year
《Data warehouse
and it application in
Telecom》,2005
year
《Big data and Big
analysis 》, 2015 year
《Internet thought in Big
data 》, 2015 year
5. [Introduction]Application of Data Mining Model
in China Mobile
Model type
Algorithm
classification
Mining subject Common mining algorithm Application direction Application field example Application model
Models based
on mining
Time series
analysis
Autoregressive Integrated Moving
Average (ARIMA) model
Scientific decision-making KPI forecast
Telephone traffic forecast model
Revenue forecast model, etc.
Regression Forecast on customers
Linear regression Scientific decision-making Telephone traffic forecast and off-network tendency
Telephone traffic forecast model
Revenue forecast model, etc.
Logistic regression
Targeted customer retention
and marketing
Customer early warning and segmentation
Core customer off-network early warning model
Student customer identification model
Classification
forecast
Classification analysis on
customers
Decision-making tree
Targeted customer retention
and marketing
Data traffic marketing User package matching model
Bayes classification Text classification, etc. Webpage content classification model
Neural network Stock retaining Customer off-network early warning model
Support Vector Machine (SVM) CMCC market forecast CMCC customer market share measurement
KNN (K-Nearest Neighbor) Text classification Content classification model
Hidden Markov Model Semantic identification Text mining, webpage keyword classification
Clustering Finding out different groups
K-Means clustering Fine management and
targeted customer
marketing
Channel early warning Card maintenance early warning model
K-Modes clustering Market segment customer identification Customer segmentation model, etc.
Expectation Maximization Algorithm
Targeted customer retention
and marketing
Natural language processing (e.g. word
segmentation)
Chinese word segmentation model
Social network Finding out social network Social network
Targeted customer
marketing
Market segment customer identification and
marketing
Student customer identification model
Association rules
mining
Mining of association among
subscribed services by
customers, cross-selling of
products, etc.
Association rules, with Aprior as typical
algorithm Targeted customer retention
and marketing
Cross business marketing Customer service association model
Collaborative filtering Cross marketing Application precision recommendation model
Sequence
analysis
Sequence pattern mining
Targeted customer retention
and marketing
Cross marketing and shopping cart analysis
Customer service association model
Application precision recommendation model
6. [Introduction]Application of Data Mining Model
in China Mobile-2
Model type
Algorithm
classification
Mining subject
Common mining
algorithm
Business goal Application direction Application field example Application model
Models based
on rules
Business rule
definition
Principal component
analysis
Assessment
monitoring
Targeted customer
retention and
marketing
Comprehensive assessment
Customer credit evaluation and
analysis model
Analytic Hierarchy
Process
Existing customer early
warning and SMS spam
control
Core customer off-network early
warning model
SMS spam number identification
model
Index discriminance Data traffic marketing
Identification model for users
upgrading data traffic package
Comprehensive index
discriminance
Market segment, stock
retaining and attracted
customer market
Customer segmentation model
CMCC customer health
assessment model
Fine management
Channel management, and
network and market
collaborative planning and
optimized analysis support
Card maintenance early warning
model
WLAN hotspot health identification
model
Others Link analysis
Link importance
mining
PageRank
Ranking
Targeted customer
retention and
marketing
Search results ranking
(inverted index)
Webpage ranking model
HITS
Search results ranking
(inverted index)
Webpage ranking model
7. [Introduction] The architecture of Big Data system
in China Mobile"Logically integrated but physically distributed": China Mobile's enterprise-level big data platform is established in a logically integrated but physically
distributed architecture consisting of control clusters and computing clusters. Control clusters are built and maintained by the headquarters. The
computing clusters are constructed and maintained by the headquarters and provincial companies, accordingly.
Cloud resource pool
Department
Professional
base
Provincial
application
Professional
company
Protocols, such as the
development SDK and
RESTful APIs
Professional
company
data
External data
External
application
Computing cluster node of
Province A
External
data
BSS data
MSS data
OSS data
Cloud resource pool
Computing cluster of province B
Cloud resource pool
Headquarters computing cluster
External
data
BSS data
MSS data
OSS data
Control cluster
Unified
metadata
Access
cluster
Scheduling
cluster
...
The enterprise-level big data platform consists of the headquarters big data platform and
provincial-level big data platform.
8. There are more than 17000 x86 machines in 31 provinces,
including 15000 x86 machines for Hadoop , and 2000 x86
machines for MPP.
A total of 160 small minicomputers in 31 provinces
The system capacity is more than 500PB
31%
5%
34%
23%
7%
Distribution of existing Hadoop
devices
spark
kafka/flum
Hive
Hbase
Other
The system use of the 31 provinces is as follows:
The number of daily active users of the system is about
27100. The number of clicks per day is about 368509
times. A total of about 81717 people 。
27100
368509
81717
0
100000
200000
300000
400000
日活跃用户数(人) 每日点击数(次) 总使用人数(人)
System usage
[Introduction ] The Big Data system in China Mobile
9. [Introduction]Technical Architecture of the Enterprise-level
Provincial Big Data Platform
The technical architecture of the enterprise-level provincial big data platform is classified into the following layers: data collection, data storage and
computing layer, open framework, application center, and unified O&M management. In the architecture, data storage and computing are closely
related to each other.
Data source
BSS
OSS
MSS
Service
platform
Internet
External data
Stream processing
Stream
collection
Stream
computing
Batch collection
(cloud-based
ETL)
Web crawler
Storage and computing
Distributed storage
and computing
cluster
List-based data
processing
MPP
Deep analysis
Data mining
Data warehouse
Ad hoc query
KPI
Open framework
Data openness
Data service
management
Service
developer
management
Tool openness
Data mining tool
Data display tool
Data processing
tool
Application center
Industry data
products
User
Operation
analyst
Decision-maker
Product planning
manager
Channel
manager
Regional
manager
Enterprise
customer manager
Operator
VIC manager
Designer
...
Developer
O&M personnelData governance
Metadata
management
Data quality
management
Data asset
management
Data standard
management
Data security
management
Platform O&M
Multi-tenantmanagementschedulingandmanagement
O&M data
collection
Alarm analysis
Security
management
System O&M
External user
Resource
openness
Storage
resources
Computing
resources
Unified data
collection
Unified DC
Platform data
products
Other products
10. ■ Part 1: background and platform introduction
■ Part 2:data acquisition of vehicle network
■ Part 3:Risk assessment algorithm
■ Part 4:Platform case
Contents
11. Project background
As the integrated ecosystem of future people, cars, roads, platforms and applications, the vehicle networking has a wide
application field and market prospects. However, the development of the car network has not yet been found and the
development of the market is not yet ideal. Expanding the application of vehicle networking and integrating the related
resources of the vehicle network can promote the better development of the vehicle networking market.
terminal
platform Application
Data acquisition
person / car / road interaction
High efficiency data processing Mass service application
network
12. Market demand background
• According to statistics of the Ministry of communications, vehicle ownership reached 310 million at the end of 2017.
And in this year, 28 million new cars were registered in the public security traffic management department. With the
popularization of vehicles in people's daily life, the automobile industry has become an important area in the financial
industry. The income of car insurance accounts for more than 70% of the total property insurance revenue, and the
premium income is over 700 billion Yuan in 2017. However, 40 of the 53 insurance companies suffered losses in vehicle
insurance business throughout 2017. The high channel cost and the loss ration are the causes of the deficit. The channel
costs is 40% and the loss cost is 50%.
• Financial institutions have a strong demand for reducing the cost of receiving customers and paying the cost, but the
financial institutions have limited access of the users, lack of understanding of the users. It can not effectively evaluate
the user's credit and risk, and can not carry out safety monitoring to users.
13. Platform introduction
China Mobile has worked with XX insurance
company Hunan branch, and Explored the
financial services of vehicle networking in the
following aspects :
Carry out big data analysis and differentiated
insurance services on account of user driving behavior,
travel habits, driving environment,etc.
Provide driving behavior analysis service,
continuously supervise and improve driving behavior,
and provide real-time safety monitoring for insurance
companies or motorcade;
Monitor the accident situation after the customer is
out of danger and reconstruct the accident, so as to
prevent fraud accurately.
Internet of vehicles
platform
Driving data
Internet users External data
Transportation,
meteorology, GIS,
etc.
The
insurance
company
DSS transition.
Reduce the cost
Differential
car insurance
Risk early-
warning
Anti fraud Vehicle financial
services platform
Driving
behavior
Driving
habits
Driving
environment
OBD equipment
ADAS equipment
……
Real-time
monitoring
Accident
reconstruction
An actuarial
model
14. ■ Part 1: background and platform introduction
■ Part 2:data acquisition of vehicle network
■ Part 3:Risk assessment algorithm
■ Part 4:Platform case
Contents
15. Internet of vehicles data collection
As a booming area, the Internet of vehicle has emerged with various data acquisition devices, which are pre-installed by
automobile manufacturers or installed by external manufacturers. It also has data collected from the vehicle itself, and also
monitors the environment inside and outside the vehicle.. Based on the requirements of the platform and the existing
conditions, the platform currently collects data mainly through the following means:
1. Collect data of vehicle operation data and driving behavior through the rear mounted OBD equipment or mobile
gyroscope ;
2. Collect the vehicle trajectory through GPS or signaling;
3. Collect driver status data through the Driver Fatigue Monitor System;
4. Collect vehicle driving environment through ADAS System;
16. Recognizes Vehicle Driving Behavior Through OBD
Using OBD or gyroscope equipment to collect data information such as vehicle speed, driving time and other data,
the starting time, end time, duration, maximum acceleration value and average acceleration value of the
transmission are analyzed.
When the acceleration reaches the set maximum acceleration threshold, the acceleration process begins to
calculate. If and only when the acceleration time exceeds a certain value, it is a rapid accelerated event. And the
acceleration process will be completed when the acceleration is less than the maximum acceleration threshold for a
period of time.
Variable speed behavior recognition criteria
behavior acceleration duration speed
Rapid acceleration a ≥ 3m/s2 0.25s≤T≤3s
Rapid deceleration a ≤-3m/s2 0.25s≤T≤3s
slam the brakes on a ≤-4m/s2 0.1s≤T≤3s V≤0.5m/s
17. Recognizes Vehicle Driving Behavior Through OBD
The red dotted line in the above figure indicates that the acceleration
decision threshold is 3m/s2, and the black line shows the driver's
acceleration in the course of the moving of the vehicle. The area of the
purple dots indicates that the variable behavior in the time section is
judged to be an acceleration. The acceleration of the vehicle exceeds the
decision threshold and the duration is 2.5s, which satisfies the criteria for
the acceleration.
The following figure shows the effect map of the quick deceleration behavior. The red dotted
line indicates the decision threshold of the accelerated deceleration decision acceleration, and
the purple dot area indicates that the variable speed behavior in this period is a quick
deceleration behavior, and the black real line indicates the acceleration of the driver in the
course of the moving of the vehicle. Although the motion acceleration in the following figure
exceeds the critical Braking threshold, the vehicle does not stop running, and can not be judged
as an emergency braking action. Because the motion acceleration of the vehicle exceeds the
decision threshold of the rapid deceleration, and the duration is 1.5s, it is in line with the
decision conditions of the rapid acceleration behavior, so it is judged to be a brakes.
18. Sharp Turn Identification
The blue dotted line indicates the angular velocity threshold of the sharp turn behavior, the black real line indicates the
curve of the motion angular velocity, the red dotted line represents a sudden turn, and the blue dotted line shows the
change of the azimuth.
19. Identification of drastic driving behavior
Set in 1 minutes, there are three times of consecutive emergency acceleration, sharp deceleration,
brakes and sharp turns, the driver is set as a violent driving behavior.
20. The application of visual analysis in driver
behavior and driving environment identification
Answer the phone
Operate air conditioning,
acoustics , etc.
Drink water, smoke, etc
……
Lane keeping
lane change
Car distance
Light intensity
……
Through the existing vehicle networking equipment such as fatigue driving warning system and ADAS device, the
driver and the external environment can be identified.
1. Monitoring and analysis of drivers
Based on the driver's physiological image response, the driver's fatigue state is deduced from the driver's facial features,
eye signals and head motility.
2. Identification of the external environment
Recognition of lanes, pedestrians and other vehicles based on vehicle external image recognition.
Fatigue driving warning equipment Advanced Driver Assistance System
21. Using Of Other External Data
Weather data
Including visibility on the day, wet conditions on the ground and wind power, etc, which have a direct impact on driving
behavior;
traffic police data
Including vehicle violation data and vehicle accident data, etc.
GIS data
Including road width, slope, camber, etc.
22. ■ Part 1: background and platform introduction
■ Part 2:data acquisition of vehicle network
■ Part 3:Risk assessment algorithm
■ Part 4:Platform case
Contents
23. AEW-AHP Algorithm
Using the improved AEW-AHP algorithm. The entropy weight method relies on the objective data to calculate the index
weight and is in good agreement with the actual situation. The judgment matrix in the analytic hierarchy process is
determined by the experience value, and it can handle the multi-objective decision problem flexibly. EW-AHP calculates
the objective weight obtained by the entropy weight method AI and the analytic hierarchy process. The subjective weight I
is derived from the comprehensive consideration, and the final index weight is obtained by combining the entropy weight
method and the analytic hierarchy process.
Step 1: Supposing there are M standard layers and n index layers, each standard layer corresponds to t1,t2,……,tk
indexes, meanwhile, satisfying t1+t2+……tk=n,and judgment matrix is Bij。 According to the formula, calculate the
weight of the standard layer ABi(i=1,2,……,m)and the weight of each index ASj(j=1,2,……,m) based on
the analytic hierarchy process 。AWj(j=1,2,……,n) Represents the weight value of the analytic hierarchy process,
which is the product of the standard layer weight Abj and the index layer weight Asj.
AWi=
Where Mi represents the product of each row element of the matrix, that is: Mi= (i=1,2,……,k)
24. AEW-AHPAlgorithm
Step 2:
It is assumed that the score target of the driving behavior score system is Ai(i=1,2,……,m), and the index set
is Xj(j=1,2,……,n),AXij represents the original data value of the j index in the i group data in the driving
behavior score data ,represents the data value of AXij after it has been normalized. According to the following formula,
we can calculate the index weight based on entropy weight method. (j=1,2,……,n)
AXij=
EWj=
Pij represents the proportion of the j index in the i data of the driving behavior scoring data, that is :
Pij= (i=1,2,……,m;j=1,2,……,n)
25. AEW-AHP Algorithm
Step 3:
The index weight Asj and the weight index EWj of the entropy method are combined according to the following formulas,
then we can deduce the weight of the comprehensive index γj(j=1,2,……,n). After that, according to the standard
layer, the index layer is normalized again, and we can obtain the weight of corresponding index layer Ωij(i=1,2,……,
m;j=1,2,……,n). Finally, the Ωij is multiplied by the standard layer weight ABj in the analytic hierarchy, then
getting the AEW-AHP Driving behavior score index weight Wj(j=1,2,……,n)
γj =
26. Risk Assessment Model
Hunan Mobile cooperated with
Hunan XX Insurance Co., Ltd, to
conduct follow-up analysis on more
than 4,000 motor vehicles, analyzed
the correlation between various
driving behaviors and accidents, then
found the corresponding accident
probability, established a vehicle risk
assessment model, and finally
achieved risk assessment for vehicles.
The standard
layer
Index layer AHP weight EW weight AEW-AHP weight
Driving
behavior
Accelerated frequency 0.096
0.6
0.072
0.468
0.101798
0.701498
Deceleration frequency 0.089 0.069 0.090443
frequency of sharp turns 0.083 0.073 0.089235
Sudden braking frequency 0.091 0.071 0.095156
Fatigue driving frequency 0.127 0.092 0.172079
irregularities 0.114 0.091 0.152786
Driving
environment
Road type 0.015
0.2
0.021
0.22
0.004639
0.132344
The width of the road 0.021 0.041 0.012681
Road conditions 0.012 0.012 0.002121
topography 0.015 0.015 0.003314
The frequency of dangerous
road sections
0.072 0.058 0.061503
weather 0.05 0.062 0.045656
Vehicle daily traffic 0.015 0.011 0.00243
Driving
habits
The length of driving 0.022
0.1
0.019
0.088
0.006156
0.033844
The time of driving 0.025 0.024 0.008837
average speed 0.03 0.035 0.015464
route familiarity 0.023 0.01 0.003387
Vehicle
properties
State of the vehicle 0.021
0.1
0.021
0.094
0.006495
0.045803Operation or self-use 0.035 0.035 0.018042
Commercial or household 0.038 0.038 0.021267
Driver
attribute
sex 0.031
0.1
0.031
0.13
0.014153
0.086511age 0.047 0.047 0.032534
Been driving 0.052 0.052 0.039824
27. ■ Part 1: background and platform introduction
■ Part 2:data acquisition of vehicle network
■ Part 3:Risk assessment algorithm
■ Part 4:Platform case
Contents
28. Risk Assessment Of Driving Behavior
By analyzing the user's driving
behavior, driving environment, driver's
analysis, vehicle condition and driving
habits, the vehicle risk score is given.
Hunan Mobile cooperated with Hunan
XX Insurance Company to collect user
driving data by installing OBD devices
or installing mobile phone software for
4243 customers, and built a car
network financial risk assessment
model.
29. User Risk Assessment Cases
The following figure shows the risk assessment scores of 100 users randomly selected. In the red part of the figure, there
are 14 persons with a score lower than 60 points. One of the drivers’ driving habits is very bad. And there are 6 persons
with a score higher than 90 points. Most risk assessments are between 70-90 minutes.
30. User Risk Assessment Cases
The 100 drivers in the past year have been claiming compensation. Among these drivers, the risk assessment of the users
below 60 points has basically appeared in the history of claims, and the drivers with a risk assessment score higher than
80 have basically no claim.
31. Full Range Vehicle Risk Comparison
The vehicle network financial service platform provides a full range of vehicle risk comparison, including five
dimensions of driver risk, driving risk, driving habit risk, driving environment risk and vehicle risk. The analysis results
provide a comparison and ranking of vehicle risk overall indices in five cities, including the scores and ranking of each
sub risk. It is necessary to strengthen the management to identify Shaoyang as a high risk market.
32. Vehicle Risk Index Cluster Analysis
Through the cluster analysis of the vehicle risk index, the
vehicle network financial service platform identifies the
customers with different risk indices as the data basis for
the vehicle customization insurance cooperation with the
insurance companies. The analysis results show that
vehicle risk index can be regarded as an excellent
customer of insurance company before 56, and it can be
an important target for insurance companies. The
clustered vehicles 56-60, 66-70 and 86-90 are risk
transferable customers. They can use insurance risk
control strategies to monitor risk status as a follow-up
development strategy. A risk index of more than 90 is a
high-risk customer, which is a precaution for insurance
companies to retain customers.
33. Vehicle Safety Monitoring
Monitor the driver's real-time status, and timely
detect drivers' racks, fatigue and other conditions;
Provide drivers with warnings such as road
conditions, weather, and reminders to rest and
reduce the probability of accidents;
Help car owners to obtain timely accident
information and effectively reduce accident
damage;
Provide claims for vehicle insurance, identify
fraudulent behaviors, and reduce claims risk;
Provide vehicle maintenance/road assistance
agencies with data on the first time of the accident
and provide timely service to the owner.
34. Insurance Fraud
In 2017, there were 2,724 valid complaints received by the Hunan insurance regulatory bureau, of which 847 were
involved in auto insurance claims, among which the liability dispute was the largest and the dispute of auto insurance
claims became the focus. Our system can provide abundant accident data, such as vehicle trajectory, vehicle brake data,
vehicle steering, driver's mental state, etc., through accident reconstruction, provide strong evidence for car insurance
claims, also can effectively prevent the occurrence of insurance fraud case.
In the case of drunk driving, Insurance companies can be exempt from compensation by
rule .The driver lied that the others were driving.
In 34 vehicle insurance fraud cases, there are 6 cases based on drunk driving and the owner try
to cheat in various ways. The accident caused by drunk driving for MR Zhang, results in serious
damage to the vehicle. To obtain insurance claims, he will call up a person to the scene, falsely
claiming that he was the driver. The car network financial service platform can redeem the loss for
the insurance company through finding the evidence to identify the deceit based on user signaling
data and phone records.
case