PITSTOP_05 Project Summary

Содержание

Слайд 2

Project Summary
Explanation of the Brake Model
Nissan’s Data Set
Summary of existing

Project Summary Explanation of the Brake Model Nissan’s Data Set Summary of
dataset
Applying Pitstop Brake Model & how it works
Success & Validation
Conclusion
Next steps / Phase 2 to further prove out the model

03
04
05
06
08
10
11
11

Слайд 3

Brake model is working
xxx
xxx
Comparison to mileage based shows a distinct advantage
xxx
xxx
Clear next

Brake model is working xxx xxx Comparison to mileage based shows a
steps to Achieve…. _____
Next steps / Phase 2 to further prove out the model

Слайд 4

TL;DR the existing dataset can be used for a brake model

Secret Sauce:

TL;DR the existing dataset can be used for a brake model Secret
Combining telematics, service records with big data and machine learning for example: (i) reliably detect all braking events, (ii)  manage cohorts to create correct statistical distributions for energy and for brake maintenance records (iii) Validating the model against maintenance records and known replacements 

From the existing list of Pitstop prognostic models, it seems that the brake model would be the most applicable to the Nissan dataset as it stands.

How The Brake Model Works 

Problem: If brakes wear out it is a safety and regulatory issue, but inspections mean downtime and expense

Em = kinetic energy of motion, where m = vehicle mass and V = speed of vehicle
Brakes wear because vehicles must dissipate (convert to heat) their energy of motion Em
The vehicles dissipating the most energy are wearing out their brakes fastest and should be targeted for inspection

Слайд 5

Detect when braking events occur.
Calculating a metric of brake usage per vehicle

Detect when braking events occur. Calculating a metric of brake usage per
- energy dissipation per unit distance driven (called the dissipation value).
Creating a frequency distribution of the above metric
Creating a distribution of brake services as a function of mileage driven
Mapping between the distributions to get an estimated mileage for brake
Replacement given the dissipation value

Слайд 6

The data has good attributes for Brake Predictions

High resolution data from a

The data has good attributes for Brake Predictions High resolution data from
small volume of vehicles (Engineering test fleet)

Measurements of physical components every week/month (brakes, tires)
CAN bus data including detailed attributes like brake pressure 
GPS & Acceleration data at high frequencies  (~1s or faster)
Speed, power terrain parameters; torque, coolant, engine oil temp, temp throttle position amongst others (~1s)
High mileage in short periods of time

Consistent datastreams from large volumes of vehicles (Customer vehicles)

GPS & Acceleration data at low frequencies (~30s)
Maintenance records includes brake measurements
Big Data Volume!  Thousands of vehicles with more than 2 brake measurements. 

Слайд 7

High resolution data from a small volume of vehicles (Engineering test fleet)

Trip

High resolution data from a small volume of vehicles (Engineering test fleet)
data does not add up to the total mileage driven. Ex. CTB531 has 10,000 km of accumulated mileage between the first brake measurement and last but there is only ~5000 km’s worth of trip data
There is not enough data volume, both length of time or number of vehicles to perform any meaningful accuracy/validation calculations
There are cases where either dates, or pad measurements are inconsistent. ex. brake pads increase in thickness over time based on the data

The data has some challenges for Brake Predictions

Consistent data streams from large volumes of vehicles (Customer vehicles)

30 second sampling frequency can miss out on relevant brake events, making the dissipation calculation less accurate
Service data dates and odometers don’t match up always. Sometimes we see reducing mileage over 1 year which signals incorrect data entry. 

Слайд 8

Applying the brake model - exploration on FET data

Expectation is satisfied with

Applying the brake model - exploration on FET data Expectation is satisfied
engineering test fleet which is that more energy dissipation in brakes => more wear between measurements (seen in pad thickness measurement) (CTB546)

Слайд 9

Front Left Inner
Front left Outer
Front Right Inner
Front Right Outer
Rear Left Inner
Rear left

Front Left Inner Front left Outer Front Right Inner Front Right Outer
Outer
Rear Right Inner
Rear Right Outer

Green line is the expected slope

Note: Higher dissipation values are to the left (dissipation is negative by convention)
Note: Data Timespan ~4 months

0.7

0.6

0.5

0.4

0.3

0.2

0.1

0

-7E=09

-6E=09

-5E=09

-4E=09

-3E=09

-2E=09

-1E=09

0

Слайд 10

The brake model is showing Success & validation

Showcase accuracies and strong signs

The brake model is showing Success & validation Showcase accuracies and strong
of success with the available dataset
Improvements of the model are better described as reliability rather than accuracy, since it means the model can be adjusted to avoid incorrect assumptions about different vehicle cohorts. However, if we think of accuracy as an average measure of agreement, such as R2, it will amount to the same thing.
Accuracy is not the same as precision. For example, it does not matter if measurements are made to the nearest 100 μ if the standard deviation of the measurement is 1.0 mm.

Слайд 11

Next steps to further prove out the brake model

High resolution data helps

Next steps to further prove out the brake model High resolution data
create accurate dissipation models. However to take advantage of the cohorts via big data there is not enough cases (< 20). This serves as a great start to show that energy dissipation directly correlates with brake wear (slide 7). 
However to be statistically relevant a validation test needs to incorporate more cases. The low resolution UIO data helps to put vehicles in cohorts and then plot them on a distribution. An R^2 measure can be made between each vehicle and the “average”. The average is defined as the mileage suggested brake replacement that is provided to every customer. 
The accuracy will be the error between the algorithms estimated brake replacement and the average case. 

Слайд 12

Steps to validate the model

Step 1: calculate the dissipation for each vehicle

Steps to validate the model Step 1: calculate the dissipation for each
and assign it to a cohort

Cohort distribution

Expected brake wear at mileage for =-1800

Table 1.

Table 2.

Table 3.

Expected brake wear at mileage for =-1000

Step 2: Each cohort will have a wear pattern which can estimate when a brake pad replacement will be needed. Note: vehicles can change between cohorts as additional data is captured

Слайд 13

Validate the model

Step 3: Comparison between each cohort (blue dotted line) and

Validate the model Step 3: Comparison between each cohort (blue dotted line)
the average (orange dotted line) will provide an accuracy measure. Cohorts that experience more wear will benefit from safety whereas those that experience less wear will benefit from receiving an accurate suggestion.

Слайд 14

Alert would be early. This leads to customer trust issues. “ The

Alert would be early. This leads to customer trust issues. “ The
dealer just wants me to do service that I don't need”.

Unsafe suggestion that would be too late. Could lead to an accident because of low brakes

64,000 KMs brake replacement suggested

0

-2

-4

-6

-8

-10

-12

-14

Слайд 15

Summary: Expected conclusion of phase 2

We expect phase 2 will prove that

Summary: Expected conclusion of phase 2 We expect phase 2 will prove
the brake model works on the UIO data and be able to showcase a percentage accuracy.
We will use the validation technique described in figure 9 (slide 9).
Based on Pitstops current brake model it seems the accuracy should be within this range x-y% which would be the target.

Слайд 16

Nissan Roadmap to Additional Predictions 

Nissan Roadmap to Additional Predictions

Слайд 17

Pitstop’s current Models
How the Pitstop data engine works
Current Pitstop Models / Data

Pitstop’s current Models How the Pitstop data engine works Current Pitstop Models
Requirements 
Custom Models - to solve specific problems
What Data Nissan Has today:
Positive attributes and what can be done with it today
Challenges & Gaps
Recommendations Priorities for how to fill the data gap
Suggested Road Map

18
19
20
21
22
23
24
25
26

Слайд 18

Time series sensor data

Repair order data

Pitstop insights

Time series sensor data Repair order data Pitstop insights

Слайд 19

Battery

Remove no start scenarios
Reduce electrical failures Examples include: Battery, Alternator, Starters, Parasitic

Battery Remove no start scenarios Reduce electrical failures Examples include: Battery, Alternator,
loads etc..

Engine Control

Improve Fuel Efficiency
Manage Engine Fault Priorities
Examples include: Spark plug, Wires, Injectors, Timing, Crank sensor, O2 sensor, Exhaust, Water-pump etc..

Emissions

Reduce Diesel Lockouts
Maintain emissions system before catastrophic failures
Examples include: DEF, DPF, EGR, Air filter, Hose leaks, Pressure leaks, EVAP issues, Turbo leaks etc..

Brakes

Improve vehicle safety
Brake wear analysis across entire fleet
Examples include: Brake pads, Rotors, hydraulic, pneumatic etc..

Слайд 20

Recommendation to extract further value

Underinflation &
Leakage

Load & Utilization
Monitoring

Pad Wear
Insights

Tread Wear

Recommendation to extract further value Underinflation & Leakage Load & Utilization Monitoring

insight

Tire
Tracking

Alignment
Monitoring

Sensata Technologies (NYS; ST) became a strategic investor (September 2020) is a leader in sensing solutions and the global market share leader in TPMS 
Strategic initiatives include a brake and tire prediction solution for the transportation industry. 

Слайд 21

Additional Algorithm Details 

Problem
Delivery Van Sliding Door was not intended to open and

Additional Algorithm Details Problem Delivery Van Sliding Door was not intended to
close 100’s of times per day - causing bracket failure and eventually body panel damage

Solution:
Utilizing a couple readily available telematics PIDs and repair order information, Pitstop can create a custom algorithm to predict when this failure will occur -avoiding a significant body panel repair cost

Слайд 22

The Nissan data has good attributes for models

High resolution data from a

The Nissan data has good attributes for models High resolution data from
small volume of vehicles (Engineering test fleet
Measurements of physical components every week/month (brakes, tires)
CAN bus data including detailed attributes like brake pressure, Speed, power terrain parameters;torque, coolant, engine oil temp, temp throttle position amongst others 
GPS & Acceleration data at high frequencies  (~1s or faster)
High mileage in short periods of time
Consistent datastreams from large volumes of vehicles (Customer vehicles)
GPS & Acceleration data at low frequencies (~30s)
Maintenance records as long as the customer arrives at the dealer
Big Data Volume!  10’s of thousands of vehicles

Слайд 23

The dataset overall does have challenges & gaps

The dataset consists of telematics

The dataset overall does have challenges & gaps The dataset consists of
generated and service data
acceleration, gps at 30 second intervals and odometer
Service records from 30K or so vehicles. 
With the current state of telematics data alone solutions related to route optimization and driver risk can be implemented. 
With service data alone can assist with getting ahead of defects or looking at inventory and service lane statistics. You can build mileage based prediction models as well.
A value item to be extracted from both data sets is a brake model! 
Additional models that maybe extracted include brake and tire wear. These will require extensive analysis and research before being certain that the reliability and accuracy of the models are suitable. 

Слайд 24

Recommendation to extract further value

Start by asking what types of value propositions

Recommendation to extract further value Start by asking what types of value
are most important to the market. 
For example if it’s clear that Nissan wants to have models for as many components as possible, then the strategy requires deep edge to cloud implementation. This is capability Pitstop has in the market. 
If Nissan decides they want to focus on brakes, batteries and tires then the roadmap will just require specific time-series sensors to be enabled in the data stream. 
Pitstop suggests taking a fully integrated approach in order to take advantage of rapid software and data science iteration cycles. New problems will emerge that you cannot currently predict, and hence you need a flexible infrastructure to quickly build new models. This will payback returns as customer satisfaction will improve as well as reduction of recall and warranty costs. 

Слайд 25

TL;DR the existing dataset can be used for a brake model

From the

TL;DR the existing dataset can be used for a brake model From
existing list of Pitstop prognostic models, it seems that the brake model would be the most applicable to the Nissan dataset as it stands.

How The Brake Model Works 

Problem: If brakes wear out it is a safety and regulatory issue, but inspections mean downtime and expense

Secret Sauce:
Combining telematics, service records with big data and machine learning for example: (i) reliably detect all braking events, (ii)  manage cohorts to create correct statistical distributions for energy and for brake maintenance records (iii) Validating the model against maintenance records and known replacements 

Em = kinetic energy of motion, where m = vehicle mass and V = speed of vehicle
Brakes wear because vehicles must dissipate (convert to heat) their energy of motion Em
The vehicles dissipating the most energy are wearing out their brakes fastest and should be targeted for inspection

Слайд 26

Steps required to track Brake Wear

Detect when braking events occur.
Calculating a metric

Steps required to track Brake Wear Detect when braking events occur. Calculating
of brake usage per vehicle - energy dissipation per unit distance driven (called the dissipation value).
Creating a frequency distribution of the above metric
Creating a distribution of brake services as a function of mileage driven
Mapping between the distributions to get an estimated mileage for brake
Replacement given the dissipation value

For more in depth information: Paper on Brake Wear Model

Слайд 27

Custom Algorithm Example  

Problem: Delivery Van Sliding Door was not intended to open

Custom Algorithm Example Problem: Delivery Van Sliding Door was not intended to
and close 100’s of times per day - causing bracket failure and eventually body panel damage

Solution: Utilizing a couple readily available telematics PIDs and repair order information, Pitstop can create a custom algorithm to predict when this failure will occur -avoiding a significant body panel repair cost

Слайд 28

Start by asking what types of value propositions are most important to

Start by asking what types of value propositions are most important to
the market. 
For example if it’s clear that Nissan wants to have models for as many components as possible, then the strategy requires deep edge to cloud implementation. This is capability Pitstop has in the market. 
If Nissan decides they want to focus on brakes, batteries and tires then the roadmap will just require specific time-series sensors to be enabled in the data stream. 

Recommendation to extract further value

Pitstop suggests taking a fully integrated approach in order to take advantage of rapid software and data science iteration cycles. New problems will emerge that you cannot currently predict, and hence you need a flexible infrastructure to quickly build new models. This will payback returns as customer satisfaction will improve as well as reduction of recall and warranty costs. 

Слайд 29

Start by asking what types of value propositions are most important to

Start by asking what types of value propositions are most important to
the market. 
For example if it’s clear that Nissan wants to have models for as many components as possible, then the strategy requires deep edge to cloud implementation. This is capability Pitstop has in the market. 
If Nissan decides they want to focus on brakes, batteries and tires then the roadmap will just require specific time-series sensors to be enabled in the data stream. 

Recommendation to extract further value

Слайд 30

Recommendation to extract further value

67%

Recommendation to extract further value 67%

Слайд 31

Start by asking what types of value propositions are most important to

Start by asking what types of value propositions are most important to
the market. 
For example if it’s clear that Nissan wants to have models for as many components as possible, then the strategy requires deep edge to cloud implementation. This is capability Pitstop has in the market. 
If Nissan decides they want to focus on brakes, batteries and tires then the roadmap will just require specific time-series sensors to be enabled in the data stream. 

Recommendation to extract further value

Слайд 32

Recommendation to extract further value

Recommendation to extract further value
Имя файла: PITSTOP_05-Project-Summary.pptx
Количество просмотров: 42
Количество скачиваний: 0