# Project plan
> Fil Rouge Project

**Contents**
[TOC]
1. Context
1. Project context
1. Project objectives
1. Implementation methodologies
1. Planning
1. Technical resources
1. Human resources
1. Risk analysis
1. Conclusion
1. References

References

## Introduction:low_brightness:
Currently enrolled in a one-year master’s degree in Artificial Intelligence at Telecom Paris, a "fil rouge" project is at the core of this training, which lasts seven months. Composed of four students, our group is working on a major subject for Alstom: causality.
This document aims to present our plan to meet the objectives set by Alstom. The goal is to discover causalities on train mobility data.
To do this, we will first recall the context and the objectives of the project, then the methodologies to meet them, including the presentation of the schedule, technical and human resources, and finally a risk analysis.
We have chosen a collaborative work method within the group, which will allow us to increase our skills on the causality subject and on the handling of the Tigramite tool. Our work on Tigramite will require continuous iterations throughout the project.

### Context

This part presents the collaborative environment for Fil Rouge on the theme of Causal discovery using mobility data.

#### Project context :bulb:

As part of this collaboration, we will tackle causality applied to real subsystem event data from the Class 390 high speed train from the UK (i.e., the West Coast Main Line).
The Train Control Management System (TCMS) is an onboard computer that logs all the subsystem events. For example, when the driver starts a journey, the Traction subsystem logs the event of movement, and when the driver stops the train, the Brake subsystem logs the requested braking effort.

### Data :floppy_disk:

The “_tcms_stripped” file is an SQLite database that contains some events from the Class 390 WCML fleet collected in 2016 (4.9M records).
It contains two tables: “service” and “event”, which are described as follows:

• “service” table:
◦ ID: index.
◦ Date: timestamp, e.g., “2016-01-18 00:38:33”
◦ Unit: train number ID, e.g., “390001”
◦ Code: event number ID, e.g., “8356”
◦ Lon/Lat (Longitude/Latitude) : GPS location, e.g., “-2.20131993293762”, “53.4622993469238”
• “event” table:
◦ Code: event number ID, e.g., “8356”
◦ Name: Name of the event, e.g., “Saloon Door Interlock Proved On All Cars - Car A”
◦ Function: subsystem related to the event, e.g., “Door System”

## Causal Research

The different subsystems in the train are related.

For example:
* Car type (e.g., motor vs trailer) explains the degradation of brake pads.
* Temperature affects the degradation of pantograph carbon strips (especially in the winter when the catenary may be frozen).

From these relationships there are many questions that could have a causal explanation:

* If HVAC uses the heater in the winter and the AC in the summer, how does this affect the energy consumption? Are there any other variables to be accounted for?
* Excessive braking could cause a wheel flat, and therefore the Wheel Slip Protection subsystem (WSP) prevents this from happening.

Could we see counterfactuals here? (Would the train have slipped had the driver been easier on the brakes?)

## Project objectives :dart:

> Correlation Does Not Imply Causation

Traditional Machine Learning approaches essentially model the correlation of variables, and therefore, they lack the representational capacity to explain intricate interrelationships like the ones described above.
Following the adage in statistics that “correlation does not imply causation”, Causal Inference (CI) is expected to shed light into this topic.

The environments to frame this CI research are described as follows:
* Locomotion: Traction + Brakes (blended operation)
* Indoors: HVAC + Door (there are humidity and temperature swings when the doors get opened)
* Bogie (axle box + gearbox + tilt + wheel + panto): Tilt system + WSP
* Energy: transformer + auxiliary converter

##### Hypothesis:
Could these causal relationships be observed in the available event data? Could they be modelled? And exploited? How would can this be approached?
To tackle this research, the Tigramite tool is proposed [1].
Tigramite provides several causal discovery methods for timeseries data. Causal discovery aims to create graphical models that display causal associations [2]. The Fil Rouge collaboration aims to learn and understand this tool, and to apply it to the provided TCMS data. As a result, different causal models that explain the same data will be generated.

## Deliverables :card_index:

The following objectives must be completed through:
* a written technical report,
* software (e.g., notebooks…),
* usable trained models.
* Structure of causal (sub)graphs based on the TCMS data:
* The delivered (sub)graphs for each environment will be compared and discussed.
* The different discovery methods should result in different models.
* The reasons to explain their differences should be understood.
* Their correctness and/or usefulness will be evaluated with the “event/name” field. (Initially, this table column will be empty. By the end of the collaboration, its details will be provided to understand the underlying subsystem relationships.)
* To conclude the collaboration, the team should be able to assess which of the tools under evaluation best explains the data from a causal perspective.
* Additionally, the scope of the collaboration can be extended to other Causal Inference tools like DoWhy [3], which may serve as a baseline for the comparisons (this tool does not account for the temporal causal precedence criteria).
* Implementation methodologies

## Planning :calendar:

The project started on 28 November 2022 and delivery is scheduled for 22 June 2023, which is just under 7 months. To achieve the objectives of the project, we decided to divide the project into two phases.
The first phase named “Data exploration and analysis” is composed of four sub-phases:
* Understanding the data
* Data preparation
* Milestone: Analysis report
* Adjustment after the feedback meeting

Thus, to understand the data, we will load the data and made some observations. This will allow us the make simple analysis between the data. After, we plan to prepare the data by cleaning the data and generate different time-series. During the cleaning phase, we want to identify spurious data and verify with Alstom if this spurious data is a common way of dealing with missing data at Alstom. In any case, we will take this value into account by generating different time-series with and without spurious data. To gain time, we will divide the data by train, thus each of us will be assign to a certain amount of train. Thus, we think that generating different time-series, for example by train, by week, by month, will help us to better understand causality at different scale. Once completed, we plan to report our analysis to Alstom during a milestone. This will give Alstom an overview of our work and allow us to get their feedback. Finally, to consider their feedback, we will allow some time after the milestone to implement them.

The second phase named “Causal discovery in data” is an iteration phase composed of five sub-phases:

* Model implementation
* Generating a causal graph
* Graph interpretation
* Milestone: Analysis report
* Adjustment after the feedback meeting

We choose to iterate on this phase, because we think there are a lot of way to do causal discovery. Thus, we decide to dedicate one method by iteration. Firstly, we will decide on which aspect of the data we want to focus on, short-time causal effect, long-term causal effect and so on. After choosing the aspect, we will generate a causal graph and interpret it by checking whether this causal graph represents the overall reality on several trains or whether this causal graph represents a possible anomaly in the normal operation of a train. Then, we will plan to report our progress to Alstom during a milestone. This will give Alstom an overview of our work and allow us to get their feedback. Finally, to consider their feedback, we will allow some time after the milestone to implement them.

To achieve all these phases, we plan to read some bibliography on causal discovery by J. Runge and from other sources. So, each of us will take a turn to read a scientific paper and made a short report to present to the others. These reports will be submitted to Alstom to facilitate understanding of the methods used.

To conclude, we plan to start writing the final report from mid-January so that it is not written at the end of the project.

#### Resources

##### Technical resources: :computer:
In order to successfully carry out this project, we will require certain technical resources. One of the most important of these is computational resources, specifically a cluster of computers that is capable of handling intense calculation applications or a serverless calculation service such as Databricks Workstation.

In addition to computational resources, we will also need a local system for all team members.

##### Human resurce :construction_worker:

We also need to consider the human resources required for this project. Our team consists of four data/ML engineers, each with a commitment of six hours of work per week. We also have two data scientist/faculty members and one R&D executive, each with a workload of one hour per week.

In summary, the necessary technical resources for this project include computational resources, a local system for all team members, and a team of data/ML engineers, data scientists/faculty members, and an R&D executive.

##### Estimation :heavy_dollar_sign:

| Human Resource | Number of People | Per Hour | Work Load Estimation per Month | Cost Estimation per Month | For Six Month |
|----------------|-----------------|----------|-------------------------------|---------------------------|----------------|
| Data/ML Engineer | 4 | €100.00 | €24.00 | €9,600.00 | €57,600.00 |
| Data Scientist/Faculty Member | 2 | €300.00 | €3.00 | €1,800.00 | €10,800.00 |
| R&D Executive (Alexander) | 1 | €300.00 | €3.00 | €900.00 | €5,400.00 |
| Calculation Cluster | | || €1,081.00 | €6,486.00 |
| Working Space | 4 | | €40.00 | €160.00 | €960.00 |
| **Total Cost** | | || €13,541.00 | **€81,246.00** |

#### Risk analysis :red_circle:
We have established a risk analysis that could impede the achievement of the objectives.

| Identified risk | Proposal for reducing the risk 2 | Comment
| -------- | -------- | -------- |
| Poor assumption | <br> • Scientific documentation: study of articles, <br> • Regular meetings with our internship tutors. | - |
|Data security|<br> • keep the database local on our computers, even if we use the school's resources.|
|Workload| <br> • Compliance with the GANTT schedule <br> • Working time basis: 6 hours per week.

## Conclusion
In conclusion, this project aims to apply causality to real subsystem event data from the Class 390 high-speed train in the UK. While this field is new, there is a reach literature that needs to be reviewed to inform our approach. As part of our master's program, we will be working on this project alongside other commitments, with a minimum of six hours per week dedicated to this project.
Given the nature of this R&D project and the fact that it is a new area of study for us, we expect there to be a high degree of uncertainty. In order to effectively navigate this uncertainty, we plan to adopt an iterative approach to project organization, balancing a clear timeline with the need for flexibility and testing.
We are aware of the limitations we face regarding resources but are highly motivated and excited for this learning journey.

---

References :books:

[1] https://github.com/jakobrunge/tigramite/
[2] Naser, M. Z., “Causality, Causal Discovery, and Causal Inference in Structural Engineering”. 2022
[3] https://www.pywhy.org/dowhy/v0.8/