The AI Black Box
Designing for Trust, Transparency and Action.
Types of AI Machine Learning
Unsupervised
The algorithm analyzes the internal structure and mathematical distribution of the data.
Supervised
The algorithm learns direct mapping function between the input and the result.
Topic to pick
In Machine Learning. Supervised models offer a clear path to results because they rely on predefined answers. however, unsupervised Learning remains a "Black Box" a Challenge where it is notoriously difficult for users to understand why a specific anomaly was flagged.
My project embraces this challenge by focusing on Unsupervised Learning, specifically aiming to pull back the curtain on its complexity. I am developing a framework that translates these 'Black Box' outputs into transparent, intuitive insights that even a non technical stakeholder can trust and understand.
Anomaly Detection
The Topic
Why
Reducing unplnned downtime in the supply chain directly impacts the bottom line by preventing production bottlenecks aand delivery delays. Moving away from "fix it when it breaks" (which is costly and disruptive) to "fix it before it fails"
What
Anomalies are data points that stand out amongst other data points in the dataset and not confirm the normal behavior in the data. This data points or observation deviate from the dataset's normal behavior patterns
Project Focus
The Stakes
Bhopal Disaster
Background Examples
Warnings were ignored, safety systems were disabled, and catastrophic delays ensued not due to lack of data, but a tragic failure of system
The Trigger Anomaly: High pressure and temperature inside tank E610, far exceeding safe limits
Alarm Fatigue: The primary pressure alarms were known to frequently sound false warnings, leading operators to ignore or silence them routinely
Motivation
Why UX is the Missing Link
2.
Operators often override or ignore poorly explained alerts, leading to delayed or incorrect decisions.
3.
User experience has been neglected in the "last mile" intelligence delivery layer
1.
Trust and transparency are critical safety enablers in high stakes environments.
Overall Problem Statement
Users cannot easily choose which data to include or define specific timeframes for analysis (CRUD and Configuration).
Lack of Control
The "Data Mystery"
There is no guidance or "hints" to help non experts select the right sensor data to train an effective model
Short title
Bhopal Disaster
In modern industrial settings, Anomaly Detection is vital for keeping assets running. However, most systems fail because their AI models are "Black Boxes" they tell a maintenance engineer that something is wrong, but never why.
When engineers don't understand the "WHY," they experience alarm fatigue and lose trust in the system. This leads to ignored alerts, delayed repairs, and increased operational risk.
The Trust Gap
Engineers have no clear way to validate results, making it impossible to distinguish between a real machine failure and simple data "noise"
Static Models
Systems often fail to update as machine conditions change, making the AI's insights outdated almost immediately.
Insights often stay stuck in a dashboard rather than automatically triggering a "Closed Loop" action (like a Work Order) in the System
Broken Loops
Pain Points
Goal and Scope
The core goal is to define, prototype and validate a foundational Human AI Interaction Framework that can build user trust in automated systems. To pioneer a human centered design approach that makes AI explainable, Controllable and Actionable
Universal Impact
This research is designed to go beyond the supply chain. the resulting AI Experience (AIX) patterns are domain agnostic, providing a blueprint for transparency and trust in any high stakes field, form industrial assets to medical diagnostic and scientific data analysis
Critically, this research is designed to cover the entire operational journey
From the initial design and creation of the model > review the model > feedback > retrain > use the Model > feedback > Retrain > Continuous use of feedback. this establishes a truly End-to-End, Closed loop system.
Testbed
We use APM as the rich, high stakes testing ground, working with multi modal data (sensor, visual, reports) and real world operational experts.
Three Interaction Subjects
The Project is organized around the three critical phases of human interaction with an anomaly detection system
Control in Model Creation
and Configuration
Trust in Decision Review
Continuous Learning
