Back to Home
UX Design • 2019

Transforming Industrial Operations with AI-Powered Insights

Cogn8 is a leading industrial AI platform -” A comprehensive redesign focused on data liberation, anomaly detection, and iterative intelligence to support operational excellence.

77%
Fewer False Positives
40%
Time Saved
92%
Detection Accuracy
Role
Lead Product Designer
Timeline
1 Year (2019)
Focus Area
Industrial IoT / AI

The Challenge: Detecting the "Silent" Failures

In wind energy operations, standard monitoring systems often miss subtle anomalies that indicate looming failures. A classic example is "Sudden Low Turbine Speed" despite "Steady Wind Conditions".

The Problem: Traditional threshold-based alarms trigger only when values cross extreme limits. By then, it's often too late. We needed a system that understands the relationship between variables-”recognizing that low RPM is only a problem when the wind is blowing strong.

Unwanted Behavior Graph

Unwanted Behavior

Anomaly: Turbine speed (Black) drops suddenly while Wind speed (Orange) remains steady.
Result: Potential gearbox failure or efficiency loss.

OK Behavior Graph

OK Behavior

Normal Operation: Turbine speed drops, but it correlates with a drop in Wind speed.
Result: System behaving as expected.

Strategic Vision: Human-Machine Collaboration

The Gap

Machine learning excels at monitoring vast data streams for abnormalities but lacks context. Humans understand context but cannot monitor 24/7.

The Strategy

"Glass Box" Collaboration. The AI acts as a transparent "Watchdog" flagging issues with evidence, while the Engineer acts as the "Judge" validating findings to train the system.

Automated Observability Human-in-the-Loop Trustworthy AI

Wind Turbine Data Analysis

Our AI analyzes three critical metrics ”RPM, Wind Speed, and Torque”to classify operational states and detect anomalies before they cause failures.

Core Data Components

1

RPM

Rotations Per Minute - measures turbine rotational speed

Critical for detecting performance anomalies and mechanical stress

2

Wind Speed

Environmental wind velocity affecting turbine operation

Key input for expected performance benchmarking

3

Torque

Rotational force applied to the turbine drivetrain

Indicates mechanical load and potential component stress

Anomaly Classification Framework

Normal State

RPM & Low Wind

Expected operational conditions without system stress

Continuous monitoring

Turbine Issues

High Torque, RPM & Low Wind

Abnormal load or control issues within turbine system

Medium Priority

Parts Issues

High Torque & Wind, Low RPM

Wear or malfunction in bearings, gearboxes, or moving parts

High Priority

Blade/System Issues

High Wind, Low RPM & Torque

Blade damage, system faults, or sensor misalignments

High Priority

Research & Discovery

Research Strategy

Directing Discovery. I orchestrated a multi-method research strategy to uncover the "why" behind the resistance to AI. By combining ethnographic field studies with quantitative analysis, I identified that the core barrier was not "usability" but "trust."

15

In-depth interviews across roles

3

Industrial facility observations

60+

Hours of contextual inquiry

120

Survey responses analyzed

Key Research Insights

Weekly Time Loss Per Engineer (Hours)

Analysis of how industrial engineers spend non-productive time due to system inefficiencies

Impact: An average of 47 hours per week wasted across non-productive activities ”representing 118% productivity loss per engineer.

User Satisfaction with Current Systems

Prioritized Pain Points

Critical

Cannot trust data accuracy across systems

Critical

Alert fatigue from excessive false positives

High

No visibility into why AI flags anomalies

High

Complex navigation between 10+ systems

Medium

Lack of historical context for decisions

Key Research Insight

"We don't need more data we're drowning in it. What we need is confidence that what the system is telling us is actually true. I need to understand why it's flagging something, see the evidence, and make the call myself."

-” Senior Maintenance Engineer, Offshore Energy Platform (15 years experience)

This insight revealed a fundamental shift needed: from building "smarter AI" to building "more transparent AI" that empowers human decision-making rather than attempting to replace it.

Heuristic Evaluation & Remote Research

In parallel with field studies, I conducted comprehensive heuristic evaluations of existing industrial SCADA systems and remote research sessions with operators unable to participate in on-site studies due to shift schedules and geographic constraints.

Heuristic Evaluation Findings

Severity 1
Visibility of System Status

No clear indicators of AI confidence or data quality.

Severity 1
Error Prevention

No warnings before critical threshold changes.

Severity 2
Recognition vs Recall

Users must memorize sensor IDs instead of meaningful names.

Severity 2
Flexibility

No customization options for different user roles.

Remote Research Insights

Conducted 8 remote sessions via video calls with screen sharing.

Reached offshore platform workers and night-shift operators.

Observed real-time workflows during actual alert investigations.

Uncovered mobile usage patterns and connectivity challenges.

Ideation & Design Strategy

Based on research insights, I established design principles grounded in explainable AI, human-centered design, and cognitive load theory. Each principle directly addresses a documented user need.

Core Design Principles

Proactive Intelligence

Shift from reactive alerts to predictive warnings giving teams time to prevent issues rather than scrambling to fix them

Predictive warnings reduced emergency responses by 35%

Trust Through Verification

Enable users to validate AI conclusions with source data and confidence metrics-”building trust incrementally

Research showed only 23% initially trusted automated detection

Human-in-the-Loop Design

Design feedback mechanisms that let users teach the system creating a collaborative intelligence partnership

Systems with feedback loops showed 68% better long-term accuracy

Progressive Disclosure

Surface critical information first, allow deep-dives on demand reducing cognitive load while preserving access to detail

Users reported 40% faster decision-making with layered information

Context Over Data

Prioritize actionable insights over raw metrics answer 'what should I do' before 'what happened'

Field workers prefer action recommendations 4:1 over data dashboards

Explainable AI

Make AI reasoning transparent and understandable show users why the system flagged an anomaly, not just that it did

Based on finding that 73% of users override unexplained AI recommendations

Feature Mapping & Prioritization

I organized potential features into thematic groups using affinity mapping with stakeholders. Each category addresses specific pain points from research, with features prioritized using the RICE framework (Reach, Impact, Confidence, Effort).

Transparency & Trust

  • Confidence scores for every AI prediction
  • Visual reasoning paths showing detection logic
  • Source data traceability for verification
  • Historical accuracy metrics per anomaly type

Contextual Intelligence

  • Asset relationship visualization
  • Temporal pattern analysis
  • Cross-system data correlation
  • Maintenance history integration

Actionable Insights

  • Recommended actions with impact assessment
  • Priority-based anomaly triage
  • Risk-level classification
  • Time-to-action estimates

The Solution: Agentic Automation

We evolved from "Human-in-the-Loop" to "Human-on-the-Loop". The Agentic AI autonomously detects anomalies, identifies the root cause, and executes corrective actions in milliseconds notifying the engineer only for critical oversight.

Turbine #42: Autonomous Correction Active

ID: AUTO-ACT-2024-892 - Confidence: 99.8%

Demo Telemetry: Real-Time Analysis

Wind Speed (m/s)
Turbine RPM
Torque (kNm)
Cogn8 Autonomous Agent

10:42:05 AM

Anomaly Detected: RPM drop (-72%) in steady wind.

Root Cause Identified

Diagnosed Gearbox Slippage (99.8% confidence).

10:42:06 AM

Action Executed:

  • Adjusted blade pitch to -5°
  • Increased torque load by 12%
  • Stabilized RPM output
System Stable

Performance Comparison

Quantitative Impact: Before vs After

Key metrics 3 months before vs 6 months after implementation

After
Before

User Satisfaction Analysis

Multi-dimensional satisfaction scores (n=43 users)

After
Before

AI Performance Metrics

Anomaly Detection Performance Over Time

AI system learning from human feedback-”declining false positives demonstrate improving precision.

-37.5%

False Positives

+72.7%

Precision

Data Quality Evolution

Multi-dimensional quality metrics improving as AI clustering and structuring mature.

58% - 92%

Overall Quality

+59%

Improvement

AI Capabilities & Techniques

Machine Learning

Unsupervised Clustering

K-means and hierarchical clustering to identify patterns in operational data and group similar anomaly types

Natural Language Processing

LLM-Powered Structuring

Large Language Models parse unstructured maintenance logs, sensor notes, and documentation into structured knowledge graphs

Deep Learning

Real-Time Anomaly Detection

Time-series analysis with LSTM networks and statistical methods to identify deviations from normal operational patterns

Graph AI

Knowledge Graph Integration

Semantic connections between assets, sensors, maintenance history, and operational context for enriched insights

Data Science

Multi-Source Data Fusion

Bayesian networks combine data from SCADA, IoT sensors, ERP, and maintenance systems with confidence weighting

Reinforcement Learning

Self-Aware Learning Loop

System monitors its own prediction accuracy and adapts thresholds based on human feedback and outcome validation

AI-Powered Classification

10,000+

Historical Incidents analyzed

94%

Classification accuracy

3.2min

Average diagnosis time

Understanding Self-Aware AI

The term "self-aware" in this context doesn't imply consciousness-”it refers to an AI system's ability to monitor its own performance, recognize its limitations, and adapt its behavior based on feedback loops. This includes:

  • Detection: Identifying when predictions may be uncertain or unreliable
  • Reflection: Analyzing which types of anomalies it detects accurately vs. those requiring human validation
  • Adaptation: Adjusting detection thresholds and confidence levels based on historical accuracy

Financial Impact

Value derived from: reduced downtime ($4.2M), productivity gains ($2.8M), maintenance optimization ($1.2M), and improved asset lifespan ($0.5M)

$8.7M
Annual Value
4.2x
ROI
18
Weeks to ROI

Results & Impact

The implementation of Cogn8's AI-driven solution delivered measurable improvements in efficiency, safety, and data reliability across global operations.

False Positives
-77%

Reduction in noise

Time Savings
40%

Per engineer / week

Detection Accuracy
92%

True positive rate

Unplanned Downtime
-35%

Avoided shutdowns

User Satisfaction
89%

Up from 35%

Annual Value
$8.7M

Generated per facility

Design Showcase

Full-Stack Observability Architecture

Global Operations Dashboard

A centralized command center providing real-time visibility into global asset health, operational efficiency, and critical alerts across all facility locations.

Automated Root Cause Analysis Flow

Individual Turbine AI Analysis

AI-driven diagnostics overlaying predictive insights on turbine schematics, highlighting potential failure points before they impact proactive maintenance.

AI-Driven Insight Wireframes

Individual Turbine Manual Analysis

Detailed telemetry view allowing engineers to drill down into specific sensor data, comparing historical benchmarks against real-time performance metrics.

Smartscape Visualization

Individual Turbine Manual Analysis Hover on Chart

Interactive data visualization enabling precise cursor tracking on time-series graphs to inspect exact values and correlate anomalies with specific timestamps.

Cognitive AI Dashboard

Individual Turbine Manual Analysis Take Action

Direct action interface for maintenance teams to schedule repairs, order parts, or trigger emergency protocols directly from the analysis dashboard.

Problem Evolution View

Individual Turbine Manual Analysis Anomaly Classification

A feedback loop interface where experts classify detected anomalies, enriching the training dataset and improving the model's future accuracy.

Log & Trace Explorer

Individual Turbine Manual Analysis Marked

Visual confirmation of user-validated anomalies, creating a clear audit trail of human-ai collaboration and decision making.

Reinforcement Learning Loop

Individual Turbine Manual Analysis Model Trained

Post-training verification view showing how the updated model interprets historical data with improved precision and reduced false positives.

Conclusion

Through the cycle of detection, correction, and reinforcement, Cogn8 has laid the foundation for a self-aware AI. This intelligent system not only identifies anomalies but reflects on its own performance, improving over time and empowering the workforce to act with unprecedented confidence.

Project Summary

Role Lead Product Designer
Timeline 8 Months
Tools Figma, Miro

Implementation Lessons

Adopting AI-driven observability is a journey. Here are the key takeaways from our implementation process.

Data Quality is Paramount

AI is only as good as the data it's fed. Ensuring full stack coverage with unified data collection was critical for accurate root cause analysis.

Culture Shift Required

Moving from reactive firefighting to proactive optimization requires a mindset shift. Trusting the AI's answers took time but paid off.

Automation is the Goal

Observability without automation is just noise. Connecting Causal AI to remediation workflows for auto-remediation was the game changer.

Security at Speed

Integrating App Security into the observability platform allowed us to detect vulnerabilities in runtime without slowing down releases.

Skills Demonstrated

Agentic AI Autonomous Systems Causal Inference Self-Healing Architectures User Research & Testing Design Systems Interaction Design Data Analysis Accessibility (WCAG)