System Architecture

FracAdapt Complete System Architecture

System Overview

FracAdapt is designed as a cloud-native, scalable predictive maintenance system for military vehicles, integrating terrain complexity with vehicle operational data. The system ingests diverse data sources, processes them through specialized analytical modules, and generates actionable insights for proactive maintenance and operational planning. Its architecture emphasizes modularity, scalability, and robust data security, leveraging AWS services for infrastructure and computational power.

Detailed Phases

1

Data Ingestion

Raw data, including satellite terrain elevation models (DEMs) and Condition-Based Maintenance (CBM) telemetry from military vehicles, is securely ingested into the system. This phase ensures data integrity and prepares it for initial processing.

• Satellite Data: SRTM, ASTER Global DEMs

• CBM Data: Vehicle sensor readings, fault codes, operational logs, GPS data

2

Data Pre-processing & Normalization

Ingested data undergoes cleaning, transformation, and normalization. Terrain data is georeferenced and segmented based on vehicle operational areas. CBM data is timestamp-aligned and filtered to remove anomalies.

• Georeferencing & Segmentation

• Time-series alignment & filtering

• Feature engineering for CBM data

3

Fractal Analysis Module

The pre-processed terrain elevation data is fed into the fractal analysis module. This module computes the fractal dimension (complexity score) for specific operational areas, quantifying terrain roughness.

• Box-counting method for fractal dimension calculation

• Complexity score assignment (2.0–3.0 scale)

• Spatial indexing of fractal scores

4

Predictive Modeling (LSTM & GAN)

Integrated LSTM neural networks predict component failure risks based on combined CBM data and terrain fractal scores. A Generative Adversarial Network (GAN) augments the training dataset with synthetic terrain profiles, improving model generalization.

• LSTM for time-series prediction

• GAN for synthetic data generation

• Dynamic model retraining based on new data

5

Output & Reporting

Predicted failure risks and associated insights are presented through a user-friendly dashboard and accessible via secure API endpoints for integration with existing DoD systems. This includes predictive maintenance alerts and operational recommendations.

• Interactive dashboard for visualization

• RESTful API for system integration

• Automated alert generation

Phase Transitions and Technical Implementation Details

Data transitions seamlessly between phases using asynchronous messaging queues and serverless functions for event-driven processing. This ensures high throughput, low latency, and fault tolerance across the system.

• Data Lake: Amazon S3 serves as the central data lake for raw and processed data.

• Orchestration: AWS Step Functions orchestrates multi-step data pipelines.

• Real-time Processing: AWS Lambda functions handle event-driven triggers for pre-processing and fractal calculations.

• Compute: Amazon SageMaker for model training and inference, utilizing GPU instances for LSTM and GAN. Amazon EC2 instances for specialized compute tasks.

API Endpoint Specifications

FracAdapt exposes a set of RESTful API endpoints, secured with AWS IAM and API Gateway, to allow external systems to query predictive maintenance insights and integrate FracAdapt's capabilities. All data returned is in JSON format.

• GET /predictive-maintenance/vehicle/{vehicleId} — Retrieves current and predicted failure risks for a specific vehicle.

• GET /predictive-maintenance/area/{lat}/{lon} — Provides aggregated risk assessment for vehicles operating in a given geographic area.

• GET /terrain-complexity/{lat}/{lon} — Returns the fractal dimension for a specified terrain coordinate.

• GET /alerts — Lists active predictive maintenance alerts across the fleet.

Dashboard Components

The FracAdapt dashboard, built using AWS Amplify and React, provides an intuitive interface for visualizing vehicle health, operational risks, and terrain insights. Key components include:

• Fleet Overview: Heatmap of vehicle health statuses across operational areas.

• Vehicle Detail View: Component-level failure probability, historical data, and maintenance recommendations for individual vehicles.

• Terrain Complexity Map: Interactive map displaying fractal dimensions for different regions.

• Alerts & Notifications: Real-time display of critical alerts and scheduled maintenance flags.

• Reporting & Analytics: Customizable reports on fleet readiness and operational efficiency improvements.

Technical Implementation (AWS Infrastructure)

FracAdapt is built entirely on Amazon Web Services (AWS) to ensure scalability, reliability, and security. The chosen services optimize for cost-efficiency while providing robust computational and storage capabilities.

Category	Services	Description
Data Storage	Amazon S3, Amazon RDS	Scalable object storage for raw data and data lake; Relational database for metadata and application data.
Compute & ML	AWS Lambda, Amazon SageMaker, Amazon EC2	Serverless functions for event-driven tasks; Managed service for ML model training and deployment; Virtual servers for specialized computational tasks.
Networking & Security	Amazon VPC, AWS API Gateway, AWS IAM	Isolated cloud resources; Secure API access; Fine-grained access control and authentication.
Messaging & Orchestration	Amazon SQS, AWS Step Functions	Decoupled microservices communication; Workflow management for complex data pipelines.

Deployment Phases

1

Pilot Deployment

Initial deployment to a limited scope, typically a single vehicle type or operational area, to validate core functionality and gather early feedback.

2

Feature Expansion

Phased rollout of additional features and integration with more data sources based on pilot results and stakeholder priorities.

3

Fleet-wide Rollout

Expansion to the entire military vehicle fleet, scaled to handle large volumes of data and concurrent users.

4

Continuous Optimization

Ongoing monitoring, performance tuning, and model retraining to adapt to new data and evolving operational requirements.

Complete System Flow Diagrams

The following diagram illustrates the integrated system flow, showcasing how raw data is transformed into actionable intelligence through FracAdapt's core technologies.

LSTM Predicts

GAN Augments

Fractal Score

Real Terrain

How Our Three Technologies Work Together

FRACTAL ANALYSIS

• Takes terrain elevation data from satellites

• Calculates complexity score (2.0–3.0 fractal dimension)

• Simple number that describes "how rough" terrain is

• Example: Plains = 2.1, Mountains = 2.8

LSTM NEURAL NETWORKS

• Learns from US Army CBM historical data (200+ vehicles)

• Discovers patterns: e.g., "When terrain=2.7 AND vehicle_age=40 months → 68% failure risk"

• Predicts future component failures for military vehicles based on terrain + vehicle state

• Trained on real military vehicle sensor data and fault codes

GENERATIVE AI (GAN)

• Generates synthetic terrain profiles that augment the training dataset

• Purpose: Enable the LSTM to learn from a more diverse set of terrain-vehicle interactions including edge cases not present in historical data

• Enables predictions for new military deployment areas globally

Data Limitations and Methodology

Important Note: repository data limitations

The repository data contains sensor readings from VIN00738 and base names from 8 Afghanistan locations, but lacks GPS coordinates or explicit location-sensor correlation. Our terrain analysis approach uses engineering assumptions based on Afghanistan geography and operational context, rather than direct terrain-sensor correlation from the repository data.

Terrain Integration Strategy:

• While the repository lacks GPS coordinates linking sensors to specific base locations, we utilize the 8 Afghanistan base names for their general terrain characteristics pending complete location data. This approach allows us to:

  • Apply known Afghanistan terrain types (Desert, Mountain, Desert/Mountain) to operational scenarios

  • Estimate elevation and stress factors based on geographic knowledge of these bases

  • Simulate terrain-specific sensor patterns for proof-of-concept validation

  • Demonstrate the methodology's potential when complete GPS-sensor correlation becomes available

This terrain-aware analysis serves as a bridge methodology that showcases the system's capabilities using available base context while acknowledging the need for precise location data in production deployment.

The Innovation

To our knowledge, FracAdapt represents the first systematic approach to combine military vehicle health data (CBM) with quantified terrain complexity analysis methodology. This proof-of-concept demonstrates how terrain-aware predictive maintenance could be implemented for military vehicles when complete location-sensor correlation data becomes available.