PragyaKosh: Predicting Battery Failure Before It Happens

1. Background

Batteries power everything from off-grid solar installations and robotics to uninterruptible power supplies (UPS). However, knowing the true operational health and state of a battery system in real time can be surprisingly difficult. In many everyday setups, users only realize a battery is degraded or failing when the connected device unexpectedly shuts down or suffers a drop in performance.

We started thinking about this project to bridge that visibility gap. Our goal is to design an accessible monitoring platform that gives users clear, continuous insights into their power systems. Rather than simply displaying a static charge percentage, the proposed platform will actively help users understand how their system behaves over time and alert them to potential issues before a failure occurs.

2. Where Conventional Systems Fall Short

Most entry-level monitoring setups rely on simple voltage meters. While checking voltage is straightforward, it provides an incomplete picture of battery health:

Voltage Fluctuations: Battery voltage drops under heavy load and changes with temperature, making voltage alone an unreliable metric for estimating remaining energy during active use.

• Voltage Fluctuations: Battery voltage drops under heavy load and changes with temperature, making voltage alone an unreliable metric for estimating remaining energy during active use.

Rigid Thresholds: Basic monitors trigger alerts only after a fixed safety limit is breached (for example, when a battery reaches a dangerously high temperature). They miss subtle early-stage warnings, such as a battery heating up faster than expected under light loads.

• Rigid Thresholds: Basic monitors trigger alerts only after a fixed safety limit is breached (for example, when a battery reaches a dangerously high temperature). They miss subtle early-stage warnings, such as a battery heating up faster than expected under light loads.

Limited Accessibility: Industrial-grade monitoring solutions that offer advanced diagnostic features are typically expensive, complex, and closed off from customization by makers, researchers, or small-scale developers.

• Limited Accessibility: Industrial-grade monitoring solutions that offer advanced diagnostic features are typically expensive, complex, and closed off from customization by makers, researchers, or small-scale developers.

Our proposed system aims to address these practical limitations by combining multi-parameter sensing, low-power telemetry, and smart pattern analysis into an open, adaptable hardware design.

3. Our Proposed System / Setup

We are planning an adaptable hardware monitoring module that measures core electrical and environmental variables, processes the data locally, and relays insights wirelessly to a custom mobile interface.

+-------------------------------------------------------------------+
|                          SENSING LAYER                            |
|       Current  |  Voltage  |  Temperature  |  Extra Sensors       |
+-------------------------------------------------------------------+
                                  |
                                  v
+-------------------------------------------------------------------+
|                   MICROCONTROLLER & PROCESSING                     |
|    - Sample Data & Filter Signals                                 |
|    - Calculate Battery Capacity (Current Integration)             |
|    - Run On-Device AI Anomaly Detection (Proposed)                |
+-------------------------------------------------------------------+
                       /                     \
                      v                       v
+----------------------------------+   +----------------------------+
|         LOCAL ALERTS             |   |   BLE WIRELESS TELEMETRY   |
|   - Visual Indications (LED)     |   |   - Android App Interface  |
|   - Audible Warnings (Buzzer)    |   |   - Live Data & Analytics  |
+----------------------------------+   +----------------------------+

Core Parameters

Current: The rate of electric charge flowing into or out of the battery.

• Current: The rate of electric charge flowing into or out of the battery.

Voltage: The electrical pressure across the battery terminals.

• Voltage: The electrical pressure across the battery terminals.

Temperature: Thermal readings taken at key points across the power setup.

• Temperature: Thermal readings taken at key points across the power setup.

Estimated Battery Capacity: An ongoing estimate of the total charge remaining in the battery.

• Estimated Battery Capacity: An ongoing estimate of the total charge remaining in the battery.

How Capacity Estimation Works

To measure battery capacity accurately, we plan to use a technique called current integration (often referred to as Coulomb counting). Think of a battery as a water storage tank: measuring current is like measuring how fast water flows through the discharge pipe. By continuously tracking the flow rate over time, we can calculate exactly how much charge has entered or left the tank. This provides a much more precise capacity estimate than simply reading water pressure (voltage).

Modular Design

To ensure flexibility, the core system architecture is designed to allow users to attach optional sensors—such as ambient humidity, vibration, or expansion sensors—depending on their specific application requirements.

4. Our Plan

This project is currently in the initial conceptual stage. No physical circuits, firmware, app code, or trained machine learning models have been produced or tested yet.

Our roadmap outlines the following development stages:

Hardware & Sensing Circuitry: Design the signal-conditioning circuit and select precision sensors for current, voltage, and temperature measurement.

• Hardware & Sensing Circuitry: Design the signal-conditioning circuit and select precision sensors for current, voltage, and temperature measurement.

Firmware & Data Processing: Program a low-power microcontroller to sample sensor inputs, filter noise, compute capacity via current integration, and manage system power states.

• Firmware & Data Processing: Program a low-power microcontroller to sample sensor inputs, filter noise, compute capacity via current integration, and manage system power states.

Wireless Communication: Integrate Bluetooth Low Energy (BLE) functionality to transmit live telemetry data efficiently.

• Wireless Communication: Integrate Bluetooth Low Energy (BLE) functionality to transmit live telemetry data efficiently.

Custom Android Application: Build a mobile interface to view real-time parameters, track historical trends, and receive push alerts.

• Custom Android Application: Build a mobile interface to view real-time parameters, track historical trends, and receive push alerts.

Onboard Safety & Fallback Alerts: Connect an LED and piezo buzzer directly to the hardware board for instant local feedback when mobile connections are inactive.

• Onboard Safety & Fallback Alerts: Connect an LED and piezo buzzer directly to the hardware board for instant local feedback when mobile connections are inactive.

AI-Based Anomaly Detection: Investigate lightweight machine learning algorithms capable of running on embedded hardware to flag irregular system behavior. We are currently considering two primary candidate approaches:

Autoencoder: A neural network model trained on normal operating profiles; anomalies are detected when the reconstruction error exceeds a set threshold.

• Autoencoder: A neural network model trained on normal operating profiles; anomalies are detected when the reconstruction error exceeds a set threshold.

Isolation Forest: An unsupervised algorithm that isolates unexpected data points based on feature anomalies.

• Isolation Forest: An unsupervised algorithm that isolates unexpected data points based on feature anomalies.
• AI-Based Anomaly Detection: Investigate lightweight machine learning algorithms capable of running on embedded hardware to flag irregular system behavior. We are currently considering two primary candidate approaches:Autoencoder: A neural network model trained on normal operating profiles; anomalies are detected when the reconstruction error exceeds a set threshold.Isolation Forest: An unsupervised algorithm that isolates unexpected data points based on feature anomalies.

The final choice of anomaly detection model will be determined based on processing overhead, memory constraints, and detection accuracy once hardware prototyping begins.

5. Why BLE and AI Matter

When designing a continuous monitoring tool, choosing the right communication method and analytical approach is key to usability and reliability.

Bluetooth Low Energy (BLE)

BLE allows short-range wireless communication directly between the hardware board and a smartphone without needing Wi-Fi routers, local network infrastructure, or cellular data plans. Because BLE consumes very little energy, the monitoring hardware itself draws minimal power from the battery it is monitoring.

On-Device Intelligence / AI

Standard systems rely on simple "if/then" rules (e.g., alert if temperature > 50°C). However, battery degradation often displays subtle patterns—such as a disproportionate rise in temperature relative to a very small current draw—that simple threshold checks miss. An anomaly detection model learns what normal relationship patterns look like across voltage, current, and temperature, allowing it to flag unexpected behavior before critical limits are breached.

Local Fallback Indicators

Wireless monitoring is convenient, but mobile devices are not always connected. Integrating an onboard status LED and a simple buzzer ensures that critical alerts are triggered directly on the hardware board, maintaining safety regardless of smartphone connectivity.

6. Installation and Running

This section will be completed once the hardware, firmware, and application are implemented.

7. Testing and Results

Testing and results are not available yet because implementation has not started.

Once physical prototypes are manufactured and programmed, we plan to systematically evaluate the platform across several key criteria:

Sensor Measurement Accuracy: Benchmarking voltage and current readings against calibrated reference instruments.

• Sensor Measurement Accuracy: Benchmarking voltage and current readings against calibrated reference instruments.

Capacity Tracking Precision: Assessing current-integration calculations across complete charge and discharge cycles under varying load conditions.

• Capacity Tracking Precision: Assessing current-integration calculations across complete charge and discharge cycles under varying load conditions.

BLE Performance: Measuring signal range, data throughput, reconnection speed, and mobile power consumption.

• BLE Performance: Measuring signal range, data throughput, reconnection speed, and mobile power consumption.

Anomaly Detection Performance: Evaluating model accuracy, false-positive rates, execution latency, and memory utilization on the target controller.

• Anomaly Detection Performance: Evaluating model accuracy, false-positive rates, execution latency, and memory utilization on the target controller.

Local Alert Reliability: Verifying latency between an anomaly detection trigger and the activation of visual/audible indicators.

• Local Alert Reliability: Verifying latency between an anomaly detection trigger and the activation of visual/audible indicators.

8. Wrap-up / Future Work

We have outlined the concept and system architecture for a modular, intelligent battery monitoring platform that combines multi-parameter sensing, local alert mechanisms, low-power mobile telemetry, and machine-learning anomaly detection.

Our immediate next steps will focus on finalizing component selection, designing the initial circuit schematics, writing proof-of-concept firmware, and building our first hardware prototype to begin collecting benchmark dataset cycles.

Reference & Resources

• T. M. Birlin et al., “IoT-based Smart Battery Monitoring Systems using Multi-Sensor Fusion, ” Proc. ICCBI, 2026, pp. 331–336, doi: 10.1109/ICCBI68589.2026.11619650.
• Edge AI for Real-Time Anomaly Detection: How It Works in 2026: https://www.marsdevs.com/guides/exploring-edge-ai-for-software-based-real-time-anomaly-detection
• For more information about the nRF54LM20 DK: https://www.nordicsemi.com/Products/Development-hardware/nRF54LM20-DK
• Datasheet for the nRF54LM20A / nRF54LM20B https://docs.nordicsemi.com/r/bundle/ps_nrf54lm20a/page/keyfeatures_html5.html
• Designing for Edge AIhttps://www.nordicsemi.com/Products/Technologies/Edge-AI
• Custom Neuton AI Modelshttps://www.nordicsemi.com/Products/Technologies/Edge-AI/Neuton-models?lang=en#infotabs
• Integrated AI Axon Acceleratorhttps://www.nordicsemi.com/Products/Technologies/Edge-AI/Axon-NPU?lang=en#infotabs