AXIOM LABS

• 

• 

Inspiration

$$ \text{250 Million Smart Meters} \quad \longrightarrow \quad \text{< 25% Data Used} $$

India is mid-way through the largest smart meter rollout on Earth. Yet utilities only see this:

"5 kWh consumed."

They cannot see:

• Was it a 2-star AC running all night?
• Was it an EV charging spike stressing the transformer?
• Was it a phase imbalance signalling grid failure?
• Was it a missed solar self-consumption window?

This is the Granularity Gap — and it costs India billions annually.

We are two undergraduates from Indian Institute of Science Education and Research Tirupati (IISER Tirupati) who built Axiom Labs because raw energy data without intelligence is nearly worthless.

---

What It Does

Dimension-N is our NILM engine that disaggregates a single aggregate energy signal into individual appliance-level signatures — no sub-meters needed.

The math at the core:

$$ P_{\text{total}}(t) = \sum_{i=1}^{N} P_i(t) + \epsilon(t) $$

Where \( P_i(t) \) is the power draw of appliance \( i \) at time \( t \), and \( \epsilon(t) \) is noise. Our job: recover every \( P_i \) from only \( P_{\text{total}} \).

Layer Architecture

Layer	Intelligence Level	Key Output
Layer 1	Appliance Disaggregation (NILM)	Per-device energy signature
Layer 2	Consumer Optimization	Bill forecast, solar savings
Layer 3	Industrial Intelligence	Peak load, power factor, downtime risk
Layer 4	Grid Control (DISCOM)	Transformer overload, phase imbalance

---

ML Pipeline

Trained on 40,000 datasets from UK DALE (Domestic Appliance Level Electricity)

Our ensemble uses:

$$ \hat{y} = \arg\max_k \; \left[ \alpha \cdot RF_k + \beta \cdot GB_k + \gamma \cdot \text{LSTM}_k + \delta \cdot \text{HMM}_k \right] $$

• \( RF \) — Random Forest (feature-based classification)
• \( GB \) — Gradient Boosting (residual correction)
• \( \text{LSTM} \) — Long Short-Term Memory (sequence modeling)
• \( \text{HMM} \) — Hidden Markov Model (temporal state transitions)

---

How We Built It

Bharat-NILM | MVP Cost: Rs. 2,500

Power Line (Appliance) | SCT-013 (Current) + ZMPT101B (Voltage) | Signal Conditioning → ADS1115 ADC | ESP32 Microcontroller → Feature Extraction | Cloud / Edge NILM Models → 4-Layer Dashboard

Smart Plug | MVP Cost: Rs. 500

Wall AC → ACS712 Sensor → ESP8266 → Relay (device control) → WiFi (MQTT/HTTP)

→ ML/NILM System → Dashboard

Competitive Benchmark

$$ \text{Payback Period} = \frac{\text{Installation Cost}}{\text{Annual Savings}} $$

Solution	Install Cost	Efficiency Gain	Payback
TP-Kasa Smart Plug	Rs. 2,00,000	4%	2.08 years
Sense Monitor	Rs. 55,000	6%	4.6 months
Bidgely Enterprise	Rs. 2,00,000	8%	1.04 years
Dimension-N	Rs. 20,000	12%	< 1 month

No competitor achieves payback in under a month. We do.

---

Accomplishments

• Working NILM hardware prototype for under Rs. 2,500
• 12% efficiency gains — highest in class
• Residential payback: < 1 month
• Industrial payback: < 2 days
• Full 4-layer architecture: appliance → consumer → industry → grid

---

What We Learned

The core signal model for NILM at low sampling frequency \( f_s \):

$$ \mathbf{x}(t) = [P(t),\; Q(t),\; \Delta P(t),\; \text{THD}(t),\; \sigma^2_P] $$

Where \( Q(t) \) is reactive power, \( \Delta P \) is the transient signature, and \( \text{THD} \) is total harmonic distortion — the fingerprint of each appliance.

Key insight: physics-informed features + ensemble ML > end-to-end deep learning at the low sampling rates used in Indian smart meters.

---

What's Next

• [ ] Indian appliance training data (desert coolers, submersible pumps, ceiling fans)
• [ ] DISCOM pilot under India's RDSS scheme (20+ crore sanctioned meters)
• [ ] Push NILM inference onto the ESP32 — fully offline, privacy-preserving
• [ ] Publish INDALE — India's first open Appliance Level Electricity dataset

---

Built With

Python · TensorFlow · Scikit-learn · ESP32 · ESP8266 ADS1115 · SCT-013 · ZMPT101B · ACS712 · MQTT UK DALE · HMM · LSTM · Random Forest · Gradient Boosting

---

$$ \boxed{\text{From Blind Consumption} \longrightarrow \text{Intelligent Energy Architecture}} $$

Team: BS-MS Milind & BS-ESS Darun.A — Indian Institute of Science Education and Research Tirupati

Built With

• acs712
• ads1115
• esp32
• esp8266
• hmm
• lstm
• mqtt
• python
• sct-013
• sicklit
• tensorflow
• zmpt101b