INDRA
Turns raw three-phase waveforms from a substation into plain-language guidance for the operator, in the seconds the operator has to act.
- Kind
- Real-time analytics platform
- Year
- 2026
- Area
- Grids and guards
- Recognition
- 2nd place · VISISONICS AI’26, judged by Hitachi India R&D
- Built with
- Next.js
- FastAPI
- TensorFlow
- XGBoost
- pandapower
- ChromaDB
- 3 more
The question
When a transformer misbehaves, how quickly can the person on shift know what it is and what to do?
INDRA takes the raw waveform and returns a named fault, a reason and a recommended action, in words an operator can act on.
Why it exists
Utilities collect waveform data from transformers, but operators wait minutes for analysis while the window to act is seconds long. INDRA sits on top of Hitachi’s TXpert Hub, as the VISISONICS AI’26 challenge asked, and turns the data into advice.
How it works
From a waveform to guidance
- 1
Waveform
A 20 ms window of three-phase voltage and current from the substation or a simulator.
- 2
Classification
Two models read the same window and their probabilities are blended.
- 1D CNN
- Reads the raw signal (60% weight).
- XGBoost
- Reads 31 engineered features (40% weight).
- 3
Physics and memory
A digital twin checks the grid behaviour, and the fault memory looks for similar past events.
- 4
Guidance
Gemma-4 writes the recommendation, with the cost and the compliance status attached.
- 5
Feedback
The operator confirms the root cause, which goes back to the memory and the models.
Decisions
Two models that read the signal differently
The CNN learns shapes in the raw waveform, while XGBoost works from engineered features and can explain its answer with SHAP values. Blending them covers each model’s blind spots.
An honest “uncertain”
Below 70% confidence the system does not name a fault. It reports an uncertain result instead of guessing.
Memory that fails quietly
If the ChromaDB memory is down, the search returns nothing and the rest of the pipeline carries on.
A dashboard that works without the backend
The Next.js dashboard has a layer of demo states and sample data, so it can be shown with no live feed.
Evidence
- Built
The feature extraction, the ensemble, the physics models, the fault memory, the compliance report and an eight-router API are in the repository, along with the training scripts for all three models.
Not built yet
- The trained CNN weights (they are not in the repository)
Built with
Frameworks and libraries
- Next.js16
- FastAPI
- pandapower
Data and storage
- ChromaDB
- PostgreSQL
AI and models
- TensorFlow
- XGBoost
- Gemma-4
Protocols and chains
- WebSocket