Announcing Jneopallium: Biologically-Grounded Neuron Networks with Industrial-Grade Safety for Real Plants
Hi Hugging Face community!Today I’m excited to share Jneopallium — a Java framework for building biologically-grounded, multi-timescale neuron networks that can actually run as safety-gated controllers in chemical plants, power grids, manufacturing lines, water treatment, and HVAC systems.While most of the ecosystem focuses on Python/PyTorch models for the cloud or edge inference, Jneopallium was designed from the ground up for industrial operational technology (OT) — where a single bad write to an actuator can have serious consequences.What makes it different
True biological grounding: Typed signals with independent propagation, receptor heterogeneity, dual fast/slow processing loops (regulatory vs supervisory timescales), and per-signal frequency control via ProcessingFrequency(loop, epoch). It’s not just “bio-inspired” — the architecture mirrors how natural neuron networks actually work.
Industrial OPC UA bridge (Eclipse Milo-based): Full read/write integration with any standards-compliant PLC, SCADA, or digital twin. Subscriptions, alarms & conditions, quality propagation, wall-clock timestamps from the server — all handled cleanly.
Six non-negotiable safety invariants baked into the bridge:
No raw actuator writes — everything goes through Planning → SafetyGate → Interlock → OperatorOverride → Aggregator.
Interlocks have absolute authority (fail-safe write on trip).
Operator override always wins for regulatory control.
Every single write is audited (JSONL + optional OPC UA mirror).
Quality never silently promoted.
Timestamps come from the plant, not the JVM.
Progressive autonomy done right: Per-loop SHADOW → ADVISORY → AUTONOMOUS commissioning sequence. You can run 90 % of loops fully autonomous while keeping 1 % in full shadow mode. No dangerous global “AI on/off” switch.
Built-in safety neurons: SafetyGateNeuron, InterlockNeuron, Human-harm discriminator (five-dimensional consequence model), OscillationMonitor, etc.
License: BSD 3-Clause (fully open).Quick start (industrial smoke test)
yaml
# Minimal unsecured demo config (public Milo server)
connection:
endpointUrl: "opc.tcp://milo.digitalpetri.com:62541/milo"
...
Add the single Maven dependency, drop in the YAML, run the bootstrap class, and you’re connected. Full manual (with worked temperature-loop example, cascaded loops, audit format, etc.) is included in the repo.Links
GitLab mirror + Maven packages: https://gitlab.com/rakovpublic/jneopallium
Full OPC UA Integration Manual: see JNEOPALLIUM_OPCUA_INTEGRATION.md (or the attached .docx version in releases)
Jneopallium is still early (1.0-SNAPSHOT) but already battle-tested in concept for real industrial use. The core framework is mature; the industrial module is production-ready for controlled commissioning.I’d love feedback from the HF community — especially from people working on:
Safe autonomous agents
Industrial / robotics control
Multi-timescale or biologically-plausible architectures
OT cybersecurity & functional safety
If you’re in process control, functional safety, or just curious about neuron nets that can actually touch physical hardware without blowing up the plant — come say hi!Star the repo if this resonates, try the demo, and let me know what you think.— Dmytro Rakovskyi
Discussion in the ATmosphere