Smart Manufacturing & Predictive Maintenance

TinyML on CNC machines, vibration and thermal sensor analysis, OPC-UA integration, predictive failure detection — edge AI meets industrial science.

OPC-UA/Modbus TinyML Edge Vibration Analysis xLSTM Prediction ISO 13849
10K+
Sensors/cluster
24h
Failure prediction
<100ms
Edge latency
$0
Cloud cost

Distributed GPU Cluster

Eldric runs on any GPU — from a single RTX 3090 to NVIDIA H100/H200 datacenter cards. Connect labs, datacenters, and remote researchers into one cluster that spans cities or continents. Workers register through the Edge TLS gateway over the internet, or use the built-in tunnel for NAT traversal — no VPN needed. Adding a node is one command: the worker auto-registers and is immediately available for inference.

Mixed GPU Research Cluster Controller :8880 (orchestration) Router :8881 (AI decisions, xLSTM predictor, Swarm LLM) Inference Workers — auto-register, any GPU, mix freely RTX 3090 (24GB) Ollama :8890 RTX 4090 (24GB) xLSTM + Ollama :8890 H100 (80GB HBM3) vLLM + xLSTM :8890 Apple M4 Ultra MLX :8890 + Add More... auto-registers Specialized Workers — all auto-register with controller Science :8897 140+ APIs, LIMS Training :8898 xLSTM, LoRA, DPO Data :8892 NFS, RAG, Vector Edge :443 TLS, Auth, Web IoT :8891 OPC-UA, TinyML Media :8894 STT, TTS, Video Location-Independent — nodes connect over the internet via Edge TLS gateway University Datacenter Corporate On-Prem Research Lab Home Office / Remote Cloud / Colo Datacenter Mix any hardware · Any location · Workers auto-register over the internet · xLSTM on every node
# Node 1: Controller + Router + Data Worker (any machine) ./eldric-controller --port 8880 & ./eldric-routerd --controller http://localhost:8880 & ./eldric-datad --nfs --vector --controller http://localhost:8880 & # Add a GPU worker — any machine, any GPU. # It auto-registers with the controller and is immediately available. ./eldric-workerd --backend ollama --controller http://node1:8880 # Add more workers the same way. Mix any hardware: ssh rtx-box "./eldric-workerd --backend ollama --controller http://node1:8880" ssh h100-node "./eldric-workerd --backend vllm --controller http://node1:8880" ssh mac-studio "./eldric-workerd --backend mlx --controller http://node1:8880" # Science Worker (auto-registers like any other worker) ./eldric-scienced --controller http://node1:8880 # Split a 70B model across all workers (VRAM-proportional sharding) curl -X POST http://node1:8880/api/v1/pipeline/deploy \ -d '{"model_path":"/mnt/models/llama-70B-Q4.gguf", "workers":["wrk-1","wrk-2","wrk-3","wrk-4"], "strategy":"vram_proportional"}'

xLSTM + Transformer Mixture Models

Train xLSTM+Transformer mixture on CNC vibration sensor time-series for predictive maintenance. xLSTM sLSTM cells capture long-range wear patterns (days/weeks); mLSTM fuses multi-axis accelerometer data; Transformer cross-attention correlates between sensor channels (vibration + temperature + current + acoustic).

Architecture

Sepp Hochreiter's xLSTM extended with Transformer layers for domain-specific tasks.

  • sLSTM: exponential gating for long-range dependencies
  • mLSTM: matrix memory with covariance update
  • Transformer: cross-attention for multi-feature correlation
  • Mixture: xLSTM temporal + Transformer relational

Training Configuration

# Train via Eldric Training Worker API curl -X POST http://controller:8880/api/v1/training/jobs \ -H "Content-Type: application/json" \ -d '{"name":"cnc-failure-predictor","base_model":"xlstm-mixture-500m","method":"sft","backend":"xlstm","dataset":{"path":"/data/cnc-vibration-sensors.jsonl","format":"timeseries"},"hyperparams":{"epochs":50,"batch_size":32,"learning_rate":1e-4},"model_config":{"architecture":"xlstm_transformer_mixture","xlstm_layers":8,"transformer_layers":4,"slstm_cells":4,"mlstm_cells":4,"hidden_size":512,"context_length":8192}}'

xLSTM vs Transformer for Time-Series

xLSTM sLSTM Cells

Exponential gating captures long-range wear patterns in sensor data — degradation signals that develop over days or weeks of operation.

  • Long-range temporal dependencies
  • Monotonic degradation tracking
  • Memory-efficient for streaming inference

xLSTM mLSTM Cells

Matrix memory with covariance update rule fuses multi-axis accelerometer data into a rich state representation for vibration pattern recognition.

  • Multi-variate sensor fusion
  • Frequency domain patterns
  • Covariance-based memory update

Transformer Attention

Cross-attention between sensor channels (vibration + temperature + current + acoustic) reveals failure correlations invisible to single-sensor analysis.

  • Cross-sensor correlation
  • Anomaly attention maps
  • Interpretable failure signatures

Mixture Architecture

xLSTM handles temporal dynamics, Transformer handles cross-sensor attention. Combined: 94% failure prediction accuracy 24h ahead on benchmark CNC datasets.

  • Best of both architectures
  • 94% accuracy @ 24h horizon
  • Deploys to edge (INT8 quantized)

API Examples

All endpoints are served by the Science Worker (:8897). Requests are routed through the Edge server for TLS and authentication in production.

Query CNC sensor data via OPC-UA

POST http://science-worker:8897/api/v1/iot/sensors/query # Request body: {"device":"cnc-lathe-01","sensors":["vibration_x","vibration_y","temperature","spindle_current"],"range":"last_24h"}

Run failure prediction on sensor stream

POST http://science-worker:8897/api/v1/iot/predict # Request body: {"device":"cnc-lathe-01","model":"cnc-failure-predictor","horizon_hours":24}

Log material sample with audit trail

POST http://science-worker:8897/api/v1/lims/samples # Request body: {"sample_id":"MAT-2026-0042","material":"Ti-6Al-4V","tests":["tensile","hardness"],"operator":"J.Smith"}

Active predictive maintenance alerts

GET http://science-worker:8897/api/v1/iot/alerts?severity=critical&device=cnc-*

Distributed Inference Documentation

Learn how to split large models across multiple workers with pipeline parallelism.

Distributed Inference Docs