VibroAgent: On-Device Building Vibration Diagnostics

 Qualcomm Dragonwing™ IQ-9075 EVK  

Qualcomm GenieX

STMicroelectronics STEVAL-STWINBX1 (STWIN.box)

Powered USB hub (at least 6 ports)

<html><body><h3>Making a building listen at the edge</h3>Footsteps, machinery, traffic and structural resonances leave signatures in a building's vibration field. VibroAgent explores how one edge computer can compare several locations, detect unusual episodes and present results while keeping raw sensor data local. The project connects six STMicroelectronics STWIN.box boards to a Qualcomm Dragonwing™ IQ-9075 EVK. One board provides a reference signal and five monitor target locations. The IQ-9075 runs acquisition, signal preparation, model inference and a local browser interface. <h3>Hardware and software</h3>• One Qualcomm Dragonwing IQ-9075 EVK running Linux aarch64 with Hexagon HTP access. • Six STEVAL-STWINBX1 STWIN.box boards, each running FP-SNS-DATALOG2 v3.2.0. • One powered USB hub with at least six ports and the USB connections for the boards. Each board supplies vibration data from its IIS3DWB accelerometer. Mount the sensors consistently, verify axis orientation and range, and update the physical-device-to-logical-role mapping for one reference and five targets before acquisition. The installers set up isolated Python environments, the pinned STDATALOG-PYSDK v1.3.0 acquisition stack, Qualcomm GenieX and the selected model assets. Downloaded artifacts are checked against their integrity manifests. The full applications run on Linux on the IQ-9075. <h3>Two inference paths, one sensor installation</h3>VibroAgent-Codec is the established discrete-token path. A frozen residual-vector-quantized codec converts one configured axis from each synchronized ten-second window into 250 symbols per sensor. A fine-tuned Qwen3-4B model receives these symbols and target levels relative to the reference through GenieX on the Hexagon accelerator. Its monitoring response is constrained and validated as JSON. VibroAgent-Gemma is the continuous-embedding research path. It uses synchronized X, Y and Z data. A frozen ONNX encoder produces 14 embedding rows per board, or 84 rows across all six sensors, which enter reserved positions in Gemma 4 E2B. The included GenieX patches support this embedding path and the global normal/anomaly and joint five-target decisions. This checkpoint remains research-only; its validation promotion gate has not been met. Both implementations live in separate directories and use separate environments and launchers. Run one live stack at a time because they share the USB sensor fleet. <h3>From vibration to an operator view</h3>• Native USB callbacks collect the six sensor streams. • Independent reader processes isolate sensor SDK state and prepare synchronized ten-second episodes. • Deterministic checks handle sensor identity, timing and invalid inputs such as missing, stale, flat or clipped data. • The selected encoder creates compact symbols or continuous embeddings for local model inference. • The application validates the resulting structured record before displaying the monitoring decision. The browser interface provides waveforms, spectrum and Welch PSD inspection, monitoring history, replay, chat and alerts. Raw acceleration stays on the IQ-9075; the model receives compact representations within the local runtime. <h3>Install and run</h3>Clone the combined repository on the IQ-9075 and open the interactive installer: <pre>git clone https://github.com/hobbitlv1/vibroagent-iq9075-combined.git </pre><pre>cd vibroagent-iq9075-combined </pre><pre>./install.sh </pre>The menu offers Codec live, Codec offline, Gemma live, Gemma offline and a direct Gemma LUMO demo. Select one deployment. Setup and startup are separate so USB group and udev changes can take effect; a live installation may require logging in again or rebooting. For a board-free Codec demonstration: <pre>./install.sh --codec-offline </pre><pre>./vibroagent.sh start </pre>For the Gemma offline web replay: <pre>./install.sh --gemma-offline </pre><pre>./VibroAgent-Gemma/vibroagent.sh start </pre>Use --codec-live or --gemma-live with the installer for live sensors, followed by the corresponding launcher. Open the web address printed at startup, normally port 7860. The full offline applications still require the IQ-9075, but do not require USB sensor boards. A direct Gemma fixture can also be run with: <pre>./install.sh --gemma-demo </pre><pre>./VibroAgent-Gemma/vibroagent.sh demo target_3 </pre><h3>Replay and validation</h3>Codec replays five minutes of immutable six-channel recordings. Gemma provides a sixty-second web replay with scheduled LUMO-derived events. The cover screenshot shows the offline monitor with all six recordings and a Target 3 anomaly indication. It illustrates the replay workflow, rather than a field measurement or proof of structural damage. The repository reports roughly 342–422 prefill tokens/s, 6–11 decode tokens/s and a six-sensor verdict in under 30 seconds for the measured Codec configuration. These measurements depend on the hardware, model and runtime configuration. The Gemma checkpoint's reported validation macro-F1 is 0.9265 for global normal/anomaly classification, while affected-target F1 is 0.4701 and single-target exact-set match is 0.3393. Global anomaly discrimination is stronger than source-zone localization. The repository keeps this limitation visible and includes integrity checks, recorded fixtures and hardware-independent tests. <h3>Scope and source</h3>VibroAgent is an edge-AI monitoring and integration demonstrator for operator triage. It does not identify a structural damage mechanism, certify a building or replace an engineering inspection. The local web interface has no authentication and should remain on a trusted network. Source code and complete setup guide (https://github.com/hobbitlv1/vibroagent-iq9075-combined) include both deployments, installers, model manifests, recorded demonstrations and verification instructions. The project was developed by Niks Kordjukovs in 2026 during the internship period at System Research and Applications, STMicroelectronics Italy, under the supervision of Danilo Pau (danilo.pau@st.com). Original repository material is licensed under CC BY-NC-SA 4.0. Third-party software, models, datasets, images, and trademarks retain their respective licenses. </body></html>

VibroAgent: On-Device Building Vibration Diagnostics

Monitor building vibrations with six STWIN.box (IIS3DWB) sensors and local AI inference on Qualcomm Dragonwing IQ-9075.