LLM Edge Predictive Maintenance — AI Agents for Vibration Diagnostics

See it in action

Live Video Demo

Sample output

Download the Hairdryer Diagnostic Report

See the same style of evidence based analysis shown in the demo, including fault indicators, severity assessment and recommended actions.

📄 Download Sample Report

Why this is different

Not Just an LLM Guessing

The added value is not “asking Claude to guess the fault”. Claude operates with real diagnostic tools exposed through MCP servers and dedicated Skills.

✘

Traditional LLM Approach

“Based on the description you gave me, it might be a bearing fault…”

Text-only, no real data, no measurements, no repeatability.

✔

This Project — Tool-Augmented Diagnostics

“Possible outer-ring bearing wear detected around 89.4 Hz (expected 89.8 Hz for bearing 6205 at 1470 RPM). Reliability score: 72%. Vibration level: medium (ISO 10816 Zone B).”

Real sensor data, clear signal-processing steps, and internationally recognised vibration guidance. Evidence-based and repeatable.

Real-time edge diagnostics: from sensor to insight in seconds

📊

Evidence-Based

Every diagnosis is backed by measured frequency spectra, computed fault frequencies, and quantified severity, not guesses.

🔄

Repeatable

Same sensor, same tools, same analysis pipeline. Results are consistent and auditable across operators and sessions.

🏭

Industrial-Grade

Vibration severity rated with ISO 10816 guidance, bearing checks for outer ring / inner ring / rolling element / cage faults, and standardised reporting for maintenance teams.

36

MCP Tools

3

Expert Skills

ISO

10816 Compliant

How it works

Talk to Your Machine: Sensor → AI → Action

📡

Acquire

Connect the STWIN.box via USB. Claude controls the IIS3DWB wideband accelerometer at 26.7 kHz and acquires vibration data with a single chat message.

datalog2_start_acquisition(duration_s=5)

📊

Analyze

Complete signal-analysis pipeline (from frequency spectrum to envelope checks), with bearing fault detection and ISO 10816 severity guidance in plain language.

diagnose_vibration(rpm=1470, bearing="6205")

📋

Report

Claude generates operator-friendly reports with traffic-light severity, fault evidence, confidence levels and actionable maintenance recommendations.

🟢 Good 🟡 Monitor 🟠 Alarm 🔴 Danger

Natural Language Interface

Talk to Your Machine

No scripting, no dashboards. Just describe what you need and Claude handles the rest.

💬

Quick Health Check

Connect to my STWIN.box and check the vibration on the pump

Connected to STEVAL-STWINBX1 ✓

Acquiring 5 s from IIS3DWB at 26,667 Hz...

Loaded 133,335 samples ✓

🟢 ISO 10816: Zone A

RMS velocity: 0.71 mm/s · No faults detected

🔍

Bearing Diagnosis

Run a full bearing diagnosis. The motor runs at 1470 RPM with 6205 bearings.

Running envelope analysis with Hilbert transform...

Checking bearing frequencies (outer ring, inner ring, rolling element, cage)...

🟡 Outer race defect suspected

Outer-ring bearing frequency found at 89.4 Hz (expected 89.8 Hz) · Reliability: 72%

📋

Maintenance Report

Generate a maintenance report for the operations team

📋 Diagnostic report ready:

Executive summary with severity level
Sensor configuration and measurement details
Findings with evidence and confidence
Recommended actions and timeline

📊

Signal Exploration

Show me the FFT spectrum of the last acquisition, focused on 0 to 1000 Hz

FFT computed (133,335 points, Hann window)

Dominant peaks identified:

· 24.5 Hz (1×RPM, 0.042 g)

· 49.0 Hz (2×RPM, 0.018 g)

· 89.4 Hz (BPFO, 0.011 g)

Get Started

Quick Start

1

Clone and install

git clone https://github.com/LGDiMaggio/claude-stwinbox-diagnostics.git
cd claude-stwinbox-diagnostics

# Install both MCP servers
cd mcp-servers/stwinbox-sensor-mcp && uv venv && uv pip install -e . && cd ../..
cd mcp-servers/vibration-analysis-mcp && uv venv && uv pip install -e . && cd ../..

2

Configure Claude Desktop

// claude_desktop_config.json
{
  "mcpServers": {
    "stwinbox-sensor": {
      "command": "/path/to/stwinbox-sensor-mcp/.venv/Scripts/python.exe",
      "args": ["-m", "stwinbox_sensor_mcp"]
    },
    "vibration-analysis": {
      "command": "/path/to/vibration-analysis-mcp/.venv/Scripts/python.exe",
      "args": ["-m", "vibration_analysis_mcp"]
    }
  }
}

3

Install Skills and Go

Upload the 3 skill .zip files from skills-zips/ in Claude Desktop settings under Capabilities, then Skills. Restart Claude Desktop and start chatting!

Built with

Technology Stack

Python 3.10+ NumPy SciPy Pandas FastMCP MCP 1.0 Claude STDATALOG SDK USB-HID PnPL STEVAL-STWINBX1 FP-SNS-DATALOG2 ISO 10816 vibration guide

Predictive maintenance AI agents

Three Specialised AI Agents,
One Diagnostic Workflow

Three focused agents cover acquisition, diagnosis and reporting. Each one maps to a Claude Skill, so the workflow stays modular, reliable and easy to scale.

1

Monitoring Agent

machine-vibration-monitoring

Connects to STWIN.box, configures IIS3DWB, captures data and checks baselines.

Sensor control USB-HID Baseline

2

Diagnosis Agent

vibration-fault-diagnosis

Runs frequency and envelope analysis, checks bearing-specific frequencies, and rates overall vibration severity with ISO 10816.

Frequency analysis Envelope ISO 10816 guide

3

Reporting Agent

operator-diagnostic-report

Builds operator-ready reports with severity, evidence, confidence and recommended actions.

Reports Severity Actions

Monitoring Agent → Diagnosis Agent → Reporting Agent

Every agent is a Claude Skill backed by MCP tools, orchestrated directly from chat.

Toolkit highlights

Diagnostic Capabilities at a Glance

From raw signals to actions with 36 purpose-built MCP tools.

🔬

Frequency Pattern Discovery

Identify frequency signatures with spectrum analysis, windowing and peak detection.

🎯

Early Defect Detection

Reveal early bearing defects with Hilbert-envelope spectra.

⚙️

Bearing Knowledge Library

Use built-in bearing models or enter custom geometry or frequencies.

🩺

Clear Fault Identification

Classify unbalance, misalignment, looseness and bearing faults with confidence scores.

📏

Standards Aligned Severity Ratings

Apply ISO 10816 severity zones (A-D) across machine groups, explained in operator-friendly terms.

🧠

Expert Guided Diagnostic Flows

Three Claude Skills encode monitoring, diagnosis and reporting best practices.

💾

Efficient Data Management

Store high-volume samples server-side while sending compact summaries to the AI.

⏱️

Reliable Measurement Timing

Use server-side timing for precise acquisitions from milliseconds to minutes.

🔌

Complete Sensor and Analysis Toolkit

22 sensor tools + 14 analysis tools provide full USB-based control and diagnostics.

Under the hood

Architecture

Frequently Asked Questions

FAQ

What is LLM-based predictive maintenance? +

LLM-based predictive maintenance uses large language models (like Claude) as the reasoning engine for machine condition monitoring. Instead of dedicated monitoring software, the operator describes what they need in natural language, and the LLM orchestrates sensor data acquisition, signal processing (FFT, envelope analysis), fault detection, and report generation through connected tools. This makes diagnostics conversational, accessible and extensible.

Can AI agents diagnose bearing faults from vibration data? +

Yes. The Diagnosis Agent performs envelope analysis via Hilbert transform and checks for characteristic bearing fault frequencies: BPFO (outer race), BPFI (inner race), BSF (rolling elements), and FTF (cage). It compares detected peaks against computed theoretical frequencies for any given bearing model and RPM, providing confidence levels for each finding. This is the same approach used by professional vibration analysts.

Edge sensors and LLM: what runs where? +

The STEVAL-STWINBX1 edge sensor handles data acquisition at the physical layer (IIS3DWB accelerometer at 26.7 kHz). The MCP servers run locally on your machine and handle sensor communication (USB-HID) and signal processing (FFT, PSD, envelope). Claude (the LLM) runs in the cloud and acts as the reasoning engine, coordinating the agents through natural language. Raw data never leaves your machine — only compact summaries reach the LLM.

How does MCP connect Claude to industrial sensors? +

The Model Context Protocol (MCP) is an open standard that allows LLMs to call external tools. This project provides two MCP servers: one for sensor control (22 tools for USB-HID communication with the STWIN.box) and one for vibration analysis (14 tools for FFT, PSD, envelope, bearing fault detection, ISO 10816). Claude calls these tools programmatically during the conversation, just like a human analyst would use specialised software.

What is agentic maintenance? +

Agentic maintenance describes a paradigm where AI agents autonomously handle parts of the maintenance workflow. In this project, three specialised agents (Monitoring, Diagnosis, Reporting) each encapsulate domain expertise as Claude Skills. The agents coordinate through the conversation to deliver a complete diagnostic cycle, from sensor acquisition to actionable maintenance report, without requiring the operator to know DSP or vibration analysis.

What is an MCP Server and why does this project need two? +

An MCP (Model Context Protocol) Server is a lightweight service that exposes tools an LLM can call programmatically during a conversation. Think of it as giving Claude hands: instead of only generating text, Claude can read sensors, run computations, and interact with hardware. This project uses two MCP servers: one for sensor communication (USB control, data acquisition from the STWIN.box) and one for vibration analysis (FFT, envelope analysis, bearing fault detection, ISO 10816). Separating them keeps each server focused and makes the architecture extensible.

What are Claude Skills and how do they enable predictive maintenance AI agents? +

A Claude Skill is a set of instructions, packaged as a simple folder, that teaches Claude how to handle specific tasks or workflows. If MCP servers are the professional kitchen (access to tools, data, equipment), Skills are the recipes: step-by-step instructions on how to create something valuable. In this project, three Skills turn raw MCP tool access into reliable diagnostic workflows: the Monitoring Skill guides sensor acquisition and baseline checks, the Diagnosis Skill encodes vibration analysis expertise (fault pattern matching, bearing-specific checks), and the Reporting Skill produces operator-friendly maintenance reports. Together, they form the predictive maintenance AI agents that make the system work end to end.

What faults can the system detect? +

The system can identify: unbalance (1×RPM dominant), misalignment (2×RPM and axial components), mechanical looseness (sub-harmonics and broadband noise), and bearing defects (outer race, inner race, rolling element, cage faults) with characteristic frequency matching. Each finding includes a confidence level and ISO 10816 severity classification.

Is this a replacement for professional vibration analysis? +

No. This is a support tool and a proof of concept, not a replacement for qualified vibration analysts. This release (v0.3.0) was developed as an early-stage proof of concept with intensive assistance from Claude and has not yet been validated extensively on real industrial machinery. The analysis algorithms implement well-established signal processing techniques (FFT, envelope analysis, ISO 10816), but their integration with the STWIN.box hardware and the MCP protocol should be considered experimental. Real-world testing, calibration, and refinement will follow in subsequent versions. Do not use this as the sole basis for critical maintenance decisions without independent verification.

LLM EdgePredictive Maintenance