Predictive Maintenance MCP Server

What Predictive Maintenance MCP Server does

Predictive Maintenance MCP Server supports vibration-based condition monitoring for industrial machinery. It works with signal files stored under data/signals/, loaded signals held in an in-memory repository, machine manuals in resources/machine_manuals/, and a local verified bearing catalog. The server keeps one signal_id as the handle from loading through analysis, diagnosis, and report generation.

Its analysis surface covers frequency and time domains. FFT, power spectral density, and STFT tools expose frequency content, spectral energy, and changes over time. Envelope analysis screens for bearing-related frequency lines. Statistical analysis reports RMS, crest factor, kurtosis, and peak-to-peak values, while sliding-window extraction produces 17 time-domain features for downstream modeling.

How it works

Use list_signals to find files or inspect signals already cached in memory. load_signal accepts individual files or atomic batches and reads sampling-rate and unit information from explicit arguments or companion metadata. CSV, NumPy, and other self-describing formats can be loaded directly; headerless binary files require a declared sample format and sampling rate. Raw integer data remains in ADC counts unless a scale factor is supplied.

After loading, call analysis tools with the returned signal_id. FFT and envelope analysis use a deterministic leading one-second segment by default, with options for a full-signal or seeded random segment. Bearing checks accept a verified catalog bearing, explicit expected frequencies, or complete bearing geometry. The server calculates BPFO, BPFI, BSF, and FTF from supplied geometry and reports the provenance of the frequencies used.

diagnose_vibration combines FFT, PSD, STFT, bearing checks, and ISO severity assessment. generate_diagnostic_report turns the integrated results into a self-contained HTML or PDF document, optionally comparing the signal with a known-good baseline. A separate DOCX tool accepts caller-supplied report sections.

Setup and configuration

The project requires Python 3.11 or newer. A typical source installation clones the repository, creates a Python environment with uv, installs the package, and points an MCP client at the module predictive_maintenance_mcp. Claude Desktop setup uses an absolute project path and the PDM_PROJECT_DIR environment variable so the server can locate signals and reports. The repository includes a Windows setup script for creating the environment and writing the Claude Desktop entry.

Optional extras provide additional functionality: predictive-maintenance-mcp[vector-search] enables FAISS and sentence-transformers for semantic retrieval, [pdf] enables PDF diagnostic rendering, and [docx] enables Word report generation. Without the vector-search extra, documentation search uses TF-IDF keyword retrieval.

Tools and capabilities

• Generate reproducible synthetic bearing-fault, gear-fault, imbalance, or normal signals.
• Inspect signal metadata, clear cached signals, and load single files or validated batches.
• Match envelope-spectrum peaks against bearing fault frequencies and calculate characteristic frequencies.
• Assess ISO 20816-3 vibration zones for supported machines above 15 kW, using declared units or direct RMS velocity.
• Train OneClassSVM or LocalOutlierFactor models on healthy signals and predict bounded anomaly summaries.
• Read manuals, extract machine specifications, search documentation, and create interactive HTML, PDF, or DOCX reports.

Limitations and notes

ISO severity tools require a declared signal unit and adequate sampling coverage; units are never inferred from amplitude. The ISO scope is for machines rated above 15 kW, and declared power below that threshold is rejected. Statistical values are screening indicators rather than standalone diagnoses.

Bearing geometry must come from a manual, the verified catalog, or the user. A catalog miss is treated as a legitimate negative result, so the server does not invent geometry. Fault checks require exactly one expected-frequency route, and envelope bands above the signal's Nyquist frequency are rejected rather than silently adjusted.

Anomaly training is unsupervised or semi-supervised: fault data can assist hyperparameter tuning, but OneClassSVM and LocalOutlierFactor are trained on healthy data only. Generated reports return file paths and metadata rather than embedding all HTML content in tool responses.