A vibration sensor on the spindle of a machining centre, read out over OPC UA with Python. Getting the raw signal out of the hardware was the difficult part. A passive subscription delivered nothing, so I built a stateful command-and-retrieve acquisition routine.
Independent Master's thesis on one machine, in healthy condition only.
Real machine faults were not tested, and no fault detection is claimed.
In a healthy-machine test, an artificial impact applied to the sensor produced a kurtosis excursion from about −0.1 to above 230, and the alert logic responded.
The chain at the top stays fixed. The drawing below it changes with each stage. All drawings are original schematics, not measured data.
The subject is a Spinner VC850 vertical machining centre with a Heidenhain TNC 640 control, in a university laboratory.
An IFM VSA004 piezo accelerometer is stud-mounted on the spindle housing. It feeds an IFM VSE100 module, which an IFM VOS050 OPC UA server (with a VOD001 device licence) makes available to a Python client. IFM provided the licences.
The first approach was a passive OPC UA subscription to the sensor's data. The subscription did not deliver raw data.
Without raw data there was nothing to process, so this became the hard part of the project.
I used UaExpert, a third-party OPC UA client, to browse the server and inspect its methods (StartRecording, Open, Read and Close) and their arguments. From that I built a stateful command-and-retrieve routine in Python: command a recording, wait until it is finished, open it, read it, close it, then decode the data and queue it. The routine matched the manufacturer's manual.
Acquisition runs in a producer thread. A FIFO queue passes each buffer to the consumer, which does the analysis and drives the dashboard. The result is a near-real-time acquisition cycle.
Each buffer is detrended, band-pass filtered with a 4th-order Butterworth filter, and Z-score normalised. Z-score normalisation rescales a signal by its own mean and standard deviation.
From the processed signal the system computes RMS, Peak, Crest Factor and Kurtosis in the time domain, and an FFT in the frequency domain.
Baselines were recorded on the healthy machine, thermally stabilised and running free without a tool, at discrete spindle speeds.
Alert and alarm thresholds were defined in the thesis from baseline statistics using μ+2σ and μ+3σ. Rule-based diagnostic logic is applied to the features. Features are appended to persistent CSV files for trending.
Impact test. In a healthy-machine test, an artificial impact applied to the sensor produced a kurtosis excursion from about −0.1 to above 230, and the alert logic responded.
Load comparison. One machine, one light facing cut in aluminium 6061, one spindle speed (2000 RPM).
| Condition | Raw RMS | Normalised kurtosis |
|---|---|---|
| No load | 0.03 g | 26.29 |
| Light cut (facing, aluminium 6061) | 0.03 g | 26.13 |
Kurtosis values were similar between the two states and are not interpreted here; one machine, one light cut, one spindle speed. This is not a validation of load-independent monitoring.
One place to check exactly what this thesis is, and is not.
A public snapshot of the code from August 2025. It contains the Streamlit dashboard, command-and-retrieve acquisition, signal processing, baselines and diagnostics, and trending. It predates the thesis submission and differs from the thesis implementation in some details. It is not the final thesis implementation.