AI detects. An engineer verifies. You get evidence.
A monitoring station running continuously produces far more data than anyone reads. The result is that most acoustic data is archived rather than used — the problem was visible in the record weeks before anyone noticed it.
Automated classification changes that. Events get detected and sorted as they occur, anomalies get flagged, and the pattern becomes visible while it still matters. But an automated classification is not a finding. It is a hypothesis produced by an algorithm, and on its own it will not survive a regulator, a tribunal or a competent opposing expert.
So the chain does not end there. What the system detects, an acoustic engineer verifies — and what you receive carries a name on it.
What automated analysis genuinely adds
- Event detection — impulsive events, tonal components, level excursions and night-time incidents, isolated from continuous data
- Source classification — sorting events by likely origin: plant, traffic, aircraft, construction, weather, wildlife, human activity
- Anomaly detection — the machine that has started sounding different, before it has started sounding wrong
- Trend extraction — slow drift that no single measurement reveals and no human reading a chart reliably spots
- Correlation — acoustic events matched against production records, weather, traffic counts and complaint logs
- Prioritisation — surfacing the 40 events out of 400,000 data points that a person actually needs to look at
What it does not do
An algorithm classifies. It does not take responsibility. It cannot confirm the microphone was calibrated that week. It cannot recognise that a station drifted, that a windscreen degraded, or that a reflecting surface was erected nearby last month. It cannot tell you that the index it reported is not the index your limit is expressed in. It cannot judge whether the measurement answered the question that is actually being asked. And it cannot sign a document that a regulator or a court will accept. Those are the parts that make a result usable, and they are the parts a person does.
The chain
- Sensor — Class 1 instrumentation to IEC 61672, positioned by an acoustic engineer, calibration verified and documented
- Data — continuous logging, transmitted and retained, with system health monitored so gaps are detected rather than discovered
- Detection — automated classification and anomaly flagging against thresholds set to your obligation
- Verification — an acoustic engineer reviews what the system found, confirms or rejects each classification, and checks it against instrument condition and site conditions
- Report — a compliance or investigation report, signed by a named engineer, with the automated analysis as an input rather than the conclusion
Each link is ordinary. The value is that none of them is missing
Most systems sold in this market are missing one of these steps. It is usually the last one.
Event attribution — the question that is actually asked
The most common dispute in continuous monitoring is not whether a level occurred. It is what caused it. A boundary station records an exceedance at 02:40. The complainant attributes it to your plant. It may have been your plant. It may equally have been a heavy vehicle on the road behind the microphone, a delivery at the neighbouring site, thunder, or a dog. Classification across frequency content, temporal signature, level history and correlated records makes that question answerable rather than arguable — and the engineer's verification is what converts an answer into something you can rely on when it is challenged.
Machine condition and predictive use
Sound and vibration carry information about mechanical condition well before failure. Bearing degradation, imbalance, cavitation, misalignment and looseness all change the acoustic signature. Where monitoring is already deployed for compliance, the same data supports condition monitoring at low marginal cost. It is not a replacement for a dedicated vibration programme on critical assets, and we will say so where that is the honest answer — but a plant already instrumented for boundary compliance is holding condition information it is currently discarding.
Audio capture and personal data
Automated classification improves when the system retains audio, and audio recording in environments where people are present carries personal data obligations under Thai law that logged sound levels do not. This is configured deliberately, with your instruction, and raised at scoping rather than after installation. Most deployments do not need audio at all. Where event identification requires it, configurations exist that classify events without retaining intelligible speech. We would rather have this conversation at the design stage than have you discover the obligation during an audit.
Standards & method
IEC 61672-1 · IEC 61260 · ISO 1996-1 · ISO 1996-2 · ISO 20906 · ISO 10816 · ISO 13373 · relevant Thai notifications on environmental and industrial noise · Thai personal data protection requirements where audio is captured
FAQ
Is an AI classification accepted as evidence?
Not on its own, and we would not present it that way. The automated analysis is an input. What is issued is a report verified and signed by a named acoustic engineer, with the instrument calibration, method and limitations documented. That is what makes it defensible.
How accurate is the automated classification?
It varies with the acoustic environment, the sources present and how much site-specific tuning has been done. It is reliable enough to prioritise what a person examines, which is its actual job. Any classification that matters — an exceedance, a disputed event, anything going into a report — is verified before it is relied on.
Can it tell whether the noise came from our site or the road?
Frequently, yes, using frequency content, temporal signature and correlation with other records. Where the evidence genuinely does not separate the two, we say so rather than assigning a cause the data does not support.
Do we need new hardware?
Not always. Where compliant Class 1 instrumentation is already deployed, we can often work with the existing data stream. Where it is not, the monitoring service supplies it.
Does the system record conversations?
Only if audio capture is deliberately configured, and most deployments do not need it. Where event identification requires audio, there are configurations that classify without retaining intelligible speech. Any audio capture is agreed with you in advance and covered by the personal data obligations that apply.
Can it warn us before we breach a limit?
Yes. Thresholds are set below your limit and alerts go to whoever you nominate, while there is still time to act. Automated classification also reduces false alarms, which is what stops alerting being ignored.
Can we use the same data for machine condition monitoring?
Often, yes, at low marginal cost. For genuinely critical rotating assets we would tell you honestly whether it is sufficient or whether a dedicated vibration programme is warranted.
Who owns the data and the analysis?
You do. You keep the full record and the analysis output, and both stay yours if you change provider.
Is this available now or is it a roadmap item?
The monitoring, the detection and the engineer verification chain are deliverable. The depth of automated classification depends on the site and how much tuning the environment justifies — we will tell you at scoping what is realistic for your specific situation rather than promising a capability generically.