AI Classification of Vibration Impacts using Waveforms

SiteHive presented research at INTER.NOISE 2026 on using AI models to automatically classify vibration impacts. Here we summarise the research, and explain why it provides significant benefits for the way construction teams triage vibration alerts.

Environmental vibration monitoring on construction sites relies on metrics like Peak Particle Velocity (PPV) to monitor exceedances. However, interpreting why an exceedance happened has been a slow, manual and expert-dependent process.

SiteHive’s new AI framework classifies vibration events automatically, distinguishing genuine construction activity from false alarms. Waveform data captured by SiteHive Hexanode Vibration devices is analysed, labelled and presented in SiteHive Enviro Pro, providing immediate and easy-to-understand information about every exceedance.

Authors

Dr. Aryan Sharma (1), Michael-Manraj Darroch (1), Ben Cooper-Woolley (1) and Dr. Benjamin Halkon (2)

1. SiteHive, Surry Hills, Australia
2. School of Mechatronic and Mechanical Engineering, University of Technology Sydney, Sydney, Australia

Understanding AI classification of vibration impacts: why it matters

Vibration monitoring is a fundamental requirement for risk mitigation on construction and infrastructure projects in Australia and New Zealand. Activities such as excavation, pile driving and demolition generate ground-borne vibration that can damage nearby structures, with potential impacts from minor cracking through to significant structural compromise. They can also disrupt the comfort and wellbeing of nearby residents and sensitive receivers.

In Australia, these impacts are managed against standards such as AS 2436 and DIN 4150-3, most commonly addressing three different types of impact:

• Structural Damage – assessed against peak particle velocity (PPV) measured in mm/s, which is the instantaneous maximum disturbance, and the industry standard for assessing risk of damage to buildings and infrastructure.
• Human Comfort – assessed against vibration dose value (VDV) measured in m/s^1.75, which is a cumulative metric that accounts for the duration and frequency of vibration events, reflecting human annoyance and comfort over a work day or night.
• Sensitive Equipment – assessed against Vibration Criterion (VC) curves measured in RMS velocity mm/s, which is the average measured velocity and is used to assess potential impact on sensitive spaces and equipment like laboratory equipments, hospitals, etc.

However, the raw data does not identify what caused an exceedance. A PPV threshold breach could have been caused by anything from a pile driver to someone bumping the sensor. Telling these apart using waveform data is today still largely a manual, context-dependent task that requires specific and significant expertise.

Why vibration exceedances are so hard to interpret

Although the construction industry now has access to lots of data through real-time monitoring devices, it faces a persistent bottleneck: interpreting what an exceedance actually means. Traditional systems trigger an alert without context, with no way to distinguish a high-impact construction activity from external or benign sources. For example, is it a roadheader working or just a heavy vehicle passing by?

This ambiguity creates real costs:

• Unnecessary site stoppages. Every exceedance may trigger a manual investigation, even when the true cause is a false alarm.
• Slower, less useful reporting. Without automated classification, contractors are limited to reporting raw peaks rather than evidence of exactly which activities contributed to an exceedance.
• Reliance on expert interpretation. Understanding waveform patterns has traditionally required deep domain expertise that isn’t available on every project, every day.

SiteHive had already tackled a closely-related version of this problem for construction noise, developing an industry-first AI audio classifier to automatically identify noise exceedances. This research extends that same philosophy to vibration.

A new approach: real-time, AI vibration waveform classification

SiteHive set out to build a framework that could analyse vibration waveforms and automatically predict the source of a vibration exceedance, separating genuine construction activity from non-construction events.

The framework has three parts: a data acquisition system built on the SiteHive Hexanode Vibration device; a hybrid deep-learning model that classifies vibration waveforms; and a validation process that uses real-world construction data collected from across Australia.

This research has resulted in the SiteHive Waveform Classifier, a new, industry-first capability that is available now within SiteHive Enviro Pro software for construction teams across Australia and New Zealand.

How SiteHive built the classification framework

Data acquisition and signal processing

The SiteHive Hexanode Vibration device measures raw acceleration across three tri-axial MEMS accelerometers sampling at 1kHz. The signal is band-limited between 1Hz and 80Hz using filters compliant with DIN-45669, then integrated over 1-second windows to obtain instantaneous velocity.

A hybrid model: spectrograms plus Crest Factor

The classification model combines two complementary signal properties:

• Time-frequency spectrograms. A spectrogram, functionally similar to a RGB colour image, is created for each event. This spectrogram is then fed into a convolutional neural network (CNN).
• Crest Factor. This is a physical feature that is calculated directly from the raw waveform, and is the ratio of peak vibration magnitude to its root-mean-square (RMS) value. Sustained construction machinery produces a high-RMS, low Crest Factor signal, while impulsive events – such as a kick or a knock – produce sharp peaks with a distinctly high Crest Factor.

Initial validation and testing

The initial dataset was comprised of labelled samples selected from over 500 SiteHive Hexanode Vibration monitors running on active Australian construction sites, which provided an overall database of 1 million+ waveforms. These captured genuine construction activity (piling, hammering, drilling, vehicle operation, etc) alongside non-construction related events.

Rapid training

This dataset was used to train the model. After an early F1 score (a metric used to evaluate classification models. F1 = 1.0 means perfect precision and recall) of 0.76, the model attained a perfect 1.000 across F1, precision and recall.

Real-time field evaluation

The trained model was then deployed live inside SiteHive Enviro, receiving continuous waveform data from a Hexanode Vibration device.

To stress-test it, the device was subjected to 150 controlled trials across three physical actions: kicking the device (simulating accidental impact), hammering its resting surface (simulating genuine construction activity), and picking it up and moving it (simulating routine handling).

Physical activity	True class	Trials	Accuracy
Hammering	Construction	50	100.0%
Horizontal kick	Non-construction	50	98.0%
Vertical pick-up	Non-construction	50	82.0%

The model performed perfectly on hammering and near-perfectly on kicks. The pick-up motion proved the hardest to identify, as lifting and moving a sensor produces sustained, multi-axis acceleration that lowers the Crest Factor in a way that mimics continuous machinery. This learning was used to directly shape the next phase of development.

Field validation and model optimisation

Following this controlled testing, the vibration waveform classifier was deployed on three live construction sites where SiteHive Hexanode Vibration devices were monitoring vibration levels continuously. These sites provided a wide range of vibration impacts, including both construction and ambient data. This real-world data was used to identify gaps and optimise the process.

Baseline performance

On the sample set (approximately 72% construction, 28% non-construction) the baseline model achieved 87.9% accuracy, 95.0% precision, and 89.1% construction recall.

This provided a good starting point, but still had only limited exposure to the diverse ambient and disturbance signatures found in continuous, real-world deployments.

Expanding and rebalancing the training data for non-construction vibration

To close that gap, SiteHive expanded the training by:

1. Incorporating labelled field recordings from construction project sites, and
2. Generating and adding synthetic non-construction samples.

A class-weighted loss function (weighting non-construction events at 2.0) further corrected for the real-world scarcity of genuine non-construction samples relative to construction ones. The final training and validation pool grew over 5 times, with a roughly even split between construction and non-construction.

The results

The optimised model achieved perfect construction recall, eliminating false negatives entirely. This was an important achievement as this is the most operationally costly failure mode, since a missed construction event means an undetected exceedance.

The overall F1 score also improved across the board. The small precision trade-off reflects the deliberate weighting toward catching every genuine construction event.

Model	Accuracy	Precision	Recall	F1
Baseline	87.9%	95%	89.1%	0.920
Optimised	91.5%	90.2%	100%	0.949

Why this matters for construction teams and projects

Automated AI vibration classification is important because it changes how vibration alerts are triaged and reported during live projects. It provides a number of significant advantages for construction teams including:

Fewer unnecessary site stoppages. Genuine false alarms, such as a knocked sensor or a moved device, are identified automatically. Teams are able to continue working without having to manually investigate every minor exceedance. For busy construction sites this can provide significant productivity and cost savings.

Evidence-based, transparent reporting. Rather than a simple list of peaks, reports can show exactly which activities contributed to an exceedance, streamlining compliance and reducing litigation risk.

Faster, more confident decisions. Project teams can act on vibration alerts immediately, with automatic context on likely cause, rather than waiting on manual expert review.

Better community relations. Clear, data-driven explanations of vibration impacts support more transparent, trust-building conversations with regulators, residents and sensitive receivers.

Conclusion

This research demonstrates a proof-of-concept AI framework that automatically classifies environmental vibration exceedances captured using SiteHive Hexanode Vibration devices.

The process distinguishes construction machinery from impulsive, non-construction events with high accuracy. Importantly, it also provides perfect construction recall on real-world field data following optimisation.

Read or download the full Research Paper

This research was presented as a full paper at INTER-NOISE 2026. You can read or download the full paper from below.