Future Fibre Technologies: Solving the nuisance alarm dilemma

Tired of nuisance alarms? Jim Katsifolis, PhD, Chief Scientist, Future Fibre Technologies explains why deep learning is a gamechanger for perimeter intrusion detection systems (PIDS).

Ask anyone in the PIDS game and they’ll tell you one of the most persistent challenges is the battle between probability of detection (POD) and nuisance alarm rate (NAR).

It’s an age-old struggle: How do you ensure the system is sensitive enough to detect genuine intrusions without being overwhelmed by false alarms? In high-sensitivity systems, the slightest disturbance can set off an alarm, flooding operators with notifications.

When nuisance alarms pile up, operator trust starts to erode.

When trust is lost, it can lead to dangerous consequences and, in the worst-case scenarios, operators may even disable the system to avoid dealing with constant nuisance alarms. It’s known as “alarm fatigue”.

For critical infrastructure, reliable, accurate PIDS is essential. However, the problem of nuisance alarms remains one of the hurdles to ensuring a secure perimeter.

This is where the combination of fibre-optic distributed acoustic sensors (DAS) and advanced deep learning techniques is poised to transform the industry.

The evolution of fibre-optic DAS in security systems

Fibre-optic DAS technology – capable of continuously monitoring acoustic signals and vibrations over long distances with high sensitivity – has become a popular choice for perimeter security.

Fibre optics has several advantages, including being passive and requiring no power in the field, immunity to electromagnetic interference, long-range sensing capabilities and precise detection accuracy.

These features make it an excellent choice for monitoring perimeters in challenging environments.

Although fibre-optic PIDS systems have been around since the 1990s, advancements in hardware, optical design and signal processing have allowed them to reach new levels of commercial viability.

These systems offer continuous monitoring and can detect even the slightest disturbances. However, in “real-world” scenarios outside the lab, achieving high sensitivity POD without being overwhelmed by irrelevant events is no easy feat.

The ability to accurately detect legitimate intrusions while ignoring nuisance alarms is the single most critical factor in determining the reliability of any intrusion detection system.

A system that triggers frequent false alarms can lead to alarm fatigue, where operators become desensitised, risking timely responses to actual security breaches.

It is incredibly difficult to design a system that detects every intrusion while ignoring events like wind, wildlife or environmental vibrations such a trainline that runs nearby or a plane taking off.

The key to achieving this balance lies in how the system processes the vast amount of data it collects. This is where deep learning enters the picture.

Balancing sensitivity and reliability

The terms artificial intelligence (AI), machine learning (ML) and deep learning (DL) are often used interchangeably, but they represent different aspects of how machines can simulate human intelligence.

AI refers to machines or computers that can perform tasks typically requiring human intelligence, such as recognising patterns, learning from data and making decisions.

ML, a subset of AI, is about training machines to learn from data and improve performance over time without explicit programming.

DL, however, is a specific form of ML that involves multi-layered neural networks designed to mimic the structure of the human brain.

These deep neural networks (DNNs) can process vast amounts of data, extract features and classify events without human intervention, leading to much higher accuracy.

In the context of PIDS, DL has the potential to shift the balance in favour of higher POD, while minimising the NAR. Historically, there has been a trade-off between these metrics.

A more sensitive system might improve the POD but raises the NAR. DL breaks this paradigm by allowing systems to become more intelligent in their decision-making.

A game-changer for PIDS

To understand why DL is such an advancement for PIDS, it’s important to break down how it works. Traditional ML systems rely on humans to define features that the system should look for in the data.

In contrast, DL systems can learn these features themselves.

By feeding the system vast amounts of data, it can automatically identify patterns, extract relevant features and classify events with greater precision than human operators or traditional algorithms could achieve.

The application of DL to PIDS allows systems to learn from real-world data, recognise complex patterns and make decisions that differentiate between genuine intrusions and benign events.

This not only increases the POD but also reduces the NAR allowing operators to focus on real threats without being overwhelmed by irrelevant alerts.

Data is king

One of the most critical factors for the success of a DL-based PIDS is the diversity and volume of data used to train the system.

For companies like Future Fibre Technologies (FFT), which has deployed DAS PIDS systems globally, this data advantage is key. FFT has built an extensive library of data from a range of environments, site conditions and intrusion events.

This dataset enables the development of generalised and customised DL models that can be tailored to specific operating conditions.

These models are then deployed to FFT’s Aura Ai-X system through encrypted file transfers.

Once integrated, the DL engine within the Aura Ai-X processes real-time data from the fibre sensors, using the model to classify events with high accuracy.

This results in a system with the highest possible POD and minimal nuisance alarms, ensuring that operators can trust the system.

Real-world testing in real-world environments

The effectiveness of DL for PIDS is not theoretical; it has been tested and proven in real scenarios.

FFT conducted a comparison between traditional signal processing methods and DL at a seaport in the Middle East.

The site had a 32-kilometer perimeter and was prone to a variety of disturbances, including environmental noise and vibrations from infrastructure. The results were striking.

Traditional signal processing methods, while effective, struggled to filter out all nuisance alarms, especially in an environment like a seaport where factors such as waves, wind and vehicle traffic can trigger false alerts.

However, with the application of DL, the system was able to drastically reduce the NAR while maintaining a high POD.

Such results demonstrate the potential for DL to transform the security landscape, especially in high-stakes environments where trust in the system is paramount.

Trust matters

Trust in a security system is directly tied to its reliability.

When a system is plagued by nuisance alarms or false positives, operator trust erodes. This is especially true when system reliability falls below 90%.

DL models, like those employed by FFT, have the power to restore trust by delivering systems that are sensitive to genuine intrusions and intelligent enough to ignore irrelevant events.

As DL continues to evolve and improve, the future of PIDS looks promising.

With the ability to learn from diverse data and adapt to changing conditions, these systems will become more accurate, reliable, and user-friendly.

Operators will be able to trust their systems to provide accurate, actionable alerts, ensuring the protection of critical infrastructure and assets.

Looking ahead, I see a future where nuisance alarms are dramatically reduced, allowing security teams to focus solely on real threats.

This shift will boost operational efficiency and lower the risk of breaches at critical infrastructure.

Nuisance alarms have long been a challenge for security professionals, but with systems like FFT’s Aura Ai-X harnessing artificial neural networks and vast data, we can achieve high sensitivity with minimal false alarms.

The future of PIDS is intelligent, reliable and, in my opinion, powered by DL.