Training an AI model to recognize pollutants without live data
In recent years, the recognition of global warming and climate change as significant threats to humanity has become more evident. While technology and modernization are often held responsible for these issues (and for good reasons), it is essential to recognize that technology can also play a crucial role in finding solutions. One such tool at our disposal is Artificial Intelligence (AI).
AI offers a range of applications that can aid in the fight against climate change. It has the potential to enhance climate predictions, facilitate intelligent decision-making for decarbonizing industries, ranging from construction to transportation, and optimize the allocation of renewable energy resources.
At Vision Elements, we have had the privilege of contributing to the development of a system aimed at reducing pollution from one of the most significant environmental hazards: industrial chimneys. Although industrial chimneys typically operate within regulated standards, occasional faults can occur due to various factors. Even brief malfunctions lasting mere seconds can result in the release of pollutants exceeding permissible limits and causing considerable environmental harm. Detecting such faults promptly is of paramount importance as it significantly mitigates the pollution caused by chimneys.
To address this challenge, we have developed an AI-based monitoring system that identifies irregular and hazardous chimney emissions. The system gathers information through round-the-clock camera surveillance of the chimney and other sensors measuring physical parameters like air pressure, temperature, and more.
How do we obtain data to train our AI model?
The main challenge of developing such an AI model is obtaining data, specifically data on abnormal emissions. It’s true that in every machine learning project the main challenge is always data, but here the situation is even more complicated. To teach the system how a pollutant emission looks like, we need to create such an emission. We don’t want to do that! Our mission to prevent air pollution and not to create it. So, how can you teach the AI model to detect pollution without showing a live model? One solution is to train the system to identify any anomalies. Several methods specialize in anomaly detection using machine learning models, and they can be effective even with a limited number of anomalies.
Another approach involves developing an AI model that continually learns and improves over time. Our system is designed to operate in an incremental learning framework, constantly receiving feedback from human operators involved in the process (the human part of machine learning). This enables the system to incorporate new information and enhance its performance iteratively.
Furthermore, the system's 24/7 operation poses an additional challenge. It must effectively monitor air pollution under poor visibility conditions, such as mist, rain, clouds, and darkness. Moreover, it should remain reliable across various weather conditions, maintaining consistent accuracy from summer to winter.
Tackling environmental and pollution challenges using AI
In order to overcome these challenges, our data scientists employed multiple cameras that captured the chimney from various angles and distances. Additionally, thermal cameras were utilized to provide visibility in adverse conditions. The system is capable of learning and adapting to the background environment, ensuring seamless transitions between day and night.
By leveraging AI technology in this manner, Vision Elements strives to advance the monitoring and detection of chimney emissions, making significant strides towards reducing air pollution and protecting the environment.