Automating Fire Watch in Industrial Environments through Machine Learning-Enabled Visual Monitoring

Nuclear power plants (NPPs) are experiencing significant cost challenges to remain competitive with other energy-generation industries. Unlike other industries, the cost of operations and maintenance (O&M) activities at an NPP is mostly attributed to workforce cost. The NPP industry has therefore resorted to automating manually intensive tasks to reduce O&M costs, especially for monitoring activities. One monitoring function that is visually demanding and can occur frequently in order to meet the requirements of the fire protection program in an NPP is fire watch. A fire watch is a worker physically stationed at a given location (where normal fire protection measures are challenged) with the sole responsibility of observing a given area to ensure a fire is detected and mitigated promptly. This effort targets migrating fire watch from a manual model to an automated model. An automated visual monitoring approach using machine learning was pursued to closely resemble the actual manual visual monitoring performed by a fire watch.

Multiple machine-learning methods were evaluated for automated visual recognition of fire; long short-term memory (LSTM) neural networks, coupled to a color-based feature-extraction method to extract fire regions, resulted in the most accurate results. In addition to developing the method, a training data set for fire in industrial environments was developed. Because machine-learning methods are sensitive to the training data used, using only industrial fire videos was essential. Various sources of publicly available datasets containing scenes approximating industrial settings with and without fire were investigated and aggregated. They consist of YouTube-8M datasets that were downloaded and used to train and validate the neural network modeling approaches used in this analysis. Additional video datasets available from previously published literature sources and an online repository provided by Yahoo (YFCC100m and FiSmo) were also downloaded and aggregated to test the color feature extraction efforts and isolate fire regions in a frame.

The developed method was trained and validated using 1,000 industrial-only fire videos extracted from YouTube-8M datasets. After the developed method was trained and validated using the YouTube-8M data, it was tested using a set of 62 videos, split evenly between fire and no-fire scenarios, from a Yahoo videos repository. The accuracy achieved by the developed method was 0% missed positives (i.e., fire occurred but not detected) and 8% false positives (i.e., system indicates fire though there is no fire). Though smoke was not targeted in this effort, the method is expected to perform with similar accuracy for smoke detection, assuming a smoke training dataset is developed. The integration of this developed method with a suite of fire and smoke detection technologies (being researched or sold as products) is planned next to reduce false-positive events.

Additionally, using a machine-learning method in a process to meet the licensing requirement necessitates unboxing the machine learning “black box,” i.e., explaining the rationale behind the decision-making and providing extensive validation to the regulator of the method’s ability to replace the human fire watch. This is also planned for future research.