SBIR Phase I: Intelligent Fire Detection for Indoor Settings

这个小企业创新研究（SBIR）第一阶段项目将通过(1)计算实验和(2)在公司的火灾探测器原型上进行广泛的真实火灾和滋扰实验来验证公司新颖的火灾探测原理和软计算技术的使用。这将使公司能够解决软件和硬件集成到最小可行的火灾探测器中可能出现的任何技术障碍，预计将大大优于当前市场解决方案。该公司利用NIST进行的火灾和滋扰实验的真实数据集来测试和基准测试这项创新。例如，利用温度数据，该公司的算法检测火灾的速度比目前的技术（固定阈值）快10倍。此外，所有NIST实验均未发生漏检，未触发虚警。虽然这些结果令人印象深刻，但由于它们仅基于69个实验，因此不足以使创新商业化。第一阶段提出的活动旨在降低技术风险，并在商业化之前增强对公司技术的信心。这一点尤其重要，因为火灾探测领域的监管负担和相关责任很高。通过利用ML， erli。Ai正在申请专利的技术将处理来自热/烟雾传感器的流数据，以检测最早的火灾迹象，并将火灾和滋扰事件分开。我们引入了一种新的基于传感器数据时间分析的报警触发原理，取代了现有的固定阈值方法。我们建议进一步发展我们的计算架构，该架构采用深度长短期记忆（LSTM）神经网络和变分自编码器，为我们的算法产生的卓越性能提供理论基础。我们将使用动态时间翘曲和相似矩阵技术开发一种用于时间流聚类的异常检测评估方案。该奖项反映了美国国家科学基金会的法定使命，并通过使用基金会的知识价值和更广泛的影响审查标准进行评估，被认为值得支持。

This Small Business Innovation Research (SBIR) Phase I project will validate the company’s novel fire detection principle and the use of soft computing techniques by means of (1) computational experiments, and (2) a wide range of real fire and nuisance experiments on the company’s fire detector prototypes. This will allow the company to resolve any technical hurdles that may arise from the integration of the software and hardware into a minimum viable fire detector expected to significantly outperform current market solutions. The company has utilized real-world datasets of fire and nuisance experiments conducted by NIST to test and benchmark the innovation. Using temperature data, for example, the company’s algorithm detected fires up to 10 times faster than current technology (fixed-threshold). In addition, no missed detection occurred, and no false alarms were triggered in all NIST experiments. While these results are impressive, they are not sufficient to commercialize the innovation given that they are based on 69 experiments only. The activities proposed in phase I are designed to reduce the technical risk and strengthen the confidence in the company’s technology prior to commercialization. This is especially important because of the high regulatory burden and associated liabilities in the fire detection space.By leveraging ML, erli.ai’s patent-pending technology will process streaming data from heat/smoke sensors to detect the earliest signs of fire and separate fire and nuisance events. We introduce a new alarm triggering principle rooted in the temporal analysis of sensor data instead of the existing fixed-threshold approach. We propose to further develop our computational architecture, which employs deep Long-Short Term Memory (LSTM) neural networks and a variational autoencoder, by developing a theoretical basis for the outstanding performance our algorithm produced. We will develop an anomaly detection evaluation scheme for time streaming clustering using Dynamic Time Warping and Similarity Matrix technologies.This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.