Sensei: Self-Supervised Sensor Name Segmentation

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1017–1027 August 1–6, 2021. ©2021 Association for Computational Linguistics 1017 Sensei: Self-Supervised Sensor Name Segmentation Jiaman Wu, Dezhi Hong, Rajesh K. Gupta, Jingbo Shang Department of Computer Science and Engineering, University of California San Diego, CA, USA Halıcıo˘glu Data Science Institute, University of California San Diego, CA, USA {j4wu, dehong, gupta, jshang}@eng.ucsd.edu Abstract Sensor names as alphanumeric strings typi- cally encode their key contextual information such as their function or physical location. We focus here on sensors used in smart build- ing applications. In these applications, sen- sor names are curated in a building vendor- specific manner using different structures and esoteric vocabularies. Tremendous manual ef- fort is needed to annotate sensor nodes for each building or even to just segment these sensor names into meaningful chunks for in- telligent operation of buildings. We propose here a fully automated self-supervised frame- work, Sensei, that can learn to segment sensor names without any human annotation. We em- ploy a neural language model to capture the underlying structure in sensor names and then induce self-supervision based on information from the language model to build the segmen- tation model. Extensive experiments on five real-world buildings comprising thousands of sensors demonstrate the superiority of Sensei over baseline methods. 1 Introduction Sensor name segmentation seeks to partition a sen- sor name string into semantic segments. It is an essential task to enable smart building applica- tions (Weng and Agarwal, 2012) that critically de- pend upon understanding the context of sensory data. For example, to increase the flow of air to improve air quality, such an application will need to identify and locate the airflow controller for the specific area. Such a controller is identified as a control point associated with a specific actuator such as a variable air valve. This is typically done manually by searching appropriate strings in the name that refer to specific sensor or actuator type and/or its location. This information is encoded as a concatenation of segments. Correct segmentation of sensor names is a key first step. Note, we use Figure 1: Sensor names adopt distinctive structures and vocabularies in different buildings, thus requiring man- ual effort to interpret. sensor to refer to both sensors and actuators in a building application. Figure 1 shows examples of sensor names con- sisting of multiple segments, each encoding key context about the sensor (building name, loca- tion, sensor type, etc). Thus, the sensor name SODA4R731 ASO should be segmented as SOD (building name), A4 (equipment id), R731 (room id), and ASO (measurement type – area tempera- ture setpoint). Note that the meanings of the same punctuation may vary; for example, ‘ ’ can be a delimiter or part of a segment. Currently, sensor name segmentation requires domain knowledge and tedious manual effort due to its building-specific nature. Sensor names are created by building vendors, and as we see from Figure 1, in different buildings they usually adopt distinctive structures and vendor-specific esoteric vocabularies. Thus, …

Content truncated to 500 words