Scalable Building Efficiency Applications Using Normalizedof Metadata Automatic collection Building Metadata Arka Bhattacharya, David Culler (UC Berkeley) Dezhi Hong, Kamin Whitehouse (University of Virginia), Jorge Ortiz (IBM Research) Synthesis of Transformation Programs Motivation Problem : Approach: 1. Analytics and Control Applications written for buildings depends on quality of metadata. i. Semantic Relationships between sensors typically not available for buildings. 2. Metadata is i. imprecise and hard-to-understand (e.g SODA1R465__ART). ii. varies across vendors, and across deployments. 3. Hard to write analytics and control applications that scale across millions of sensor network deployments, without constant input and intervention of facilities manager. 4. Facilities managers often only people who understand the metadata. Learn through model, i.e the input-output example 1. Ask an expert (e.g the facilities manager) to provide transformed metadata (Project Haystack) for a sensor. 2. Learn the metadata transformation rules, and then apply these rules to other sensors wherever applicable. 3. Ask for another example, and continue this process until most sensor metadata has been transformed / augmented. Metadata Keys Primi ve Sensor Metadata BLDA1R435_ _ART Example from expert MaxRemaining: % Keys Correctly Identified Sensor network in Building 1: 100 80 • Built in 1994 • ~1600 sensors vendor: Barrington 60 40 % Sensors Fully qualified SameRemaining: % Keys Correctly Identified MaxRemaining:% Sensors fully qualified 20 Building expert e.g Building Manager Primi ve Sensor Metadata Sensor network in Building 3: 80 60 • ~2500 sensors • 5 different schema • Technique used : Random 40 20 0 15 20 25 30 35 40 45 Overall System Architecture 5 10 15 20 Number of examples 50 # Examples 25 BMS Value/Indicator SameRemaining: % Sensors fully qualified SameRemaining: % Keys Correctly Identified MaxRemaining: % Sensors fully qualified MaxRemaining: % Keys Correctly Identified Sensor network in Building 2: site ahu ahuRef zone zoneRef Metadata Format 1 BLD A 5 R 577A BMS 100 • Built in 2009 • ~2600 sensors vendor: Siemens Percent 80 60 40 80 60 40 Can correctly Iden fy and extract learned (keyvalue) pairs in other buildings 20 20 0 0 0 0 5 10 15 20 25 # Examples 30 35 40 45 50 Most frequent 20 40 60 80 100 120 Least frequent Metadata Keys Learned in Building 1 (Keys ranked by frequency of Occurrence in Building 1 ) Driver Boost Metadata Format 2 If b1 then e1 b1 := Occurs(input, ‘ART’, 1) e1 := SubString(input, Constant(11), Constant(14) ) Common Namespace Metadata Sensor Network 2 BMS Portable Analy cs Apps Driver Boost Metadata Format 3 Common Namespace Metadata Sensor Network 1 Results on 10 buildings: Find rogue zones: thermal zones which are constantly heating / cooling. • Requires: Finding temp sensors and temp setpoints which are semantically related Inefficient Air Handling units : Air handling units which serve hot zones, and in the process over-cool other zones. • Requires: Finding air handling unit, temp setpoints and temp sensors that are semantically related Detailed Results for 1 building sensor network 30 Fig. showing how many buildings learnedmetadata-keys were applicable to, out of 60 buildings Common Namespace Metadata Sensor Network 3 Ongoing Research Agenda: 100 % True Positive Buildings Random: % Keys Correctly Identified Boost (in common namespace) • Random : %Sensors fully qualified Driver Example Scalable Applications with Transformed Metadata & Future Work • 0 10 Metadata Keys Example applications : 100 0 BLD ( c) A ( c) 1 ( v) R ( c) 435 ( v) ART ( c) For key zoneRef: If b1 then e1 b1 := OccursAtPos(input, ‘R’, 5) e1 := SubString( input, Constant(6), PrecedeSucceed(ε, _, 1) ) • SameRemaining: %Sensors fully qualified site ahu ahuRef zone zoneRef zone air temp sensor For key zone air temp sensor: Ask for another example to expert Alternatives to automatically choose next sensor to present to the expert for manual parsing: 1. Random: Choose random sensor. 2. MaxRemaining:Choose sensor with maximum amount of metadata not yet transformed. 3. SameRemaining: Choose sensor whose un-transformed metadata string occurs in most other sensors. Random: % Keys Correctly Identified Value/Indicator (in common namespace) Op mized Time Series Database SYNTHESIS OF TRANSFORMATION RULES Results on commercial building sensor networks Percent Expert-provided Transforma on Example (using transforma on programs learnt from example above) Normalizing Existing Sensor Metadata To Common Format 1. Transform/Augment existing sensor metadata to a common namespace through examples provided by an expert i. Transformation also helps identifying semantic relationships between sensors. 2. Write applications against the common namespace that can scale across these sensor deployments. 5 Example Rules formed: Automated Transforma on : Goal: 0 No delimiters to differentiate one key from another Not a regular language No a-priori knowledge of tokens. Inconsistent naming, different tokens mean same thing. Programs should not become erroneous as more and noisy examples are provided (convergence). BLDA5R577A _ARS Random: %Sensors fully qualified Transformation Language Challenges: • • Sample Web Report Generated Making language more robust, and applying technique to • geographically diverse buildings, • sensor networks commissioned by different vendors, and • sensor networks in other contexts. Efficient data scraping from these (often) challenged primitive sensor networks and databases. Implementing active and passive data-driven approaches to complement the synthesis technique. Related Work 1. Project Haystack 2. Automated String Processing in Spreadsheets using Input-Output Examples, Sumit Gulwani, POPL 2011