Plastering is a unified framework for normalization of buildings metadata. Different frameworks can be unified into a workflow and/or compared with each other in Plastering.

Getting Started


  1. Install MongoDB: instruction
  2. Install Dependencies: pip install -r requirements.txt
  3. Install Plastering package: python install
  4. Download dataset here. This link is not public yet. You may use synthesized data to test the algorithms for now. Unfortunately, UCSD does not approve publicly sharing the data. We may have a procedure to sign an agreement, but it's still under development. Until then please refer to a synthesized data as specified in an example.

Example with synthesized data.

  1. Load data: python examples/tutorial/
  2. Run Zodiac: python examples/tutorial/
    • This will print out accuracy (F1 scores) step by step.

Example with SDH data

  1. Load ata: python examples/tutorial/
  2. Run Scrabble: python examples/tutorial/
    • This produces scrabble_output.ttl.
    • There will be an update about how to produce other types of results (metrics, other files, etc.)

Other examples

  1. Run Zodiac test: python
  2. Run Workflow test: python
  3. Run Zodiac experiments: python scripts/ ap_m
  4. Produce figures: python scripts/


Data Format

Raw Metadata

  1. It is defined as RawMetadata inside plastering/
  2. Every BMS point is associated with a unique source identifier (srcid) and a building name.
  3. All BMS metadata is in the form of JSON document. A BMS point corresponds to a row with metadata possibly in multiple entries. Example:
        "srcid": "123-456",
        "VendorGivenName": "RM-101.ZNT",
        "BACnetName": "VMA101 Zone Temp",
        "BACnetUnit": 64

Ground Truth of Metadata (LabeledMetadata)

  1. It is defined as LabeledMetadata inside plastering/
  2. tagsets: Any TagSets associated with the point.
  3. point_tagset: Point TagSet among the associated TagSets. If it's not defined, one may select Point-related TagSets from tagsets.
  4. fullparsing: Each entry has parsing results. An example for 123-456's VendorGivenName:
    1. Tokenization: ["RM", "-", "101", ".", "ZN", "T"]
    2. Token Labels: ["Room", None, "leftidentifier", None, "Zone", "Temperature"]
    3. (Though Plastering by default supports the above token-label sets, different tokenization rules may apply from a framework. For example, one may want to use ZNT -> Zone_Temperature_Sensor instead. Such combinations can be extended later.)
  5. One may use a part of different label types or add a new label type if needed.

Raw Timeseries Data

  1. Every BMS point may produce a timeseries data associated with the corresponding srcid.
  2. Its data format is TODO.

Output Metadata in Brick

  1. Brick graph: Result graph in Brick (in Turtle syntax).
    ex:RM_101_ZNT rdf:type brick:Zone_Temperature_Sensor .
    ex:RM_101 rdf:type brick:Room .
    ex:RM_101_ZNT bf:hasLocation ex:RM_101 .
  2. Confidences: A map of confidence of triples.
    1. A key is a triple in string and the value is its confidence. If the triple is given by the user, it should be 1.0. E.g.,
       ("ex:RM_101_ZNT", "rdf:type", "brick:Zone_Temperature_Sensor"): 0.9,

Framework Interface

  1. Each framework should be instantiated as the common interface.

Common Procedure

  1. Prepare the data in MongoDB. Example:
  2. The number of seed samples are given and a framework is initialized with the number as well as the other configurations, which depend on the framework as different framework may require different initial inputs.
    conf = {
            'source_buildings': ['ebu3b'],
            'source_samples_list': [5],
            'logger_postfix': 'test1',
            'seed_num': 5}
    target_building = 'ap_m'
    scrabble = ScrabbleInterface(target_building, conf)
  3. Start learning the entire building's metadata with the instance.
    scrabble.learn_auto() # This function name may change in the near future.

    Each step inside learn_auto() looks like this:

    1. Pick most informative samples in the target building.
      # this code is different from acutal Scrabble code as it internally contains all the process.
      new_srcids = self.scrabble.select_informative_samples(10)
    2. Update the model
    3. Infer with the update model
      pred = self.scrabble.predict(self.target_srcids)
    4. Store the current model's performance.


  1. Each framework aligned to the interface (./plastering/inferencers/ can be a part, called Inferencer, of a workflow to Brickify a building.
  2. Workflow/Inferencer interface is defined under TODO.
  3. Workflow usage scenario:
    1. Each part is initiated with the raw data for target buildings in the format described in Data Format.
    2. In each iteration, each part runs algorithm sequentially.
    3. In each part, it receieves the result from the previous part and samples from an expert if necessary.


  1. Plastering also can be used to benchamrk different algorithms. It defines the common dataset and interactions with the expert providing learning samples.
  2. Benchmark usage scenario.


  1. Initialize data

    python -b ap_m
  2. Test with Zodiac