Presentation is loading. Please wait.

Presentation is loading. Please wait.

A Longitudinal Study of Pressure Sensing to Infer Real-World Water Usage Events in the Home Jon Froehlich, Eric Larson, Elliot Saba, Tim Campbell, Les.

Similar presentations


Presentation on theme: "A Longitudinal Study of Pressure Sensing to Infer Real-World Water Usage Events in the Home Jon Froehlich, Eric Larson, Elliot Saba, Tim Campbell, Les."— Presentation transcript:

1 A Longitudinal Study of Pressure Sensing to Infer Real-World Water Usage Events in the Home
Jon Froehlich, Eric Larson, Elliot Saba, Tim Campbell, Les Atlas, James Fogarty, and Shwetak Patel Nina Sakhnini

2 Overview Pressure sensing to infer real-world water usage events in homes Deployed pressure sensors to monitor pressure transients Deployed a ground truth sensor network to monitor valve- level water usage by fixtures and appliances Usages are classified using a probabilistic algorithm that leverages a language model, grammar, and prior probabilities

3 Deployment Deployed in three homes (H1, H2, H3) and two apartments (A1, A2) At each site, two pressure sensors were and a set of custom wireless sensors that provided ground truth labels of water usage activity for the pressure stream The deployments began February 2010 and lasted for five weeks

4 Ground Truth An array of seven external ground truth sensors was deployed Tracked when each valve was opened or closed Categorized temperature into hot only, cold only, and mixed Components: A parent sensor board to communicate water usage data XBee Pro wireless modem to transmit sensor state Transmits a signal once every four minutes to a laptop installed at each site Sensors

5 Ground Truth Sensors Faucets:
Reed switches (N=34) that react to the presence of a magnetic field Accelerometers (N=14) for faucets with a single handle to measure acceleration and interpret the handle’s position and temperature Hall effect sensors (N=3) that sense the distance between two magnets. used for sensing faucets which control temperature using planar rotation but control flow through an up/down motion Omni-directional ball switch (N=39) that is a vibration sensor to wake the parent sensor board Appliances: Power usage sensors (N=7) Power consumption patterns Push buttons (N=2) when there was no access to the power, residents push the button when using the machine Thermistors (N=3) attached to the water drain pipe to measure the temperature

6 Pressure Sensors and setup
Pace Scientific P1600s with a resolution of psi connected to a microcontroller interfaced with Bluetooth Two sensors to collect data from hot and cold- water access points simultaneously Communicates data to the laptop Communicate with the ground truth sensor network Data uploaded to a backend web server every 30 minutes The backend send notifications when a sensor was not heard from for 10 minutes Data archived locally for backup

7 Analysis of the Collected Dataset
2.9% of the data is marked as uncertain and 3.9% are unknown events Not used in classification but not removed from the dataset Even proportion of cold and hot events (40.7% for cold only, 39.2% for hot, and 20% for mixed). Four fixtures (kitchen sink, master bathroom sink and toilet, and secondary bathroom sink) are 84.7% of the events Analysis of the Collected Dataset

8 Valve Event Classification Algorithm
Classifies pressure transients using a probabilistic approach using Bayesian estimation inspired by models used in speech recognition Provides robustness against transient deformations that can occur during natural valve usage such as compound and collision events A compound valve event is a valve event that occurs while another fixture is using water Result a severe attenuation of the high frequency component of the pressure transient A collision valve event is a valve event that occurs within two seconds of one or more other valve events The two colliding transient waveforms become highly distorted Usually these events they are offset by ms

9 Valve Event Classification Algorithm
Incoming water pressure data stream is buffered, and the pressure transients are segmented using time series boundaries defined by the ground truth annotations Segmented pressure transients are compared to a library of labeled templates using similarity algorithms A language model determines the likelihood of a given sequence of valve signatures and links open and close valve events into paired tuples Features are extracted from these paired tuples and compared with smoothed probability distributions Probabilities from the previous steps are multiplied for each sequence The sequence with the highest probability is selected

10 The Language Model Assigns probabilities for possible valve sequences using bigrams Bigram analysis is commonly used to examine co-occurrences of words or letters Bigrams here are groups of two sequential valve events The language model consists of transition probabilities for every valve pair and is trained by counting the number of co-occurring valve pairs in the library These counts are smoothed using Katz smoothing that works to assign a non-zero probability to every sequence Handles transition probabilities between two valves that rarely occur in the library Transition probabilities are used to select the optimal word (valve) sequence from all possible word (valve) sequences Maintain an n-best list of sequences and dynamically reorder the most probable sequences as new events occur

11 Grammar Defines the possible ways in which valve sequences can be constructed Grammar rules are: An opening of valve must be followed by a closing of valve A valve’s closure must be after its opening The temperature state of a valve must be consistent Uses a soft grammar which applies a penalty to any valve sequence that violates a rule Sequences which contain grammatical errors but have the likeliest probabilities from the other terms are selected as correct The grammar is applied to each sequence in the n-best list

12 This is interesting, but I am not sure I am convinced
This is interesting, but I am not sure I am convinced. Are sequences of valve events equivalent to grammatically correct sentences? I understand that a sequence of valve states can be a unique permutation but can the permutations be extensive enough to form grammar of their own?  --Harish The use of language processing techniques for analysis of water events provides an interesting application of technology. –Jesse Grammar Defines the possible ways in which valve sequences can be constructed Grammar rules are: An opening of valve must be followed by a closing of valve A valve’s closure must be after its opening The temperature state of a valve must be consistent Uses a soft grammar which applies a penalty to any valve sequence that violates a rule Sequences which contain grammatical errors but have the likeliest probabilities from the other terms are selected as correct The grammar is applied to each sequence in the n-best list

13 Classification Algorithms
Template- and Feature-Based Comparison: compares the segmented unknown pressure transient with open and close valve templates in the library Template Comparison + The Language Model + The Grammar Template Comparison + The Language Model + The Grammar + Paired Valve Tuple Priors: pairing valve events to link open and close transients together and to compute classification features, such as water usage duration Classification Algorithms

14 Classification Baseline performance: chance and a majority classifier
always selects the most likely result based only on frequency Three level classification to observe performance at different granularities: Valve level: identify the correct fixture and its temperature Fixture level: ignores temperature Fixture category level: classify open/close events as bath, clothes washer, dishwasher, faucet, shower or toilet

15 Results The best performing algorithm is Template Comparison + The Language Model + The Grammar + Paired Valve Tuple Priors Average overall classification accuracy of 75.5%, 89.5%, and 95.9% for valve, fixture, and fixture- category level The most frequently used fixtures: Kitchen sink Bathroom sinks Bathroom toilets Confusions tend to occur within fixture categories and between fixtures that are situated close together Regarding to compound and collision events, the language model-based algorithms tend to perform better than the template-based algorithm

16 Results of Results Three-way repeated measures ANOVA was conducted to test whether Template Comparison + The Language Model + The Grammar + Paired Valve Tuple Priors algorithm improved significantly over the template-based algorithm Using the10-fold classification accuracies as the dependent variable Sensing resolution, number of sensors, and algorithm as within- subjects factors A significant main effect of classification algorithm was found (F(1,4)=21.76, p=.010)

17 Suggestions for Improvement
The current approach trains the language model and priors using data from the home where it is deployed Suggests the usage patterns and priors across different homes reducing system calibration An interface to allow correction of misclassifications and training of the algorithm over time

18 Suggestions for Improvement
For the water monitoring project, it will be interesting to see if a one size fits all approach for training data will be sufficient, or if different training data for different environments (home, condo, apartment, high rise, etc.) would be needed to properly filter the noise. – Jesse I feel that this paper doesn't go to the depth of generalizing sensing to a larger, more interesting, context. – Andrew The current approach trains the language model and priors using data from the home where it is deployed Suggests the usage patterns and priors across different homes reducing system calibration An interface to allow correction of misclassifications and training of the algorithm over time

19 Thank You


Download ppt "A Longitudinal Study of Pressure Sensing to Infer Real-World Water Usage Events in the Home Jon Froehlich, Eric Larson, Elliot Saba, Tim Campbell, Les."

Similar presentations


Ads by Google