Ubi Comp2009 Hydro Sense Final

Editor's Notes

• #2: Infrastructure-MediatedSingle-Point Sensing of Whole-Home Water Activity
• #3: Although energy usage and measurement has lately received a great deal of attention in the HCI and UbiComp communities, water has not. And yet, the United Nations predicts that water will be the dominating issue over the next 20 years. And, as you can tell from this map, this is a problem that affects every continent on earth including the US.----By 2025, more than 2.8 billion people living in 48 countries will face water shortages. Environmental water stress is assessed as the percentage of available water resources currently abstracted compared to environmental water requirements. Environmental water requirements are defined as the volume of water needed by a river to maintain key ecosystemfunctions and biodiversity. This is expressed as a % of the naturalised flow of a river, and can vary under different conditions, notably between arid and non-arid river systems. In Figure 2 the red depicts the systems under the highest environmental water stress.
• #4: Already, cities are going to unprecedented lengths to provide their residents with water. For example, in April of last year, Barcelona, Spain took the unprecedented step of importing water by ship from neighboring nations at a cost of $35 million a month (Crawford, 2008). Nearly 23m litres of drinking water - enough for 180,000 people for a day
• #5: The US is not immune from such problems either. According to US government estimates, 36 states will face serious water shortages in the next five years (EPA, 2008). Lake Mead, which supplies 90% of Las Vegas’ water, is estimated to drop below existing intake pipes by 2012 (Lippert and Efstathiou, 2009). The city of Las Vegas is paying people to remove their lawns at $1.50 per square foot of lawn replaced with desert-friendly plants for a maximum of $300,000 per year in rebates. (http://www.greenbiz.com/print/35667). Many utilities invest millions of dollars into conservation programs meant to educate their customers about water and cost savings.---
• #6: Most residents have no idea how much water they use nor what the most water consuming activities are in the home. So, ask yourself, what do you think the most water consuming activities are in your home?Most people cannot accurately answer this question. We have been working with our local water utility, SPU…Most customers believe they use 20 gallons a day (Al Dietemann) rather than roughly 70 and have no idea that toilets, clothes washers, and showers are the top water usage activities in the home. This is a striking disconnect considering that most of us have used modern plumbing all of our lives.To close this gap, we have been working on new sensing techniques for water usage---According to a recent survey at our local water utility in Seattle, most customers believe they use 20 gallons a day (Al Dietemann) rather than roughly 70.
• #7: We’ve been working on a water sensor called HydroSense, which is a pressure-based sensor capable of identifying water usage down to the individual fixture (e.g., dishwasher, toilet, faucet) and provide per-fixture flow estimates from a single sensing point
• #8: We use the pressure sensor to continuously look at changes in pressure when you open and close valves in the home.The key insight here is that when you use the toilet or wash your hands, a water pressure transient is generated that is unique to that particular fixture.
• #9: Our prototype HydroSense sensor implementation consists of two primary parts:a customized stainless steel pressure sensora 3d-printed encasement holding an analog to digital converter (ADC), a microcontroller, and aBluetooth wireless radio (see Figure 4), which communicates the raw pressure data to a computer.
• #10: To understand how HydroSense works, its useful to go over a brief primer of home water/pipe infrastructure.Most households obtain water from a public water supply.Public water is distributed by local utilities, relying on gravity and pumping stations to push water through major distribution pipes
• #11: Cold water enters the home through a service line, typically at 40-100 pounds per square inch (psi) depending on such factors as the elevation and proximity to a water tower or pumping station.Pressure is important to the proper functioning of HydroSense because it’s a pressure-based sensing solution.------The pound per square inch or, more accurately, pound-force per square inch (symbol: psi or lbf/in² or lbf/in²) is a unit of pressure or of stress based on avoirdupois units. It is the pressure resulting from a force of one pound-force applied to an area of one square inch:1 psi (6.894757 kPa) : pascal (Pa) is the SI unit of pressure.40 psi is 275.79 kilopascals100 psi is 689.47 kilopascals
• #12: Many homes have a pressure regulator that stabilizes the water pressure and also reduces the incoming water pressure to a safe level for household fixtures.From the regulator, most homes contain a combination of series plumbed and branched piping.
• #13: The cold water supply branches to the individual water fixtures (e.g., toilets/sinks/showers) and into the water heater.The plumbing system forms a closed loop pressure system with water held at a relatively stable pressure throughout the piping. This is why, when you open a faucet, water immediately flows out.
• #14: The hot water tank connects the cold water pipes to the hot water pipes.So when you open a hot or cold water valve, the pressure signal is transmitted through the hot water heater thus allowing for a single-point pressure sensing solution.Every hot water tank has a drain valve, which is of interest to us because it provides an easy potential installation point for our hydrosensor, particularly for locations like apartments which don’t have outdoor hose bibs.
• #15: We have investigated placing HydroSense at a variety of installation points including the drain valve on the hot water heater, the outdoor hose spigot and the 3/8” hot and cold water inlet lines to bathroom and kitchen sinks.Note: remember this is a single point sensing solution; I’m just showing single installation points that we’ve investigated.
• #16: So, now let’s see what the raw pressure waveform looks like when you use a fixture in your house. (click)Take particular notice to the pressure spike (or transient) that occurs when the valve is opened and closed.Note that the signal resembles a dampening sinusoidal wave—we take advantage of this in our classification algorithms.
• #17: The instant a valve is opened or closed (be it a bathroom faucet or a mechanical valve in a dishwasher), a pressure change occurs and a pressure wave is generated in the home plumbing system. The transient can have a positive or negative rate of change depending on whether a valve is being opened or closed.The pressure drop is indicative of how much water is being used.
• #18: So, how does this work exactly?
• #19: The raw pressure signal is smoothed using a low-pass filter. We then take the derivative of this smoothed signal to check for rapid changes in pressure (i.e., pressure transients).These rapid changes in pressure (or pressure transients) are what indicate water usage.----The low pass filter is a low-pass linear phase finite impulse response filter. The smoothed signal is the 1 Hz envelope of the signal, calculated using a discrete prolatespheroidal window filter of length 1 second.The bandpass derivative has a cutoff of 1Hz and a linearly increasing frequency response, it is also 1 second long. These parameters were chosen based on the bandwidth with the highest energy concentration for all the data we collected in all homes.
• #20: So, let’s see how this works in practice. As we run a 1 sec sliding window over the data, we flag the signal when the derivative exceeds a specified threshold relative to the home’s static pressure
• #21: Then, we look for when the “amplitude” of the derivative goes below 5% of the original magnitude
• #22: The direction of stabilized pressure change and the sign of derivative determines whether the event is an open or close event.In this case, it’s a valve open event.
• #23: Similarly, for a close event, we use the same process looking for the two critical points.
• #24: The difference here is that we have a pressure increase and a positive initial derivative.
• #25: As you might imagine, the pressure signatures differ across fixture and valve types. However, the valve’s location in the home is also a strong determinant in shaping the signal. (click) As a result we haven’t seen a lot of cross-home similarity between fixture event signals.From this point forward, the vertical-axis will always refer to psi and the horizontal-axis will refer to time in seconds----First, note that the transients are unique across fixtures. (click)---Now, I’m showing data from three different homes. Note how the signals differ for like fixtures across homes; this is because the signal is influenced both by the valve type as well as the pipe pathway to the valve.Therefore, while detection and distinguishing between open and close can be done with a drop-in system in any home, actual fixture classification and flow must be calibrated on a per home basis.
• #26: Now that we’ve automatically detected events and classified them as open or close, we need to determine where the events came from. To do this we use a template matching approach and, in particular, we rely on matched filtering to compare our signals, which provides a measure of correlation between signals.
• #27: We chose three transformations in which to use the matched filter. (click) 1. The first transformation is the waveform with trending removed2. The second transformation is the derivative waveform that we used in the open/close event detection step.3. The final transformation is the cepstrum transformation, which is popular in the signal processing community and often used to separate a filter from its source.---The last space we chose was the Cepstral space. The real cepstrum is a transformation related to the Fourier transform and has been used in the speech recognition community for decades. It’s popular in the signal processing community to separate source from filter.One property of the real cepstrum is that the space is highly orthogonal and thus is good fit for our application.
• #28: To test for similarity, we look at the transformations for each saved event in the library.If the similarity is above a certain threshold in each subspace, we add the event to a list of possible events. The thresholds are learned using a leave one out cross validation of all other homes in our database.--We also use the mean squared error of the de-trended signal. If the MSE is above a cross validation threshold, it passes the similarity test.
• #29: We repeat this process for every event in our library…
• #30: Here a toilet subspace fails a similarity test and thus it is discarded from the pool of possible events.
• #31: this occurs for each event in the library.
• #32: If more than one possible event is in our possible event pool, we perform one additional comparison.
• #33: on the derivative waveform, which is the most orthogonal subspace.
• #34: We perform a nearest neighbor search on all remaining possible events using the matched filter similarity score in the derivative space.Once the event is identified, we can assign flow, in gpm, using a calibrated fluid resistance value for the specific valve.
• #35: To estimate flow, we turn back to our example waveform and look at the stabilized pressure drop—this is delta P.
• #36: Flow rate is related to pressure change via Poiseuille’s Law, which states that the volumetric flow rate of fluid in a pipeQ is dependent on the radius of the pipe r, the length of the pipe L, the viscosity of the fluid μ and the pressure drop ΔP.This can be simplified using the fluid resistance formulation, which states that the resistance of flow is proportional to the drop in pressure divided by the volumetric flow rate.(click)Simple substitution leads to: flow rate = change in pressure / fluid resistance
• #37: So, we know the change in pressure because that’s what HydroSense measures but we want to know flow rate. To do that, we have to find values for the fluid resistance, which differ according to fixture and fixture location.---This is analogous to ohms law, which relates voltage, resistance, and current.HydroSense measures the change in pressure ΔP. However, in order to estimate flow rate we must know Rf.
• #38: One way to do this would be to measure Rf at every valve in the home and save this in a dictionary. Then, the fixture classification algorithms would look up this value and solve for Q.---To acquire Rf, one could measurethe flow rate (Q) at each individual valve as well as the corresponding pressure change (∆P) at the hydrosense installation location. The division of these two values results in Rf.We can then use the learned Rf values for each valve to estimate flow during an automatically detected water usage event.
• #39: Now, I will go over our in-home data collection and experimental validation.
• #40: We installed HydroSense in a variety of locations: 8 houses, 1 apartment, and 1 cabin with a range in size from 750 sq feet to 4,000 sq ft.We also used three different installation points: 8 hose bib, 1 water heater, and 1 utility faucet connection.
• #41: Each collection session was conducted by a pair of researchers: one would record the sensed pressure signatures to a laptop while the other activated the home’s water fixtures. We recorded five trials per valve. A sink, for example, has two valves (one hot, and one cold) and thus would receive 10 trials—5 for hot and 5 for cold. For each trial, the valve was open for ~5 seconds and then closed.
• #42: For 4 of the 10 homes we also gathered flow data. For the faucet and shower trials, we measured the time it took to fill 1 gallon of water into a calibrated bucket.This provides Q (the flow rate) and allows us to solve for the fluid resistance at that valve and pipe pathway.
• #43: In all, we conducted experiments in 10 locations resulting in 706 trials with 155 flow rate trials and 84 total fixtures tested.
• #44: We conducted a cross-validation experiment that folds our data according to the home in which it was collected.We learn the thresholds for a single home from the other nine homes in the database. Then, we use those thresholds to build the possible events library for each "unknown" event in the database. We classify each “unknown” event in the test home using a leave-one-out method.
• #45: This figure shows the accuracy of fixture-level identification of valve open and valve close events within each home as well as the aggregate 97.9% accuracy of fixture-level classification.Many homes resulted in 100% open and close fixture classification.The worst performing house, H10, was due to noise from the eleven cabins that share the same supply line at the resort.
• #46: This figure shows the accuracy of fixture-level identification of valve open and valve close events within each home as well as the aggregate 97.9% accuracy of fixture-level classification.Many homes resulted in 100% open and close fixture classification.The worst performing house, H10, was due to noise from the eleven cabins that share the same supply line at the resort.
• #47: This graph shows the same data as before but categorized by fixture. Again, nearly all fixtures are above 95% or better. Here, the worst performer is the hot water valve in showers at 89%--although this is still pretty good, we’re not sure why this is and it warrants further investigation.
• #48: For our flow estimate results, three of four houses tested (H1, H4, H5) have error rates below 8% (or approximately 0.16 GPM**), comparable to 10% error rates found in empirical studies of traditional utility-supplied water meters [1]. (click)The fourth house (H7), however, had an error rate above 20%. We believe this isdue to the installation location of the sensor. Whereas the first three homes had HydroSense installed on an exterior water bib, H7’s installation used the hot water heater drain valve.This results in two confounding pressure sources (the supply water main and the gravitational pressure of the water in the tank). It seems the cold water valves were particularly affected.
• #49: When the cold water valves are removed, H7 flow estimation results are on par with the other homes.
• #51: The previous results assumed that we had measured Rf values for every fixture but one can also obtain approximate estimates of Rf by sampling at only certain strategic valves around the home and then using this Rf value for proximal valves.
• #52: Here we have flow data using interpolated Rf. The Y-Axis is the average error in GPM and the x-axis is the number of random Rf samples needed to obtain those error values. As you can see, after randomly selecting five Rf values, the average error dropped 74% to 0.27 GPM.
• #53: We have been working lately on a 2.0 version of HydroSense, which is focused primarily on naturalistic water usage.
• #54: To do this, we’ve built a range of small, lightweight sensors to distribute around the home to collect ground truth data on water fixture usage.
• #58: Infrastructure-MediatedSingle-Point Sensing of Whole-Home Water Activity
• #61: if time, go over slide
• #62: Typical water meters only provide aggregate information on water usage and require costly pipe modifications to install. (click)In constrast, the HydroSensor can simply screws on to a hose water bib.
• #63: Match filtering allows us to compare two signals’ similarity using a pointwise correlation coefficient(click) we save the maximum correlation as the degree to which the two signals are similar
• #64: And two very similar signals will have a higher similarity value,We can perform matched filtering in a variety of signal subspaces. The more orthogonal the different fixture signals are in the subspace, the better. The reason behind this is that orthogonal signals have a correlation of zero (and, thus, a matched filter comparison close to zero), while similar signals are unaffected.
• #65: The current means that provide feedback about water usage are water meters, which tend to be outside and are not designed to be read by the average person and bills, which only provide aggregate information on consumption. This can be extremely problematic, for example, James leak.