Find the difference between plant changes and sensor noise
HomePrecisionFind the difference between plant changes and sensor noise
Precision · Signal and noise

Find the difference between plant changes and sensor noise

This paper shows you how to find the signal in your sensor data and how to ignore the random jitter. Thus you know when to adjust the room. It is equally important to know when to make no adjustment.

Precision9 diagrams9 sources~14 min to read
In this paper
01 Purpose and scope02 Definitions03 Evidence and limitations04 Sources of measurement noise05 Sampling and averaging06 Control limits and adjustment thresholds07 Eight steps to decrease noise this week08 Troubleshooting09 Expected results and limitations10 References
Start here

Purpose and scope

Weak

Limit of the data: Replace the percentages of alarm noise and the control rules from this paper with your measured rate of false positives. If you use the Western Electric rules or the Nelson rules, use the correct name for each.

Each sensor reading in your grow room is the total of two parts: the accurate information (the signal) and jitter that gives no information (the noise). Your task is to find the signal in the noise.

Sensors have a low cost, and dashboards are easy to make. But growers continue to make most decisions with no method. At this time, the task is to find the difference between the signal and the noise. This paper gives you a method for this task. The paper also gives you a method to know when to make no adjustment. This method is equally important.

One flower room can make hundreds of thousands of data points in a week[8]. Thus a method to ignore most of them safely is necessary. More data does not give more information that you can use. It is less easy to use a large quantity of low-quality data for decisions than a small quantity of good data.

The same data: the signal is in the noiseThe rough line is the sensor data. You want to adjust for each low point. The signal is the stable dryback below it.0173450670h9hVWC %
Figure 1. The figure shows one smooth trend (a usual dryback) below a rough line of sensor jitter. The numbers are the same, but the information is different. Only the smooth trend is important for your decisions.
The problem in one sentence

A signal-to-noise problem is the primary cause of most alarm fatigue in grow rooms. In this problem, the jitter is too large, and the signal is not easy to see. Before tuning, most alarms are transients. A transient stops before an adjustment has an effect.

Terms

Definitions

The terms in this paper are from radio engineering, manufacturing and statistics. Signal-to-noise ratio (SNR) is the strength of the signal that is important to you, compared with the strength of all other signals[9]. In a room with a high SNR, the decisions are easy and the room is stable. In a room with a low SNR, the decisions are not easy and you adjust the room too frequently.

SignalA change in the data that is accurate and important for your decisions. Examples are a trend that continues for many readings, a step change, or a rhythm that occurs again each day.
NoiseVariation in the data that gives no information. Examples are a spike in one reading, jitter, or a transient from a door that opens or from an HVAC cycle.
DriftA slow error of a sensor in one direction. The error has the same shape as a trend in the plant. Drift is the type of noise that causes the most problems, because it looks the same as a signal.
Averaging / aggregationThe average of many readings, or of many plants. The random differences of each reading become small, and the information that all readings have stays.
Control limitsLimits that you calculate from the recorded data of the system. Variation in the limits is usual, and variation out of the limits is unusual.
CalibrationThe task in which you compare a probe with a known reference and correct the probe. Thus the readings stay accurate for a long time. Glossary →
The table compares the shapes of signal and the shapes of noise.
Signal: do something about itNoise: ignore it
A trend that continues for many readingsA spike in one reading
A step change that staysJitter that is smaller than the accuracy of the sensor
A rhythm of day and nightA transient from a door, a vent or an HVAC cycle
Some sensors agree on the changeOne probe that does not agree with the probes near it
A trend that agrees with a referenceSlow drift from a probe with no calibration
Setpoint, target band and the readingControl the value to a band, not to one numbersetpointtarget bandChange the control only when the reading is not in the band. Do not change it for small changes in the band.
Diagram. A signal is a reading that is not in the band. Noise is a small change in the band. Change the control for a signal. Do not change it for noise.
Quality of the information

Evidence and limitations

Three types of information

We want each paper to be correct. Thus we examine each paper for information that is not solid. This information comes from the view of one grower, from a small number of sources, or from grower methods that have no controlled test. We identify this information, and we do not show it as solid.

Frequently there is no paper for the decision that you must make. Then we use the reports of other growers and the methods that work in the rooms that we know. This information can help you, but it is not the result of a laboratory test. Use the methods that work for your plants, for your room and for your meters. If a table does not agree with your crop, use the data from your crop and record the difference.

Solid
Tests of plants, specifications or many sources that agree show that the information is correct
  • Control charts and filters decrease the number of false alarms. Sensors have noise.
Grower method
Many growers use this method in their rooms. Start with it, then change it for your room.
  • The sampling interval and the deadbands for automation in the grow room.
Weak
The view of one grower, a small number of sources, one test, or a method that works in one room
  • One percentage of alerts that are noise, from standards of the process industry, with no rate that you measured.
  • The names of the Western Electric rules and the Nelson rules in one mixture, with no reference to a source.

If you find an error, make a report. We will correct it. Make the report on GitHub, with the name of the paper and the information that is incorrect. If you have a source, include it: Make a report of an accuracy problem. The regulations, labels and licenses of your area always override each procedure in these papers. The label weak in a paper shows the points where the risk of an error is highest.

Primary item 1

Sources of measurement noise

You cannot decrease a source of noise if you cannot find it. Noise in a grow room comes from six sources. Three are physical sources that are part of the measurement. They are the sensor (electronic jitter, drift, no calibration), the position (a probe near a vent or a light, or a sample of one) and the biology (one plant is different from the next). The other three are procedural sources, and thus they have the lowest cost to correct. They are the environment (doors, HVAC cycles), the operator (manual sampling that is not the same each time) and the data pipeline (gaps in the data, incorrect units, clock skew).

The six sources of the measured noise1Sensorjitter, drift,no calibration→2Positionedges, vents, 1probe→3Biologydifferencebetween plants→4Environmentdoors, HVACcycles→5Operatorsampling thatchanges→6Data pipelinegaps, units,clock skewThe first three are physical. The last three are procedural and cost less to correct.
Figure 2. The figure is an Ishikawa diagram (fishbone diagram). Six sources cause the ‘measured noise’ that you see on the dashboard. Correct the procedural sources, which cost less, before you think that the equipment has a fault.
Drift is the noise that causes the most problems

Calibrate each probe with a known reference, because calibration is the only protection from drift. Drift moves slowly in one direction, the same as a trend in the plant. Thus drift can give you incorrect information for many weeks. A pH probe or an EC probe with no calibration can have drift. You can measure the drift in weeks to a month[6]. One probe is not sufficient to make sure that the data are correct.

A calibration for each sensor is necessary. In low-cost permittivity sensors for soil moisture, a calibration for each sensor decreased the error by approximately 70%, compared with the calibration of the manufacturer[6]. When a grower thinks that the room is a problem, the noise is frequently the problem.

The six sources, with the source of lowest cost to correct first. Correct the procedural rows before the physical rows.
SourceSymptomTaskCost to correct
Data pipelineGaps, incorrect units, clock skewExamine the data pipeline. Make sure that the units and the timestamps do not change.Low
OperatorReadings that change with the person who made themWrite an SOP with the same time, the same method and the same position.Low
EnvironmentSpikes that agree with doors or HVAC cyclesMake the dead-bands wider. Filter the transients.Low or medium
SensorJitter, drift in one directionCalibrate at set intervals with a referenceMedium
PositionOne probe that is different from the probes near itMove the probes from the edges to the middle of the canopyMedium
BiologyDifferences between plantsCalculate the average of many plants (you cannot remove the differences)You cannot remove it
Primary item 2

Sampling and averaging

The noise filter with the largest effect is to calculate the average of more sensors, and not to use only one sensor. Growers frequently ignore this filter. In your room, when you calculate the average of twelve plants across the bench, the differences of each plant become small. Only the information that all plants have stays.

The name of this method is replication. The error of an average decreases with the square root of the number of readings that you use[5]. Four probes decrease the noise to approximately one half, and nine probes decrease it to one third.

The frequency of the measurements is equally important. If the interval between the samples is too long, a fast pattern changes to a slow pattern that is not in the plant. The interval between the samples causes this incorrect trend. This error has the name aliasing.

The usual method comes from the Nyquist–Shannon sampling theorem[4]. Use a sampling frequency of a minimum of two times the frequency of the fastest pattern that you must see. To see an irrigation response of 30 minutes, record the data at intervals of 10–15 minutes.

Undersampling makes an incorrect trendThe fast wave has an oscillation at each step. A sample from one step in two shows a slow drift that did not occur.017345067signal
Figure 3. A fast oscillation, when the interval between the samples is too long, gives a slow incorrect wave. This incorrect trend can make you adjust the room for a change that does not occur.
A high sampling rate has a cost

A sampling rate that is too high adds noise and storage cost. It also makes you want to adjust the room for each jitter. Use a sampling cadence that agrees with each channel.

A correct cadence for each channel. Make the sampling rate agree with the rate of change of the quantity that you measure.
ChannelCadenceBecause
Substrate VWC / EC1–5 minThe irrigation responses are fast. The cadence must show the shape of the dryback.
Air temperature / RH / VPD1–5 minThe HVAC changes quickly, and the VPD is the value for crop steering at this time.
CO₂1–5 minThe CO₂ increases and decreases when doors open and when the system injects CO₂ for short periods.
Pour-through pH / EC1×/day, same timeThe indicator of slow drift. A reading each day at the same time is better than checks at random times that add noise.
Plant morphology2–3×/weekThe growth is slow. More frequent measurements add only noise from the operator.
Primary item 3

Control limits and adjustment thresholds

The most important method in this paper is from the work on quality in manufacturing: statistical process control (SPC). Walter Shewhart, at Bell Laboratories, divided all variation into two types[1]. Common-cause variation is the jitter of a stable system. It stays in the control limits, and you must not adjust the system for it. Special-cause variation has a cause that you can find. It is out of the limits, and you must examine the cause.

You calculate the control limits from the recorded data of the system, and not from an estimate. The usual limits are the mean plus or minus three standard deviations[1]. W. Edwards Deming showed that tampering (an adjustment of the system for common-cause jitter) increases the variation of a system and does not decrease it[2]. SPC lets you do a task in cultivation that is not easy: monitor a number that changes and, correctly, make no adjustment.

Control chart: most points are usual, one point is a signalPoints in ±3σ are common-cause: ignore them. A point that is out of the limit is special-cause: find the cause.±1σ: usual jitter±3σ: control limitsspecial cause: examine30 %VWC50 %VWC70 %VWC
Figure 4. The figure shows a mean line, a zone of ±1σ and control limits at ±3σ. Most readings have jitter in the limits, and this jitter does not cause a problem. One point above the top limit is the point that you must examine.[1]
More than the limits: the Western Electric rules
  • Nelson trend (frequently 6 points) in which all the points go in the same direction. The trend shows a drift, also when the points are in the limits.
  • Western Electric: 8 points on one side of the mean show that the system changed.
  • Points too near the mean are unusual. They are frequently a sign that a filter is too strong, or that a person made the data and did not measure them.

These Western Electric rules and Nelson rules find changes that you cannot find when you compare only one point with the limits. They do not cause false alarms for usual noise[3]. A grower who adjusts a stable system many times in a day is usually the largest source of noise in the room.

Start this week

Eight steps to decrease noise this week

It is not necessary to have new capital for the steps with the largest effect. For most steps, you must only do the same work each time. Do the steps in the sequence below.

  1. 1
    Do not monitor the values at all times
    A display that shows the values at all times causes tampering. Examine the decision charts at set times, and not the raw data all day.
  2. 2
    Set a sampling cadence for each channel
    Use the table above. Use a high rate for fast channels and a low rate for slow channels. Do not use a sampling rate that is higher than the rate of your adjustments.
  3. 3
    Add a rolling average to each decision chart
    Filter the line that you use for decisions. Keep a display of the raw data that you can show in one step. Then you can see an emergency.
  4. 4
    Calculate the average of 6–12 probes
    Do not use only one source for a decision. Use the average, and let the one probe that is different have a small effect.
  5. 5
    Write SOPs for manual readings
    Use the same time, the same method and the same position each time. The SOP removes the noise from the operator at no cost.
  6. 6
    Calibrate at set intervals and keep a record
    In the next three months, set a calibration cadence with a reference. This calibration cadence is your only protection from drift.
  7. 7
    Calculate the control limits for your 3 primary KPIs
    Calculate the mean ±3σ from your recorded data. Then you know which values are unusual.
  8. 8
    Examine the positions and make dead-bands wider
    Move the probes from the edges, the vents and the lights to the middle of the canopy. Make a dead-band wider if the equipment starts and stops frequently.
The test of one sentence before you do something about a number

“Is this value out of its usual range, for more than one reading, and do my other sensors agree?” If one of the three is no, it is noise. Ignore it.

Adjust or ignore: three gates1Out of limits?out of ±3σ→2Stays?more than onereading→3Sensors agree?near sensorsagree→4All yes → ADJUSTit is signal→5One no → IGNOREit is noiseThree yes/no gates are between an alarm and a decision.
Figure 5. The figure shows the decision flow. Each gate must be yes before you adjust the controls. If one gate is no, go to ‘ignore’.

Three filters are sufficient for almost all the data from a grow room. The moving average is easy. The EWMA (exponentially weighted) gives more weight to the last readings, and thus it has less lag for the same smoothing[7]. The median filter removes spikes of one point. Start with a filter window of approximately 30–60 minutes, and then adjust it.

Raw data and rolling average: the trend is clearThe filtered trend (the middle of the line) shows a dryback that you cannot see in the spikes. The cost is a small lag.0163249650h9hVWC %
Figure 6. The figure shows the raw data as a rough line, and the filtered trend through the middle of the line. The dryback is easy to see. The only cost is a small lag of a known size. Thus you keep a display of the raw data for an emergency.
Problems to prevent

Troubleshooting

Two errors that are opposite are over-smoothing and tampering. A weak filter makes the data clear, but a filter that is too strong causes damage to the data. If the filter window is too wide, the filter removes a fast change that is not noise. Examples are a pump fault or an EC spike from a blocked dripper, and you must see these changes when they occur. The opposite error is to adjust the room for each small change. Such adjustments make the room less stable[2].

A feedback loop with data that have noise, or with a tuning that is too strong, makes corrections that are too large. The corrections go in one direction and then in the other direction, and the loop does not become stable. A thermostat that is too sensitive is an example. It starts the heater when it reads one degree low, and the temperature becomes too high. Then it starts the chiller, and the temperature becomes too low. As a result, the temperature of the room changes continuously.

This error has the name hunting. The symptom of hunting is a regular saw-tooth pattern in the temperature, the RH or the VWC. The day/night cycle does not cause this pattern. If your HVAC or fertigation system makes opposite corrections one after the other, the cause is frequently a sensor with noise. The cause can also be a dead-band that is too small. Think of an equipment fault only after you examine these two causes.

The feedback loop and the position where noise starts1Setpointyour target→2Controllermakes decisions→3Actuatorvalve / fan /heater→4Plant / roomthe change thatoccurs→5Sensor← NOISE STARTSHEREThe controller reads noise from the sensor as an error and makes a correction. Filter the input at the sensor.
Figure 7. The loop goes setpoint → controller → actuator → room → sensor and then back. Noise from the sensor causes the most problems. The controller cannot find the difference between this noise and an error in the room. Thus it makes a correction for an error that does not occur.
Filter at the source and make the dead-band wider

Filter the input of the controller at the output of the sensor, before the controller makes a decision. If you do not filter the input, each control loop in the room adjusts for noise. A wider dead-band and a filtered input frequently correct climate equipment that you think has a fault. The equipment has no fault.

Results and limits

Expected results and limitations

Facilities go through stages in a set sequence. When you know your stage, you know the next step.

stage
1 · Blind
No data and no method.
stage
2 · Logged
Data but no method: much noise, and too many adjustments.
stage
3 · Filtered
Filters and a sampling cadence. The change to this stage has the largest effect.
stage
4 · Controlled
You adjust only for special causes.
stage
5 · Tuned
Closed loop. The target is uniformity.

Most commercial rooms are at the Logged stage. A room at the Logged stage causes more stress to the grower than a room at the Blind stage, because the noise causes many alarms. The target is one stage up, and not the last stage in one step. The change from the Logged stage to the Filtered stage has the largest effect of all the steps in this paper. It uses only filters and a sampling cadence. It has almost no cost, and you must only do the same procedure each time.

Monitor a small number of KPIs that have a high signal, and not forty instruments. The indicator that is the most related to commercial success is frequently uniformity. You measure it as the coefficient of variation of the batch, and not as the peak yield. Tests show that, when the variation of the environment in space and in time decreases (a lower coefficient of variation), the growth and the quality of the crop increase[8].

Compare two batches. In the first batch, each plant has a yield of 95 g (3.4 oz). In the second batch, the average is 110 g (3.9 oz) and the range is 40 g (1.4 oz). The first batch has a higher commercial value.

The coefficient of variation measures the noise in your crop. In commercial production, uniformity is better than the peak yield.
KPIs with a high signal: monitor theseIndicators with no value for decisions: ignore these
Grams for each kWhTemperature from one sensor at one time
Dryback trendTotal quantity of data points recorded
Time in range for the VPDPeak / record readings
DLI supplied compared with the targetNumber of alarms
Batch coefficient of variationNumber of dashboards
Summary

The target is a higher signal-to-noise ratio, and not a room with no noise. One stage up is the correct target. Filter before you adjust the room, and let a number that stays in its control limits change with no adjustment.

Next, read the paper smart watering with VWC and EC to find how a clean, filtered signal controls the irrigation. Then read the paper closed-loop control to find how a closed loop operates with no hunting.

Related papers
Root-zone condition · TEROS-12
How to read the values from the sensor
Irrigation control · VRWE
How the controller makes irrigation decisions
The closed loop
Levers, signals and plant state

References

  1. Control chart. Wikipedia. Describes Walter A. Shewhart's development of statistical process control at Bell Telephone Laboratories (1924), the distinction between common-cause and special-cause variation, and the convention of setting control limits at ±3 standard deviations from the process mean. (source not from a journal, for example from a manufacturer) https://en.wikipedia.org/wiki/Control_chart
  2. The Funnel Experiment. The W. Edwards Deming Institute. Demonstrates that adjusting (tampering with) a stable process in response to individual outcomes increases its variation rather than reducing it; leaving a stable process alone (Rule 1) yields the least variation. (source not from a journal, for example from a manufacturer) https://deming.org/explore/the-funnel-experiment/
  3. Anhoej J, Wentzel-Larsen T. Sense and sensibility: on the diagnostic value of control chart rules for detection of shifts in time series data. BMC Medical Research Methodology. 2018;18:100. doi:10.1186/s12874-018-0564-0. Evaluates the Western Electric SPC control-chart rules for detecting non-random (special-cause) variation in sequential data. https://doi.org/10.1186/s12874-018-0564-0
  4. Nyquist-Shannon sampling theorem. Wikipedia. States that perfect reconstruction of a band-limited continuous signal requires a sampling rate greater than twice the highest frequency present in the signal; under-sampling causes aliasing and irrecoverable information loss. (source not from a journal, for example from a manufacturer) https://en.wikipedia.org/wiki/Nyquist%E2%80%93Shannon_sampling_theorem
  5. Blainey P, Krzywinski M, Altman N. Points of significance: Replication. Variation: use it or misuse it - replication and its variants. PMC3424707. Explains that the standard error of an estimate decreases with the square root of the number of replicates, so replication reduces measurement variance in proportion to sample size. https://pmc.ncbi.nlm.nih.gov/articles/PMC3424707/
  6. Bogena HR, Huisman JA, Schilling B, Weuthen A, Vereecken H. Effective Calibration of Low-Cost Soil Water Content Sensors. Sensors (Basel). 2017;17(1):208. doi:10.3390/s17010208. Documents sensor-to-sensor variability and drift in low-cost permittivity-based soil moisture sensors, the two-step permittivity-to-water-content calibration, and large accuracy gains (RMSE reduced ~70%) from sensor-specific calibration. https://doi.org/10.3390/s17010208
  7. Roberts SW. Control Chart Tests Based on Geometric Moving Averages. Technometrics. 1959;1(3):239-250. doi:10.1080/00401706.1959.10489860. Introduces the exponentially-weighted (geometric) moving average chart and shows it responds faster to small persistent process shifts than an ordinary moving average of equivalent smoothing. https://doi.org/10.1080/00401706.1959.10489860
  8. Story of a paper: A Study of the Effects of Enhanced Uniformity Control of Greenhouse Environment Variables on Crop Growth. Energies. 2019;12(9):1749. doi:10.3390/en12091749. Shows that reducing spatial/temporal fluctuation (improving uniformity, i.e., lower coefficient of variation) of environment variables improves crop growth and quality. https://doi.org/10.3390/en12091749
  9. Signal-to-noise ratio. Wikipedia. Defines SNR as the ratio of signal power to background-noise power, a measure that originated in electrical/communications engineering and is now applied across telecommunications, imaging, and scientific measurement. (source not from a journal, for example from a manufacturer) https://en.wikipedia.org/wiki/Signal-to-noise_ratio

Each citation in the paper, for example [1], refers to an item in this list. Each item is a primary source or an instruction from an authority, unless the item shows a different type. The results of cannabis tissue culture change much with the genotype. Before you use the dilutions, hormone doses and local regulations, make sure that they agree with the primary sources.