Real time data validation. Methodology guidelines, comments, examples of application and prototype software tool for on-line (RTC) data validation

Transcription

1 Real time data validation Methodology guidelines, comments, examples of application and prototype software tool for on-line (RTC) data validation

2 212, 213 PREPARED The European Commission is funding the Collaborative project PREPARED Enabling Change (PREPARED, project number ) within the context of the Seventh Framework Programme 'Environment'. All rights reserved. No part of this book may be reproduced, stored in a database or retrieval system, or published, in any form or in any way, electronically, mechanically, by print, photoprint, microfilm or any other means without prior written permission from the publisher

3 COLOPHON Title Real time data validation. Methodology guidelines, comments, examples of application and prototype software tool for on-line (RTC) data validation. Report number PREPARED Deliverable number D3.3.3 Author(s) Anders Lynggaard-Jensen (DHI), Hans Peter Hansen (DHI), Erling Brodersen (Aarhus Water) Quality Assurance Jean-Luc Bertrand-Krajewski, (INSA) Document history Version Team member Status Date update Comments.1 Anders Lynggaard Jensen.2 Hans Peter Hansen Erling Brodersen.3 Anders Lynggaard Jensen 1. Anders Lynggaard Jensen 2. Anders Lynggaard Jensen Draft Creating document with first methodology descriptions Draft Description of prototype software based on test examples from WWTP Viby, Aarhus Draft Updating, formatting all descriptions added data filtration Final Minor changes Final Abstract + Update incl. actions This report is: PU = Public

4 Abstract This deliverable presents a software prototype for real time data validation including the principles of the validation methods and how these are applied on a full scale wastewater treatment plant. The validation methods are divided into two groups short term and long term data validation. Short term validation includes methods for: Missing data Range check Rate of change check Running variance Long term validation includes methods for: Expected mean check Acceptable trend check Real time data validation is especially important to use when sensor signals are used for automatic process control. However, it is equally important to define and carry out automatic actions, when the validation detects something is wrong. Therefore the prototype includes methods for actions as well. PREPARED May 213

5 Contents Abstract 1 Contents 2 1 Introduction 3 2 Single data validation methods Validation Principle Short Term Methods Gap Filling Range Check Rate of Change Check Running Variance Check Overall Assessment Long Term Methods 9 3 Cross Validation Methods 12 4 Data filtration 13 5 Prototype software Configuration of data validation Configuration of data validation actions Data validation in practice 18 PREPARED May 213

6 1 Introduction The most common errors in the input data used for RTC include missing data, measurement values out of range, peaks (outliers) and constant measurement values (indicating that the sensor is out of order). It is possible to check the data for these typical errors using simple methods known as single data validation. However, even if these methods are simple it is not common that they are implemented directly in PLCs the range check might be an exception. Therefore, the single data validation methods are applied as part of the interaction between the RTC system and the SCADA. One or more methods can be applied as data is read from the SCADA, and each method as a result gives a confidence value between and 1 for each data point. If the confidence is lower than a preset threshold, different actions can be taken (avoid using data for control, suspend the control based on the RTC-algorithm and fall back to default control by local control loops, calibrate/repair sensor, etc.). The single data validation methods are based on the immediate, recent reading, and therefore other methods must also be used to assess the quality of the data in the long term, looking over hours and days (e.g. look for gradual drift ). These long term validation methods are partially based on the same principles as the short term validation methods, but are using data from a much longer period of time typically looking days back from the actual time. The long term methods also include additional validation methods to distinguish between instrument drift and a real (actual) gradual change in the process variable. These methods include cross validation methods. Data validation is an essential step, required before measured data can be used for automated control strategies. In addition to data validation, it might also be necessary to apply filtering methods before the validated data can be used in real time control, because even correctly measured data can still include a variation that is too pronounced for closed loop RTC algorithms. PREPARED May 213

7 2 Single data validation methods 2.1 Validation Principle The validation methods are based on how much confidence should be paid to measurements which have unexpected behaviour. If a measurement is within the limits of what can be expected the confidence will be 1, and if it is showing values which are highly unexpected or even impossible the confidence will be. However, between these two situations it might be difficult to judge and confidence will gradually decrease when the measurement is moving from expected values towards unexpected values (figure 2.1). Confidence 1 min. low high max. Limits Error OK Error Criterion Figure 2.1: Principle of applying confidence to a measurement. Table 2.1 gives an overview of the calculation of the confidence for different (short term) validation methods, and as can be seen the setting of the parameters for the validation is quite important, and as mentioned these should accommodate both sensor and process response. Table 2.1: Calculation of the confidence for a measurement Method Confidence function Parameters used Gap Filling (GF); N = no. of time steps Range Check (RC); x = measurement value Rate of Change Check (RCC); x = change of measurement per minute Running Variance Check (RVC); x = running variance of last n measurements = 1 * (1 n i / N); n i < N (n i = 1, 2,...N) = ; n i >= N = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min = 1; L low < x = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min Overall Assessment = Min (C GF, C RC, C RCC, C RVC) - N = 5 L max = L high = 2 L low = 5. L min = 1.25 L max = 2.5 L high = 2. L low = -2. L min = -2.5 L low =.25 L min =.1 n = 5 PREPARED May 213

8 The table also includes the parameter setting for a measurement logged every minute for 3 hours and showing an hourly variation and a trend. Further, the data used as an example has: Missing values at :15 - :2 and :3 - :32. Peak values (outliers) at 1:5 and 2:2 Constant values at 1:55 2:5. In order to illustrate the single data validation methods, each method is applied on the same data (figure ). 2.2 Short Term Methods Gap Filling Every time a single data point is missing it has to be considered if a possible control action should use a default option or if it is possible to use an estimate for the missing value and then continue normal operation based on this. Depending on the variability and knowledge of the process measured, several possibilities exists for estimation of missing data (gap filling). The most simple is of course to use the value from the last measurement or use the trend from previous values of the measurement. However, if a correlation with other measurements exists or a model is available, better estimates can be obtained. Method Confidence function Parameters used Gap Filling (GF); N = no. of time steps = 1 * (1 n i / N); n i < N (n i = 1, 2,...N) = ; n i >= N N = Gap Filling : :3 1: 1:3 2: 2:3 Figure 2.2: Gap filling. Measured values: Blue. Calculated confidence: Red. Method adjusted to give zero confidence after 5 missing values and estimate a value using the previous measured value. No matter what method used it shall be reflected that confidence in an estimate is lower than confidence in a value from a real measurement. Furthermore, confidence shall decrease more and more for each consecutive value missing - eventually resulting in zero confidence. How fast confidence shall reach zero and if the decrease in confidence shall be linear or decrease faster for each value missing depends on the belief in the estimates. PREPARED May 213

9 2.2.2 Range Check Values out of range can be related to either the measurement itself or knowledge concerning the process monitored. The measurement is having a normal working range, where if properly calibrated - values are believed to be true. This working range should fit the normal variation band of the measurement. The working range of a sensor is not necessarily the same as the full scale (often the quality of a measurement is lower at the ends of the full scale) which suggests that confidence in the measurement shall be maximum within the normal working range and decrease gradually to zero through two warning bands on each side of the working range. The result of applying these assumptions is shown in figure 2.3, where the calculated confidence shows unacceptable values in the start of the period, but after a few minutes the values are only triggering a warning. Due to the hourly variation and the drift in the measurement a warning is again issued after an hour (values too low but might be valid) and after 2.5 hours (values too high but might be valid). Method Confidence function Parameters used Range Check (RC); x = measurement value = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min L max = L high = 2. L low = 5. L min = Range : :3 1: 1:3 2: 2: Figure 2.3: Range Check. Measured values: Blue. Calculated confidence: Red. Method adjusted to accept values in the band from 5 to 2 (confidence = 1) and give zero confidence for values below 1.25 and above Rate of Change Check The rate of change in a process variable is an important indicator of the signal quality/reliability. When the rate of change is higher than the realistic variations in the measurement, and/or the noise generated by the sensor, it is usually a clear signal of a disturbance or a problem. The confidence in data that exceeds the feasible (physical, actual) rate of change shall of course be zero. In such cases, it is often reasonable to accept the previous value (before the disturbance) as a reasonable estimate of the correct value of the measurement. PREPARED May 213

10 Smaller outliers and unexpected peaks in a measurement can also be detected, but it can be difficult to distinguish between these when using a simple method. In such cases, confidence can be calculated using a warning band similar to the method used for the range check. The measurement value used in the example varies with approximately 15(units)/half hour corresponding to.5/minute and the random component is not greater than 1.5/minute. Therefore, as long as the numerical difference between two consecutive values is less than 2 the values are given a maximum confidence. If the width of the warning band is set to.5, a numerical difference between two consecutive values has to be greater than 2.5 before the calculated rate of change is declared as a peak value (peak height greater than 2.5). Figure 2.4 shows that the peaks are detected without any problem. Also the shift from estimates to real measurement values at :2 is detected, but as this is already known from the gap filling it can safely be neglected. However, this shows that a sudden shift in measurement values also will be detected by this method. Method Confidence function Parameters used Rate of Change Check (RCC); x = change of measurement per minute = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min L max = 2.5 L high = 2. L low = -2. L min = Rate of Change : :3 1: 1:3 2: 2:3 Figure 2.4: Rate of change. Measured values: Blue. Calculated confidence: Red. Method adjusted to accept peak heights of +/- 2 (confidence = 1) and give zero confidence for peak heights of +/ Running Variance Check Constant measurement values are typically a result of planned (automatic calibration) or unexpected (failure) sensor performance, because normally functioning sensing devices always have a small variation in the measurement values. None of the previous methods is able to detect if a sensor fails and locks on a fixed measurement value, if this value is within the working range. PREPARED May 213

11 Method Confidence function Parameters used Running Variance Check (RVC); x = running variance of last n measurements = 1; L low < x = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min L low =.25 L min =.1 n = Running Variance : :3 1: 1:3 2: 2:3 Figure 2.5: Running variance check. Measured values: Blue. Calculated confidence: Red. Method adjusted to use the last 5 measurement values and to accept a variance higher than.25 (confidence = 1) and give zero confidence for a variance less than.1 Therefore, the need for a validation method allowing a certain variation as the norm is obvious. The method shall decrease the confidence if the expected variation decreases and end up with a zero confidence when the variation nearly has disappeared (still allowing for bits to shift in analog to digital converters). The running variance over 5 minutes for the measurement used in the example is always higher than.5, so the method applies a max. confidence for a running variance bigger than.25 and a min. confidence for a running variance less than.1. Figure 2.5 shows that the constant values around 2: are nicely detected together with the estimates used for gap-filling Overall Assessment One or more if not all of the above methods can be applied at the same time, and an overall assessment of the confidence in the measurement value is calculated as the minimum of the involved calculated confidences (figure 2.6). However, as the resulting confidence has focus on sensor performance here and now without any comparison with reference values, it will not be able to express anything about the trueness (measured as bias) or long term drift of the values. Method Confidence function Parameters used Overall Assessment = Min (C GF, C RC, C RCC, C RVC) Resulting Confidence : :3 1: 1:3 2: 2:3 Figure 2.6: Overall assessment. Measured values: Blue. Calculated confidence: Red. PREPARED May 213

12 2.3 Long Term Methods A known bias is not a problem, as this can be compensated and an unknown bias should be avoided by proper calibration procedures. However, a bias in the measurement value can slowly develop in time, therefore, a validation method detecting long term drift in the measurement value is needed. Checking for long term drift includes two steps, first detection and then identification of the cause (instrument based or a true long term trend in the measurement). Detection can be done using two different methods, Expected Mean Check and Acceptable Trend Check, both of them demonstrated below (figure 2.7 and 2.8) using a measurement having a range of 1 and logged every 12th minute (5 times per hour). Data from a period of 14 days shows a process variable with a periodic variation of half a day, an expected mean value of approximately 5 and a drift in the measurement starting approximately after a week. Method Degree of Compliance with Limits Parameters used Expected Mean Check (EMC); x = moving average of last n measurements = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min Expected Mean = 5 L max = 55. L high = 52.5 L low = 47.5 L min = 45. n = 6 Measurement and Compliance 1, 9, 8, 7, 6, 5, 4, 3, 2, 1,, Time in days Moving Average Measurement Degree of Compliance Moving Average Figure 2.7: Expected Mean Check. The expected mean has a value of 5 and the moving average (calculated using the last 6 measurement values) is allowed to vary in the band from 47.5 to 52.5 with warning bands on each side with a width of 2.5. The Expected Mean Check method is based on a calculation of a moving average of the last n measured values. The moving average is compared to the expected mean value and is allowed to vary within a certain range, and the method is therefore a long term equivalent of the range PREPARED May 213

13 check method. It is important that n is chosen in such a way that the moving average is able to accommodate possible periodical variations in the process monitored. The Acceptable Trend Check method is based on a calculation of a moving slope. The slope is calculated from the last n values of the moving average calculated above. The moving slope is compared to an acceptable trend, and the method is therefore a long term equivalent of the rate of change check method. It is important that n is chosen in such a way that the moving slope in fact shows long term drift, and the n for this method is typically 2-5 times higher than for Expected Mean Check method. As can be seen in figure 2.8, a drift is detected and a warning given after 7 days by the moving slope. However, the moving average (figure 2.7) shows that the resulting bias (or drift in the process variable) is still within acceptable limits. During day 8 the drift becomes unacceptable (figure 2.8) and the resulting bias increases over the next two days to an unacceptable level (figure 2.7). After day 13 the drift disappears again but leaves the measurement with an unacceptable bias. Method Degree of Compliance with Limits Parameters used Acceptable Trend Check (ATC); x = moving slope (slope determined by linear regression of last n moving average values) = ; x > L max = 1 * ((L max x) / (L max L high)); L high <= x <= L max = 1; L low < x < L high = 1 * ((x L min) / (L low L min)); L min <= x <= L low = ; x < L min L max = 1. L high =.75 L low = -.75 L min = -1. n = 24 Measurem ent and Compliance 1, 9, 8, 7, 6, 5, 4, 3, 2, 1,, Time in days Moving Slope Measurement Degree of Compliance Moving Slope Figure 2.8: Acceptable Trend Check. The moving slope (calculated using the last 24 measurement values) is allowed to vary in a band from -.75 to.75 with warning bands on each side with a width of.25. PREPARED May 213

14 These methods enables an operator to assess the warnings, and even if both methods works in real time, they are (at least in this example) not so time critical as the single data validation methods. Therefore, actions taken do not necessarily have to be started automatically (even if they could), and the operator can therefore proceed to the next step, which is to determine the cause of the drift. The operator can decide to make a manual measurement (measure a level, analyze a grab sample, etc.) or he can invoke a cross validation method (see next section) and compare the result to the measurement values, and thereby judge if the drift is instrument or process based. PREPARED May 213

15 3 Cross Validation Methods A cross validation method exploits the possible correlation between measurements and is therefore a multi data validation method. If the measurements are highly correlated, it is possible to build and calibrate a model (simple or complex) describing the relation between two or more measurements. It can be a deterministic or a statistic model or a combination (known as a greybox model) and it can have one or more measurements from different sensors as input and one or more of measurement(s) as output. If such a model is running in real time giving a new estimate every time data are logged from different sensors, the estimate can be regarded as a measurement with the same properties as other measurements. Such an estimate is also known as a software sensor or a virtual sensor. Some of the most simple examples of software sensors are: a flow sensor based on a level measurement and a flow-depth (Q/H) relationship (one input, one output), a mass sensor based on a flow and a concentration measurements (two inputs, one output), a process rate sensor based on several consecutive measurements of a concentration (at least two inputs, one output), and finally the more complex examples of software sensors which becomes real time modeling of the sewer system itself (multiple inputs, multiple outputs). If a software sensor measures the same parameter at the same location as a real sensor, cross validation is quite straightforward, because the difference of the two measurements then can be regarded as a new measurement with an expected mean at zero. This new measurement can undergo a range check with narrow limits or directly give an estimate of a bias, which can be followed by the expected mean check. Such a software sensor is also very useful in the case of missing data, because the estimate of the measurement then can be used for gap filling. The simplest cross validation can of course be performed, if an application (for operational security) uses redundant (two or more of the same type of) sensors. Another simple cross validation can be set up in a situation where a flow meter, a level measurement and a Q/H relation are available. PREPARED May 213

16 4 Data filtration Data filtration is a necessary step in the process of turning on-line sensor/analyser data into useful information. On-line measurements can be rather noisy and fluctuating, which can make them difficult to overlook and difficult to use in calculation of set-points for control (closed loop or as a decision support function). Table 4.1: Data filters to be used in real time on measurements from sensors and analysers. x t: raw measurement at time t; y t: filtered measurement at time t; n: number of time steps involved Filter General expression Comment Moving average y t n 1 1 n i x t i Also called a rectangular filter. All measurements included are given the same weight. Filters out periodical events with a period equal to the time period covered. Parzen filter y t n 1 i w t i x t i Also called low-pass filter if weights calculated to decrease with age. Filters out events occurring more often than the time period covered. Robust filter y t n n i x t i x max x min Filters out the max and min value in the time period covered and calculates the mean of the rest. Exponential filter y y a x y ) t t 1 ( t t 1 An auto regressive filter remembering past measurements according to the exponential factor a. If a=.5 an abrupt change in a measurement will have reached 97% of its value after 5 steps. Four of the most commonly used (and simple) filters are described in table 4.1. Filtering gives a more smooth signal which is better suited for control purposes. However, it has to be remembered, that filtering add to the response time of an on-line measurement as shown in figure 4.1. Measurement Moving Average Parzen Filter Robust Filter Exponential Filter Timestep Figure 4.1: The response of a filtered measurement to a step change in the measured value. PREPARED May 213

17 Figure 4.2 shows two of these filters working on a suspended solids measurement (top) in the bottom outlet from a clarifier and a sludge blanket measurement (bottom) in the clarifier. Both measurements are used for real time control of the clarifier, but as can be seen fluctuations in the measurements makes it necessary to use filters before the measurements can be used for calculations of set points. The Suspended Solids measurement has a periodic behavior due to a sludge scraper passing every 16th minute. Therefore a moving average with a period of 16 minutes has been selected to filter this measurement. The sludge blanket measurement shows a more random behavior, therefore an exponential filter has been selected for this measurement. Figure 4.2: Top: suspended solids measurement in a clarifier outlet; raw signal in red, filtered signal in blue. Bottom: sludge blanket level measurement in a clarifier; raw signal in brown, filtered signal in orange. PREPARED May 213

18 5 Prototype software Prototype software for real time data validation has been developed based on the specifications in the previous chapters. However, real time data validation is more than the validation itself. As the validation is automatic and therefore a service running 24 hours a day, 7 days a week, it shall be possible to define what action shall be taken, if the validation fails. Further, this action shall be started automatically as well. Examples on configuration of the data validation and the actions to be taken using the prototype software are described below together with the results obtained with the running prototype. 5.1 Configuration of data validation Configuration of data validation consists of 3 steps: 1. Selection of validation method 2. Adjustment of parameters for the method 3. Configuration of confidence thresholds. Figure 5.1: Selection of validation method and parameters for the method for a measurement of suspended solids after secondary clarifiers at the Viby wastewater treatment plant in Aarhus. Figure 5.1 shows the dialogue box for configuring the first two steps. After selecting a validation method (left), the right dialog for parameter settings pops up (right). When this is done the confidence for the measurement will be calculated each time the measurement is logged to the database. Several validation methods may be running at the same time for the individual measurements. PREPARED May 213

19 The quality of a measurement is now defined by threshold values for the calculated confidence. Figure 5.2 shows that two threshold values are selected. These are now the limits that define the selected qualities. In this case it is decided that the measurement will be OK if confidence is over 8. Between 2 and 8 the quality is in a warning state and below 2 the quality is critical. Figure 5.2: Selection of confidence thresholds and associated qualities for a measurement of suspended solids after secondary clarifiers at the Viby wastewater treatment plant in Aarhus. Each time the thresholds are crossed the quality changes and the software also generate an event that the quality has changed. This event is used to generate an action as explained below. Figure 5.3: Extract from database showing Time stamp, Duration (seconds), Value, Quality and Confidence for the suspended solids measurement as the result of the above configuration. PREPARED May 213

20 5.2 Configuration of data validation actions Configuration of actions is done using the dialogue in figure 5.4. The dialogue includes at the top left a filter helping the user to find the measurements to define actions for as all defined measurements will appear in the drop down box at top right. Selecting a measurement from the drop down box (here TagName XM-TU-FI74-PV) will reveal the data validation method and its parameters and the quality settings for the measurement. Figure 5.4: Configuration of actions to be taken when the quality of the suspended solids measurement (TagName: XM-TU-FI74-PV) changes. Actions selected: Sending SMS messages. Alarms activated on Critical. Will only be sent in the period 7: until 22: to the mobile no. stated including the Subject and Body text in the message. Sending mails. In figure 5.4 when quality goes to critical. Mail addresses, which will receive the text written in the Subject and Body, are defined. In addition alarms are sent to the SCADA system (default). PREPARED May 213

21 5.3 Data validation in practice The number and types of sensors used in wastewater treatment plants are numerous. There are sensors measuring flow, level, pressure, temperature, NH4-N, NO3-N, PO4-P, ph, dry matter, turbidity, sludge level, sludge volume etc., and many different locations and functions are linked to the various sensors. In other words each sensor has its own unique function depending on its location in the wastewater treatment plant. In the data validation of a sensor it is therefore important to identify the parameters to be validated as they always will differ from location to location and from sensor to sensor. However, it is important to notice that sensors used today are generally very stable in operation and do not make many errors. Lacking or incorrect maintenance on the other hand will result in errors. At the Viby wastewater treatment plant in Aarhus, a number of different configurations have been set up to test the prototype software for various real time data validation methods. In the process of finding the validation method, the data history has been studied to find the weaknesses in the behaviour of the sensor in question. Data can be copied to a spread sheet to experiment with the choice of set-up constants. Setting up a validation of a measurement is fast, and the goal is to fine-tune the validation so sharply that errors are only reported when there is a reason for it (optimising on false positives). Figure 5.5: Validation during 1 days of SS/FTU sensor measurement (TagName: XM-TU-FI74-PV) located at the outlet of the secondary clarifiers at Viby wastewater treatment plant in Aarhus. PREPARED May 213

22 Measurements are set up to show the resulting confidence graphically together with the measured value. Set-ups with a confidence alarm activating a control related action is possible. On the SCADA system a master alarm is activated in connection with confidence alarms. Figure 5.5 shows the monitoring of the effluent from the 1 secondary clarifiers using a SS/FTU sensor. After 1 to 3 weeks of operation some non-visible deposits will appear on the optical window of the sensor, resulting in a slowly upward drift of the signal. This is an example of basic line operation, and the validation method Long Term Drift/Expected mean is applied. Instead of weekly routine cleaning, the validation method monitors the condition of the sensor, and when a critical quality is reached, the sensor is cleaned. The effect of cleaning the optical window is clearly seen in the figure. Figure 5.6: Validation during 12 hours of an Ammonium sensor in an aeration tank at Viby wastewater treatment plant in Aarhus. In figure 5.6 two validation methods are selected for an Ammonium measurement in an aeration tank. The rate of change method is selected to notify a momentary change in the signal, whereas the running variance is selected to find where the measuring signal is a straight line. A constant value will appear when the sensor is undergoing calibration. If the sensor release activating the sensor to measure after calibration is forgotten, the validation will detect the situation. Further, PREPARED May 213

23 sensor signals may very well freeze at a value during operation. It also results in a straight line which is detected by the validation. Figure 5.7: Validation during 12 hours of a Nitrate sensor in an aeration tank at Viby wastewater treatment plant in Aarhus. The last example from the test of the prototype software shows a range check on a nitrate sensor in an aeration tank. Figure 5.7 shows a nitrate value increasing due to a mal-functioning process in the end reaching a level which calls for an alarm. PREPARED May 213