In the system as it is designed, the signals from all of the instruments are processed from five separate parameter measures into a single value in an event monitor computer system that contains the algorithm. The signal then goes through the crucial proprietary baseline estimator. A deviation of the signal from the estimated baseline is then derived. Then a gain matrix is applied that weights the various parameters based on experimental data for a wide variety of probable threat agents. The magnitude of the deviation signal is then compared to a preset threshold level that is set by the operator. If the signal exceeds the threshold, the trigger is activated. Figure 3 shows the same data from Figure 2 processed through the algorithm.
You Might Also Like
Explore This IssueFebruary/March 2007
Also By This Author
Even with extremely noisy data, the system does not trigger at a threshold level set at one (1). Therefore, during normal operation, with no threats present, the process deviation should not be large enough to produce a trigger signal less than one (1) in this case.
However when the data for a cyanide incursion at 1 percent of the LD-50 or approximately 2.8 mg/L is superimposed on the system, the trigger level of 1 is easily exceeded (see Figure 4). Other contaminants exhibit similar results.
The deviation vector that is derived from the trigger algorithm contains significantly more data than what is needed to simply trigger the system. The deviation vector’s magnitude relates to concentration and trigger signal, while the deviation vector direction relates to the agent characteristics. Seeing that this is the case, laboratory agent data can be used to build a threat agent library of deviation vectors. A deviation vector from the water monitor can be compared to agent vectors in the threat agent library to see if there is a match within a tolerance. This system can be used to classify what agent is present (see Figure 5). Each vector results in a vector angle in n-space that, from the research conducted so far, appears to be unique to the class of agent present. The fact that the direction of the vector is unique for a given agent type allows the use of an algorithm to classify the cause of a trigger being set off. When the event trigger is set off, the library search begins. The agent library is given priority and is searched first. If a match is made, the agent is identified. If no match is found, the plant library is then searched and the event is identified if it matches one of the vectors in the plant library. If no match is found, the data is saved and the operator can enter an ID when one is determined. The agent library is provided with the system, and the plant library is learned on site.
The unknown alarm rate when the system is tracking real world data is also quite low. The system is equipped with a learning algorithm, so that as unknown alarm events occur over time, the system has the ability to store the signature that is generated during the event. The operator can then go into the program and identify that function and associate it with a known cause such as the turning on of a pump or the switching of water sources, etc. Then, the next time that event occurs; it will be recognized and identified appropriately.
Over time as the system learns, the probability of an unknown alarm that has not been previously encountered and identified will continue to decrease and will eventually approach zero. The probability of an unknown alarm due to a given event depends upon the frequency of the occurrence of such an event and the time that the algorithm has had to learn that event. Events that occur frequently will be quickly learned while rare or singular events will take longer to be learned and stored. This should result in a fairly rapid drop off in the number of unknown alarms as common events are quickly learned.