CSSM Procedure
Monitoring of Streaming Data
In many fields, such as industrial process control, monitoring of atmospheric pollutant levels, and so on, a sequential data stream is monitored continuously. The goal is to flag unexpected changes in the normal behavior of the process as soon as possible. Scoring with state space models can be quite useful in these situations. See Example 14.20: Monitoring Streaming Data Using Scoring for an illustration of how to use SSM-based scoring to detect level shift in a quarterly series of traffic accident data from the United Kingdom. The following list describes the usual steps in an SSM-based work flow in monitoring cases such as these:
Find a good model for the normal behavior of the process that is being monitored.
Create an initial score-store by using a suitably long series of data,
. At creation time, use the OUT(k)= form with a suitable data window length k, and set the desired structural break detection settings, such as the CHECKBREAK and MAXSHOCK options in the appropriate statements. After the initial score-store creation, the process monitoring starts as the new data packets
arrive and are scored in succession. Each time a new data packet arrives, it is scored, and an updated score store is created that becomes the input score store for processing the next data packet.
As seen in Example 14.20: Monitoring Streaming Data Using Scoring, a structural break that happens at a particular time point to be confirmed as a structural break often requires several subsequent observations to be scored. You must make an appropriate choice of k at the time of score-store creation to ensure that a sufficient number of past data rows are retained for detecting structural breaks.
-
If a structural break is confirmed at a time point, you can take the following types of action:
Flag the structural break, and continue the process monitoring without any further action.
Correct the model to take into account the structural break, and continue monitoring.
Pause the process, address the causes of the break, and resume monitoring (possibly after modifying the model and refitting).