Detect Anomaly Intervals with SQL

Question

My problem is simple: I have a table with a series of statuses and timestamps (for the sake of curiosity, these statuses indicate alarm levels) and I would like to query this ta

Answer 1

This has to be one of the harder questions I've seen today - thanks! I assume you can use CTEs? If so, try something like this:

;WITH Filtered
AS
(
    SELECT ROW_NUMBER() OVER (ORDER BY dateField) RN, dateField, Status
    FROM Test    
)
SELECT F1.RN, F3.MinRN,
    F1.dateField StartDate,
    F2.dateField Enddate
FROM Filtered      F1, Filtered F2, (
SELECT F1a.RN, MIN(F3a.RN) as MinRN
FROM Filtered      F1a
   JOIN Filtered F2a ON F1a.RN = F2a.RN+1 AND F1a.Status = 2 AND F2a.Status <> 2
   JOIN Filtered F3a ON F1a.RN < F3a.RN AND F3a.Status <> 2
GROUP BY F1a.RN ) F3 
WHERE F1.RN = F3.RN AND F2.RN = F3.MinRN

And the Fiddle. I didn't add the intervals, but I imagine you can handle that part from here.

Good luck.

Answer 2

I do something similar by using id that is an identity to the table.

    create table test(id int primary key identity(1,1),timstamp datetime,val int)

    insert into test(timstamp,val) Values('1/1/2013 00:00:00',1)
    insert into test(timstamp,val) Values('1/1/2013 00:00:05',2)
    insert into test(timstamp,val) Values('1/1/2013 00:00:25',1)
    insert into test(timstamp,val) Values('1/1/2013 00:00:30',2)
    insert into test(timstamp,val) Values('1/1/2013 00:00:35',1)

    select t1.timstamp,t1.val,DATEDIFF(s,t1.timstamp,t2.timstamp) 
    from test t1 left join test t2 on t1.id=t2.id-1

    drop table test

I would also make the timestamps be seconds since 1980 or 2000 or whatever. But then you might not want to do the reverse conversion all the time and so it depends on how often you use the actual time stamp.

Answer 3

Just for the sake of having an alternative. Tried to do some test on performance, but did not finish.

SELECT
  MIN([main].[Start]) AS [Start],
  [main].[End],
  DATEDIFF(s, MIN([main].[Start]), [main].[End]) AS [Seconds]
FROM
(
  SELECT
    [sub].[Start],
    MIN([sub].[End]) AS [End]
  FROM
  (
    SELECT
      [start].[Timestamp] AS [Start],
      [start].[Status] AS [StartingStatus],
      [end].[Timestamp] AS [End],
      [end].[Status] AS [EndingStatus]
    FROM [Alerts] [start],  [Alerts] [end]
    WHERE [start].[Status] = 2 
      AND [start].[Timestamp] < [end].[Timestamp]
      AND [start].[Status] <> [end].[Status]
  ) AS [sub]
  GROUP BY
    [sub].[Start],
    [sub].[StartingStatus]
) AS [main]
GROUP BY
  [main].[End]

And here is a Fiddle.

Answer 4

Finally figured out a version I was happy with. It took me remembering an answer from another question (can't remember which one though) where it was pointed out that the difference between two (increasing) sequences was always a constant.

WITH Ordered (occurredAt, status, row, grp) 
             as (SELECT occurredAt, status, 
                        ROW_NUMBER() OVER (ORDER BY occurredat), 
                        ROW_NUMBER() OVER (PARTITION BY status 
                                           ORDER BY occurredat)
                 FROM Alert)

SELECT Event.startDate, Ending.occurredAt as endDate,
       DATEDIFF(second, Event.startDate, Ending.occurredAt) as interval

FROM (SELECT MIN(occurredAt) as startDate, MAX(row) as ending
      FROM Ordered
      WHERE status = 2
      GROUP BY row - grp) Event

LEFT JOIN (SELECT occurredAt, row
           FROM Ordered
           WHERE status != 2) Ending
        ON Event.ending + 1 = Ending.row

(working SQL Fiddle example, with some additional data rows for work checking).

This unfortunately doesn't correctly deal with level-2 statuses that are end rows (behavior unspecified), although it does list them.