Methods for the cyclical pattern determination of time-series data using a clustering approach

1. A method that detects and records patterns in a data set stored in one or more physical data-storage devices within a computer system, the computer system controlled to carry out the method by execution of computer instructions stored in a memory of the computer system, the method comprising:

clustering a data set comprising multiple data-point/time-point pairs into one or more clusters, each cluster associated with a cluster index and each data point contained in a single cluster;

generating a discretized data set by replacing each data point of each data-point/time-point pair within the data set with a cluster index for the cluster that contains the data point;

for each large time interval of an ordered set of contiguous large time intervals, dividing the large time interval into small time intervals and storing, for each small time interval, an indication, for each cluster, of whether or not the discretized data set includes an index for the cluster associated with a time point in the small time interval;

clustering large time intervals into large-time-interval clusters, each large-time-interval cluster associated with a large-time-interval-cluster index;

constructing an ordered set of discretized large time intervals by replacing each large time interval in the ordered set of large time intervals with a large-time-interval-cluster index for the large-time-interval cluster that contains the large time interval;

identifying at least one periodic occurrence of large-time-interval-cluster indexes in the ordered set of discretized large time intervals; and

storing an indication of the identified periodic occurrence in a physical data-storage device.

2. The method of claim 1 where the discretized data set comprises multiple cluster-index/time-point pairs.

3. The method of claim 1 where the ordered set of contiguous large time intervals is obtained by dividing a time range, the time range equal to or including a time period that begins with the earliest time point and that ends with the latest time point, into a set of equal-length contiguous large time intervals.

4. The method of claim 1

wherein the indication for a particular cluster and for a particular small time interval has one of two values selected from among:

a first value indicating that the discretized data set includes the cluster index for the cluster in association with the time point in the small time interval,

a second value indicating that the discretized data set does not include the index for the cluster in association with the time point in the small time interval; and

wherein the second value is different from the first value.

5. The method of claim 1 , wherein the large time interval is a day and the small time interval is an hour.

6. The method of claim 1 , wherein clustering large time intervals into large-time-interval clusters further comprises assigning, to each large-time-interval cluster, a set of pairwise matching large-time-intervals.

7. The method of claim 6 , wherein a first large time interval matches a second large time interval when the small-time-interval indications within the first large time interval are identical to the small-time-interval indications within the second large time interval or when each of those small-time-interval indications within the first large time interval that are not identical to corresponding small-time-interval indications within the second large time interval have greater than a threshold similarity to one another.

8. The method of claim 1 wherein identifying at least one periodic occurrence of large-time-interval-cluster indexes in the ordered set of discretized large time intervals further comprises:

identifying cycles in the ordered set of discretized large time intervals, each cycle associated with a period value, a starting index, and a large-time-interval-cluster index;

determining sets of related cycles from the identified cycles; and

for each set of related cycles

calculating a period weight for each cycle, and

identifying a representative cycle with a highest period weight.

9. The method of claim 8 wherein identifying cycles in the ordered set of discretized large time intervals further comprises:

identifying an ordered set of a first number of large time intervals at a particular period within the ordered set of discretized large time intervals;

identifying a second number of large time intervals within the identified first number of large time intervals that have a common large-time-interval-cluster index;

computing a ratio of the second number to the first number; and

determining whether or not the computed ratio is greater than a threshold.

10. The method of claim 8 wherein one cycle is related to another cycle when

the period of one cycle is an integer multiple of the period of another cycle;

the large-time-interval-cluster indexes of the two cycles are the same; and

the starting indexes of the two cycles starting at different starting indexes are separated by one of the periods.