Condition assessment of nonlinear processes
First Claim
1. A method for monitoring a nonlinear process, the method comprising:
- acquiring a set of channel data corresponding to at least one sensor for monitoring the process;
selecting cutsets of data from the set of channel data to form a basecase;
filtering artifacts from the selected cutsets of data to produce cutsets of artifact-filtered data for the basecase;
computing from the artifact-filtered data for each cutset in the basecase a set of connected phase space (PS) data comprising at least a first PS state and a second PS state connected to the first PS state;
computing distribution functions from the connected PS data for at least five cutsets of artifact-filtered data from the basecase;
computing at least one measure of dissimilarity (V) between a) a distribution function for a first one of the cutsets of artifact-filtered data from the basecase and b) a distribution function for a second one of the cutsets of artifact-filtered data from the basecase;
repeating the computation of the measure of dissimilarity (V) between distribution functions for non-identical pairings (i,j) of cutsets in the basecase until at least ten values (Vij) are computed for the measure of dissimilarity (V);
computing an average (V) and a corresponding sample standard deviation (σ
v) from the at least ten values of dissimilarity (Vij) for the basecase;
computing a χ
2 statistic (Σ
(Vij−
V)2/σ
2) for the at least one measure of dissimilarity (V) for the cutsets of artifact-filtered data from the basecase;
identifying any outlier cutsets of artifact-filtered data from the basecase;
removing the outlier cutsets from the basecase; and
recomputing the χ
2 statistic and testing for outlier cutsets until no outlier cutset is identified.
2 Assignments
0 Petitions
Accused Products
Abstract
There is presented a reliable technique for measuring condition change in nonlinear data such as brain waves. The nonlinear data is filtered and discretized into windowed data sets. The system dynamics within each data set is represented by a sequence of connected phase-space points, and for each data set a distribution function is derived. New metrics are introduced that evaluate the distance between distribution functions. The metrics are properly renormalized to provide robust and sensitive relative measures of condition change. As an example, these measures can be used on EEG data, to provide timely discrimination between normal, preseizure, seizure, and post-seizure states in epileptic patients. Apparatus utilizing hardware or software to perform the method and provide an indicative output is also disclosed.
-
Citations
12 Claims
-
1. A method for monitoring a nonlinear process, the method comprising:
-
acquiring a set of channel data corresponding to at least one sensor for monitoring the process;
selecting cutsets of data from the set of channel data to form a basecase;
filtering artifacts from the selected cutsets of data to produce cutsets of artifact-filtered data for the basecase;
computing from the artifact-filtered data for each cutset in the basecase a set of connected phase space (PS) data comprising at least a first PS state and a second PS state connected to the first PS state;
computing distribution functions from the connected PS data for at least five cutsets of artifact-filtered data from the basecase;
computing at least one measure of dissimilarity (V) between a) a distribution function for a first one of the cutsets of artifact-filtered data from the basecase and b) a distribution function for a second one of the cutsets of artifact-filtered data from the basecase;
repeating the computation of the measure of dissimilarity (V) between distribution functions for non-identical pairings (i,j) of cutsets in the basecase until at least ten values (Vij) are computed for the measure of dissimilarity (V);
computing an average (V) and a corresponding sample standard deviation (σ
v) from the at least ten values of dissimilarity (Vij) for the basecase;
computing a χ
2 statistic (Σ
(Vij−
V)2/σ
2) for the at least one measure of dissimilarity (V) for the cutsets of artifact-filtered data from the basecase;
identifying any outlier cutsets of artifact-filtered data from the basecase;
removing the outlier cutsets from the basecase; and
recomputing the χ
2 statistic and testing for outlier cutsets until no outlier cutset is identified.- View Dependent Claims (2, 3, 4, 5, 6, 7, 8)
acquiring additional data from said channel corresponding to at least one sensor for monitoring the process;
selecting a cutset of data from said channel data;
filtering artifacts from said cutset of additional channel data to produce a cutset of artifact-filtered data;
wherein the cutset of artifact-filtered data from the additional channel data defines a testcase;
computing from the artifact-filtered data for said testcase cutset, a set of connected phase space (PS) data comprising at least a first PS state and a second PS state connected to the first PS state;
computing a distribution function from the connected PS data for said testcase cutset of artifact-filtered data;
computing a set of dissimilarity values for at least one measure of dissimilarity (V), by comparing a distribution function derived from connected PS data for the testcase with each of the distribution functions derived for all non-outlier cutsets in the basecase;
obtaining an average (Vi) of the testcase dissimilarity over said set of dissimilarity values;
renormalizing said average testcase dissimilarity compared with the basecase, based on the formula, U=|Vi−
V|/σ
v;
comparing said renormalized measure of dissimilarity to a predetermined threshold as one factor in anticipating a nonlinear event, and in response to anticipating a nonlinear event, providing a warning indication.
-
-
3. The method of claim 1, wherein the above acts are repeated for data acquired from a plurality of data channels corresponding to a plurality of sensors.
-
4. The method of claim 2, wherein the above acts are repeated for data acquired from a plurality of data channels corresponding to a plurality of sensors.
-
5. The method of claim 4, wherein at least two measures of dissimilarity (Lc, χ
- c2) are computed between a) a distribution function derived from connected PS data for the testcase and b) each of the distribution functions computed for each of the non-outlier cutsets in the basecase.
-
6. The method of claim 1, further comprising acquiring additional basecase cutsets when more than four outlier cutsets are removed from the basecase.
-
7. The method of claim 2, further comprising:
-
computing from the artifact-filtered data for each cutset in the basecase a set of unconnected phase space (PS) data comprising a d-dimensional vector, where “
d”
is a number of phase space dimensions selected to represent the process;
computing distribution functions from the unconnected PS data for at least five cutsets of artifact-filtered data from the basecase;
computing at least one measure of dissimilarity (V) by comparing the distribution function for the unconnected PS data for a first one of the cutsets of artifact-filtered data from the basecase with the distribution function for the unconnected PS data for a second one of the cutsets of artifact-filtered data from the basecase;
repeating the computation of the measure of dissimilarity (V) for the unconnected PS data between distribution functions for non-identical pairings (i,j) of cutsets in the basecase until at least ten values (Vij) are computed for the measure of dissimilarity (V);
computing an average (V) and a corresponding sample standard deviation (σ
v) from the at least ten values of dissimilarity (Vij) for the basecase for the unconnected PS data;
computing a χ
2 statistic (Σ
(Vij−
V)2/σ
2) for the measure of dissimilarity (V) for the unconnected PS data for the cutsets of artifact-filtered data from the basecase;
identifying any outlier cutsets of artifact-filtered data from the basecase;
removing the outlier cutsets from the basecase;
recomputing the χ
2 statistic for the unconnected PS data and testing for outliers until no outlier cutset is identified;
computing from the artifact-filtered data for said testcase cutset, a set of unconnected PS data;
computing a distribution function from the unconnected PS data for said testcase cutset of artifact-filtered data;
computing a set of dissimilarity values for at least one measure of dissimilarity (V) for the unconnected PS data, by comparing a distribution function derived from unconnected PS data for the testcase with each of the distribution functions derived from the unconnected PS data for all non-outlier pairs of cutsets in the basecase;
obtaining an average (Vi) of the testcase dissimilarity over said set of dissimilarity values for the unconnected PS data;
renormalizing said average testcase dissimilarity compared with the basecase, based on the formula, U=|Vi−
V|/σ
v;
comparing said renormalized measure of dissimilarity to a predetermined threshold as one factor in anticipating a nonlinear event, and in response to anticipating a nonlinear event, providing a warning indication.
-
-
8. The method of claim 7, wherein
two measures of dissimilarity (Lc, χ -
c2) are derived from distribution functions derived from the connected PS data and wherein the two measures of dissimilarity (L, χ
2)are derived from the distribution functions derived from unconnected PS data;
renormalizing values for four measures of dissimilarity (Lc, χ
c2, L, χ
2)signaling a condition change in response to the four renormalized dissimilarity measures exceeding a predetermined threshold for a predetermined number of times; and
repeating the aforementioned acts for a plurality of data channels.
-
c2) are derived from distribution functions derived from the connected PS data and wherein the two measures of dissimilarity (L, χ
-
9. A method for monitoring a nonlinear process, the method comprising:
-
acquiring a set of channel data corresponding to at least one sensor for monitoring the process;
selecting cutsets of data from the set of channel data to form a basecase;
filtering artifacts from the selected cutsets of data to produce cutsets of artifact-filtered data for the basecase;
computing from the artifact-filtered data for each cutset in the basecase a set of unconnected phase space (PS) data comprising a d-dimensional vector, where “
d”
is a number of dimensions selected to represent the process;
computing distribution functions from the unconnected PS data for at least five cutsets of artifact-filtered data from the basecase;
computing at least one measure of dissimilarity (V) between a) a distribution function for a first one of the cutsets of artifact-filtered data from the basecase and b) a distribution function for a second one of the cutsets of artifact-filtered data from the basecase;
repeating the computation of the measure of dissimilarity (V) between distribution functions for non-identical pairings (i,j) of cutsets in the basecase until at least ten values (Vij) are computed for the measure of dissimilarity (V);
computing an average (V) and a corresponding sample standard deviation (σ
v) from the at least ten values of dissimilarity (Vij) for the basecase;
computing a χ
2 statistic (Σ
(Vij−
V)2/σ
2) for the measure of dissimilarity (V) for the cutsets of artifact-filtered data from the basecase;
identifying any outlier cutsets of artifact-filtered data from the basecase;
removing the outlier cutsets from the basecase;
recomputing the χ
2 statistic and testing for outlier cutsets until no outlier cutset is identified;
acquiring additional data from said channel corresponding to at least one sensor for monitoring the process;
selecting a cutset of data from said channel data;
filtering artifacts from said cutset of additional channel data to produce a cutset of artifact-filtered data;
wherein the cutset of artifact-filtered data from the additional channel data defines a testcase;
computing from the artifact-filtered data for said testcase cutset, a set of unconnected phase space (PS) data comprising a d-dimensional vector, where “
d”
is a selected number of phase space dimensions selected to represent the process;
computing a distribution function from the unconnected PS data for said testcase cutset of artifact-filtered data;
computing a set of dissimilarity values for at least one measure of dissimilarity (V), by comparing a distribution function derived from unconnected PS data for the testcase with each of the distribution functions derived for all non-outlier pairs of cutsets in the basecase;
obtaining an average (Vi) of the testcase dissimilarity over said set of dissimilarity values;
renormalizing said average testcase dissimilarity compared with the basecase, based on the formula, U=|Vi−
V|/σ
v;
comparing said renormalized measure of dissimilarity to a predetermined threshold as one factor in anticipating a nonlinear event, and in response to anticipating a nonlinear event, providing a warning indication. - View Dependent Claims (10, 11)
computing from each cutset of artifact-filtered data, a set of connected phase space (PS) data derived from a first PS state, and from a second PS state connected to the first PS state;
computing distribution functions from the connected PS data for at least five cutsets of artifact-filtered data from the basecase;
computing at least one measure of dissimilarity (V) by comparing the distribution function for the connected PS data for a first one of the cutsets of artifact-filtered data from the basecase with the distribution function for the connected PS data for a second one of the cutsets of artifact-filtered data from the basecase;
repeating the computation of the measure of dissimilarity (V) for the connected PS data between distribution functions for non-identical pairings (i,j) of cutsets in the basecase until at least ten values (Vij) are computed for the measure of dissimilarity (V);
computing an average (V) and a corresponding sample standard deviation (σ
v) from the at least ten values of dissimilarity (Vij) for the basecase for the connected PS data;
computing a χ
2 statistic (Σ
(Vij−
V)2/σ
2) for the measure of dissimilarity (V) for the connected PS data for the cutsets of artifact-filtered data from the basecase;
identifying any outlier cutsets of artifact-filtered data from the basecase;
removing the outlier cutsets in the basecase;
recomputing the χ
2 statistic for the connected PS data and testing for outliers until no outlier cutset is identified;
computing from the artifact-filtered data for said testcase cutset, a set of connected PS data;
computing a distribution function from the connected PS data for said testcase cutset of artifact-filtered data;
computing a set of dissimilarity values for at least one measure of dissimilarity (V) for the connected PS data, by comparing a distribution function derived from connected PS data for the testcase with each of the distribution functions derived connected PS data for all non-outlier pairs of cutsets in the basecase;
obtaining an average (Vi) of the testcase dissimilarity over said set of dissimilarity values for the connected PS data;
renormalizing said average testcase dissimilarity compared with the basecase, based on the formula, U=|Vi−
V|/σ
v;
comparing said renormalized measure of dissimilarity to a predetermined threshold as one factor in anticipating a nonlinear event, and in response to anticipating a nonlinear event, providing a warning indication.
-
-
12. A method for monitoring a nonlinear process via renormalized values, the method comprising:
-
computing an average value of dissimilarity (V) among unique pairings of non-identical, non-outlier cutsets of basecase data;
computing a sample standard deviation (σ
v), corresponding to said average dissimilarity;
computing an average value of dissimilarity for a plurality of testcase cutsets and the non-outlier basecase cutsets;
computing a renormalized value, U=|Vi−
V|/σ
v, representing a number of standard deviations by which the average value of dissimilarity for the testcase varies from the average value of dissimilarity among non-outlier basecase cutsets;
repeating the previous four acts for all of the testcase cutsets for a plurality of dissimilarity measures;
signaling a condition change in response to all of the plurality of renormalized dissimilarity measures exceeding a predetermined threshold for a predetermined number of times; and
repeating the aforementioned acts for a plurality of data channels.
-
Specification