Dynamically non-gaussian anomaly identification method for structural monitoring data

Abstract

The present invention belongs to the technical field of health monitoring for civil structures, and a dynamically non-Gaussian anomaly identification method is proposed for structural monitoring data. First, define past and current observation vectors for the monitoring data and pre-whiten them; second, establish a statistical correlation model for the whitened past and current observation vectors to obtain dynamically whitened data; then, divide the dynamically whitened data into two parts, i.e., the system-related and system-unrelated parts, which are further modelled by the independent component analysis; finally, define two statistics and determine their corresponding control limits, respectively, it can be decided that there is anomaly in the monitoring data when each of the statistics exceeds its corresponding control limit. The non-Gaussian and dynamic characteristics of structural monitoring data are simultaneously taken into account, based on that the defined statistics can effectively identify anomalies in the data.

Claims

1. A dynamically non-Gaussian anomaly identification method for structural monitoring data by using a finite element model built for two-span highway bridge model to simulate structural responses, wherein responses at finite element nodes are acquired as monitoring data; there are two datasets generated including a training dataset and a testing dataset; the training dataset consisting of normal monitoring data, and part of the testing dataset being used to simulate abnormal monitoring data; the method comprising: (1) defining x(t)∈.sup.m that represents a sample at time t in the normal structural monitoring data, where m is the number of measurement variables; define a past observation vector x.sup.p(t)=[x.sup.T(t−1), x.sup.T(t−2), . . . , x.sup.T(t−τ)].sup.T (note: τ is time-lag) and a current observation vector x.sup.c(t)=x(t); (2) defining J.sup.p and J.sup.c that represents whitening matrices corresponding to x.sup.p(t) and x.sup.c(t), respectively, the whitened x.sup.p(t) and x.sup.c(t) is obtained by {tilde over (x)}.sup.p(t)=J.sup.px.sup.p(t) and {tilde over (x)}.sup.c(t)=J.sup.cx.sup.c(t), respectively; (3) dynamically modeling of structural monitoring data to establish a statistical correlation model between {tilde over (x)}.sup.p(t) and {tilde over (x)}.sup.c(t):
{tilde over (S)}.sub.pc=E{{tilde over (x)}.sup.p{tilde over (x)}.sup.cT}=PΣQ.sup.T where {tilde over (S)}.sub.pc represents a cross-covariance matrix of {tilde over (x)}.sup.p and {tilde over (x)}.sup.c; P∈.sup.mτ×mτ and Q∈.sup.m×m represent matrices consisting all left and right singular vectors of singular value decomposition, respectively; Σ∈.sup.mτ×mτ represents a singular value matrix, which contains m non-zero singular values; (4) defining a projection of {tilde over (x)}.sup.p(t) on P, termed as z(t), which is calculated by the following equation:
z(t)=P.sup.T{tilde over (x)}.sup.p(t)=P.sup.TJ.sup.px.sup.p(t)=Rx.sup.p(t) where R=P.sup.TJ.sup.p; (5) since a covariance matrix of z(t) is an identity matrix:
S.sub.zz=E{zz.sup.T}=P.sup.TE{{tilde over (x)}.sup.p{tilde over (x)}.sup.pT}P=I and the above modeling process takes into account dynamic characteristics of structural monitoring data, R is termed as dynamically whitening matrix and z(t) is termed as dynamically whitened data; (6) dividing the dynamically whitened data z(t) into two parts using the following equations:
z.sub.s(t)=R.sub.sx.sup.p(t)
z.sub.n(t)=R.sub.nx.sup.p(t) where z.sub.s(t) and z.sub.n(t) represent the system-related and system-unrelated parts of z(t), respectively; R.sub.s and R.sub.n consist of the first m rows and last m(i−1) rows of R, respectively; (7) establishing dynamically non-Gaussian models for z.sub.s(t) and z.sub.n(t) using independent component analysis:
s.sub.s=B.sub.s.sup.Tz.sub.s(t)
s.sub.n(t)=B.sub.n.sup.Tz.sub.n(t) where s.sub.s(t) and s.sub.n(t) represent system-related and system-unrelated independent components, respectively; B.sub.s and B.sub.n are solved by a fast independent component analysis algorithm; (8) let W.sub.s=B.sub.s.sup.TR.sub.s and W.sub.n=B.sub.n.sup.TR.sub.n, there exist the following equations:
s.sub.s(t)=W.sub.sx.sup.p(t)
s.sub.n(t)=W.sub.nx.sup.p(t) where W.sub.s and W.sub.n represent de-mixing matrices corresponding to the system-related and system-unrelated parts, respectively; (9) defining two statistics corresponding to s.sub.s(t) and s.sub.n(t), respectively:
I.sub.s.sup.2=s.sub.s.sup.Ts.sub.s=x.sup.pT(W.sub.s.sup.TW.sub.s)x.sup.p
I.sub.n.sup.2=s.sub.n.sup.Ts.sub.n=x.sup.pT(W.sub.n.sup.TW.sub.n)x.sup.p (10) after calculating the statistics I.sub.s.sup.2 and I.sub.n.sup.2 for all normal structural monitoring data, estimating a probability density distribution of I.sub.s.sup.2 and I.sub.n.sup.2, respectively; determining control limits I.sub.s,lim.sup.2 and I.sub.n,lim.sup.2 of the two statistics through a 99% confidence criterion; and based on the determination of the control limits I.sub.s,lim.sup.2 and I.sub.n,lim.sup.2, determining whether there exist structural anomalies in a two-span highway bridge based on newly acquired monitoring data, when each of the statistics exceeds its corresponding control limit; and (11) simulating abnormal monitoring data in the testing dataset based on the anomalies in the monitoring data.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

(1) FIG. 1 is a schematic diagram for calculating anomaly identification statistics.

DETAILED DESCRIPTION

(2) The following details is used to further describe the specific implementation process of the present invention.

(3) Take a two-span highway bridge model, with a length of 5.4864 m and a width of 1.8288 m, as an example. A finite element model is built to simulate structural responses, and the responses at 12 finite element nodes are acquired as monitoring data. There are two datasets generated: the training dataset and the testing dataset; the training dataset consists of normal monitoring data, and part of the testing dataset is used to simulate abnormal monitoring data; both datasets last for 80 s and the sampling frequency is 256 Hz. The basic idea of the present invention is shown in FIG. 1.

(4) (1) Construct the past observation vector x.sup.p(t) and the current observation vector x.sup.c(t) for each data point in the training dataset; then pre-whiten all past and current observation vectors (i.e., x.sup.p(t) and x.sup.c(t)) to obtain the whitening matrices (i.e., J.sup.p and J.sup.c) and the whitened past and current observation vectors (i.e., {tilde over (x)}.sup.p(t) and {tilde over (x)}.sup.c(t)).

(5) (2) Establish a statistical correlation model for {tilde over (x)}.sup.p(t) and {tilde over (x)}.sup.c(t) to obtain the dynamically whitening matrix R; the first 12 rows of R are used to construct R.sub.s and the others are used to construct R.sub.n; calculate z.sub.s(t)=R.sub.sx.sup.p(t) and z.sub.n(t)=R.sub.nx.sup.p(t).

(6) (3) Establish independent component analysis models for z.sub.s(t) and z.sub.n(t) to obtain matrices B.sub.s and B.sub.n; correspondingly, the de-mixing matrices can be obtained through W.sub.s=B.sub.s.sup.TR.sub.s and W.sub.n=B.sub.n.sup.TR.sub.n; calculate statistics I.sub.s.sup.2 and I.sub.n.sup.2, then determine their corresponding control limits I.sub.s,lim.sup.2, and I.sub.n,lim.sup.2; it can be decided that there exist anomalies in the data when each of the statistics exceeds its corresponding control limit.

(7) (4) Simulate abnormal monitoring data in the testing dataset, that is, the monitoring data of sensor 2 gains anomaly during time 40˜80 s; identify anomalies in the monitoring data using the two proposed statistics I.sub.s.sup.2 and I.sub.n.sup.2, results show that both I.sub.s.sup.2 and I.sub.n.sup.2 can successfully identify anomalies in the monitoring data.