This is a MATLAB code for paper "Real-time monitoring of high-dimensional functional data streams via spatio-temporal smooth sparse decomposition" with paper link

The Python version of the code is given in Python.

If you like to use the code, please cite my paper with citation information in bibtex format or endnote format.

Related work: Anomaly Detection in Images with Smooth Background Via Smooth-Sparse Decomposition with paper link and code.

Here are two examples

Data:

1. data.mat, solar flare dataset

File description:

1. bsplineBasis: Function to generate bsplineBasis with k Bspline Basis and n gridded with spline degree sd and boundary degree bd

   • By default $bd = sd - 1$
   • Normally, we use cubic B-spline, where $sd = 3$

2. chartIC: Compute the IC chart based to normalize the monitoring statistics, remember the mean in mt2 and standard deviation sd. Finally, store the normalized statistics with the largest detection power in Ttr and the index in Itr.

3. chartoc: Based on the mt2 and sd computed in chartIC , computed the normalized test statistics with the largest detection power Ttr and the index Itr

4. ewmamonit: the major function proposed in the original paper:

   • Input variables

     • Y: Tensor to be monitored, currently only support 3-mode tensor. By default, the first 2 dimensions are spatio-dimension and the last is the temporal dimension. For other dimension of Y, need to further add a case length(Bs) == 1 to make it work.
     • B: Basis for the background, should be a cell with length 2. B{1} for the 1st spatial dimension of the tensor. B{2} for the 2nd spatial dimension, etc. Currently not supporting the temporal basis.
     • Bs: Basis for the anomaly, should be a cell with length 2. Bs{1} for the 1st spatial dimension of the tensor. Bs{2} for the 2nd spatial dimension, Currently not supporting the temporal basis.
     • lambda: Should be an array of length 3, [lambdax, lambday, lambdat], which enforce Spatial and Temporal smoothness of the background in x, y, t dimension. A larger lambda enforce larger smoothness of the background, this is the para
     • allgamma : Should be an array of potential tuning parameters to be used for sparsity constraint for anomaly.
     • Other potential variables stored in varargin
       • isewma: default to 0, whether used previous yhat
       • maxiter: default to 1, but can further increase
       • type: default to s, type of thresholding in solving sparse penalty
       • L: default to 0, if set to 0, we do not setup control chart for change detection. If set a number, the monitoring system will break out when capture out-of-control sample.
       • issave: whether save the intermediate results of anomaly detection.

   • Output variables

     • T2: Full testing statistics

     • Snow: Estimated anomaly

     • Yhat: Background estimation

     • t: if L is set to nonzero, t will be returned as the first changed location

     • defect: estimated defect when the change is first detected.

     • Itr: which tuning parameter achieves the best detection power, saved the index

     • Tte: testing statistics with largest power