CSP共空间模式的介绍.docx

Common Spatial Pattern(s) algorithm算法. The CSP paradigm is based on the design of the Berlin Brain-Computer Interface (BBCI) [1], more comprehensively described in [2], which is mainly controlled by (sensori-)motor imagery. The features exploited by this paradigm in its original form are Event-Related Synchronization and Desynchronization [3] localized in the (sensori-)motor cortex, but the paradigm is not restricted to these applications. CSP was originally introduced in [5] and first applied to EEG in [6]. Due to its simplicity, speed and relative robustness, CSP is the bread-and-butter实用的 paradigm for oscillatory振荡 processes, and if nothing else, can be used to get a quick estimate of whether the data contains information of interest or not. Like para_bandpower, CSP uses log-variance features over a single non-adapted frequency range (which may have multiple peaks), and neither temporal structure时间结构 (variations) in the signal is captured捕捉, nor are interactions相互作用 between frequency bands. The major strength of the paradigm 范式is its adaptive spatial filter自适应空间滤波器, which is computed计算 using the CSP algorithm. The paradigm is implemented实施 as a standard sequence of signal (pre-)processing (spatial/spectral光谱 filtering), feature extraction, and machine learning. The first preprocessing预处理 step is frequency filtering, followed by an adaptively learned spatial filter (which is the defining propery定义的性能 of the paradigm), followed by log-variance feature extraction and finally a (usually simple) machine learning step applied to the log-variance features. The spatial filtering projects the channels of the original signal down to a small set of (usually 4-6) surrogate代理 channels, where the (linear) mapping is optimized线性映射被优化 such that the variance in these channels is maximally informative w.r.t. to the prediction预测 task. The CSP filters can be obtained from the per-class signal covariance matrices协方差矩阵 by solving a generalized eigenvalue problem广义特征值问题(of the form [V,