Preprocessing of Input Data

Raw signals from chemical sensors are rarely suitable for direct multivariate analysis. Some form of signal conditioning is always necessary before the input matrix is composed. Examples of preprocessing techniques used in the static and in the dynamic mode of multicomponent analysis are summarized in Table 10.1. They can be used as such or in combination. In higher-order sensors, where different transduction modes are used, the homogeneity of the input matrix is important. Thus, the matrix must contain data that are comparable in dimensions and that are commensurate. 

Subtraction of reference Linear drift subtraction Normalization Averaging Linearization Differential measurements Baseline correction Relative signals Signal quality redundancy Transform functions, e.g., antilog 

Steady-state extrapolation Fourier analysis Gating and signal filtering Transient signals Speeding up of response Spatial and temporal information Enhancement of selectivity Increasing order of measurement 

---

As with the supervised learning categorization networks, there are a few items that need to be discussed for the unsupervised case. Scaling or preprocessing of input data is still important. The number of output PEs is usually arbitrary unless you have reason to believe your data should fall into a certain number of categories. The issue of whether a network can be trained at the near-100% level is irrelevant because you do not know what the correct answers are. It is possible to use an unsupervised learning network to classify data whose correct classifications are known. In this case you can talk about percent correct again, see Chapter 10 of Ref. 19. 

---