The PSSWR has three levels of compression: (i) the novelty detection, (ii) wavelet decomposition and (iii) bit coding. These three levels allow the optimization of memory  space.  It ensures that long periods of measurement can be saved in a memory card. Section 2 of this paper presents a description of the adopted methodology, showing the online compression scheme and the off- line reconstruction system. In Section 3, the real time implementa- tion of the PSSWR is shown, covering the role played by FPGA and ARM devices on the system. In Section 4, a prototype developed is shown. Some results and comments about the tests realized with the prototype of the PSSWR are presented in Section 5. Finally,

some general conclusions are  stated in Section   6.

2. Proposed methodology

The proposed methodology is based on the fact that only the novelties present in the signal must  be  saved  [10].  In  this way, the signal is pided in frames that contain four cycles of the fun- damental component and the main objective is to identify which ones are novelty frames (frames that present some difference com- pared to a reference one) to be saved. Furthermore, two other stages of compression are performed: the novelty frame is com- pressed by a wavelet transform followed by a modified Lempel– Ziv coding algorithm that does a lossless   compression.

2.1. Compression system – online

The compression system is composed by five main blocks as shown in Fig. 1: (a) the novelty Detector; (b) the Frequency Estima- tion; (c) the Wavelet Compression; (d) the Builder and (e) the Coding.

As mentioned above, knowing which frame is different from the reference one is the core of this system. The parameter chosen to make this decision is the energy of the signal at each frame. Before the energy calculation, the signal passes through a high-pass filter with cutoff frequency of 720 Hz in order to suppress detections due to  fundamental  frequency  variations.  Then,  the  energy  of   each

frame is calculated in the output of this filter and subtracted from the energy of the reference frame. If this value is higher than a threshold, a novelty is  detected.

The frequency estimation is necessary in the reconstruction algorithm since only the novelty frames are stored and the entire signal may be desired to be reconstructed. The non-novelty frames are then reconstructed based on the shape of the reference frame (the last frame of novelty) and on the averaged frequencies. The frequency estimation algorithm is the PLL (Phased Locked Loop) presented in [11,12]. The frequency is estimated instantaneously and averaged each cycle. Thus, four frequency  values are  stored for each frame. It is worth to mention that slow frequency variation does not trigger the novelty detector, however rapid or abrupt changes in the frequency estimation   do.

When a novelty is detected, the corresponding frame of the sig- nal must be stored. However, instead of storing the points them- selves, a wavelet compression is performed in order to eliminates the redundant signal information. The filters of the wavelet were obtained from a Daubechies 3 (db3) mother wavelet and a decom- position in three levels was adopted, as shown in Fig.    2.

In this figure v½n] is the signal to be decomposed, Hd  and Ld  are

the highpass and lowpass filter respectively, CD1; CD2 and CD3 are the detail coefficients and CA3 is  the  approximation  coefficient. The decomposition is performed in parallel with the other pro- cesses and a threshold is applied to the detail coefficients aiming at eliminating the low energy information. If they are lower than this threshold they are replaced with zero. The approximation coefficients are all remained since the processed signals have more information in the lower   frequencies.