A SIGNAL PROCESSING MODULE FOR NON-DESTRUCTIVE TARGETING OF TERMITE ACTIVITY USING THE SPECTRAL KURTOSIS AND THE DISCRETE WAVELET TRANSFORM J.J. González de-la-Rosa 1,2 , A. Moreno Muñoz 1,3 and C.G. Puntonet 4 1 Research Group PAI-TIC-168. Computational Instrumentation and Industrial Electronics 2 Electronics Area, University of Cádiz, EPSA, Av. Ramón Puyol S/N. E-11202, Algeciras-Cádiz, Spain Phone-Fax: +34 956 028020, Fax: +34 956 028001. E-mail: juanjose.delarosa@uca.es 3 Electronics Area, University of Córdoba, Campus Rabanales, E-14071, Córdoba, Spain 4 University of Granada. ESII. E-18071, Granada, Spain ABSTRACT In this paper we present the operation results of a portable computer-based measurement equipment conceived to perform non-destructive testing of suspicious termite in- festations. Its signal processing module is based in the Spec- tral Kurtosis (SK), with the de-noising complement of the Discrete Wavelet Transform (DWT). The SK pattern allows the targeting of alarms and activity signals. The DWT com- plements the SK, by keeping the successive approximations of the termite emissions, supposed more non-gaussian (less noisy) and with less entropy than the detail approximations. For a given mother wavelet, the maximum acceptable level, in the wavelet decomposition tree, which preserves the in- sects’ emissions features, depends on the comparative evolu- tion of the approximations details’ entropies, and the value of the global spectral kurtosis associated to the approximation of the separated signals. The paper explains the detection criterion by showing real-life recordings. 1. INTRODUCTION This paper deals with the performance of a final-version equipment for termite detection, whose previous prototype’s performance was described in [1]. The measurement method is mainly based in the interpretation of the spectral kurtosis graph, along with the wavelet analysis. We use the sound card, which simplifies the hardware and the criterion of de- tection. The instruments for plague detection are thought with the objective of decreasing subjectiveness of the field operator. At the same time, they should be conceived to perform an early targeting of the plague, in order to treat the infestation before serious economic damage occurs. On-site monitoring implies reproducing the natural phenomenon of insect emis- sions with high accuracy. As a consequence it is imperative the use of a deep storage device, and high sensitive probes. These features make the price paid very high, and still do not guarantee the success of the detection. Besides, the expert’s subjectiveness plays a crucial role. The methods in which the instruments are based are very much dependent on the detection of excess of power in the signals; these are the so-called second-order methods, e.g. the RMS calculation, which does not provide information regarding the time fluctuations of the amplitudes. Another handicap of the second-order principle, e.g. the classical power spectrum, attends to the preservation of the energy during data processing. Consequently, the eradication of ad- ditive noise lies in filter design and sub-band decomposition. As an alternative to improve noise rejection and complete characterization of the signals, in the past ten years, a myr- iad of higher-order methods are being applied, in scenarios which involve signal separation and characterization of non- Gaussian signals. The main handicap of applying higher- order statistics is the amount of data which they generate, and that have to be stored in the measurement unit. An examina- tion of the multi-dimensional data structures (tensors) reveals redundant (symmetrical) information; so relevant directions have to be selected within the tensor data structure. Secondly, the interpretation of higher-order cumulants and poly-spectra are reduced to a set of catalogued noise processes, and only a few attempts have been made in order to characterize the processes via HOS. This paper describes a method based in the spectral kur- tosis (a modification of the method described in [2] and [1]) to detect infestations of subterranean termites in a real-life scenario (Southern Spain). Wavelet decomposition is used as an extra tool to aid detection from the preservation of the approximation of the signal, which is thought to be more Gaussian than the details. The measurement site was selected by our partner plague eradication company, to be a suspicious location. Speech and the typical urban background sounds clearly bury the ter- mites’ emissions, which came from the soil, under all of us. Sounds were not audible, except from alarms signals, only produced when termites are clearly disturbed. The interpretation of the results is focussed on the peakedness of the statistical probability distribution associ- ated to each frequency component of the signal, to measure the distance from the Gaussian distribution. The spectral kur- tosis serves as a twofold tool. First, it enhances non-Gaussian signals over the background. Secondly, it offers a more com- plete characterization of the transients emitted by the insects. The paper is structured as follows: in Section 2 a re- view on termite detection and relevant HOS experiences sets the foundations. In Section 3 we make a brief report on the definition of kurtosis; we use an unbiased estimator of the spectral kurtosis, successfully used in [1], using a higher measurement bandwidth. Results are presented in Section 5. Wavelets are summarized in Section 4. Finally, conclusions are drawn in Section 6.