Gruppo Misure Elettriche ed Elettroniche Electrical and electronic measurements group Spectrum Analyzers Gruppo Misure Elettriche ed Elettroniche Introduction The Spectrum Analyzer (SA) is the fundamental tool for the analysis of signals in the frequency domain. This device, which can be analog or digital, may base its operation on different techniques; the most commonly employed devices are however only two: swept-tuned analyzers (Analogue spectrum analyzer) and Fast Fourier Transform analyzers (Digital spectrum Analyzer). On the market today you can find dedicated spectrum analyzers or allin-one instruments that integrate other features, proper of an oscilloscope or a data acquisition device. 2 Gruppo Misure Elettriche ed Elettroniche With a spectrum analyzer you can visualize the harmonic content of a signal, characterize circuits of various types, measure signal- to-noise ratios, verify the emission levels of EM disturbances of electronic devices and appliances etc. 3 Gruppo Misure Elettriche ed Elettroniche History of spectral analysis The origin of the spectral analysis can be traced back to Bunsen and Kirchhoff who, in 19th century, developed the idea that the spectrum of the luminous radiation emitted by a substance (for example when heated) depends on its chemical composition and its properties. Although at that time no one probably realized the great relevance of this "invention", some time later it gave rise to what is now commonly called spectral analysis, whose mathematical expression is undoubtedly the Fourier Transform (developed before the idea of Bunsen and Kirchhoff). 4 Gruppo Misure Elettriche ed Elettroniche A significant contribution to this subject was given, in 1958, by two brilliant researchers: Blackman and Tukey; they published the famous autocorrelation method which allowed to derive the power spectrum of a signal by an estimation of the autocorrelation function of the observed data. It is however the introduction, in 1965, of the Fast Fourier Transform (FFT) accredited to Cooley and Tukey which marks the advent of direct estimation methods of the spectrum of signals. 5 Gruppo Misure Elettriche ed Elettroniche The reason for the frequency analysis... The most immediate way to analyze signals, characterize devices and circuits, compare quantities and parameters of various types, is undoubtedly to use a visualization in Time Domain. The most commonly used instrument in these cases is the oscilloscope. The time domain, certainly the most immediate, is convenient in many circumstances, such as for measurements of rise or settling times etc. On the other hand, it is easy to verify that the visualization in time provides an “overview" of the signal, i.e. all the different harmonic components are added together and displayed simultaneously. 6 Gruppo Misure Elettriche ed Elettroniche The Fourier analysis, on the other hand, produces the decomposition of the entire waveform into spectral components, each with a certain amplitude, frequency, and phase value. This passage from the time domain to the frequency one does not imply any loss of information on the signal but only a different representation of it. Therefore it can be said that the frequency domain, while perhaps less congenial than time domain, is in many cases more useful for the study of signals and devices. 7 Gruppo Misure Elettriche ed Elettroniche So, for example, the oscilloscope display (time domain) of a signal with many harmonics will hardly ever be able to highlight the different components, even more so their respective amplitudes: A [V] 2.5 2 A [V] 1.5 0 -2 0.5 0 0.02 0.04 t [s] 0.06 0 100 200 300 400 f [Hz] Instead the use of a spectrum analyzer (frequency domain) will produce complete signal information: x ( t ) = 2 sin ( 2 50 t ) +1 sin ( 2 150 t ) +1 sin ( 2 250 t ) 8 Gruppo Misure Elettriche ed Elettroniche In the following example, a high frequency disturbance is shown: A [V] 35 20 25 A [V] 15 -20 5 0 0.02 0.04 t [s] 0.06 0 400 800 1200 f [Hz] by using frequency domain analysis, you can write: x ( t ) = 30 sin ( 2 50 t ) + 2 sin ( 2 1000 t ) 9 Gruppo Misure Elettriche ed Elettroniche Even in the case of a small sinusoidal signal drowned in noise: x ( t ) = 5 sin ( 2 50 t ) + n ( t ) A [V] 30 4.5 A [V] 3.5 Harmonic component at invisible in the time domain 10 50 Hz 2.5 -10 1.5 -30 0.5 0 0.04 0.08 0.12 t [s] 0.16 0.2 0 1000 2000 3000 4000 f [Hz] only the frequency analysis allows to shed light on the various components present. 10 Gruppo Misure Elettriche ed Elettroniche ... and some of its applications Therefore, applications of spectral analysis may cover multiple disciplinary sectors (not only measurements) in which frequency components matter; examples are telecommunications, digital signal processing, electrical drives, geophysics, medicine etc. In cellular radio systems it is very important to control the spectral content of transmitted signals to avoid interference with other systems operating on the same frequency bands. As another example, it is essential to carry out signal-to-noise ratio measurements in the characterization of devices or, for electronic manufacturers, to verify the emission levels at various frequencies (EM disturbances) in compliance with regulations that are more and more stringent. 11 Gruppo Misure Elettriche ed Elettroniche Analog and digital spectrum analyzers Basically a spectrum analyzer can obtain the frequency representation of a signal in one of two different ways: • Swept-tuned analog analyzers • Fast Fourier Transform Real-time digital analyzers The first method, typically implemented with the super-heterodyine technique, basically consists in translating in frequency (heterodyning) a signal in a band different from that of the original spectral components, where they will then be filtered by a suitable band-pass filter and demodulated. This technique is used for the realization of analog spectrum analyzers. 12 Gruppo Misure Elettriche ed Elettroniche The second method digitizes the signal in the time domain and produces its frequency representation using the discrete Fourier transform for a finite number of samples. The practical result of this operation is a simultaneous parallel filtering over the entire useful frequency band. A technique of this kind allows the realization of digital spectrum analyzers. 13 Gruppo Misure Elettriche ed Elettroniche The advantages of the digital analyzer over the former analog one are, in general, 1) a significant improvement in speed, so much so that the analyzers that employ it are also called "real-time", 2) the possibility of characterizing single-shot signals (a consequence of point 1) 3) the possibility to measure phase and amplitude of each single component (for this reason they are also called vector spectrum analyzers). Disadvantages of the digital technique are 1) a smaller bandwidth and 2) smaller sensitivity and dynamic range. Today, digital spectrum analyzers combine both heterodyning (frequency shifting) and FFT, hence they offers advantages of both methods. 14 Gruppo Misure Elettriche ed Elettroniche 1. Analog spectrum analyzers It was the first type of spectrum analyzer to appear on the market (early ‘60s) before the widespread diffusion of digital spectrum analyzers. However, it had some advantages that first digital spectrum analyzers did not had (very high bandwidth, wide dynamic range etc.), which made them irreplaceable for measurements on high frequency signals (order of tens of GHz) recurring in cellular and satellite radio applications. The most used analog analyzer is based on super-heterodyine which is a frequency conversion method that permits the selection and display of different signal components. More precisely, we'll see later that the displayed powers refer, in reality, to more or less narrow subbands, not to individual components. 15 Gruppo Misure Elettriche ed Elettroniche If you have well understood the duality of the representation of a signal in the time and frequency domains, you will also understand that the easiest and immediate way to perform the analysis of the spectral content of a signal would be building a sufficiently selective bandpass filter with a variable and electronically controllable central frequency. Before continuing, however, it is essential to give some details on signal filters. 16 Gruppo Misure Elettriche ed Elettroniche Analog and Digital filters A filter is an electronic device able to select (filter) the spectral components of a signal, strongly attenuating some and leaving almost unchanged (or amplifying) others. Filters can be either Analog (i.e. implemented by means of a suitable connection of circuit elements such as capacitors, inductors and resistors) or Digital, i.e. implemented by means of appropriate digital processing techniques operated on signals that have been digitized. The most common filters used in analogue spectrum analyzers are themselves analogue, although, in the most sophisticated hybrid models, there are also digital filtering stages that allow to increase performance considerably. 17 Gruppo Misure Elettriche ed Elettroniche Filters can be classified according to four basic types: Low- pass, High Pass, Band-pass and Band-stop. Their names originate from the residual spectral content that they leave in the output signal. Here is a representation of the transfer functions of ideal filters: G( j ) G( j ) C1 C 2 filtro passa-basso ideale Ideal low-pass filter filtro passa-alto ideale Ideal high-pass filter G( j ) G( j ) C1 C 2 filtro passa-banda Ideal band-passideale filter C1 C 2 filtro sopprimi-banda Ideal band-stop ideale filter 18 Gruppo Misure Elettriche ed Elettroniche The frequency band where 𝐺 𝑗𝜔 the filter; the band where 𝐺 𝑗𝜔 ≠ 0 is called the passband of = 0 is called, instead, stopband. The pulsations c, called cutoff (or break, or corner) frequencies, are in correspondence with the transitions between passband and stopband. Actual physically realizable filters, however, have not such net transitions between passband and stopband; they have, instead, intermediate bands of non-zero width, called transition bands. In addition, actual filters also present a certain variability of the transfer function in passband as well as stopband. 19 Gruppo Misure Elettriche ed Elettroniche There are several classes of analogue filters, with different advantages and drawbacks. The main classes are: • Chebyshev filters of type I, with equiripple oscillations in passband; • Chebyshev filters of type II, with equiripple oscillations in stopband; • Cauer, a.k.a. elliptic, filters, with equiripple oscillations in passband and stopband, but has smallest transition band. H ( j ) [dB] • Butterworth filters, with a maximally flat response; 𝐻(𝑗𝜔)/𝐻0 (dB) normalizzata) di trasferimento funzione dellaof (modulo transfer function normalized Module 0 -20 Butterworth -40 Chebychev tipo II Ellittico (Cauer) -60 -80 Chebychev tipo I -100 -120 -140 0.1 1 pulsazione normalizzata Normalized pulsation 𝜔(/ 𝜔𝐶/ C ) 10 The figure shows the module of the typical frequency response of several filters, normalized to maximum amplitude 𝐻0 and expressed in dB. The frequency axis is also normalized to the cutoff frequency 𝜔𝑐 . Please note the different transition 20 band widths. Gruppo Misure Elettriche ed Elettroniche The specifications on the frequency response of a filter are given through masks that indicate to the designer the limits within which the transfer characteristic should stay. For example, a design mask for a low pass filter might be the following: massima variazione di guadagno ammessa in banda passante -20 -40 -60 -80 banda passante banda di transizione H ( j ) [dB] 𝐻(𝑗𝜔)/𝐻 0 (dB) normalizzata) di trasferimento funzione dellaof (modulo transfer function normalized Module 0 banda oscura attenuazione minima richiesta in banda oscura -100 -120 -140 pulsazione di taglio 𝑐𝑢𝑡𝑜𝑓𝑓 𝑝𝑢𝑙𝑠𝑎𝑡𝑖𝑜𝑛 In more advanced applications it is often necessary to meet well-defined constraints also for the phase response of the filter, i.e. the behaviour of the phase of the 21 transfer function. Gruppo Misure Elettriche ed Elettroniche For our purposes (spectral analysis), however, it would be necessary to create a very selective band-pass filter with a central frequency variable and electronically controllable in a wide range. B2 B1 (banda passante) 0 An essential feature of a band-pass filter is its selectivity or quality factor, which have several more or less equivalent definitions. For our purposes, they can be defined as the ratio between resonance frequency 𝜔0 and bandwidth (where bandwidth is defined within a given attenuation). -20 H ( j) (dB) -40 -60 -80 𝜔0 selectivity 𝑄 = 𝐵1 Δ -100 -120 C1 (pulsazione di taglio 1) 𝜔0 C2 (pulsazione di taglio 2) However, this proved to be a rather difficult technological challenge, hence all spectrum analyzers use different techniques to achieve the same end result, i.e. the selection of small portions of the frequency spectrum of the signal to be analyzed. 22 Gruppo Misure Elettriche ed Elettroniche The most widely used analysis method in commercial analog spectrum analyzers is the so-called super-heterodyne, which has the following working principle: attenuatore di ingresso filtro passa basso di preselezione mixer sensibilità verticale filtro IF rivelatore di picco ingresso segnale filtro video VCO Display CRT generatore di rampa 23 Gruppo Misure Elettriche ed Elettroniche Let's briefly describe each functional block: Input attenuator Permits to adapt the amplitude of the input signal to the sensitivity of later stages. Its presence reduces the distortion that would be introduced by the mixer for high power signals, thus increasing the spurious free dynamic range of the analyzer which, in commercial products, can be several tens of dBs. Low-pass preselection filter Filters the input signal to eliminate signals outside the analyzer's sensitivity band; given the particular principle of operation of the spectrum analyzer, the presence of these signals would distort measurement results. VCO (Voltage Controlled Oscillator) It is a sine waveform signal generator (sine oscillator) with frequency directly proportional to the voltage applied on a control input. 24 Gruppo Misure Elettriche ed Elettroniche Mixer It is perhaps the main component of an analog spectrum analyzer; it shifts in frequency the input signal permitting the use of a fixed frequency Intermediate Frequency band-pass filter. Hence it overcomes the difficulty in implementing highly selective band-pass filters with variable center frequency. We will see soon why the mixer is able to shift in frequency input signals. Ramp generator It is a signal generator of a saw tooth waveform, virtually identical to what you can find in analog oscilloscopes; it provides the input to the VCO and, in old spectrum analyzers, it is the signal that pilots the horizontal deflection of the cathode ray tube (CRT). Vertical sensitivity It is an adjustable gain amplifier that (often) operates "in sync" with the input attenuator, in the sense that, by "default“, its gain compensates exactly the attenuation of the input attenuator. However, the operator is free to set any other gain among those available. 25 Gruppo Misure Elettriche ed Elettroniche IF filter (intermediate frequency filter) It's a band-pass filter with a fixed central frequency but with a bandwidth that can be controlled by the operator. The output signal at this stage has very small bandwidth compared to the bandwidth under examination (frequency span). Indeed, the bandwidth of the IF filter is called Resolution Bandwidth (RBW). Peak detector Allows you to measure the peaks of its input signal; if the IF filter is selective enough, this stage receives, ideally, a sine signal and, if it is well designed, it outputs a constant signal with very low residual ripple. Video filter It is basically a low-pass filter and is inserted here to improve the quality of the signal displayed on the CRT; it strongly attenuates the noise that adds to the signal at the various intermediate stages of the analyzer. Display It is used to visualize the result of the spectral analysis and, in old analog analyzers, had a structure very similar to that of the Vector scan CRT (Catod Ray Tube which use electrostatic deflection) of analog oscilloscopes. Showing instrument settings and menus can only be obtained with Raster CRT displays (which use magnetic deflection), which in modern digital analyzers are replaced by LCDs. 26 Gruppo Misure Elettriche ed Elettroniche Operating principle To make clear the principle of operation of a super-heterodyne system, few observations are sufficient: 1) the product of two signals in the time domain is equivalent to the convolution of the respective spectra in the frequency domain (a fundamental result of Fourier's theory). In formulas: ℑ +∞ 𝑥1 (𝑡) ⋅ 𝑥2 (𝑡) ֎ 𝑋1 (𝑓) ∗ 𝑋2 (𝑓) = න 𝑋1 (𝜑) ⋅ 𝑋2 (𝑓 − 𝜑) 𝑑𝜑 −∞ 27 Gruppo Misure Elettriche ed Elettroniche 2) the frequency spectrum of the function cos 𝜔0 𝑡 is real and consists of two Dirac delta distributions with area 1/2, at frequencies ±𝜔0 : T0 = 1 1 1 ( − 0 ) + ( + 0 ) 2 2 ⎯⎯ → ⎯ ⎯ cos( 0t ) 1 ( + 0 ) 2 1 2 = f 0 0 F 0 t 1 ( − 0 ) 2 1 2 − 0 1 2 0 + 0 -1 28 Gruppo Misure Elettriche ed Elettroniche 3) the convolution of any function with a Dirac delta gives the same function translated onto the delta; if we see this operation directly in the frequency domain, we have: X ( ) ( − 0 ) = X ( − 0 ) X ( ) X ( − 0 ) 0 0 0 ( −0 ) 0 0 29 Gruppo Misure Elettriche ed Elettroniche Apart from some side effects due to the inherent nonlinearity of some of its components (especially the mixer), the superheterodyne system of an analog spectrum analyzer uses precisely these three properties to produce a copy of the signal to be analyzed translating in frequency its spectrum. Afterwards, the intermediate frequency filter (a very selective band-pass filter at a fixed central frequency) selects a small portion of the translated signal spectrum and feeds it to the next stages, until it reaches the vertical deflection stage of the CRT. Since the (sinusoidal) signal provided by the VCO has a variable frequency, controlled by the ramp generator which also pilots the horizontal deflection of the CRT, the end result is to get a more detailed representation of the signal spectrum the more selective the IF filter is. 30 Gruppo Misure Elettriche ed Elettroniche Let [fmin,SA, fmax,SA] be the useful range of the spectrum analyzer. We should design the analyzer by choosing both the frequency of the IF filter, fIF , as well as the frequency range of the signal generated by the VCO. It should be considered that the mixer is a non-linear device which outputs the sum of: 1) the analyzer input signal (without frequency shift); 2) the signal coming from the VCO; 3) components at frequencies which are the sum and the difference of mixer inputs. f SA f SA fVCO In a real mixer, several components are produced at frequencies ±𝑛 ∙ 𝑓𝑆𝐴 ±𝑚 ∙ 𝑓𝑉𝐶𝑂 , where n and m are integers ≥ 0. Stronger components are obtained for small n and m. fVCO Gruppo Misure Elettriche ed Elettroniche 1.Choosing the Intermediate Filter frequency First, you must use an intermediate filter frequency that is greater than the spectrum analyzer range (fIF > fmax,SA), because part of the input signal passes through the mixer (pass through) without being shifted in frequency and would appear directly at the IF filter exit regardless of fVco. Example: Suppose the frequency range of the spectrum analyzer is [10 kHz - 4 GHz] and that, by mistake, it is fIF 2 GHz. In this case, as mentioned above, by choosing fIF within the instrument band, it cannot be identified whether the harmonic component selected by the IF filter is obtained by means of frequency shift or is an input signal component at frequency 2 GHZ. In addition, to properly analyze the input signal component located at 2 GHz, the signal should be shifted by 0 Hz, hence the VCO should generate a 0 Hz, which is not possible. Gruppo Misure Elettriche ed Elettroniche 2.Choosing the VCO frequency range Remember that real signals have symmetrical spectrum and therefore it is enough to analyze only half of the spectrum. When multiplying the input signal with the VCO sinewave output, the spectrum of the signal is shifted by fVco. At this point, you need to choose which of the two subbands analyze, namely which subband should be shifted into the band of the intermediate frequency filter. 𝑓𝐼𝐹 Lower Side band Upper Side band 0 f min f max Lower Side band 𝑓𝑉𝐶𝑂 − 𝑓max fVCO Upper Side band 𝑓𝑉𝐶𝑂 + 𝑓max 𝑓𝑉𝐶𝑂 − 𝑓min 𝑓𝑉𝐶𝑂 + 𝑓min 33 Gruppo Misure Elettriche ed Elettroniche There are two options for choosing the VCO band: Lower side band analysis Upper side band analysis 𝑓𝑉𝐶𝑂 − 𝑓max ≤ 𝑓𝐼𝐹 ≤ 𝑓𝑉𝐶𝑂 − 𝑓min 𝑓𝑉𝐶𝑂 + 𝑓min ≤ 𝑓𝐼𝐹 ≤ 𝑓𝑉𝐶𝑂 + 𝑓max f IF + f min fVCO f IF + f max f IF − f max fVCO f IF − f min You choose the solution that analyzes the lower side band to exclude that products of the mixer at harmonic frequencies of 𝑓𝑉𝐶𝑂 fall into the filter band. So the smallest frequency of the VCO is 𝑓min_𝑉𝐶𝑂 = 𝑓𝐼𝐹 + 𝑓min . Signal band f IF VCO range Choice of 𝑓𝑉𝐶𝑂 𝑓𝑉𝐶𝑂 For analyzing the upper side band f min f max f IF − f max For analyzing the f IF + f min lower side band f IF − f min f IF + f max Gruppo Misure Elettriche ed Elettroniche In conclusion the following conditions must be verified: 1. f max_ SA f IF 2. fVCO f IF + f min_ SA ; f IF + f max_ SA A consequence of point 2 is that 𝑓𝐼𝐹 ≤ 𝑓min_𝑉𝐶𝑂 . X ( f − f LO ) 0 X( f ) f f LO mixer HIF ( f ) 0 f 0 sensibilità verticale 0 f LO f0,IF f LO f filtro IF f 35 Gruppo Misure Elettriche ed Elettroniche Under these conditions it is not possible for super-heterodyne spectrum analyzers to measure the continuous component. In fact, to analyze the component at zero frequency (𝑓min_SA = 0), you should have fmin_VCO=fIF (in accordance with condition 2). In this case, due to the mixer behavior, the same signal generated by the VCO passes through the IF filter and is then displayed as if it was part of the input signal. A further observation can be made regarding the condition 𝑓max_𝑆𝐴 < 𝑓𝐼𝐹 . To enforce this condition, you use the low pass preselection filter that filters the incoming signal in the spectrum analyzer band and avoids the problem due to image frequencies. 36 Gruppo Misure Elettriche ed Elettroniche Example: Consider a SA designed to operate in the range [5 kHz-10 MHz] having an IF filter at 20 MHz. As seen above, the VCO should generate in the [20.005-30] MHz frequency range. There are no problems if you want to display a sine signal having frequency fIN1=5 MHz (which complies to fIN1<fmax_SA), just use FVCO=25 MHz. In this case, at the exit of the mixer you will get two pulses: one at 20 MHz frequency that falls into the IF filter band and the other at 30 MHz frequency that is not displayed. -5 −𝑓𝐼𝑁1 5 𝑓𝐼𝑁1 20 𝑓𝑉𝐶𝑂 − 𝑓𝐼𝑁1 25 𝑓𝑉𝐶𝑂 MHz 37 Gruppo Misure Elettriche ed Elettroniche Let's see, instead, what happens to an input component of the SA at frequency fIN2=45 MHz (fIN2>fmax_SA) in the case that component is not filtered by preselection filter. At fVCO=25 MHz, two pulses are obtained at the mixer output: one at a 20 MHz frequency that falls into the IF filter band and the other at a 70 MHz frequency that is not displayed. In conclusion, the analog spectrum analyzer in the absence of the preselection low pass filter, at the same VCO setting, would simultaneously measure the component associated with the 5 MHZ signal and the component at 45 MHz, which is therefore called image frequency. -5 −𝑓𝐼𝑁1 5 𝑓𝐼𝑁1 20 𝑓𝑉𝐶𝑂 − 𝑓𝐼𝑁1 𝑓𝐼𝑁2 − 𝑓𝑉𝐶𝑂 25 𝑓𝑉𝐶𝑂 MHz 45 𝑓𝐼𝑁2 image frequency of 5 MHz 38 Gruppo Misure Elettriche ed Elettroniche 2. Digital Spectrum analyzers The scheme of a digital spectrum analyzer can be traced back to 5 key blocks: Sensor Or Transducer Analog preconditioning (amplifiers, filters) ADC Processing Display Transduction: It happens (if necessary) at the probe level Preconditioning: Consisting of electronic amplifiers, anti-aliasing filters, low-pass filters (for noise elimination) and high-pass filters for the elimination of DC. Conversion: Fast or very fast ADCs (on the order of GSa/s) with low resolutions (typical value: 8 bit); 39 Gruppo Misure Elettriche ed Elettroniche Processing: possible application of a weighing window (Hann, Hamming, Kaiser, Flat-Top etc.); Fast Fourier Transform (FFT); calculations on the spectrum (RMS of individual harmonics, RMS of noise floor, THD, etc.) Display: representation of the result of the spectral analysis It is therefore clear that the "heart" of the instrument is the processing block, that is, the block that deals with the implementation of the Discrete Fourier Transform (DFT). To better understand the working principle of the digital spectrum analyzer it is worth recalling some fundamental properties of the Fourier Transform (FT). It is known that the FT of a continuous time signal 𝑥(𝑡) is given by: +∞ 𝑋 𝑓 =න 𝑥 𝑡 ⋅ 𝑒 −𝑗⋅2⋅𝜋⋅𝑓⋅𝑡 𝑑𝑡 −∞ 40 Gruppo Misure Elettriche ed Elettroniche while the inverse Fourier transform is: +∞ 𝑥 𝑡 =න 𝑋 𝑓 ⋅ 𝑒 𝑗⋅2⋅𝜋⋅𝑓⋅𝑡 𝑑𝑓 −∞ The implementation of the FT In the discrete time domain (signal sampled with period Ts) as the sum of an infinite number of samples leads to the definition of the DTFT (Discrete Time Fourier Transform): +∞ 𝑋 𝑓 = 𝑥 𝑛𝑇𝒔 𝑒 −𝑗⋅2⋅𝜋⋅𝑓⋅𝑛⋅𝑇𝒔 X (f) is periodic with period 1/Ts , as a consequence of sampling 𝑛=−∞ Clearly, since the memory of any real instrument is finite, it is necessary to limit the length of the sample; this leads to the definition of Discrete Fourier Transform (DFT): 𝑁−1 𝑋 𝑚𝐹 = 𝑥 𝑛𝑇𝒔 ⋅ 𝑛=0 𝑁−1 𝑒 −𝑗⋅2⋅𝜋⋅𝑛⋅𝑚⋅𝐹⋅𝑇𝒔 = 𝑥 𝑛𝑇𝒔 ⋅ 𝑛=0 2⋅𝜋 −𝑗⋅ 𝑁 ⋅𝑛⋅𝑚 𝑒 , 1 𝐹= 𝑁𝑇𝑆 41 Gruppo Misure Elettriche ed Elettroniche where N is the number of samples, Ts the sampling period, To= N Ts the window duration and F the frequency resolution given by: 𝐹𝑠 𝟏 𝟏 𝐹= = = 𝑁 𝑵𝑻𝒔 𝑻𝑶 Also in this case it will be possible to derive the time-discrete signal of origin by means of the Inverse Discrete Fourier Transform (IDFT): 𝑥 𝑛𝑇𝒔 = 𝑁−1 𝑁−1 𝑚=0 𝑚=0 1 1 𝑋 𝑚𝐹 ⋅ 𝑒 𝑗⋅2⋅𝜋⋅𝑛⋅𝑚⋅𝐹⋅𝑇𝒔 = 𝑋 𝑚𝐹 ⋅ 𝑒 𝑗⋅2⋅𝜋⋅𝑛⋅𝑚/𝑁 𝑁 𝑁 In Practice the DFT establishes the relationship between the signal samples in the time domain and their representation in the frequency domain. 42 Therefore if you consider N samples of x (t): Gruppo Misure Elettriche ed Elettroniche 𝑥(0), 𝑥(𝑇𝒔 ), 𝑥(2𝑇𝒔 ), . . . , 𝑥((𝑁 − 1)𝑇𝒔 ) The DFT, which produces the frequency representation of the signal, is also made up of N samples: 𝑋(0), 𝑋(𝐹), 𝑋(2𝐹), . . . , 𝑋((𝑁 − 1)𝐹) X ( mF ) x ( nTs ) 0.8 0.9 0.4 0.7 DFT 0 -0.4 0.5 0.3 -0.8 0.1 0 10 20 30 0 Time Domain 10 20 30 Frequency domain 43 Gruppo Misure Elettriche ed Elettroniche where the distance between two successive samples in time will be equal to the sampling period Ts while the distance between two samples in frequency will be equal to the spectral resolution 𝐹 = 1 𝑁𝑇𝑠 . The sequence notation 𝒙 𝒏 = 𝒙𝒏 = 𝒙(𝒏𝑻𝒔 ) and 𝑿 𝒎 = 𝑿𝒎 = 𝑿(𝒎𝑭) is also used. To fully understand the operation of the digital Spectrum Analyzer it is now useful to recall some properties of the DFT. 1. The discrete signal 𝑥(𝑛𝑇𝑠) and its DFT 𝑋(𝑚𝐹) can both be considered periodic with periods 𝑁𝑇𝑠 and 𝑁𝐹 = 1/𝑇𝑠 , respectively: 𝑥( 𝑛 + 𝑘 ⋅ 𝑁 𝑇𝒔 ) = 𝑥(𝑛𝑇𝒔 ) 𝑋( 𝑚 + 𝑘 ⋅ 𝑁 𝐹) = 𝑋(𝑚𝐹) 2. The DFT of a signal x(nTs) is, generally, a complex function of 𝑚𝐹: 𝑋(𝑚𝐹) = 𝑋𝑟𝑒 (𝑚𝐹) + 𝑗 ⋅ 𝑋𝑖𝑚 (𝑚𝐹) 44 Gruppo Misure Elettriche ed Elettroniche 3. If the signal x (nTs) is real, the module of DFT has even symmetry, while phase has odd symmetry. Equivalently, we can say that the real part of the DFT is an even function of 𝑚𝐹 while the imaginary part is an odd function of 𝑚𝐹: X re ( m) 𝑋𝑟𝑒 (−𝑚𝐹) = 𝑋𝑟𝑒 ( 𝑁 − 𝑚 𝐹) = 𝑋𝑟𝑒 (𝑚𝐹) X im ( m) 𝑋𝑖𝑚 (−𝑚𝐹) = 𝑋𝑖𝑚 ( 𝑁 − 𝑚 𝐹) = −𝑋𝑖𝑚 (𝑚𝐹) 45 Gruppo Misure Elettriche ed Elettroniche 4. If the signal x(nTs) is imaginary, its DFT will have a real part with odd symmetry and an imaginary part with even symmetry: 𝑋𝑟𝑒 (−𝑚𝐹) = 𝑋𝑟𝑒 ( 𝑁 − 𝑚 𝐹) = −𝑋𝑟𝑒 (𝑚𝐹) 𝑋𝑖𝑚 (−𝑚𝐹) = 𝑋𝑖𝑚 ( 𝑁 − 𝑚 𝐹) = 𝑋𝑖𝑚 (𝑚𝐹) 5. The DFT of the linear combination of two signals over time is the linear combination of the individual DFTs (i.e. the DFT is a linear operator): 𝑥(𝑛𝑇𝒔 ) = 𝑎1 ⋅ 𝑥1 𝑛𝑇𝑺 + 𝑎2 ⋅ 𝑥2 𝑛𝑇𝒔 ⇔ 𝑋(𝑚𝐹) = 𝑎1 ⋅ 𝑋1 (𝑚𝐹) + 𝑎2 ⋅ 𝑋2 (𝑚𝐹) 6. A delay (cyclic shift) of the signal in the time domain is equivalent to a phase shift of the DFT: 𝑦(𝑛𝑇𝒔 ) = 𝑥( 𝑛 − 𝑘 𝑇𝒔 ) ⟺ 𝑌(𝑚𝐹) = 𝑋(𝑚𝐹) ⋅ 2⋅𝜋 −𝑗 𝑁 𝑘𝑚 𝑒 46 Gruppo Misure Elettriche ed Elettroniche 7. A phase shift in the time domain introduces a corresponding delay (cyclic shift) in the DFT : 𝑥 𝑛𝑇𝒔 ⋅ 𝑒 +𝑗 2⋅𝜋 𝑘𝑛 𝑁 ⟺𝑋 ( 𝑚 − 𝑘 𝐹) 8. The DFT of the cyclic convolution of two signals in the time domain is given by the product of the individual DFT In the frequency domain: 𝑁−1 𝑦 𝑘𝑇𝑠 = 𝑥 𝑘𝑇𝑠 ∗ ℎ 𝑘𝑇𝑠 = 𝑥 𝑛𝑇𝑠 ⋅ ℎ 𝑘 − 𝑛 𝑇𝑠 ⟺ 𝑛=0 ⟺ 𝑌(𝑚𝐹) = 𝑋(𝑚𝐹) ⋅ 𝐻(𝑚𝐹) 9. The product of two signals in the time domain is equivalent to the cyclic convolution of their respective DFTs: 𝑧(𝑛𝑇𝒔 ) = 𝑥 𝑛𝑇𝒔 ⋅ 𝑦 𝑛𝑇𝒔 ⟺ 𝑍(𝑚𝐹) = 𝑋(𝑚𝐹) ∗ 𝑌(𝑚𝐹) 47 Gruppo Misure Elettriche ed Elettroniche Returning now to the problem of the elaboration of the FT in the digital spectrum analyzer and using the previously seen DFT properties, two important conclusions can be drawn: • The digital spectrum analyzer will only process real signals over time, its DFT will be complex and contain information of both amplitude and phase: 𝑋 𝑚𝐹 = 𝑋 𝑚𝐹 𝑒 𝑗∡𝑋 𝑚𝐹 • Moreover, since the signal will certainly be real, its DFT will be symmetric with respect to the index N/2: 𝑋(𝑚𝐹) = 𝑋( 𝑁 − 𝑚 𝐹) ∡𝑋[𝑚𝐹] = −∡𝑋[ 𝑁 − 𝑚 𝐹] 48 Gruppo Misure Elettriche ed Elettroniche Hence the DFT of a real signal contains redundant information in its N samples, and the use of N/2 samples is sufficient for a full representation in the frequency domain of the signal. It is now easy to describe how the DFT algorithm works. It has been seen that a sampling interval of Ts corresponds to a frequency resolution 𝐹 = 𝐹𝑠 /𝑁 = 1/𝑁𝑇𝑠 . This means that the k-th output of the DFT corresponds to the harmonic frequency component 𝑓𝑘 = 𝑘 · 𝐹 ; the properties of the DFT ensure that only the first half of its samples carry independent information because the remaining N/2 are redundant and contain information about the negative frequencies. We’ll see things in more detail... 49 Gruppo Misure Elettriche ed Elettroniche The bin X (0·F) = X [0] always represents the continuous 𝑁=32 3 component of the signal. If N is even, bins X [1].... X [N/2-1] 2 contain useful spectral information, bin X[N/2] shows the component (cosine only, not 1 sine) at the folding frequency and bins from N/2 + 1 to N-1 represent components at 0 5 15 25 31 negative frequencies, which will have amplitude (module) equal to the corresponding components at positive frequency, e.g. 𝑋 1 = 𝑋 𝑁 − 1 ∗ . Continuous Component Positive Frequencies Folding Frequency f = 16 F Negative Frequencies 50 Gruppo Misure Elettriche ed Elettroniche The only difference when N is 3 an odd number, is that the folding component is not 2 visible because in this case N/2 isn’t a whole number. 1 0 Component Go on 5 15 Frequencies Positive 25 30 Frequencies Negative 51 Gruppo Misure Elettriche ed Elettroniche It is worth noting that to get a significant time saving in digital processing, the implementation of DFT is done efficiently through the FFT. This algorithm (which, as mentioned above, was originally proposed by Cooley and Tukey), is effective when N is a power of 2, and exploits many symmetries in DFT calculation. The computational saving produced in this case is the reduction of the number of complex value operations from N2 to about N·log2N (e.g. when N = 4096, you pass from almost 17 million to just about 49’000 operations!). In case N is not a power of 2 several artifices can be used to allow the use of the FFT; the most widespread artifice, useful when the signal is not periodic, is the zeropadding, that is the addition at the end of the sequence of zero-value samples, which speeds up processing and "fictitiously" increases frequency resolution. This operation actually produces only a frequency interpolation of the spectrum of a signal which is zero outside the observation window; it increases the number of DFT bins while maintaining the shape of the spectrum almost intact but producing a more detailed view of it (reduction of the picket-fence effect). 52 Gruppo Misure Elettriche ed Elettroniche 0.7 0.8 0.5 0.4 0 0.3 -0.4 0.1 -0.8 0 4 8 12 16 20 0 4 8 12 16 20 0.7 0.8 0.5 0.4 0 0.3 -0.4 0.1 -0.8 0 10 20 30 0 10 20 30 53 Gruppo Misure Elettriche ed Elettroniche ampiezze 3 The Power spectrum 2 It has been seen that the DFT measures not only amplitudes but, 1 compared to the analogue SA, also the phases of the individual signal components. In all those cases where 0 5 15 25 35 25 35 80 potenza you are not interested in the signal 60 phase, it may be convenient to consider the Power Spectral Density 40 (PSD). The PSD, proportional to the square of the individual components of the DFT, shows the power of the signal components at the various frequencies. 20 0 5 15 54 Gruppo Misure Elettriche ed Elettroniche For time continuous signals 𝑥 𝑡 , power is 1 +𝑇/2 𝑃𝑥 = lim න 𝑥 𝑡 𝑇→∞ 𝑇 −𝑇/2 2 𝑑𝑡 Power Spectral Density 𝑆𝑥 𝑓 is 2 𝑋𝑇 𝑓 𝑆𝑥 𝑓 = lim 𝑇→∞ 𝑇 𝑇 where 𝑋𝑇 (𝑓) is the Fourier transform of 𝑥𝑇 𝑡 = ൝ 𝑇 𝑥 𝑡 , for − 2 < 𝑡 < 2 0, elsewhere . 𝑆𝑥 𝑓 has the dimensions of power divided frequency, hence it is measured in W/Hz , or, for voltage signals, V 2 /Hz. 𝑓 Power in a band between 𝑓1 and 𝑓2 is given by 𝑓2 𝑆𝑥 𝑓 𝑑𝑓 . 1 That definitions are developed further in the context of random signals, however deterministic signals will be considered here. 55 Gruppo Misure Elettriche ed Elettroniche • For finite sequences of samples we can calculate the DFT 𝑋(𝑚𝐹), instead of the Fourier transform 𝑋(𝑓) , and the Periodogram is defined as: 𝑋(𝑚𝐹) 𝑁 2 • Power Spectral Density of the original continuous signal can be estimated from the Periodogram as: 𝑋(𝑚𝐹) 𝑆𝑥 𝑚𝐹 = 𝑇𝑠 × 𝑁 2 2 2 𝑋𝑟𝑒 (𝑚𝐹) + 𝑋𝑖𝑚 (𝑚𝐹) = 𝑇𝑠 × 𝑁 units W Hz • Power integrated in a frequency bin of the DFT, which has resolution 𝑓𝑠 /𝑁, is: 𝑓𝑠 𝑿(𝒎𝑭) 𝟐 𝑆𝑥 𝑚𝐹 × = 𝑁 𝑵×𝑵 units W 56 Gruppo Misure Elettriche ed Elettroniche 𝑓𝑠 = 10 Hz Example 𝑁 𝑁 = 16 𝑇𝑂 = 𝑓 = 1.6 s 𝑠 Resolution 𝐹 = 1/𝑇𝑂 = 0.625 Hz Synchronous sampling of a signal at frequency 𝑓𝑥 = 3/𝑇𝑂 = 1.875 Hz 𝑦 𝑡 = 𝐴 cos 2𝜋𝑓𝑥 𝑡 Power is 𝐴2 /2 = 1/2 V 2 𝐴=1V 8 𝑗2𝜋𝑓 𝑡 𝑥 𝑒 16 8 −𝑗2𝜋𝑓 𝑡 𝑥 𝑒 16 Power is DFT 3𝐹 2 𝑁2 11 + DFT −3𝐹 2 𝑁2 = 8 2 16 + 8 2 16 1 1 1 = 4 + 4 = 2 V2 1𝑁 Density at 𝑓𝑥 is 4 𝐹 = 4 𝑓 = 0.4 V 2 /Hz , changes with 𝑁 and 𝑓𝑠 ! 𝑠 58 Gruppo Misure Elettriche ed Elettroniche Of course, thanks to the symmetry property of the DFT of a real signal, the power spectrum at the positive ampiezze 3 frequencies is the same of negative 2 frequencies and, even though the 1 number of its components is 0 always N, just as in the case of 5 15 25 35 80 potenza DFT, only N/2 components are sufficient to describe the PSD of the signal. Again, the component at the folding frequency is visible only 60 40 20 when N is even. 0 5 15 25 35 59 Gruppo Misure Elettriche ed Elettroniche Leakage and windowing problems It is known that the sampling theorem ensures the perfect reconstruction of band-limited signals (note: a band-limited sig. cannot be also time-limited) : - for an aperiodic signal, an infinite number of samples is required; - for a periodic signal, a finite number of samples is required, but with an observation window which is a whole multiple of the signal period. In any other case an inherent discontinuity on the sampled signal will be introduced. Because of the limited memory of the spectrum analyzer, there is an inherent restriction of the observation interval, and therefore on the number N of samples of the DFT. In DFT, it is assumed a “periodization” of the signal in time with period N. Indeed, the FT of a 𝑇0 periodic signal has only frequency components (Dirac deltas) at multiples of 1/𝑇0 . 60 Gruppo Misure Elettriche ed Elettroniche This periodization, which is applied in both cases of periodic and aperiodic signals, produces, in general, a discontinuity a mpiezza ampiezza between the successive repetitions of the acquired signal. discontinuità tempo tempo finestra di osservazione 61 Gruppo Misure Elettriche ed Elettroniche The result of this operation in the frequency domain is the emergence of spectral components that are not actually present in the spectrum of the starting signal, so that the final spectrum is a "smeared" version of the original spectrum. A [V] 3.5 3.5 A [V] 2.5 2.5 1.5 1.5 0.5 0 40 80 N 120 0.5 0 40 80 N 120 62 Gruppo Misure Elettriche ed Elettroniche This can be understood by considering that using a finite sequence corresponds to multiplying the original signal 𝑥𝑜 𝑛𝑇𝑠 by a rectangular window 𝑤 𝑛𝑇𝑠 , which in frequency domain corresponds to convolve with the transform of a rectangular signal. 𝑥 𝑛𝑇𝑠 = 𝑥𝑜 𝑛𝑇𝑠 ∙ 𝑤(𝑛𝑇𝑠 ) ⇔ 𝑋(𝑓) = 𝑋𝑜 𝑓 ∗ 𝑊 𝑓 In general, this will cause: • Truncation error, observed in time domain. • Leakage (or spectral dispersion) observed in frequency domain •A possible aliasing error, because truncation and leakage can produce frequency components higher than folding frequency 𝑓𝑠 /2, which will be folded into the base band. Indeed, because of possible abrupt discontinuities between the successive repetitions of the periodized signal, high frequency components may be introduced. 63 Gruppo Misure Elettriche ed Elettroniche To reduce these adverse effects, windows different from the rectangular one can be used. The aim is to attenuate the discontinuities introduced by the signal truncation and periodization using windows with more suitable weights 𝑤 𝑛𝑇𝑠 that show a more "gentle" trend at the extremes of the observation interval. These new weights smooth the signal near window borders. The time behavior of the window will define the spectral characteristics of the lobes of the corresponding Fourier Transform and, consequently, the final effect on the reconstructed spectrum. All this is explained by keeping in mind that windowing implies the multiplication in time of the data record with a function and, consequently, the convolution in the frequency domain of the signal spectrum and the window spectrum. 64 Gruppo Misure Elettriche ed Elettroniche There is no “perfect" window because each window makes a compromise between two factors: • Frequency resolution (width of the main lobe). It affects the accurate estimation of the frequency of a pure tone. • Spectral dispersion (amplitude of the side lobes). High dispersion means that higher-power components can leak in frequency and hide smaller components. The width of the main lobe and the amplitude of the side lobes cannot be minimized at the same time. It is worth highlighting that only when the data record contains a whole number of cycles there is no leakage and the rectangular window should be used, however this is a condition that can not be perfectly realized in practice. 65 Gruppo Misure Elettriche ed Elettroniche Usefulness of windowing should be discussed in relation to the signal to be analyzed and applications. 1) Periodic signals (e.g. a single tone) can be measured without leakage errors if the rectangular window contains an integer number of signal cycles (a condition called synchronized sampling, which will be discussed later). If not, windows choice matters. 2) Deterministic aperiodic signals that extend outside the observation window, are subjected to unrecoverable information loss, independently on window choice. 3) Deterministic aperiodic signals, which are entirely inside the window and zero outside (such as transient response of a system): windowing may be useful. In this case, signal is not bandlimited, hence in some circumstances sampling may produce interference from spectral replicas, which can be mitigated by proper widowing. Windowing can be also used to reduce measurement noise contribution where the signal fades out. 4) For random stationary signals, statistical properties of the entire signal may be estimated inside an observation window: in this case window choice matters. 66 Gruppo Misure Elettriche ed Elettroniche Class of cosine windows 2 ln w ( nTc ) = ( − 1)l a l cos N l =0 L Rectangular Hann (Hanning) Blackman Cosine window of order L Normalization condition: σ𝐿𝑙=0 𝑎𝑙 = 1 Flat Top Rectangular Hann Blackman Sample index Flat Top Properties 𝑁 Odd derivatives, at the edges 𝑛 = ± 2 are null because the window is a sum of cosines. 𝑁 Even derivatives of order 𝑗, at the edges 𝑛 = ± 2 , are given by 𝐿 𝑤 𝑗 ±𝑁/2 = −1 𝑙 𝑙 𝑗 𝑎𝑙 𝑙=0 67 Gruppo Misure Elettriche ed Elettroniche Spectrum (DTFT) of the cosine window: sin ( ) L al l W( f )= − 1 ( ) 2 − l2 ao l =0 Flat Top Rectangular Where 𝛿 = 𝑁𝑇𝑠 𝑓 Hann (Hanning) Blackman Rectangular Hanning Blackman Flat Top The asymptotic frequency decay (spectral dispersion) depends on the cancellation of the derivatives at the edges. w ( N / 2 ) = w (t ) 0 t → N /2− The decay is 1/f when the window is not zero at the edges (the derivative of order 𝑗 = 0 is not zero) 68 Gruppo Misure Elettriche ed Elettroniche 𝑤 ±𝑁/2 = 𝑤 𝑡 =0 𝑤 ′′ ±𝑁/2 = 𝑤 ′′ 𝑡 ≠0 𝑤 ±𝑁/2 = 𝑤 𝑡 =0 𝑡 →𝑁/2− 𝑡 →𝑁/2− 𝑡 →𝑁/2− 𝑤 ′′ ±𝑁/2 = 𝑤 ′′ 𝑡 𝑡 →𝑁/2− ⇒ The decay Is 1/𝑓 3 If the window is zero at edges, but it’s 2-nd derivative is not zero. = 0 ⇒ frequency decay 1/𝑓 5 𝑤 𝐼𝑉 ±𝑁/2 = 𝑤 𝐼𝑉 𝑡 ≠ 0 𝑡 →𝑁/2− 69 Gruppo Misure Elettriche ed Elettroniche The equivalent noise bandwidth of the window, 𝑬𝑵𝑩𝑾, is defined in terms of DTFT, 𝑊(𝑓), of the sampled window or, equivalently, in terms of its DFT, 𝑊 𝑚𝐹 , or also in time domain (remember Parseval’s theorem): 𝑓𝑠 2 𝑊 2 0 ⋅ 𝐸𝑁𝐵𝑊 = න 𝑓 − 2𝑠 𝑁−1 𝑓𝑠 𝑊 𝑓 2 𝑑𝑓 = 𝑊 𝑚𝐹 𝑁 𝑁−1 2 𝑚=0 = 𝑓𝑠 × 𝑤 2 𝑛𝑇𝑠 𝑛=0 It is the bandwidth 𝐸𝑁𝐵𝑊 of a rectangular filter with fixed gain 𝑊(0) having the same power of the window. 𝐸𝑁𝐵𝑊 can be calculates also in time domain, as 𝑁−1 𝐸𝑁𝐵𝑊 = 𝑓𝑠 × 𝑛=0 2 𝑁−1 𝑤2 𝑛𝑇𝑠 / 𝑤 𝑛𝑇𝑠 𝑛=0 𝑁𝑃𝐺 = 𝑓𝑠 × 𝑃𝑆𝐺 2 Δ 2 where we have defined Noise Power Gain 𝑁𝑃𝐺 = σ𝑁−1 𝑛=0 𝑤 𝑛𝑇𝒔 Δ and Peak Signal Gain 𝑃𝑆𝐺 = σ𝑁−1 𝑖=0 𝑤 𝑛𝑇𝑐 = 𝑊(0). For the rectangular window, 𝑬𝑵𝑩𝑾 = 𝒇𝒔 /𝑵 =frequency resolution. Indeed, for any window, 𝐸𝑁𝐵𝑊 is also called frequency resolution of the window. The normalized 𝐸𝑁𝐵𝑊 is defined, by comparison with a rectangular window, as 𝐸𝑁𝐵𝑊 . 𝑓𝑠 /𝑁 70 Gruppo Misure Elettriche ed Elettroniche Starting from the previously reported expressions for the discrete signal, its DFT and its PSD, obtained for the rectangular window case 𝑁−1 Sampled signal: 1 𝑥 𝑛𝑇𝒔 = 𝑋 𝑚𝐹 ⋅ 𝑒 𝑗⋅2⋅𝜋⋅𝑛⋅𝑚⋅𝐹⋅𝑇𝒔 𝑁 𝑚=0 𝑁−1 DFT: 𝑋 𝑚𝐹 = 𝑥 𝑛𝑇𝒔 ⋅ 𝑒 −𝑗⋅2⋅𝜋⋅𝑛⋅𝑚⋅𝐹⋅𝑇𝒔 𝑛=0 PSD: Power in a bin: 𝑋 𝑚𝐹 𝑆𝑥 𝑚𝐹 = 𝑇𝑠 × 𝑁 𝑋 𝑚𝐹 2 𝑁×𝑁 2 units W Hz units W 71 Gruppo Misure Elettriche ed Elettroniche we can introduce equivalent expressions for the case a generic window is used: Sampled and windowed signal: 𝑥𝑤 𝑛𝑇𝒔 = 𝑥 𝑛𝑇𝒔 ⋅ 𝑤 𝑛𝑇𝒔 𝑁−1 DFT (convolution): 𝑋𝑤 𝑚𝐹 = 𝑥 𝑛𝑇𝒔 𝑤 𝑛𝑇𝒔 𝑒 −𝑗⋅2⋅𝜋⋅𝑛⋅𝑚⋅𝐹⋅𝑇𝒔 = 𝑋 𝑚𝐹 ∗ 𝑊 𝑚𝐹 𝑛=0 PSD: 𝑆𝑥𝑤 𝑋𝑤 𝑚𝐹 𝑚𝐹 = 𝑇𝑠 × 𝑁𝑃𝐺 2 where NPG is the Noise Power Gain of the window: Δ 2 𝑁𝑃𝐺 = σ𝑁−1 𝑛=0 𝑤 𝑛𝑇𝒔 , which for a rectangular window is equal to N. Power in a bin: 𝑋𝑤 𝑚𝐹 2 𝑁𝑃𝐺 𝟐 72 Gruppo Misure Elettriche ed Elettroniche Basically we can't give universal rules about the type of window to be used in signal analysis; it will change from time to time and the correct choice depends on a prior knowledge of the signal to be examined. It can certainly be said that: • The rectangular window gives maximum frequency resolution (narrower main lobe but lateral lobes decaying more slowly) but amplitude estimates have errors that can reach 36%. • The flat-top window gives the most accurate amplitude estimates but its frequency selectivity is very low (flat main lobe, but wider among all windows and very small side lobes) ; • The Hann window is the most used for its globally good properties (main lobe larger than the rectangular window but lateral lobes decaying faster). 73 Gruppo Misure Elettriche ed Elettroniche The estimation of spectral components Let’s recall that the DFT has N outputs spaced apart by a range equal to frequency resolution: 𝐹= 1 𝑓𝑠 = 𝑁𝑇𝑠 𝑁 Among the N output elements of the DFT, only the first N/2 will be significant: they represent the continuous component 𝑋(0𝐹) and components 𝑋(𝑖𝐹)|, with i=1:N/2-1. 1. Estimation of amplitudes To obtain the correct value of the individual amplitudes, it is necessary to divide the DFT by the number N of samples, and keep in mind that for components other than continuous, it is necessary to consider both the positive and negative components, or equivalently, multiply by 2 the DFT at positive frequencies and ignore negative frequencies. 74 Gruppo Misure Elettriche ed Elettroniche N • For a rectangular window synchronous sampling, it is: 𝑋(0𝐹) 𝐴0 = ; 𝑁 𝑋(𝑘𝐹) 𝐴𝑘 = 2 𝑁 and A3 N k=1...N/2−1 It can be demonstrated by considering that component 𝐴𝑘 has power 𝐴2𝑘 2 , which should be equal to power in frequency domain in bins +𝑘𝐹 𝑋 𝑘𝐹 2 , 𝑁2 𝑋(𝑘𝐹) and −𝑘𝐹, which is 2 hence 𝐴𝑘 = 2 . 𝑁 For windowed signals and synchronous sampling it is: X ( 0F ) A0 = ; PSG X ( kF ) Ak = 2 PSG k=1...N/2-1 where 𝑃𝑆𝐺 = σ𝑁−1 𝑖=0 𝑤 𝑛𝑇𝑐 = 𝑊(0) is the Peak Signal Gain , which is N for a rectangular window. 75 • In the case of not synchronous sampling the above formula results in an error in estimating the amplitude components that depends on the type of the window and on the fractional bin of spectral offset, 𝛿 = (𝑓 − 𝑘𝐹)/𝐹 ∈ −0.5,0.5 N Amplitude error () The error is proportional to 𝑆𝐿 𝛿 and can be corrected as follows: A0 = X ( 0F ) SL ( ) PSG ; Ak = 2 X ( kF ) SL ( ) PSG k=1...N/2-1 where with 𝑆𝐿 𝛿 it has been indicated Scalloping Loss of the window. For cosine windows, it is 𝑆𝐿 𝛿 = 𝛿⋅sin 𝜋𝛿 𝜋⋅𝑎0 σ𝐿𝑙=0 −1 𝑙 𝑎𝑙 𝛿 2 −𝑙 2 Since 𝛿 is, in general, unknown, an alternative that gives good estimations consists in interpolating between adjacent spectral lines (interpolated DFT techniques). 76 2. Frequency estimation Frequency estimation requires no special precaution except to multiply the single output of the DFT by the frequency resolution F. • In the case of synchronous sampling the generic component Ak is located at frequency: f = kF ; k=0....N/2-1 • In the case of not synchronous sampling the above formula results in an error 𝛿𝐹 in frequency estimation, therefore: 𝑓 = 𝑘 + 𝛿 𝐹; k=0....N/2−1 77 3. Phase estimation A cosine signal with frequency 𝑘𝐹 and phase 𝜑 is defined as 𝑥 𝑡 = A cos 2𝜋 𝑘𝐹 𝑡 + 𝜑 = 𝐴 Its DFT has only components 𝑋 𝑘𝐹 = 𝑒𝑖 𝑘𝐹𝑡+𝜑 𝐴𝑒 𝑖𝜑 𝑁 2 + 𝑒 −𝑖 2 𝑘𝐹𝑡+𝜑 and 𝑋 −𝑘𝐹 = 𝐴𝑒 −𝑖𝜑 𝑁 . 2 When performing a DFT, 𝑋 𝑘𝐹 can be expressed, in general, as 𝑋 𝑘𝐹 = 𝑋𝑟𝑒 𝑘𝐹 + 𝑗𝑋𝑖𝑚 𝑘𝐹 = 𝑋 𝑘𝐹 𝑒 𝑗⋅∠𝑋 𝑘𝐹 By comparison: 𝐴𝑒 𝑖𝜑 𝑁 = 𝑋 𝑘𝐹 𝑒 𝑗⋅∠𝑋 2 𝑘𝐹 Hence, a cosine signal can be estimated as 𝐴 = 𝜑 = ∠𝑋 𝑘𝐹 = arg 𝑋 𝑘𝐹 2 𝑁 𝑋 𝑘𝐹 𝑋𝑖𝑚 𝑘𝐹 𝑎𝑟𝑐𝑡𝑔 , 𝑋𝑟𝑒 𝑘𝐹 = 𝑋𝑖𝑚 𝑘𝐹 𝑎𝑟𝑐𝑡𝑔 + 𝜋, 𝑋𝑟𝑒 𝑘𝐹 and 𝑖𝑓 𝑋𝑟𝑒 𝑘𝐹 ≥ 0 𝑖𝑓 𝑋𝑟𝑒 𝑘𝐹 < 0 78 Instead, a sine signal with phase 𝜑, is defined as By 𝜋 𝑥 𝑡 = A sin 2𝜋 𝑘𝐹 𝑡 + 𝜑 = A cos 2𝜋 𝑘𝐹 𝑡 + 𝜑 − 2 using the previous result on cosines, we obtain 𝜋 2 𝜑 − = arg 𝑋 𝑘𝐹 , hence 𝜑 = 𝑎𝑟𝑔 𝑋 𝑘𝐹 𝜋 + 2 When sampling is not synchronous, 𝑓 = 𝑘 + 𝛿 𝐹, −0.5 ≤ 𝛿 ≤ 0.5, but frequency is known, phase can be approximated as 𝜑 ≅ arg 𝑋 𝑘𝐹 + 𝜋δ, for a cosine signal 𝜑 ≅ arg 𝑋 𝑘𝐹 + + 𝜋δ, for a sine signal 𝜋 2 Please note that the term 𝜋δ may be significant in the above formulas but, unfortunately, it can not be calculated when frequency is not accurately known. 79 Gruppo Misure Elettriche ed Elettroniche Digital Spectral Analysis Summary ➢ At the heart of the digital spectrum analyzer is FFT processing; it is an algorithm for calculating the DFT that reduces the number of numerical operations without producing any loss of information. ➢ The FFT of a real signal is a complex function with even amplitude symmetry and odd phase symmetry; only half outputs carry independent information. ➢ The spectrum returned by the DFT will be exact only if the signal under consideration is periodic with limited band; in any other case, discontinuities will be introduced at the extremes and the signal will be periodized (with a period equal to the number of samples); this operation produces a leakage effect in the frequency domain. 80 Gruppo Misure Elettriche ed Elettroniche ➢ To reduce leakage, you can use appropriate windows to "smooth" the data record at the extremes, thus eliminating the abrupt truncation; there is no universal window but you will have to choose one according to the signal and quantity of interest (amplitude or frequency). 81 Gruppo Misure Elettriche ed Elettroniche Specifiche e considerazioni pratiche Un analizzatore di spettro è completamente caratterizzato quando si conoscono: • il suo range di frequenza utile • la sua risoluzione • la sua distorsione • il suo range dinamico • la sua incertezza (sia in frequenza che in ampiezza) • la sua sensibilità Prima di passare a esaminare queste quantità è bene fare qualche precisazione sulle unità di misura utilizzate da questo strumento. 82 Gruppo Misure Elettriche ed Elettroniche • assi orizzontale e verticale Il CRT dell’analizzatore di spettro è costituito da un reticolo costituito da 10 divisioni orizzontali e 8 (o 10) divisioni verticali. L’asse orizzontale è calibrato direttamente in frequenza [Hz] ed essa aumenta linearmente da sinistra a destra. L’asse verticale è calibrato in volt o volt2/hertz e c‘è la possibilità di utilizzare una scala lineare o logaritmica [dB]. Per segnali che si differenziano di poche decine di dB si può utilizzare la scala lineare (si tenga presente che 20-30 dB di differenza producono rapporti di tensione da 10 a circa 30), mentre per segnali che si differenziano di molte decine di dB si rende necessario l’uso della scala logaritmica (calibrata in dBV, dBmV, dBV, dBW, dBm, dBW). Infatti differenze di 70-100 dB corrispondono a rapporti di tensione da 3.000 a 100.000. 83 Gruppo Misure Elettriche ed Elettroniche Il dBm è definito come: 𝑃|𝑑𝐵𝑚 𝑃 = 10 ⋅ 𝑙𝑜𝑔 1 mW Il valore in volt corrispondente alla potenza espressa in dBm può essere calcolato considerando una tensione di riferimento 𝑉𝑟𝑒𝑓 : 𝑃|𝑑𝐵𝑚 𝑉𝑟𝑚𝑠 = 20 ⋅ 𝑙𝑜𝑔 𝑉𝑟𝑒𝑓 dove Vrms è il valore efficace della tensione misurata e Vref è la tensione misurata ai capi di una resistenza da 50 ohm (nel settore elettromagnetico) o 600 ohm (nel settore audio) quando essa dissipa una potenza di 1 milliwatt. 𝑉𝑟𝑒𝑓 = 𝑃 ⋅ 𝑅 = (1 ⋅ 10−3 ) ⋅ 50 = 0.2236 V 𝑉𝑟𝑒𝑓 = 𝑃 ⋅ 𝑅 = (1 ⋅ 10−3 ) ⋅ 600 = 0.7746 V Nel caso di misura espressa in dBmV, si usa il valore 𝑉𝑟𝑒𝑓 = 1 mV. 84 Gruppo Misure Elettriche ed Elettroniche • range di frequenza E' l'estensione della banda di frequenza nella quale lo strumento riesce ad eseguire misure con l‘incertezza specificata; per gli analizzatori di spettro analogici a super-eterodina il range di frequenza va da pochi kilohertz (limite inferiore di banda) fino a diverse decine di gigahertz (limite superiore). gli analizzatori di spettro digitali coprono un intervallo di frequenza che va dalla continua a qualche gigahertz. Di recente realizzazione sono i cosiddetti analizzatori di spettro ibridi che racchiudono i vantaggi dell’analizzatore analogico (bande ampie) e quelli dell’analizzatore digitale (elevate risoluzioni): ad es. modelli di fascia alta hanno banda di 1 GHz e limite superiore di frequenza di 50 GHz. 85 Gruppo Misure Elettriche ed Elettroniche • risoluzione in frequenza E' la capacità dello strumento di distinguere (risolvere) armoniche dello spettro del segnale in esame che sono distinte ma più o meno prossime tra loro. Analizzatore di spettro analogico: la risoluzione è in massima parte determinata dalla "forma" della risposta in frequenza del filtro IF, ossia dalla sua ampiezza di banda di risoluzione (RBW) e dalla sua selettività. Per il principio stesso di HIF ( f ) funzionamento di un 3dB analizzatore di spettro a 0 f super-eterodina, infatti, se si analizza un RBW risposta del filtro IF segnale costituito da X( f ) poche e distinte componenti, quello che appare sullo schermo 0 f spettro del segnale spettro visualizzato 86 Gruppo Misure Elettriche ed Elettroniche è la sovrapposizione di varie repliche della risposta in frequenza del filtro IF disposte a cavallo delle singole componenti. Se due componenti hanno ampiezza uguale o confrontabile, si verifica che la minima distanza in frequenza che le rende distinguibili utilizzando un filtro IF con una definita RBW è proprio f=RBW. X( f ) spettro del segnale di ingresso 0 f HIF ( f ) HIF ( f ) HIF ( f ) risposta filtro IF 0 f 0 f 0 f visualizzazione 87 Gruppo Misure Elettriche ed Elettroniche I problemi di risolvibilità di due armoniche sono però tanto più seri quanto più l'ampiezza di un'armonica è ridotta in confronto a quella dell'armonica ad essa più prossima. In questi casi, infatti, le armoniche più piccole possono letteralmente scomparire sotto le code laterali delle repliche della risposta del filtro IF dovute alle armoniche più grandi. Semplicisticamente si potrebbe pensare che per raggiungere una qualsivoglia risoluzione in frequenza sia sufficiente aumentare a piacere la selettività del filtro IF. Tuttavia è necessario tenere conto che il filtro IF è pur sempre un dispositivo dinamico, quindi che la sua uscita a un definito istante t0 dipende dagli stati che l'ingresso e l'uscita hanno assunto per t < t0. Questo fa si che l'uscita del filtro IF si porti a regime solo dopo un certo tempo dall'applicazione di un fissato ingresso; questa "memoria" del filtro IF (che può essere quantificata mediante il tempo di salita, detta rise time) è, in prima approssimazione, inversamente proporzionale alla RBW: 𝑇𝒓 ቚ 𝑛𝑠 ≅ 350 𝑅𝐵𝑊 |𝑀𝐻𝑧 88 Gruppo Misure Elettriche ed Elettroniche E' allora necessario commisurare opportunamente la velocità di sweep del VCO al valore della RBW utilizzata: tanto più piccola è la RBW utilizzata (quindi maggiore è la risoluzione dell'analisi) tanto più tempo ci vorrà per completare l'analisi in un fissato frequency span. Se non si aspetta abbastanza a lungo, l'analizzatore visualizzerà una versione non perfetta dello spettro del segnale; in questo caso infatti, il filtro non ha il tempo di rispondere completamente e la sua risposta sarà distorta: le ampiezze visualizzate saranno più basse e le frequenze più alte (ritardo del filtro). velocità di sweep corretta velocità di sweep eccessiva 89 Gruppo Misure Elettriche ed Elettroniche Il firmware degli analizzatori di spettro normalmente imposta automaticamente lo sweep time in funzione della RBW utilizzata ma lascia comunque libero l'operatore di agire manualmente su questo parametro. In ogni caso l'analisi eseguita in questo modo non è mai in "tempo reale". Analizzatore di spettro digitale: la risoluzione in frequenza di un analizzatore di spettro digitale coincide ovviamente con la risoluzione in frequenza della FFT da esso implementata: 𝐹= 𝑓𝑠 1 = 𝑁 𝑁𝑇𝒔 Un'attenta analisi di questa espressione mette in evidenza che la risoluzione può essere aumentata i) riducendo la frequenza di campionamento o ii) aumentando il numero di campioni da elaborare. Entrambe queste operazioni possono essere regolate in modo manuale sull’analizzatore e non sono naturalmente a costo zero. 90 Gruppo Misure Elettriche ed Elettroniche Infatti: • un aumento del numero di campioni N rallenta l’elaborazione FFT; • una riduzione della frequenza di campionamento 𝑓𝑠 può comportare la comparsa dell’aliasing. 91 Gruppo Misure Elettriche ed Elettroniche • distortion E' un indice dell'alterazione introdotta dallo strumento sullo spettro del segnale sotto esame. E' generata dalle nonlinearità dei circuiti che il segnale attraversa. Per l’analizzatore di spettro analogico lo stadio più critico a questo riguardo è il mixer, mentre nell’analizzatore di spettro digitale le non linearità sono introdotte dagli stadi di precondizionamento analogici e dall'ADC . Il modo più comune di dare le specifiche di distorsione è quello di quantificare le ampiezze della seconda e della terza armonica spurie introdotte nello spettro ed è normalmente data in dBc (decibel riferiti alla fondamentale). Per distinguere le armoniche (e i prodotti di distorsione) del segnale da quelle introdotte dallo strumento si può agire sui due stadi di attenuazione e guadagno. Incrementando ugualmente sia l’attenuazione che il guadagno (ad es. di 10 dB), la fondamentale e tutte le armoniche appartenenti al segnale in ingresso rimangono invariate, invece le armoniche introdotte dal mixer si riducono (questo perché l’ampiezza della seconda armonica prodotta dal mixer dipende dal quadrato dell’ampiezza della fondamentale, mentre quella della terza armonica dipende dal cubo). 92 Gruppo Misure Elettriche ed Elettroniche • Dynamic range (DR) Rappresenta l’intervallo di sensibilità dello strumento, ossia la differenza in dB tra la potenza della componente di segnale più grande e più piccola che lo strumento è in grado di rilevare contemporaneamente. Le piccole componenti di segnale sono mascherate dal tappeto di rumore e/o dalla distorsione introdotta dallo strumento, pertanto il DR è la differenza tra la componente della fondamentale ed il livello più grande tra noise floor e distorsione. Quando la fondamentale è a bassa potenza o è attenuata, il DR dipende dal noise Amplificando il floor. segnale, le distorsioni del secondo e terzo ordine introdotte dal mixer superano il noise floor. Quindi è possibile ottimizzare variando il guadagno. il DR 93 Gruppo Misure Elettriche ed Elettroniche Per l’analizzatore digitale di spettro il range dinamico è limitato dalle caratteristiche del convertitore A/D cui si aggiungono gli effetti negativi di rumore e non linearità dell’oscilloscopio (strumento all-in-one). Così ogni tecnica che sia in grado di ridurre il rumore di fondo permette di incrementare il range dinamico; ad esempio un incremento del numero di campioni per la FFT, ridurrà il noise floor che si spalmerà su un numero maggiore di uscite FFT (mantenendo costante la potenza di rumore). 94 Gruppo Misure Elettriche ed Elettroniche • incertezza di misura Così come per un oscilloscopio l’incertezza di misura è specificata per le misure di tempo e ampiezza, per un analizzatore di spettro l‘incertezza di misura deve essere specificata sia per le misure di frequenza sia per quelle di ampiezza; in ambedue i casi le specifiche sono date sia per misure assolute che per misure relative. Le prime sono quelle dirette, eseguite con un singolo cursore e sono, ad esempio, le misure di potenza di una portante RF, le misure di frequenza della stessa ecc. Le seconde sono quelle indirette, eseguite utilizzando due cursori contemporaneamente (misure di distorsione o di distanza di armoniche o subarmoniche dalla portante). Un’attenta analisi delle diverse componenti di incertezza può portare a una riduzione dell’incertezza complessiva dello strumento nelle misure relative (senza dover ricorrere ad analizzatori di spettro con bande di errore minore e quindi più costosi) perché in questo caso alcune componenti d’errore si elidono a vicenda. 95 Gruppo Misure Elettriche ed Elettroniche • sensibilità E' una misura della capacità dello strumento di rilevare piccoli segnali. Essa è normalmente specificata in dBm e può assumere valori anche < -140 dBm, ossia esistono strumenti in grado di rilevare segnali di potenza pari a 10-14 milliwatt! La sensibilità è in massima parte limitata dal rumore generato dalla circuiteria interna dell'analizzatore. 96 Gruppo Misure Elettriche ed Elettroniche Measurement of distortion and noise A common application of spectral analysis is the evaluation of performance in terms of distortion and noise. Ideal amplifiers, filters, signal generators, measuring instruments, sensors and transducers, ADCs and DACs (neglecting intrinsic quantization), have an output 𝑦 which is proportional to its input 𝑥: 𝑦 =𝑘∙𝑥 and, in frequency domain: 𝑌(𝑓) = 𝑘 ∙ 𝑋(𝑓) 97 Gruppo Misure Elettriche ed Elettroniche However real-life components have: 1) Offset 𝑂: 𝑦 =𝑘∙𝑥+𝑂 2) Linear distortion (for linear time-invariant dynamic systems): 𝑌 𝑓 =𝐻 𝑓 ∙𝑋 𝑓 𝑥 𝑡 = 𝐴𝑐𝑜𝑠 2𝜋𝑓𝑡 + 𝜙 → 𝑦 𝑡 = 𝐻 𝑓 𝐴 cos 2𝜋𝑓𝑡 + 𝜙 + ∡𝐻 𝑓 3) Harmonic distortion (which is a type of nonlinear distortion of nonlinear timeinvariant dynamic systems): +∞ 𝑥 𝑡 = 𝐴1 𝑐𝑜𝑠 2𝜋𝑓1 𝑡 + 𝜙1 → 𝑦 𝑡 = 𝐵𝑖 cos 2𝜋𝑚𝑓1 𝑡 + Φ𝑚 𝑚=0 Moreover, when the input contains two components at frequencies 𝑓1 and 𝑓2 , many components at frequencies 𝑙 ∙ 𝑓1 + 𝑚 ∙ 𝑓2 appear (intermodulation distortion), with 𝑙 and 𝑚 integers. 98 Gruppo Misure Elettriche ed Elettroniche 4) For nonlinear time-varying systems, there are different analysis formulae, according to the particular system and application. For an input sinusoid, the output can be an amplitude and phase modulated output sinusoid with modulated harmonics. 5) Interference: the output contains signals due to external disturbances, which are not related to the input. 6) Noise: the output contains an unpredictable component (random process) 𝑛(𝑡) which may be correlated or uncorrelated with the input. For example, for a noisy nonlinear time-invariant system: +∞ 𝑥 𝑡 = 𝐴1 𝑐𝑜𝑠 2𝜋𝑓1 𝑡 + 𝜙1 → 𝑦 𝑡 = 𝐵𝑖 cos 2𝜋𝑚𝑓1 𝑡 + Φ𝑚 + 𝑛(𝑡) 𝑚=0 99 Gruppo Misure Elettriche ed Elettroniche It is clear, then, that frequency analysis of a system output gives important information on its characteristics. In a typical characterization experiment, a system is stimulated by a pure tone 𝑥 𝑡 = 𝐴1 𝑐𝑜𝑠 2𝜋𝑓1 𝑡 + 𝜙1 and the output 𝑦 𝑡 is measured. We will assume that the system is nonlinear time-invariant and noisy, hence the following output is expected +∞ 𝑦 𝑡 = 𝐵𝑖 cos 2𝜋𝑚𝑓1 𝑡 + Φ𝑚 + 𝑛(𝑡) 𝑚=0 Frequency analysis allows one to estimate the power of each component. The fundamental, at frequency 𝑓1 (first harmonic), has power 𝑃1 = The second harmonic, at frequency 2𝑓1 , has power 𝑃2 = 𝐵2 , 2 𝐵1 . 2 and so on: 𝑃𝑚 = 𝐵𝑚 . 2 100 Gruppo Misure Elettriche ed Elettroniche Harmonics are considered up to a finite order 𝑀, which varies according to the application (from 5 to hundreds), and express the nonlinear distortion of the system. Their power is 𝑃𝐻 = 𝑃2 + 𝑃3 + ⋯ + 𝑃𝑀 The other components that differs from the 𝑀 harmonics (𝑓 ≠ 𝑓1 , 2𝑓1 , … , 𝑀𝑓1 ) constitute: - noise if they are unpredictable (e.g. thermal noise and, under some approximation, also quantization noise); - non-harmonic distortion, or spurious components, if they are related to the input or to the working of the system (e.g. a clock signal of a digital system); - interference, if they can be related to external sources which makes them distinguishable from random noise. 101 Gruppo Misure Elettriche ed Elettroniche The total power of the three previously described components will be referred, for simplicity, as noise power 𝑃𝑁 . The continuous component (frequency zero) is generally excluded from 𝑃𝑁 computation. Assuming that system output has been sampled synchronously (integer number of cycles 𝑳) at frequency 𝑓𝑠 in a window of 𝑁 samples and that a DFT (resolution 𝑓𝑠 /𝑁) has been performed, we have that the fundamental frequency coincides exactly with a DFT bin of index 𝐿 𝑓𝑠 𝑓1 = 𝐿 𝑁 and harmonics are at DFT indexes 2𝐿, 3𝐿, … , 𝑀𝐿. Hence 𝑃𝑁 is the sum of powers at DFT indexes different from 0, 𝐿, 2𝐿, … 𝑀𝐿. 102 Gruppo Misure Elettriche ed Elettroniche 𝑃1 , 𝑃𝐻 and 𝑃𝑁 are used to express the degree of ideality of the system. Signal-to-Noise ratio is defined as 𝑆𝑁𝑅 = 𝑆𝑁𝑅𝑑𝐵 𝑃1 𝑃𝑁 𝑃1 = 10 log10 = 𝑃1,𝑑𝐵 − 𝑃𝑁,𝑑𝐵 𝑃𝑁 Total Harmonic Distortion is defined as 𝑇𝐻𝐷 = 𝑃𝐻 /𝑃1 Noise and Distortion is defined as 𝑁𝐴𝐷 = 𝑃𝐻 + 𝑃𝑁 A good system has high SNR, low THD, low NAD. 103 Gruppo Misure Elettriche ed Elettroniche A synthetic figure of merit is the Signal to Noise and Distortion, which is found with two different definitions: Most common definition LabVIEW’s Distortion Express VI) 𝑆𝐼𝑁𝐴𝐷 = (used also in Definition used in standards for testing ADCs Measurements and DACs (IEEE STD 1241-2010 , IEEE STD 1658-2011) and in MATLAB’s sinad() function. 𝑃1 + 𝑁𝐴𝐷 𝑃1 + 𝑃𝐻 + 𝑃𝑁 = 𝑁𝐴𝐷 𝑃𝐻 + 𝑃𝑁 𝑆𝐼𝑁𝐴𝐷 = 𝑃1 𝑃1 = 𝑁𝐴𝐷 𝑃𝐻 + 𝑃𝑁 SINAD, NAD, THD and 𝑃1 can be related as follows 𝑆𝐼𝑁𝐴𝐷 = 1 𝑃𝐻 𝑃𝑁 + 𝑃1 𝑃1 +1= 𝑁𝐴𝐷 = 1 𝑇𝐻𝐷 + 𝑃1 𝑆𝐼𝑁𝐴𝐷 − 1 1 𝑆𝑁𝑅 +1 𝑆𝐼𝑁𝐴𝐷 = 1 𝑃𝐻 𝑃𝑁 + 𝑃1 𝑃1 𝑁𝐴𝐷 = 1 = 𝑇𝐻𝐷 + 1 𝑆𝑁𝑅 𝑃1 𝑆𝐼𝑁𝐴𝐷 104