Оценка ошибок квантования

Quantization errors in digital filters can be classified as:

•

Round-off errors derived from internal signals that are quantized before or after more down additions;

•

Deviations in the filter response due to finite word length representation of multiplier coefficients; and

•

Errors due to representation of the input signal with a set of discrete levels.

A general, digital filter structure with quantizers before delay elements can be represented as in Figure 2.23, with the quantizers implementing rounding for the granular quantization and saturation arithmetic for the overflow nonlinearity.

FIGURE 2.23. Digital Filter Including Quantizers at the Delay Inputs

The quantization error (e) or noise tends to have a random behavior, and they could be mathematically represented using statistical variables. Power of a random variable with a probability density function of p(e) could be obtained by computing the second-order statistics of variance, and it is denoted by

σ2=∫−q/2q/2e2p(e)de

A good assumption for p(e) is a uniform probability density function which will have a value of 1/q over the range of −q/2 to q/2.

=∫−q/2q/2e2·1qde=q212

The quantization error is present because the analogue value of each sample is transformed into a digital word. This A-to-D conversion entails a quantization of the recorder’s measuring range into a number of bands or code bins, each represented by its central value which corresponds to a particular digital code or level. The number of bands is given by 2^N, where N is the resolution of the A-to-D converter. The digital output to analogue input relationship of an ideal digitizer is shown diagrammatically in Fig. 3.49. For any input in the range (iΔV_av – 0.5 * ΔV_av to iΔV_av + 0.5 * ΔV_av), where iΔV_av is the voltage corresponding to the width of each code bin, or one least significant bit (LSB), and iΔV_av is the centre voltage corresponding to the i th code, an ideal digitizer will return a value of I_i. Therefore, the response of an ideal digitizer to a slowly increasing linear ramp would be a stairway such as that shown in Fig. 3.50. A quick study of these figures reveals the character of the quantization error associated with the ideal A-to-D conversion process. The maximum error possible is equivalent to a voltage corresponding to ±(½) of an LSB. For an ideal digital recorder, this quantization would be the only source of error in the recorded samples. For a real digital recorder, this error sets the absolute upper limit on the accuracy of the readings. In the case of an 8-bit machine, this upper limit would be 0.39 per cent of the recorder’s full-scale deflection. The corresponding maximum accuracy (lowest uncertainty) of a 10-bit recorder is 0.10 per cent of its full-scale deflection.

Figure 3.49. Analogue input to digital output relation of an ideal A/D converter

Figure 3.50. Response of an ideal A/D converter to a slowly rising ramp

The error caused by discrete time sampling is most easily demonstrated with reference to the recording of sinusoidal signals. As an example we can look at the discrete time sampling error introduced in the measurement of a single cycle of a pure sine wave of frequency f, which is sampled at a rate of four times its frequency. When the sinusoid and the sampling clock are in phase, as shown in Fig. 3.51, a sample will fall on the peak value of both positive and negative half-cycles. The next closest samples will lie at π/2 radians from the peaks. As the phase of the clock is advanced relative to the input sinusoid the sample points which used to lie at the peak values will move to lower amplitude values giving an error (Δ) in the measurement of the amplitude (A) of

Figure 3.51. Sample points with sinusoid and sampling clock in phase. (Error in peak amplitude = 0)

(3.93)Δmax=I−cos(πfts)

where t_s is the recorder’s sampling interval and f the sinewave frequency.

The maximum errors obtained through quantization and sampling when recording a sinusoidal waveform are shown in Fig. 3.53. The plotted quantities were calculated for an 8-bit 200-MHz digitizer.

Figure 3.53. Sampling and quantization errors of an ideal recorder

The constructed indicator WMQE is further input into the RUL prediction module. In this module, a PF-based prediction algorithm is utilized to predict RUL of the rotating machinery whose degradation processes are described using a variant of Paris–Erdogan model. The Paris–Erdogan model is formulated as

(6.96)dxdt=c(Δδ)γ, Δδ=mx

where x represents the semicrack length, t is the number of stress cycles (i.e., the fatigue life), c, γ, and m are material constants which are determined by tests, and Δδ is amplitude of stress intensity factor roughly proportional to the square root of x.

It is seen from Eq. (6.96) that there are several model parameters in the Paris–Erdogan model, that is, c, γ, and m, which are difficult to measure during the operation process of the rotating machinery. For convenient application, the Paris–Erdogan model is transformed into the following format with α=cmγ and β = γ/2.

(6.97)dxdt=αxβ.

Then, the above function is rewritten into the following state space model.

(6.98)xk=xk−1+αk−1xk−1βΔtkαk=αk−1zk=xk+νk,

where α_k−1 is a random variable following a normal distribution of Nμα,σα2, β is a constant parameter, Δtk=tk−tk−1, z_k is the measured WMQE value at t_k and ν_k is the measurement noise following the normal distribution of N0,σν2. With the transformation of the Paris–Erdogan model, the model parameters are more convenient to estimate according to the measurements. In addition, the state space model inherits the superiority of the Paris–Erdogan model in describing the general degradation processes. Therefore, it is supposed to be a good model for a general degradation process.

After the transformation, the unknown model parameters are changed to be Θ=μα,σα2,β,σν2′, where (·)′ denotes the vector transposition. Then, the measured WMQE values constructed from vibration signals are input into the model, and the model parameters are initialized using MLE. It is assumed that there are a series of measurements z0:M=z0,…,zM′ at ordered times t₀,…,t_M. According to Eq. (6.98), z_k is formulated as follows:

(6.99)zk=xk−1+αkxk−1βΔtk+νk.

The degradation state x_k−1 has the following relationship with the measurement z_k−1.

(6.100)xk−1=zk−1−νk−1.

The degradation state x_k−1 is hard to be acquired in real applications. If the measurement noise ν_k−1 is small enough compared with the measurement itself, it is negligible and x_k−1 is approximated by z_k−1. Let T=z0βΔt1,…,zM−1βΔtM′. z1:M=z1,…,zM′ is multivariate normally distributed, which is denoted as follows:

(6.101)z1:M∼Nz0:M−1+μαT,σα2TT′+σν2IM,

where I_M is an identity matrix of order M.

Let Δz1:M=z1−z0,…,zM−zM−1′, and the log-likelihood function of the unknown parameters based on the measurements is expressed as

(6.102)ℓΘ|z0:M=−M2ln2π−12lnσα2TT′+σν2IM     −12Δz1:M−μαT′σα2TT′+σν2IM−1Δz1:M−μαT     =−M2ln2π−M2lnσα2−12lnTT′+σ˜ν2IM     −12σα2Δz1:M−μαT′σα2TT′+σν2IM−1Δz1:M−μαT,

with σ~ν2=σν2/σα2. The first partial derivatives of ℓΘ|z0:M with respect to μ_α and σα2 are calculated and formulated with

(6.103)∂ℓΘ|z0:M∂μα=1σα2T′TT′+σ~ν2IM-1Δz1:M−μαT,

(6.104)∂ℓΘ|z0:M∂σα2=−M2σα2+12σα4Δz1:M−μαT′TT′+σ~ν2IM−1Δz1:M−μαT.

Let ∂ℓΘ|z0:M/∂μα=0 and ∂ℓΘ|z0:M/∂σα2=0. The MLE results of μ_α and σα2 are

(6.105)μα=T′TT′+σ~ν2IM−1Δz1:MT′TT′+σ~ν2IM−1T,

(6.106)σα2=Δz1:M−μαT′TT′+σ~ν2IM−1Δz1:M−μαTM.

With Eqs. (6.105) and (6.106) substituted into Eq. (6.102), the log-likelihood function is reduced into a two-variable function about β and σ~ν2, which is denoted by

(6.107)ℓΘ|z0:M=−M2ln2π−M2lnσα2−12lnTT′+σ~ν2IM−M2.

After parameter initialization, the model parameters are further updated and the RUL is predicted using a PF-based prediction algorithm. Based on the initialized parameters, a series of initial particles y0nn=1:Ns are sampled from the initial PDF of the system state p(y0n|Θ0)∼N(y0,Q0) with

(6.108)y0=x0μα and Q0=000σα2.

Ns is the number of particles and the weight of each particle is set to be 1/Ns. Then new particles ykni=1:Ns are obtained following

(6.109)ykn=xknμαn=xk−1n+μαnxk−1nβΔtkμαn.

When the new measurement z_k at t_k is available, each particle weight is updated and normalized by

(6.110)wkn=wk−1npzk|ykn, w~kn=wkn∑n=1Nswkn,

where

(6.111)pzk|ykn=12πσνexp−12zk−xknσν2.

The particles are resampled according to the particle weights and their weights are reset to be 1/Ns. After that, the RUL is predicted based on the resampled particles. The RUL l_k at t_k is defined as

(6.112)lk=inflk:xlk+tk≥λ|x1:k,

where λ is a prespecified failure threshold. Each particle is transmitted following the transition function of Eq. (6.98) from current state until the state value exceeds the failure threshold, and the RUL lknn=1:Ns predicted using each particle is acquired. Then the PDF of the RUL is approximated by

(6.113)plk|z0:k=∑n=1Nsw~knδlk−lkn.

A typical model of the random error d˜ in Eq. (8.4), which includes quantization errors, is the linear continuous-time stochastic state equation

(8.7)x˜˙(t)=A˜x˜+G˜w˜d˜(t)=C˜x˜+D˜w˜E{w˜(t)}=0,E{w˜(t)w˜T(t+τ)}=S˜w2δ(τ)E{x˜(0)}=x˜0,E{(x˜(0)−x˜0)(x˜(0)−x˜0)T}=P˜0≥0E{(x˜(0)−x˜0)w˜T(0)}=0,

which is similar to the DT Eq. (4.159) of Section 4.8.1. Eq. (8.7) being continuous-time, the eigenvalues of the state matrix A˜ are assumed to lie on the imaginary axis and when equal to zero may be multiple. The statistics in Eq. (8.7) assumes that w˜ is a zero-mean second-order stationary white noise with constant spectral density S˜w2, and impulsive covariance, where δ(τ) denotes a Dirac delta (see Sections 13.2.1 and 13.7.3Section 13.2.1Section 13.7.3). The initial state may be modeled as a random vector with mean value x˜0 and covariance matrix P˜0, but is uncorrelated from any simultaneous white noise as expressed by the last identity in Eq. (8.7). This uncorrelation has been already referred to as the causality constraint. In principle, Eq. (8.7) may be unobservable from the output d˜ and uncontrollable by the noise w˜, because the output d˜ may include polynomial and trigonometric components (deterministic signals) just driven by the initial state x˜0. For instance, a trigonometric component tuned to the angular frequency ω˜ corresponds to a second-order subsystem with eigenvalues ±jω˜. A first-order polynomial corresponds to a second-order subsystem with a pair of zero eigenvalues and a single eigenvector. The mixed case of stochastic processes and deterministic signals can be simplified by assuming that trigonometric and polynomial components are the free response of the equations driven by w˜, and that Eq. (8.7) is observable and controllable.

The simplest model of the class in Eq. (8.7), which is common to inertial sensors (accelerometers in Section 8.4 and gyroscopes in Section 8.5), is the scalar first-order random drift [32]:

(8.8)x˜˙(t)=w˜x,x˜(0)=x˜0d˜(t)=x˜+w˜dE{x˜(0)}=x˜0,var{x˜}=σ02,E{(x˜−x˜0)w˜T(0)}=0w˜=[w˜xw˜d],E{w˜(t)}=0,E{w˜(t)w˜T(t+τ)}=[S˜wx200S˜wd2]δ(τ),

where, if [unit] denotes the unit of measurement of x˜, we find S˜wx2 in [(unit/s)2Hz−1] and S˜wd2 in [unit2Hz−1]. The initial state x˜0 accounts for a constant bias and is uncorrelated with any simultaneous noise; the scalar input noise w˜x and the output noise w˜d in Eq. (8.8) are uncorrelated with each other. The output process d˜ is nonstationary and the autocorrelation is given by

(8.9)R˜d(t,t+τ)=S˜wx2min(t,t+τ)+S˜wd2δ(τ).

Although x˜ is nonstationary, the spectral density S˜x2(f) can be defined through the AS equation x˜˙=−εx˜+w˜x, where ε > 0 must be sufficiently smaller than the cutoff frequency f₀ to be defined in the next paragraph. We can write the following identities:

(8.10)S˜d2(f)=S˜x2(f)+S˜wd2=S˜wx2(2πf)2+S˜wd2,f>ε2π=fεσ˜x2=limf→∞∫0fS˜x2νdν<∞,

where if, for f < ε, S˜x2(f) is bounded, also the variance σ˜x2 is bounded. The Bode plot of S˜d(f) is approximately flat for f>f0=2π−1S˜wx/S˜wd>fε and has a −20 dB/decade slope for f_ε < f < f₀. The first PSD in the first row of Eq. (8.10) is a first-order random walk, which is known, in the realm of inertial sensors, as the (long-term) bias instability of the sensor. The name is appropriate because it describes the long-term fluctuations—bounded because of ε—around the mean sensor bias x˜0. The square root of the second term S˜wd corresponds to the minimum-valued profile of the overall spectral density. Let us call it, as already anticipated, noise floor, although the name sometimes refers to the whole instrument noise (here referred to as the background noise). In the realm of inertial sensors, it is known as the velocity random walk in the case of linear accelerometers, the unit being [m/(s2Hz)], and the angular random walk (ARW) in the case of gyroscopes, the unit being[rad/(sHz)]. The name, which may cause same confusion, is justified by the fact that when either of the two measurements (linear acceleration and angular rate) is time integrated for generating either velocity or attitude measurements, the integrated noise floor becomes a random walk. By restricting to gyroscopes, the ARW unit [rad/(sHz)] is usually simplified to [rad/s], de facto to the non-SI unit [degree/hour]. Indeed, the simplified unit is at the same time the unit of S˜wd and the unit of the root mean square (RMS) σ¯w(t,Δt) of the random walk increment w¯d(t,Δt)=∫tt+Δtw˜d(τ)dτ, namely:

(8.11)σ¯w(t,Δt)=E{1Δt(∫tt+Δtw˜d(τ)dτ)2}=S˜wd.

(8.23)Δ=dynamic range of signal2b=2max|x(t)|22

where b = 2 is number of bits of the code assigned to each level. The bits assigned to each of the levels uniquely represents the different levels [-2Δ,-Δ,0,Δ]. As to how to approximate the given sample to one of these levels, it can be done by rounding or by truncating. The quantizer shown in Figure 8.12 approximates by truncation, i.e., if the sample kΔ≤x(nTs)<(k+1)Δ, for k = −2, −1,0,1, then it is approximated by the level kΔ.

To see how quantization and coding are done, and how to obtain the quantization error, let the sampled signal be

x(nTs)=x(t)|t=nTS

The given four-level quantizer is such that if the sample x(nT_s) is such that

(8.24)kΔ≤x(nTs)<(k+1)Δ⇒xˆ(nTs)=kΔk=-2,-1,0,1

The sampled signal x(nT_s) is the input of the quantizer and the quantized signal xˆ(nTs) is its output. So that whenever

-2Δ≤x(nTs)<-Δ⇒xˆ(nTs)=-2Δ-Δ≤x(nTs)<0⇒xˆ(nTs)=-Δ0≤x(nTs)<Δ⇒xˆ(nTs)=0Δ≤x(nTs)<2Δ⇒xˆ(nTs)=Δ

To transform the quantized values into unique binary 2-bit values, one could use a code such as

xˆ(nTs)⇒binary code-2Δ10-Δ110Δ00Δ01

which assigns a unique 2 bit binary number to each of the 4 quantization levels. Notice that the first bit of this code can be considered a sign bit, “1” for negative levels and “0” for positive levels.

If we define the quantization error as

ε(nTs)=x(nTs)-xˆ(nTs)

and use the characterization of the quantizer given by Equation (8.24) as

xˆ(nTs)≤x(nTs)≤xˆ(nTs)+Δ

by subtracting xˆ(nTs) from each of the terms gives that the quantization error is bounded as follows

(8.25)0≤ε(nTs)≤Δ

i.e., the quantization error for the four-level quantizer being considered is between 0 and Δ. This expression for the quantization error indicates that one way to decrease the quantization error is to make the quantization step Δsmaller. Increasing the number of bits of the A/D converter makes Δ smaller (see Equation (8.23) where the denominator is 2 raised to the number of bits) which in turn makes smaller the quantization error, and improves the quality of the A/D converter.

In practice, the quantization error is random and so it needs to be characterized probabilistically. This characterization becomes meaningful when the number of bits is large, and when the input signal is not a deterministic signal. Otherwise, the error is predictable and thus not random. Comparing the energy of the input signal to the energy of the error, by means of the so-called signal to noise ratio (SNR), it is possible to determine the number of bits that are needed in a quantizer to get a reasonable quantization error.

Estimates of bounds on the quantization error in Subsection 4.3 are derived from the next two lemmas [5].

Figure 12.15 compares a classical ring oscillator structure to one that utilizes a multi-path topology. The key idea behind the multi-path topology is that one can use the transition information of previous delay stages to predict when a current delay stage should begin its transition and thereby reduce the effective delay time through each delay stage. Using this technique, a significant increase in operating frequency has been demonstrated for ring oscillators [7,8], and a multi-path GRO implementation has been presented which achieves roughly a five times reduction of effective gate delay [9].

Figure 12.15. Reduction of effective delay per stage using multi-path oscillator implementation.

Reduction of gate delay time improves TDC resolution, but carries the cost of higher operating frequencies and therefore higher power. Utilizing counters at each delay stage as shown in Figure 12.13 is particularly problematic for power dissipation when such high operating frequencies are involved. Given this issue, it is worthwhile considering alternative methods of recording transition information within the oscillator such as the phase sampling method shown in Figure 12.16. In this approach, each delay stage is connected to a simple register rather than a counter. These registers sample the transition states of the ring oscillator at the conclusion of each measurement, which yields the quantized phase state of the oscillator. One of the delay stages is also connected to a counter which keeps track of phase wrapping in the oscillator. By combining the phase state and wrapping information from the counter, the number of transitions of the oscillator within a given measurement window can be computed at much lower power dissipation than with the parallel counter method.

Figure 12.16. Phase sampling of GRO TDC state for reduced power dissipation.

Using the above techniques, a prototype GRO TDC was developed as shown in Figure 12.17 [9]. Here a 47-stage multi-path oscillator structure is utilized which achieves 6 ps delay per stage in 130 nm CMOS. The phase sampling operation is performed by seven interleaved channels, the total range input range is 12 ns, total active area is 0.04 mm², and power consumption varies from 2.2 to 21 mW depending on the input time between Start(t) and Stop(t) signals. Additional performance results for this structure are shown in the following sub-section.

Figure 12.17. Example GRO implementation.

Quantization errors in digital filters can be classified as:

•

Round-off errors derived from internal signals that are quantized before or after more down additions;

•

Deviations in the filter response due to finite word length representation of multiplier coefficients; and

•

Errors due to representation of the input signal with a set of discrete levels.

A general, digital filter structure with quantizers before delay elements can be represented as in Figure 2.23, with the quantizers implementing rounding for the granular quantization and saturation arithmetic for the overflow nonlinearity.

FIGURE 2.23. Digital Filter Including Quantizers at the Delay Inputs

The quantization error (e) or noise tends to have a random behavior, and they could be mathematically represented using statistical variables. Power of a random variable with a probability density function of p(e) could be obtained by computing the second-order statistics of variance, and it is denoted by

σ2=∫−q/2q/2e2p(e)de

A good assumption for p(e) is a uniform probability density function which will have a value of 1/q over the range of −q/2 to q/2.

=∫−q/2q/2e2·1qde=q212

The quantization error is present because the analogue value of each sample is transformed into a digital word. This A-to-D conversion entails a quantization of the recorder’s measuring range into a number of bands or code bins, each represented by its central value which corresponds to a particular digital code or level. The number of bands is given by 2^N, where N is the resolution of the A-to-D converter. The digital output to analogue input relationship of an ideal digitizer is shown diagrammatically in Fig. 3.49. For any input in the range (iΔV_av – 0.5 * ΔV_av to iΔV_av + 0.5 * ΔV_av), where iΔV_av is the voltage corresponding to the width of each code bin, or one least significant bit (LSB), and iΔV_av is the centre voltage corresponding to the i th code, an ideal digitizer will return a value of I_i. Therefore, the response of an ideal digitizer to a slowly increasing linear ramp would be a stairway such as that shown in Fig. 3.50. A quick study of these figures reveals the character of the quantization error associated with the ideal A-to-D conversion process. The maximum error possible is equivalent to a voltage corresponding to ±(½) of an LSB. For an ideal digital recorder, this quantization would be the only source of error in the recorded samples. For a real digital recorder, this error sets the absolute upper limit on the accuracy of the readings. In the case of an 8-bit machine, this upper limit would be 0.39 per cent of the recorder’s full-scale deflection. The corresponding maximum accuracy (lowest uncertainty) of a 10-bit recorder is 0.10 per cent of its full-scale deflection.

Figure 3.49. Analogue input to digital output relation of an ideal A/D converter

Figure 3.50. Response of an ideal A/D converter to a slowly rising ramp

The error caused by discrete time sampling is most easily demonstrated with reference to the recording of sinusoidal signals. As an example we can look at the discrete time sampling error introduced in the measurement of a single cycle of a pure sine wave of frequency f, which is sampled at a rate of four times its frequency. When the sinusoid and the sampling clock are in phase, as shown in Fig. 3.51, a sample will fall on the peak value of both positive and negative half-cycles. The next closest samples will lie at π/2 radians from the peaks. As the phase of the clock is advanced relative to the input sinusoid the sample points which used to lie at the peak values will move to lower amplitude values giving an error (Δ) in the measurement of the amplitude (A) of

Figure 3.51. Sample points with sinusoid and sampling clock in phase. (Error in peak amplitude = 0)

(3.93)Δmax=I−cos(πfts)

where t_s is the recorder’s sampling interval and f the sinewave frequency.

The maximum errors obtained through quantization and sampling when recording a sinusoidal waveform are shown in Fig. 3.53. The plotted quantities were calculated for an 8-bit 200-MHz digitizer.

Figure 3.53. Sampling and quantization errors of an ideal recorder

The constructed indicator WMQE is further input into the RUL prediction module. In this module, a PF-based prediction algorithm is utilized to predict RUL of the rotating machinery whose degradation processes are described using a variant of Paris–Erdogan model. The Paris–Erdogan model is formulated as

(6.96)dxdt=c(Δδ)γ, Δδ=mx

where x represents the semicrack length, t is the number of stress cycles (i.e., the fatigue life), c, γ, and m are material constants which are determined by tests, and Δδ is amplitude of stress intensity factor roughly proportional to the square root of x.

It is seen from Eq. (6.96) that there are several model parameters in the Paris–Erdogan model, that is, c, γ, and m, which are difficult to measure during the operation process of the rotating machinery. For convenient application, the Paris–Erdogan model is transformed into the following format with α=cmγ and β = γ/2.

(6.97)dxdt=αxβ.

Then, the above function is rewritten into the following state space model.

(6.98)xk=xk−1+αk−1xk−1βΔtkαk=αk−1zk=xk+νk,

where α_k−1 is a random variable following a normal distribution of Nμα,σα2, β is a constant parameter, Δtk=tk−tk−1, z_k is the measured WMQE value at t_k and ν_k is the measurement noise following the normal distribution of N0,σν2. With the transformation of the Paris–Erdogan model, the model parameters are more convenient to estimate according to the measurements. In addition, the state space model inherits the superiority of the Paris–Erdogan model in describing the general degradation processes. Therefore, it is supposed to be a good model for a general degradation process.

After the transformation, the unknown model parameters are changed to be Θ=μα,σα2,β,σν2′, where (·)′ denotes the vector transposition. Then, the measured WMQE values constructed from vibration signals are input into the model, and the model parameters are initialized using MLE. It is assumed that there are a series of measurements z0:M=z0,…,zM′ at ordered times t₀,…,t_M. According to Eq. (6.98), z_k is formulated as follows:

(6.99)zk=xk−1+αkxk−1βΔtk+νk.

The degradation state x_k−1 has the following relationship with the measurement z_k−1.

(6.100)xk−1=zk−1−νk−1.

The degradation state x_k−1 is hard to be acquired in real applications. If the measurement noise ν_k−1 is small enough compared with the measurement itself, it is negligible and x_k−1 is approximated by z_k−1. Let T=z0βΔt1,…,zM−1βΔtM′. z1:M=z1,…,zM′ is multivariate normally distributed, which is denoted as follows:

(6.101)z1:M∼Nz0:M−1+μαT,σα2TT′+σν2IM,

where I_M is an identity matrix of order M.

Let Δz1:M=z1−z0,…,zM−zM−1′, and the log-likelihood function of the unknown parameters based on the measurements is expressed as

(6.102)ℓΘ|z0:M=−M2ln2π−12lnσα2TT′+σν2IM     −12Δz1:M−μαT′σα2TT′+σν2IM−1Δz1:M−μαT     =−M2ln2π−M2lnσα2−12lnTT′+σ˜ν2IM     −12σα2Δz1:M−μαT′σα2TT′+σν2IM−1Δz1:M−μαT,

with σ~ν2=σν2/σα2. The first partial derivatives of ℓΘ|z0:M with respect to μ_α and σα2 are calculated and formulated with

(6.103)∂ℓΘ|z0:M∂μα=1σα2T′TT′+σ~ν2IM-1Δz1:M−μαT,

(6.104)∂ℓΘ|z0:M∂σα2=−M2σα2+12σα4Δz1:M−μαT′TT′+σ~ν2IM−1Δz1:M−μαT.

Let ∂ℓΘ|z0:M/∂μα=0 and ∂ℓΘ|z0:M/∂σα2=0. The MLE results of μ_α and σα2 are

(6.105)μα=T′TT′+σ~ν2IM−1Δz1:MT′TT′+σ~ν2IM−1T,

(6.106)σα2=Δz1:M−μαT′TT′+σ~ν2IM−1Δz1:M−μαTM.

With Eqs. (6.105) and (6.106) substituted into Eq. (6.102), the log-likelihood function is reduced into a two-variable function about β and σ~ν2, which is denoted by

(6.107)ℓΘ|z0:M=−M2ln2π−M2lnσα2−12lnTT′+σ~ν2IM−M2.

After parameter initialization, the model parameters are further updated and the RUL is predicted using a PF-based prediction algorithm. Based on the initialized parameters, a series of initial particles y0nn=1:Ns are sampled from the initial PDF of the system state p(y0n|Θ0)∼N(y0,Q0) with

(6.108)y0=x0μα and Q0=000σα2.

Ns is the number of particles and the weight of each particle is set to be 1/Ns. Then new particles ykni=1:Ns are obtained following

(6.109)ykn=xknμαn=xk−1n+μαnxk−1nβΔtkμαn.

When the new measurement z_k at t_k is available, each particle weight is updated and normalized by

(6.110)wkn=wk−1npzk|ykn, w~kn=wkn∑n=1Nswkn,

where

(6.111)pzk|ykn=12πσνexp−12zk−xknσν2.

The particles are resampled according to the particle weights and their weights are reset to be 1/Ns. After that, the RUL is predicted based on the resampled particles. The RUL l_k at t_k is defined as

(6.112)lk=inflk:xlk+tk≥λ|x1:k,

where λ is a prespecified failure threshold. Each particle is transmitted following the transition function of Eq. (6.98) from current state until the state value exceeds the failure threshold, and the RUL lknn=1:Ns predicted using each particle is acquired. Then the PDF of the RUL is approximated by

(6.113)plk|z0:k=∑n=1Nsw~knδlk−lkn.

A typical model of the random error d˜ in Eq. (8.4), which includes quantization errors, is the linear continuous-time stochastic state equation

(8.7)x˜˙(t)=A˜x˜+G˜w˜d˜(t)=C˜x˜+D˜w˜E{w˜(t)}=0,E{w˜(t)w˜T(t+τ)}=S˜w2δ(τ)E{x˜(0)}=x˜0,E{(x˜(0)−x˜0)(x˜(0)−x˜0)T}=P˜0≥0E{(x˜(0)−x˜0)w˜T(0)}=0,

which is similar to the DT Eq. (4.159) of Section 4.8.1. Eq. (8.7) being continuous-time, the eigenvalues of the state matrix A˜ are assumed to lie on the imaginary axis and when equal to zero may be multiple. The statistics in Eq. (8.7) assumes that w˜ is a zero-mean second-order stationary white noise with constant spectral density S˜w2, and impulsive covariance, where δ(τ) denotes a Dirac delta (see Sections 13.2.1 and 13.7.3Section 13.2.1Section 13.7.3). The initial state may be modeled as a random vector with mean value x˜0 and covariance matrix P˜0, but is uncorrelated from any simultaneous white noise as expressed by the last identity in Eq. (8.7). This uncorrelation has been already referred to as the causality constraint. In principle, Eq. (8.7) may be unobservable from the output d˜ and uncontrollable by the noise w˜, because the output d˜ may include polynomial and trigonometric components (deterministic signals) just driven by the initial state x˜0. For instance, a trigonometric component tuned to the angular frequency ω˜ corresponds to a second-order subsystem with eigenvalues ±jω˜. A first-order polynomial corresponds to a second-order subsystem with a pair of zero eigenvalues and a single eigenvector. The mixed case of stochastic processes and deterministic signals can be simplified by assuming that trigonometric and polynomial components are the free response of the equations driven by w˜, and that Eq. (8.7) is observable and controllable.

The simplest model of the class in Eq. (8.7), which is common to inertial sensors (accelerometers in Section 8.4 and gyroscopes in Section 8.5), is the scalar first-order random drift [32]:

(8.8)x˜˙(t)=w˜x,x˜(0)=x˜0d˜(t)=x˜+w˜dE{x˜(0)}=x˜0,var{x˜}=σ02,E{(x˜−x˜0)w˜T(0)}=0w˜=[w˜xw˜d],E{w˜(t)}=0,E{w˜(t)w˜T(t+τ)}=[S˜wx200S˜wd2]δ(τ),

where, if [unit] denotes the unit of measurement of x˜, we find S˜wx2 in [(unit/s)2Hz−1] and S˜wd2 in [unit2Hz−1]. The initial state x˜0 accounts for a constant bias and is uncorrelated with any simultaneous noise; the scalar input noise w˜x and the output noise w˜d in Eq. (8.8) are uncorrelated with each other. The output process d˜ is nonstationary and the autocorrelation is given by

(8.9)R˜d(t,t+τ)=S˜wx2min(t,t+τ)+S˜wd2δ(τ).

Although x˜ is nonstationary, the spectral density S˜x2(f) can be defined through the AS equation x˜˙=−εx˜+w˜x, where ε > 0 must be sufficiently smaller than the cutoff frequency f₀ to be defined in the next paragraph. We can write the following identities:

(8.10)S˜d2(f)=S˜x2(f)+S˜wd2=S˜wx2(2πf)2+S˜wd2,f>ε2π=fεσ˜x2=limf→∞∫0fS˜x2νdν<∞,

where if, for f < ε, S˜x2(f) is bounded, also the variance σ˜x2 is bounded. The Bode plot of S˜d(f) is approximately flat for f>f0=2π−1S˜wx/S˜wd>fε and has a −20 dB/decade slope for f_ε < f < f₀. The first PSD in the first row of Eq. (8.10) is a first-order random walk, which is known, in the realm of inertial sensors, as the (long-term) bias instability of the sensor. The name is appropriate because it describes the long-term fluctuations—bounded because of ε—around the mean sensor bias x˜0. The square root of the second term S˜wd corresponds to the minimum-valued profile of the overall spectral density. Let us call it, as already anticipated, noise floor, although the name sometimes refers to the whole instrument noise (here referred to as the background noise). In the realm of inertial sensors, it is known as the velocity random walk in the case of linear accelerometers, the unit being [m/(s2Hz)], and the angular random walk (ARW) in the case of gyroscopes, the unit being[rad/(sHz)]. The name, which may cause same confusion, is justified by the fact that when either of the two measurements (linear acceleration and angular rate) is time integrated for generating either velocity or attitude measurements, the integrated noise floor becomes a random walk. By restricting to gyroscopes, the ARW unit [rad/(sHz)] is usually simplified to [rad/s], de facto to the non-SI unit [degree/hour]. Indeed, the simplified unit is at the same time the unit of S˜wd and the unit of the root mean square (RMS) σ¯w(t,Δt) of the random walk increment w¯d(t,Δt)=∫tt+Δtw˜d(τ)dτ, namely:

(8.11)σ¯w(t,Δt)=E{1Δt(∫tt+Δtw˜d(τ)dτ)2}=S˜wd.

(8.23)Δ=dynamic range of signal2b=2max|x(t)|22

where b = 2 is number of bits of the code assigned to each level. The bits assigned to each of the levels uniquely represents the different levels [-2Δ,-Δ,0,Δ]. As to how to approximate the given sample to one of these levels, it can be done by rounding or by truncating. The quantizer shown in Figure 8.12 approximates by truncation, i.e., if the sample kΔ≤x(nTs)<(k+1)Δ, for k = −2, −1,0,1, then it is approximated by the level kΔ.

To see how quantization and coding are done, and how to obtain the quantization error, let the sampled signal be

x(nTs)=x(t)|t=nTS

The given four-level quantizer is such that if the sample x(nT_s) is such that

(8.24)kΔ≤x(nTs)<(k+1)Δ⇒xˆ(nTs)=kΔk=-2,-1,0,1

The sampled signal x(nT_s) is the input of the quantizer and the quantized signal xˆ(nTs) is its output. So that whenever

-2Δ≤x(nTs)<-Δ⇒xˆ(nTs)=-2Δ-Δ≤x(nTs)<0⇒xˆ(nTs)=-Δ0≤x(nTs)<Δ⇒xˆ(nTs)=0Δ≤x(nTs)<2Δ⇒xˆ(nTs)=Δ

To transform the quantized values into unique binary 2-bit values, one could use a code such as

xˆ(nTs)⇒binary code-2Δ10-Δ110Δ00Δ01

which assigns a unique 2 bit binary number to each of the 4 quantization levels. Notice that the first bit of this code can be considered a sign bit, “1” for negative levels and “0” for positive levels.

If we define the quantization error as

ε(nTs)=x(nTs)-xˆ(nTs)

and use the characterization of the quantizer given by Equation (8.24) as

xˆ(nTs)≤x(nTs)≤xˆ(nTs)+Δ

by subtracting xˆ(nTs) from each of the terms gives that the quantization error is bounded as follows

(8.25)0≤ε(nTs)≤Δ

i.e., the quantization error for the four-level quantizer being considered is between 0 and Δ. This expression for the quantization error indicates that one way to decrease the quantization error is to make the quantization step Δsmaller. Increasing the number of bits of the A/D converter makes Δ smaller (see Equation (8.23) where the denominator is 2 raised to the number of bits) which in turn makes smaller the quantization error, and improves the quality of the A/D converter.

In practice, the quantization error is random and so it needs to be characterized probabilistically. This characterization becomes meaningful when the number of bits is large, and when the input signal is not a deterministic signal. Otherwise, the error is predictable and thus not random. Comparing the energy of the input signal to the energy of the error, by means of the so-called signal to noise ratio (SNR), it is possible to determine the number of bits that are needed in a quantizer to get a reasonable quantization error.

Estimates of bounds on the quantization error in Subsection 4.3 are derived from the next two lemmas [5].

Figure 12.15 compares a classical ring oscillator structure to one that utilizes a multi-path topology. The key idea behind the multi-path topology is that one can use the transition information of previous delay stages to predict when a current delay stage should begin its transition and thereby reduce the effective delay time through each delay stage. Using this technique, a significant increase in operating frequency has been demonstrated for ring oscillators [7,8], and a multi-path GRO implementation has been presented which achieves roughly a five times reduction of effective gate delay [9].

Figure 12.15. Reduction of effective delay per stage using multi-path oscillator implementation.

Reduction of gate delay time improves TDC resolution, but carries the cost of higher operating frequencies and therefore higher power. Utilizing counters at each delay stage as shown in Figure 12.13 is particularly problematic for power dissipation when such high operating frequencies are involved. Given this issue, it is worthwhile considering alternative methods of recording transition information within the oscillator such as the phase sampling method shown in Figure 12.16. In this approach, each delay stage is connected to a simple register rather than a counter. These registers sample the transition states of the ring oscillator at the conclusion of each measurement, which yields the quantized phase state of the oscillator. One of the delay stages is also connected to a counter which keeps track of phase wrapping in the oscillator. By combining the phase state and wrapping information from the counter, the number of transitions of the oscillator within a given measurement window can be computed at much lower power dissipation than with the parallel counter method.

Figure 12.16. Phase sampling of GRO TDC state for reduced power dissipation.

Using the above techniques, a prototype GRO TDC was developed as shown in Figure 12.17 [9]. Here a 47-stage multi-path oscillator structure is utilized which achieves 6 ps delay per stage in 130 nm CMOS. The phase sampling operation is performed by seven interleaved channels, the total range input range is 12 ns, total active area is 0.04 mm², and power consumption varies from 2.2 to 21 mW depending on the input time between Start(t) and Stop(t) signals. Additional performance results for this structure are shown in the following sub-section.

Figure 12.17. Example GRO implementation.