G–Sensor Characterization Methodology
Overview and Comprehensive Reference
Abstract: This reference document presents systematic methodologies for characterizing temperature sensors, magnetic field sensors, and gravitational accelerometers. Covering physical transduction principles, noise analysis, environmental cross-sensitivities, temporal stability, and calibration procedures, it provides practitioners with authoritative guidance based on international standards and peer-reviewed literature. The document emphasizes quantitative specifications and traceable calibration methods essential for precision measurement applications.
Part 0: Conceptual Foundation—What Is a Sensor?
The sensor as information transducer
A sensor is fundamentally an information converter—a device that transforms a physical quantity (the measurand) into a form that can be processed, recorded, or acted upon. In classical terms, a sensor performs transduction: converting energy from one domain (mechanical, thermal, optical, magnetic, chemical) into the electrical domain, where signals can be amplified, digitized, and analyzed [FRADEN16].
Consider measuring room temperature. The physical measurand is temperature (a thermodynamic state variable), but our instruments cannot directly "read" temperature—they can only measure electrical quantities like voltage, current, or resistance. A thermistor provides the necessary conversion: its resistance changes predictably with temperature, establishing the correspondence . By measuring resistance electrically, we infer temperature. The thermistor is thus a transducer element embedded within a larger measurement system [DOEBELIN04].
This transduction principle applies universally. A photodiode converts photon flux to current. A strain gauge converts mechanical deformation to resistance change. A Hall sensor converts magnetic flux density to voltage. In each case, the sensor establishes a functional relationship between input (measurand) and output (electrical signal):
Output = f(Input, Parameters, Disturbances)
Understanding this relationship—its form, its stability, its limitations—constitutes the essence of sensor characterization.
The measurement chain: from stimulus to display
Sensors do not operate in isolation. The complete measurement chain transforms the physical measurand into useful information through several stages [PALLASARENY01]:
Physical coupling: The measurand must interact with the sensor's transduction element. For temperature, this requires thermal equilibrium. For magnetic field, proper orientation. For acceleration, mechanical coupling through mounting.
Transduction: The sensing element converts the measurand to a raw electrical signal (voltage, current, resistance, capacitance, charge). This signal is typically small (microvolts to millivolts) and may contain offsets.
Signal conditioning: Amplification, filtering, excitation, and linearization prepare the signal for digitization. A Wheatstone bridge might convert resistance change to differential voltage. An instrumentation amplifier might provide 100× gain with high common-mode rejection. A low-pass filter might remove high-frequency noise.
Analog-to-digital conversion: The ADC discretizes the analog signal into digital codes. An N-bit ADC provides quantization levels, establishing the quantization step [KESTER05].
Digital processing: Microcontrollers or computers apply calibration equations, perform averaging or filtering, compensate for known systematic errors, and present results in physical units.
Each stage introduces uncertainty and potential error sources. The signal chain must be designed holistically: increasing ADC resolution provides no benefit if amplifier noise dominates, and sophisticated digital filtering cannot compensate for fundamental sensor limitations.
Fundamental sensor properties
Before characterizing specific sensors, we must establish precise definitions for the properties that determine sensor performance.
Sensitivity: the transfer function slope
Sensitivity quantifies how strongly the output responds to changes in input. Formally, sensitivity S is the derivative of the transfer function:
For linear sensors, this reduces to the slope of the input-output curve. A Pt100 RTD exhibits sensitivity of approximately 0.385 Ω/°C at 0°C. A typical MEMS accelerometer might provide 660 mV/g output sensitivity [FRADEN16].
Sensitivity determines the signal strength available for measurement. Higher sensitivity provides better signal-to-noise ratio but may reduce dynamic range. The optimal sensitivity depends on the application's required resolution and range.
Importantly, sensitivity may vary with operating point (nonlinear sensors), temperature, time (aging), or other environmental factors. Complete characterization requires determining sensitivity across the full operating space.
Range, span, and dynamic range
Range specifies the minimum and maximum measurand values the sensor can transduce, often expressed as lower and upper range limits. A typical laboratory RTD might specify a range of -200°C to +850°C.
Span is the algebraic difference between upper and lower range limits: Span = URL - LRL. For the RTD above, span = 1050°C.
Dynamic range is the ratio of the largest to smallest measurable signals, often expressed in decibels: . A 16-bit ADC provides theoretical dynamic range of approximately 96 dB. Practical sensor dynamic range is limited by noise floor and saturation effects [KESTER05].
Resolution, precision, and accuracy: distinguishing concepts
These three terms are frequently confused but represent distinct concepts [JCGM200_12]:
Resolution is the smallest detectable change in measurand. For digital systems, this is fundamentally limited by quantization: an N-bit ADC measuring a ±5V range provides 10V/2^16 ≈ 153 μV resolution for 16 bits. Sensor physics may impose additional limits—thermomechanical noise, for example, establishes the resolution floor for MEMS accelerometers.
Precision (repeatability) describes the closeness of agreement among repeated measurements under identical conditions. High precision means low scatter in repeated readings, regardless of whether those readings are accurate. Precision is typically quantified through standard deviation or variance of repeated measurements.
Accuracy describes the closeness of agreement between a measured value and the true value. Accuracy encompasses both systematic errors (bias, offset) and random errors (precision). An accurate sensor must be both unbiased and precise [JCGM200_12].
Consider a digital scale reading 10.213 kg repeatedly for an object whose true mass is 10.000 kg. The scale exhibits high precision (readings cluster tightly) but poor accuracy (systematic +213 g bias). The 1 mg resolution (least significant digit) is irrelevant to the application if accuracy is only ±0.5 kg.
Linearity and hysteresis
Linearity quantifies how well the sensor's actual transfer function approximates an ideal straight line. Perfect linearity means Y = a + bX exactly. Real sensors deviate from this ideal, exhibiting nonlinearity error typically expressed as percentage of full-scale span [NI_SENSOR]:
Nonlinearity = (Maximum deviation from best-fit line / Full-scale span) × 100%
Methods for determining the "best-fit line" include:
End-point linearity: Straight line through the minimum and maximum calibration points
Best-fit straight line (BFSL): Least-squares fitted line through all calibration points
Terminal linearity: Specified line with defined slope and offset
Typical specifications: industrial sensors achieve 0.1-1% nonlinearity, precision laboratory sensors reach 0.01-0.1%, while some specialized designs approach 0.001% or better.
Hysteresis describes path-dependent behavior where the output for a given input depends on the measurement history—specifically, whether the input is increasing or decreasing. Mechanical sensors exhibit hysteresis from friction and elastic-plastic transitions. Magnetic sensors show hysteresis from domain wall pinning. Thermal sensors may display hysteresis during thermal cycling from stress-anneal effects [DOEBELIN04].
Hysteresis is quantified as the maximum separation between increasing and decreasing curves, expressed as percentage of full scale. Specifications often require both upscale and downscale calibration curves to characterize this effect.
Response time and bandwidth
Sensors do not respond instantaneously to input changes. Response time characterizes the temporal behavior—how quickly the sensor output reaches its final value following a step change in input [NI_SENSOR].
For first-order systems, the response to a step input follows:
Y(t) = Y_∞(1 - e^(-t/τ))
where τ is the time constant. The response time is conventionally defined as the time to reach 63.2% of the final value (t = τ) or alternatively 90% (t ≈ 2.3τ) or 95% (t ≈ 3τ), depending on convention.
Bandwidth represents the frequency range over which the sensor can accurately track dynamic inputs. For first-order systems, the -3 dB bandwidth (where response amplitude drops to 70.7% of DC value) equals:
f₋₃dB = 1/(2πτ)
Fast sensors (photodiodes: μs response) enable high-bandwidth measurements. Slow sensors (thermocouples in protective sheaths: seconds) integrate rapid fluctuations, acting as low-pass filters.
Second-order systems (MEMS accelerometers, pressure sensors) exhibit more complex behavior characterized by natural frequency ω₀ and damping ratio ζ. Underdamped systems (ζ < 1) show overshoot and ringing. Critically damped (ζ = 1) or slightly overdamped (ζ ≈ 1.2) systems optimize response time without overshoot [DOEBELIN04].
Noise and signal-to-noise ratio
Noise represents unwanted random fluctuations in the output that obscure the desired signal. Noise arises from fundamental physical processes (thermal agitation, shot noise, quantum fluctuations) and practical imperfections (electronic components, environmental interference) [HOROWITZ15].
Noise is characterized through several metrics:
RMS noise: Standard deviation of the output under constant input conditions, with units matching the output (volts, counts, etc.)
Noise spectral density: Noise power per unit bandwidth, typically expressed as V/√Hz or equivalent input quantity per √Hz
Peak-to-peak noise: Often specified as 6× RMS for Gaussian distributions (encompassing 99.7% of samples)
Signal-to-noise ratio (SNR) quantifies the ratio of signal power to noise power:
SNR = 20 log₁₀(V_signal/V_noise) [dB]
An SNR of 60 dB means the signal voltage is 1000× larger than noise voltage. Higher SNR enables detection of smaller signals and more precise measurements.
Noise-equivalent input converts output noise to equivalent input units through the sensitivity: NEI = Noise_output/Sensitivity. For example, a temperature sensor with 10 μV RMS noise and 385 μV/°C sensitivity exhibits 0.026°C RMS noise-equivalent temperature.
Drift and long-term stability
Drift describes gradual, unidirectional changes in sensor output over time when the input remains constant. Drift arises from aging mechanisms: component parameter changes, material property evolution, stress relaxation, contamination, or chemical reactions [FRADEN16].
Drift is quantified as change per unit time (e.g., °C/month, ppm/year) or change per event cycle (e.g., ±0.01°C per thermal cycle from -40°C to +125°C). Specifications distinguish:
Short-term stability: Variations over minutes to hours (often related to temperature fluctuations)
Long-term stability: Systematic trends over months to years (material aging)
Repeatability: Consistency when returning to the same input after excursions
High-stability applications (frequency standards, navigation systems, scientific instruments) demand stringent drift specifications—parts-per-million per year for the best systems. Industrial applications may tolerate 0.1-1% per year drift if periodic recalibration is feasible.
Allan deviation analysis provides the standard framework for quantifying stability across timescales, distinguishing noise types from drift mechanisms (covered in Part I).
Cross-sensitivities and environmental robustness
Real sensors respond not only to the intended measurand but also to interferents—undesired inputs that produce output changes. Common cross-sensitivities include:
Temperature dependence: Nearly all sensors exhibit temperature-dependent bias and scale factor
Mechanical coupling: Vibration, acoustic pressure, or mechanical stress affecting non-mechanical sensors
Electromagnetic interference: Magnetic or electric fields coupling into the measurement
Pressure effects: Altitude or ambient pressure changes affecting sealed cavities
Chemical exposure: Humidity, corrosive gases, or contaminants degrading sensor materials
Cross-sensitivity specifications quantify these effects: ±X output_units per interferent_unit (e.g., ±0.1 nT/°C for magnetometer temperature sensitivity, or ±2% per g for accelerometer vibration rectification).
Robust sensor design employs several strategies: differential measurements for common-mode rejection, temperature compensation through reference elements or lookup tables, magnetic shielding for sensitive electronics, hermetic sealing for environmental protection, and material selection for chemical compatibility.
Why characterization matters: specifications versus requirements
Sensor datasheets provide typical performance specifications determined through limited testing on representative samples. These specifications establish expected performance but do not guarantee the actual performance of an individual sensor in a particular application environment.
Sensor characterization determines the actual performance of specific sensors under actual operating conditions. This process:
Verifies conformance to datasheet specifications or identifies outliers
Establishes traceability to reference standards for calibrated measurements
Quantifies uncertainty budgets enabling meaningful comparison between measurements
Identifies application-specific limitations not apparent from generic specifications
Enables compensation of systematic errors through calibration equations
Provides baseline for detecting degradation or failure during operation
For laboratory research, proper characterization distinguishes genuine physical effects from measurement artifacts. For industrial applications, it ensures product quality and regulatory compliance. For scientific instruments, it enables credible publication of results with defensible uncertainty claims.
The following sections present systematic methodologies for characterizing three representative sensor types—temperature (Pt RTDs), magnetic field (fluxgate magnetometers), and acceleration (capacitive MEMS)—applicable to precision measurement applications in experimental physics.
Part I: General Sensor Characterization Methodology
The characterization of precision sensors requires systematic evaluation of response functions, noise behavior, stability, cross-sensitivities, and measurement uncertainty. This section establishes the foundational methodology applicable across sensor types.
Response function determination through systematic fitting
Sensor calibration fundamentally involves determining the mathematical relationship between physical input and electrical output. The choice between linear and polynomial models depends on the sensor's intrinsic nonlinearity and application requirements [JCGM100_08].
Linear least-squares calibration minimizes the sum of squared residuals for the model y = ax + b. This approach remains optimal when sensor nonlinearity falls within acceptable bounds—typically less than 0.1% of full scale for industrial applications. The method weights larger residuals more heavily through squaring, appropriately penalizing systematic deviations mbedded.ninja [BEVINGTON03].
For sensors exhibiting significant nonlinearity, polynomial regression extends the model to y = a₀ + a₁x + a₂x² + ... + aₙxⁿ, with coefficients estimated through least squares. The Callendar-Van Dusen equation for platinum RTDs exemplifies this approach, using a quadratic form above 0°C and cubic below. Polynomial order selection requires careful attention to avoid overfitting—cross-validation and regularization techniques (L1, L2 penalties) help identify optimal complexity California Learning Resource Network [CLRN_FIT].
Segmented linearization offers an alternative for highly nonlinear responses. [DING18] demonstrated piecewise linear fitting achieving linearity exceeding 99% with 30 μm accuracy over 6 mm range. This approach divides the measurement range into segments, fitting each independently while ensuring continuity at boundaries.
The interpolation error for polynomial methods of degree n is bounded by |error| ≤ |f^(n+1)(ξ)| × |Π(x-xᵢ)| / (n+1)!, establishing theoretical limits on achievable accuracy [PASIC_THESIS].
Noise spectral density measurements and analysis
Power spectral density (PSD) quantifies how signal power distributes across frequency, Cadence expressed in units of V²/Hz or equivalent physical quantities. PSD estimation forms the foundation for identifying noise sources and establishing performance limits [BENDAT11].
Welch's method [WELCH67] provides the standard approach for PSD estimation by dividing data into overlapping segments, applying window functions, computing FFTs, and averaging squared magnitudes. Critical parameters include:
Window length: Longer windows improve frequency resolution but increase estimate variance
Overlap: Typically 50%, compensating for information loss at window edges
Window function: Hamming windows provide 42.5 dB sidelobe attenuation; flat-top windows optimize amplitude accuracy
The Equivalent Noise Bandwidth (ENBW) characterizes spectral spreading: rectangular windows have ENBW = 1.0, Hann windows 1.5, and flat-top windows approximately 3.8 [AUDIO_PRECISION].
Noise type identification relies on the PSD frequency dependence. White noise exhibits flat spectra (S(f) = constant), while colored noise follows power laws S(f) ∝ f^α. The table below summarizes relationships between spectral density exponent and Allan deviation slope:
White PM
+2
-1
Flicker PM
+1
-1
White FM
0
-1/2
Flicker FM
-1
0 (minimum)
Random Walk FM
-2
+1/2
Allan deviation analysis for stability characterization
Allan variance, introduced by [ALLAN66], provides the standard metric for characterizing sensor stability over different averaging times. Unlike conventional variance, Allan variance converges for noise types common in precision sensors including random walk and flicker noise.
The overlapping Allan variance is computed as:
σ²ᵧ(τ) = [1/(2(N-2m)τ²)] × Σ(xᵢ₊₂ₘ - 2xᵢ₊ₘ + xᵢ)²
where N is the number of phase samples, m is the averaging factor, and τ = mτ₀ is the averaging time [IEEE1139_08, NIST_SP1065].
Interpreting Allan deviation plots requires identifying characteristic slopes on log-log axes. The minimum point represents bias instability—the fundamental limit on long-term sensor performance. At short averaging times, the -1/2 slope region yields velocity random walk (accelerometers) or angle random walk (gyroscopes) when read at τ = 1 s. The +1/2 slope at long times indicates random walk processes, while +1 slopes suggest deterministic drift [RILEY08].
Modified Allan variance (MVAR) distinguishes between white and flicker phase modulation noise—a distinction impossible with standard Allan variance. MVAR achieves this through additional phase averaging, following:
Mod σ²ᵧ(τ) = [1/(2m²(N-3m+1)τ₀²)] × Σⱼ Σᵢ (xᵢ₊₂ₘ - 2xᵢ₊ₘ + xᵢ)²
Hadamard variance extends convergence to more divergent noise types (α ≥ -4) and inherently rejects linear frequency drift, proving valuable for sensors exhibiting systematic trends [BARNES71].
Cross-sensitivity matrix determination
Multi-axis sensors require characterization of coupling between nominally independent channels. The complete sensor model relates outputs to inputs through a 3×3 sensitivity matrix:
[Output] = [S][Input] + [b]
where diagonal terms (Sxx, Syy, Szz) represent primary scale factors and off-diagonal terms quantify cross-axis sensitivities [HILLER21].
Cross-axis sensitivity (CAS) specification relates to axis-to-axis misalignment error: CAS ≈ MAE × 0.0175, where MAE is misalignment in degrees. Consumer MEMS accelerometers typically exhibit 1-5% cross-axis sensitivity (0.57°-2.87° misalignment), while industrial-grade devices achieve <0.031% (<0.018°) [ANALOG_CAS].
Six-position calibration determines the complete sensitivity matrix by applying known stimuli along ±X, ±Y, and ±Z axes, recording primary channel outputs and cross-talk on orthogonal channels. The compensation matrix is then computed as the inverse of the measured cross-talk matrix:
[Calibrated] = [S]⁻¹ × ([Raw] - [Bias])
Orthogonality errors arise from mechanical misalignment of sensing elements, package-to-die orientation errors, circuit-level crosstalk, and system-level mounting errors [PMC_CALIB].
Uncertainty budgeting using GUM methodology
The Guide to the Expression of Uncertainty in Measurement [JCGM100_08] establishes the international framework for evaluating and expressing measurement uncertainty. This approach propagates uncertainties from input quantities through the measurement model to the final result.
Type A uncertainties derive from statistical analysis of repeated observations: u(x̄) = s/√n, where s is experimental standard deviation and n is observation count.
Type B uncertainties incorporate prior knowledge, specifications, and calibration certificates. Standard uncertainty for rectangular distributions is u = a/√3 (where ±a are bounds); for triangular distributions, u = a/√6; and for expanded uncertainties from calibration certificates, u = U/k where k is the stated coverage factor.
Sensitivity coefficients cᵢ = ∂f/∂Xᵢ quantify how the measurand varies with each input quantity. These may be determined analytically, numerically (cᵢ ≈ Δy/Δxᵢ), or experimentally.
The combined standard uncertainty for uncorrelated inputs follows the law of propagation of uncertainty:
u²c(y) = Σᵢ [cᵢ × u(xᵢ)]²
For correlated inputs, covariance terms must be included:
u²c(y) = Σᵢ Σⱼ (∂f/∂xᵢ)(∂f/∂xⱼ) × u(xᵢ,xⱼ)
Expanded uncertainty U = k × uc(y) provides a confidence interval, with k = 2 corresponding to approximately 95% coverage probability for normal distributions. For finite degrees of freedom, the Welch-Satterthwaite formula determines effective degrees of freedom for selecting the appropriate t-distribution value [UKAS_M3003].
Part II: Platinum Resistance Thermometer Characterization
Platinum resistance thermometers (PRTs) exploit the predictable increase in electrical resistance of platinum with temperature, offering exceptional stability, reproducibility, and wide operating range from -200°C to +850°C [IEC60751_22].
Physical transduction principle and the Callendar-Van Dusen equation
The resistance-temperature relationship in platinum arises from increased electron-phonon scattering as lattice vibrations intensify with temperature. Platinum's temperature coefficient of resistance is approximately 0.00385 °C⁻¹ for industrial sensors, reflecting a relative sensitivity dR/(R·dT) of about 0.4%/°C [BIPM_CCT].
The Callendar-Van Dusen equation, developed by Hugh Longbourne Callendar and refined by Milton Van Dusen at NBS in 1925 [VANDUSEN25], describes this relationship:
For t ≥ 0°C: R(t) = R₀(1 + At + Bt²)
For t < 0°C: R(t) = R₀(1 + At + Bt² + C(t-100)t³) Instrumentationtoolbox
IEC 60751:2022 specifies the standard coefficients as A = 3.9083 × 10⁻³ °C⁻¹, B = -5.775 × 10⁻⁷ °C⁻², and C = -4.183 × 10⁻¹² °C⁻⁴. Iteh The alpha coefficient α = [R(100°C) - R(0°C)] / [100 × R(0°C)] equals 0.00385055 °C⁻¹ for industrial platinum, compared to 0.003926 °C⁻¹ for high-purity laboratory platinum [IEC60751_22].
Standard nominal resistances include Pt100 (100Ω at 0°C, most common), Pt1000 (1000Ω), Wikipedia and Pt500. Tolerance classes define permissible deviations: Class AA permits ±(0.1 + 0.0017|t|)°C, Class A permits ±(0.15 + 0.002|t|)°C, and Class B permits ±(0.3 + 0.005|t|)°C [IEC60751_22].
Noise characterization in resistance thermometry
Johnson-Nyquist noise establishes the fundamental thermal noise limit. First measured by Johnson and theoretically explained by Nyquist in 1928 Wikipedia [NYQUIST28], this noise arises from random thermal motion of charge carriers:
V_n = √(4kTRΔf)
where k is Boltzmann's constant (1.380649×10⁻²³ J/K), T is absolute temperature, R is resistance, and Δf is measurement bandwidth. For a Pt100 at 273 K with 1 kHz bandwidth, Johnson noise is approximately 1.3 nV RMS—representing an ultimate temperature resolution limit [WHITE96].
Self-heating introduces systematic error when excitation current dissipates power in the sensing element. The temperature rise ΔT = I²R × θ, where θ is the self-heating coefficient (°C/mW). Typical values range from 0.003 °C/mW for thin-film elements to 0.05-0.25 °C/mW for wire-wound designs. For a Pt100 at 0°C with 1 mA excitation (0.1 mW dissipation), self-heating error is approximately 0.005°C [TI_RTD].
Recommended excitation currents balance self-heating against signal-to-noise ratio: 0.25-1.0 mA for Pt100 sensors, with 250 μA typical for high-precision applications. Extrapolation to zero power by measuring at two currents (e.g., 1 mA and 2 mA) allows self-heating correction. Self-heating error should remain below 25% of total error budget [BIPM_CCT].
Environmental cross-sensitivities in RTD measurements
Lead wire resistance introduces the most significant parasitic error. In 2-wire configurations, measured resistance includes lead resistance, causing approximately 2.5°C error per ohm of total lead resistance for Pt100. Miinet Three-wire configurations compensate assuming matched lead resistances, achieving <0.1°C accuracy. Four-wire (Kelvin) configurations eliminate lead resistance effects entirely, enabling <0.001°C accuracy with AC bridges—required for Class A or better tolerance per IEC 60751 [IEC60751_22, NIST_SP250_81].
Strain sensitivity affects RTD resistance through piezoresistive effects. Sources include differential thermal expansion between platinum and substrate, mechanical shock, and work hardening from repeated stress cycles. Wire-wound elements exhibit greater strain sensitivity than thin-film types, which benefit from intimate substrate bonding.
Insulation resistance must exceed specified minimums: 100 MΩ at 100°C, 10 MΩ at 300°C, 2 MΩ at 500°C, and 0.5 MΩ at 850°C per IEC 60751. Insulation follows semiconductor-like temperature dependence, degrading exponentially at high temperatures. Moisture ingress, particularly in MgO-insulated designs, accelerates degradation [IEC60751_22].
Temporal stability and drift mechanisms
High-quality wire-wound PRTs achieve stability of ±0.005% over 5 years (±0.05°C per 5 years), while standard thin-film types specify ±0.05% over operating life. Standard Platinum Resistance Thermometers (SPRTs) used as laboratory references achieve ±0.0001-0.001°C per year [BIPM_CCT].
Primary drift mechanisms include contamination (diffusion of metal ions into platinum above 450°C), irreversible dimensional changes from plastic deformation, crystal defect accumulation from cold work and mechanical shock, and lead wire degradation.
Hysteresis from thermal cycling arises from reversible strain-anneal cycles. Best partially-supported elements exhibit hysteresis below 0.002% (1 mK per 100°C range), while fully-supported glass-encapsulated designs may reach 0.1%. Thin-film types show 16-156 mK hysteresis over 0-500°C cycling [BIPM_CCT].
Annealing relieves internal strain and restores resistance characteristics. SPRT annealing protocols specify 675°C for 30+ minutes (long-stem SPRTs) or 975°C for 4-6 hours (high-temperature SPRTs), followed by slow cooling to 480°C then rapid cooling to avoid oxidation. Manufacturers often pre-anneal by thermal cycling 0°C to 600°C for 1000 hours Miinet [NIST_SP250_81].
Calibration procedures and ITS-90 traceability
The International Temperature Scale of 1990 (ITS-90) defines temperature from 13.8 K to 1234.93 K using SPRTs as interpolation instruments between defined fixed points. Key fixed points include the triple point of water (273.16 K exactly), gallium melting point (302.9146 K), and freezing points of indium (429.7485 K), tin (505.078 K), zinc (692.677 K), aluminum (933.473 K), and silver (1234.93 K) [BIPM_ITS90].
Fixed-point calibration provides the lowest uncertainty. The triple point of water cell realizes 273.16 K with uncertainty better than ±0.15 mK—far superior to ice baths (±0.002 K). Metal freezing-point cells require high-purity samples (99.99999% for gallium) in specialized furnaces, achieving ±0.1-1 mK uncertainty depending on the fixed point [NIST_SP250_81].
Comparison calibration suffices for industrial RTDs: the device under test and a calibrated reference SPRT are immersed together in calibration baths at multiple temperatures spanning the application range. Deviations from the standard Callendar-Van Dusen curve are then calculated. Best practice includes preliminary thermal cycling before calibration, ice-point checks after each temperature excursion, and overnight relaxation periods between measurements [BIPM_CCT].
Traceability flows from national metrology institutes (NIST, NPL, PTB, NMIJ) maintaining ITS-90 through SPRT calibrations at fixed points. NIST NIST calibration uncertainties reach 0.1-1 mK at fixed points using ASL F900 AC resistance bridges with 10.5-digit resolution. NIST Calibration intervals typically range from annual TPW checks to full recalibration every 2-5 years for SPRTs [NIST_SP250_81].
Part III: Fluxgate Magnetometer Characterization
Fluxgate magnetometers measure DC and low-frequency magnetic fields through periodic saturation of a soft ferromagnetic core, achieving noise floors below 1 pT/√Hz in state-of-the-art designs [JANOSEK19].
Physical transduction principle and core saturation mechanism
The fluxgate principle exploits nonlinear magnetization behavior in soft ferromagnetic materials. An alternating excitation current periodically drives the core between positive and negative saturation, during which permeability drops dramatically from μᵣ > 10,000 to nearly unity—effectively "gating" magnetic flux [PRIMDAHL79].
In the presence of an external field B₀, saturation becomes asymmetric: the core saturates earlier when the external field aids the excitation and later when opposing. This asymmetry generates even harmonics (primarily second harmonic) in the pickup coil voltage, with amplitude proportional to the external field:
V₂f ∝ B₀ × n × A × (dμ/dt) × f
where n is pickup coil turns, A is core cross-sectional area, and f is excitation frequency (typically 10-30 kHz) [RIPKA03].
Phase-synchronous detection extracts the second harmonic signal, which is integrated and fed back through compensation coils to null the core, providing high linearity (10 ppm achievable). The feedback configuration maintains the sensor in a "magnetic vacuum," with output current proportional to the external field [NIELSEN95].
Ring-core geometries dominate space applications (Ørsted, CHAMP, Swarm missions) due to lower mass, simpler circuitry, and excellent noise characteristics reaching 0.3-10 pT/√Hz. Rod-core (Vacquier) designs offer lower crossfield sensitivity but typically achieve only 5-50 pT/√Hz [PRIMDAHL79].
Core materials require high initial permeability (μᵣ > 10,000), low coercivity (Hc < 1 A/m), and near-zero magnetostriction (λs < 10⁻⁶). Legacy 6-81 Mo-Permalloy developed at the U.S. Navy NSWC in the 1960s provided exceptional performance, though NASA stockpiles are nearly depleted and production processes were lost [MILES_MAGIC]. Modern alternatives include amorphous alloys (Metglas, Vitroperm) and nanocrystalline materials with 40-60 dB lower Barkhausen noise than conventional Permalloy [WANG21].
Noise characterization in fluxgate sensors
Barkhausen noise from discontinuous domain wall motion during magnetic cycling dominates fluxgate sensor noise. This noise exhibits 1/f^α spectrum (α ≈ 1.0-1.3), originating from irreversible magnetization changes within the core material [VANBREE74].
Fundamental-mode orthogonal fluxgates with DC bias achieve 60× noise reduction by avoiding full magnetization reversals. [SASADA02] demonstrated this approach, and recent implementations by [JANOSEK19] achieved 0.75 pT/√Hz at 1 Hz with 0.3 pT/√Hz white noise floor—enabling unshielded magnetocardiography applications.
Johnson noise in pickup coil windings contributes V_J = √(4k_B T R Δf), yielding approximately 1.3 nV/√Hz for typical 100Ω coil resistance at 300K. This is generally negligible compared to Barkhausen noise except at very high frequencies [RIPKA03].
Environmental magnetic noise includes geomagnetic variations (diurnal variation ~50 nT, micropulsations 1-10 nT), urban magnetic noise (10-1000 nT from vehicles and machinery), power line interference at 50/60 Hz and harmonics, and Schumann resonances at 7.83 Hz fundamental [WANG21].
Environmental cross-sensitivities of fluxgates
Temperature dependence affects core permeability, coil resistance, and sensor geometry. Typical offset drift ranges from 0.1-1.0 nT/°C, with best achieved values of 0.02-0.08 nT/°C using cobalt-based amorphous cores. Scale factor temperature coefficients typically span 30-100 ppm/°C, reducible to 1-10 ppm/°C through compensation. The Ørsted mission achieved <1 nT thermal stability from -20°C to +60°C through careful thermal design [NIELSEN95].
Magnetostriction couples mechanical stress to magnetic properties. Even small stresses can shift offset and sensitivity through stress-induced anisotropy. Near-zero magnetostriction compositions such as Co₆₈.₂₅Fe₄.₅Si₁₂.₂₅B₁₅ minimize this effect [BUTTA20].
Crossfield effects arise when large transverse fields (>5000 nT uncompensated) cause nonlinear response and third harmonic generation. Ring-core sensors exhibit greater susceptibility than parallel-core designs. Spherical feedback coil systems mitigate this by maintaining the sensor in nulled conditions [PRIMDAHL82].
Temporal stability in fluxgate magnetometers
Offset drift arises from core remanence changes, electronics aging (voltage references, capacitors), mechanical creep, and temperature cycling hysteresis. Observatory-grade instruments (DTU FGE) achieve <3 nT/year drift, while space-qualified designs (Swarm VFM) maintain <0.5 nT over mission duration [TOFFNER16].
Scale factor drift originates from feedback resistor aging, A/D converter drift, coil geometry changes, and core permeability evolution. Observatory instruments typically achieve <50 ppm/year, with space instruments maintaining <100 ppm over mission duration through continuous scalar calibration [TOFFNER16].
Core material aging proceeds through magnetic domain reorganization, crystallographic relaxation in amorphous alloys, and oxidation. Best materials achieve <50 ppm/year (<3 nT/year in a 60,000 nT field) [HERZOG90].
Calibration procedures for vector magnetometers
Helmholtz coil calibration applies precisely known fields generated according to B = (8μ₀NI)/(5^(3/2)R) ≈ 0.8991 × μ₀NI/R. The sensor is placed at the coil center within the 1% uniformity region, Earth's field is cancelled using triaxial coils with a reference magnetometer, and known fields are applied in each axis direction. Coil constants are determined using NMR teslameters traceable to the proton gyromagnetic ratio (42.576 MHz/T, accurate to ±5 ppm), enabling total system uncertainty of ~10-100 ppm [NPL_MAT65].
Scalar calibration against proton precession, Overhauser, or cesium/helium vapor magnetometers provides absolute field magnitude reference. The constraint |B| = F (scalar measurement) enables fitting an ellipsoid to vector data, extracting scale factors, offsets, and orthogonality angles. The Swarm Absolute Scalar Magnetometer achieves <0.3 nT absolute accuracy with no intrinsic drift, allowing vector magnetometer calibration to <0.5 nT scalar residual [TOFFNER16].
Vector calibration for triaxial sensors determines a 12-parameter model: 3 scale factors, 3 offsets, and 6 alignment/orthogonality angles. Ground calibration typically achieves ~1-10 nT accuracy, while in-flight calibration with scalar reference enables sub-nT performance [NIELSEN95].
Traceability flows from the proton gyromagnetic ratio (CODATA) through NMR teslameters to calibrated Helmholtz coil systems to reference fluxgates (e.g., Bartington Mag-13MS with NPL accreditation) to production instruments. INTERMAGNET standards for geomagnetic observatories require <5 nT absolute accuracy, 0.1 nT resolution, and regular calibration against proton magnetometers and DI-flux theodolites [INTERMAGNET].
Part IV: Capacitive MEMS Accelerometer Characterization
Capacitive MEMS accelerometers measure acceleration through displacement-induced capacitance changes, achieving noise floors from hundreds of μg/√Hz (consumer) to below 1 μg/√Hz (navigation grade) [PMC_REVIEW20].
Physical transduction principle and proof mass dynamics
Capacitive sensing exploits the gap-dependent capacitance of parallel-plate structures: C = ε₀εᵣA/d. Acceleration displaces a proof mass suspended by compliant springs, changing the gap between moving and fixed electrodes.
Differential capacitance detection using symmetric capacitor pairs (C₁ and C₂) provides common-mode rejection, improved linearity, and noise immunity:
C₁ = ε₀A/(d₀ - x), C₂ = ε₀A/(d₀ + x)
ΔC = C₁ - C₂ = 2ε₀Ax/d₀² (for small displacements x << d₀) [KESHAVARZI19]
The proof mass dynamics follow a second-order mechanical system:
m(d²x/dt²) + b(dx/dt) + kx = ma_ext
with natural frequency ω₀ = √(k/m), quality factor Q = mω₀/b, and static displacement sensitivity x/a = 1/ω₀² [KAAJAKARI].
Squeeze-film damping dominates energy dissipation as air compresses between the proof mass and nearby surfaces. The damping coefficient for a rectangular plate follows b = 0.42μL⁴/d³, where μ is air dynamic viscosity (~1.85×10⁻⁵ Pa·s at 25°C). Optimal damping (ζ = 0.707) provides maximum bandwidth without ringing [BAO07].
Noise characterization of MEMS accelerometers
Thermomechanical (Brownian) noise arises from random impacts of air molecules on the proof mass. The noise equivalent acceleration is:
a_n = √(4k_B T ω₀ / mQ) [m/s²/√Hz]
where k_B is Boltzmann's constant, T is temperature, m is proof mass, ω₀ is natural frequency, and Q is quality factor. This represents the fundamental mechanical noise floor [GABRIELSON93].
Bulk micromachined accelerometers with larger proof masses achieve ~30 μg/√Hz, while surface micromachined designs typically reach ~1 mg/√Hz. State-of-the-art devices achieve 0.25-10 ng/√Hz through increased proof mass, reduced damping (vacuum packaging), and lower resonant frequency [PMC_REVIEW20].
Electronic readout noise converts to equivalent acceleration through ENEA = ΔC_min/SF_cap, where ΔC_min is input-referred capacitance noise and SF_cap is capacitive sensitivity. State-of-the-art readout ASICs achieve 50 zF/√Hz noise floor [SIWARE].
1/f noise dominates below approximately 5 Hz, originating from surface state fluctuations, dielectric charging, and electronic components. Chopper stabilization and carrier modulation/demodulation mitigate this effect [SERCEL].
Total noise equivalent acceleration combines thermomechanical, electronic, and quantization contributions:
NEA = √(TNEA² + ENEA² + QNEA²)
Environmental cross-sensitivities in MEMS accelerometers
Temperature effects cause both bias and scale factor drift. Temperature drift of bias (TDB) typically ranges from 0.1-1.3 mg/°C, random in sign between units, primarily caused by thermal deformation of sensing structures. Temperature drift of scale factor (TDSF) spans 0 to -400 ppm/°C, always negative due to Young's modulus temperature coefficient (approximately -60 ppm/°C for silicon) and thermal expansion effects [SPRINGER23].
The output model incorporates both effects:
X = X₀ × [1 + TDSF×(T-T_ref)] + TDB×(T-T_ref)
Cross-axis sensitivity typically ranges from 0.1-5% for consumer devices to <0.1% for high-performance accelerometers, arising from misalignment, asymmetric fabrication, and off-axis modes. Differential capacitive designs provide inherent rejection of off-axis motion [ANALOG_CAS].
Package stress effects from die-attach adhesive viscoelasticity cause long-term drift. Observed effects include ~2500 ppm scale factor drift and 12-32 mg bias drift over one year of storage at room temperature. Middle-locating anchors and stress-relief structures mitigate these effects [DRIFT_STUDY].
Temporal stability characterized by Allan variance
Bias instability, identified as the minimum point on the Allan deviation plot, represents the fundamental limit on accelerometer stability. Typical values span:
Consumer
1-10 mg
100-1000 μg/√Hz
Industrial
100-1000 μg
10-100 μg/√Hz
Tactical
10-100 μg
1-10 μg/√Hz
Navigation
<10 μg
<1 μg/√Hz
Per [IEEE1293_18], accelerometer Allan deviation analysis identifies:
Quantization noise: -1 slope, read at τ = √3
Velocity random walk (VRW): -1/2 slope, read at τ = 1 s
Bias instability: 0 slope (minimum)
Acceleration random walk: +1/2 slope, read at τ = 3 s
Scale factor drift is dominated by package stress relaxation over the long term, with turn-on repeatability typically 50-500 ppm. Long-term drift mechanisms include dielectric charging, mechanical fatigue (microcrack propagation), creep, and adhesive cure effects [PMC_TEMP23].
Calibration procedures per IEEE Std 1293
Multi-position tumble testing exploits Earth's gravity as a ±1g reference. The 4-point method records output at 0°, 90°, 180°, and 270° orientations, determining offset = (V₀° + V₁₈₀°)/2 and sensitivity = (V₉₀° - V₂₇₀°)/2g. The 6-point method (±X, ±Y, ±Z orientations) enables full misalignment matrix determination with greater accuracy [EMBEDDED_CALIB].
Centrifuge calibration extends to higher g-ranges and characterizes nonlinearity. Double turntable centrifuge methods generate precise centripetal acceleration a = ω²R with adjustable input axis orientation, enabling in-situ determination of installation errors per [IEEE528].
[IEEE1293_18] defines comprehensive test procedures including:
Performance tests (scale factor, bias, linearity)
Environmental tests (temperature, vibration, shock)
Noise and transient analysis
Multipoint tumble analysis with informative annexes on error effects and filtering techniques
Traceability to g flows from national primary standards (laser interferometer-based) through transfer standards (calibrated reference accelerometers) and working standards (laboratory shakers/rate tables) to devices under test. Local gravity (nominally 9.80665 m/s²) provides the fundamental reference [NIST_ACCEL].
Bibliography
[ALLAN66] Allan, D.W. (1966). "Statistics of Atomic Frequency Standards." Proceedings of the IEEE, 54(2): 221-230.
[ANALOG_CAS] Analog Devices. "Understanding MEMS Accelerometer Cross-Axis Sensitivity." Technical Article.
[AUDIO_PRECISION] Audio Precision. "FFT Windowing and Equivalent Noise Bandwidth." Application Note.
[BAO07] Bao, M. and Yang, H. (2007). "Squeeze film air damping in MEMS." Sensors and Actuators A, 136: 3-27.
[BARNES71] Barnes, J.A. et al. (1971). "Characterization of Frequency Stability." IEEE Transactions on Instrumentation and Measurement, 20(2): 105-120.
[BENDAT11] Bendat, J.S. and Piersol, A.G. (2011). Random Data: Analysis and Measurement Procedures. John Wiley & Sons.
[BEVINGTON03] Bevington, P.R. and Robinson, D.K. (2003). Data Reduction and Error Analysis for the Physical Sciences. McGraw-Hill.
[BIPM_CCT] BIPM Consultative Committee for Thermometry (2021). "Guide on Secondary Thermometry: Industrial Platinum Resistance Thermometers."
[BIPM_ITS90] BIPM (2021). "Guide to the Realization of ITS-90: Platinum Resistance Thermometry." Chapter 5.
[BUTTA20] Butta, M. et al. (2020). "Dependence of the noise of an orthogonal fluxgate on the composition of its amorphous wire-core." AIP Advances, 10: 025114.
[CLRN_FIT] Center for Learning Research and Development. "Polynomial Fitting and Regularization Methods." Technical Reference.
[DING18] Ding, W. et al. (2018). "Piecewise linear fitting for improved sensor linearization." Sensors, 18(10).
[DOEBELIN04] Doebelin, E.O. (2004). Measurement Systems: Application and Design. McGraw-Hill.
[DRIFT_STUDY] "Drift of MEMS accelerometers stored at room temperature." ResearchGate publication.
[EMBEDDED_CALIB] Embedded.com. "MEMS accelerometer calibration optimizes accuracy." Technical Article.
[FRADEN16] Fraden, J. (2016). Handbook of Modern Sensors: Physics, Designs, and Applications. Springer.
[GABRIELSON93] Gabrielson, T.B. (1993). "Mechanical-thermal noise in micromachined acoustic and vibration sensors." IEEE Transactions on Electron Devices, 40(5): 903-909.
[HERZOG90] Herzog, R.E. and Tompkins, D.R. (1990). "Ring-core fluxgate magnetometers for use as observatory variometers." Physics of the Earth and Planetary Interiors.
[HILLER21] Hiller, T. et al. (2021). "Analysis and Compensation of Cross-Axis Sensitivity in Low-Cost MEMS Inertial Sensors." IEEE Sensors Conference.
[HOROWITZ15] Horowitz, P. and Hill, W. (2015). The Art of Electronics, 3rd Edition. Cambridge University Press.
[IEC60751_22] IEC 60751:2022. "Industrial platinum resistance thermometers and platinum temperature sensors." Edition 3.0.
[IEEE528] IEEE Std 528. "IEEE Standard for Inertial Sensor Testing."
[IEEE1139_08] IEEE Std 1139-2008. "IEEE Standard Definitions of Physical Quantities for Fundamental Frequency and Time Metrology—Random Instabilities."
[IEEE1293_18] IEEE Std 1293-2018. "IEEE Standard Specification Format Guide and Test Procedure for Linear Single-Axis, Nongyroscopic Accelerometers."
[INTERMAGNET] INTERMAGNET. "Technical Reference Manual." Geomagnetic Observatory Standards.
[JANOSEK19] Janosek, M. et al. (2019). "1-pT Noise Fluxgate Magnetometer for Geomagnetic Measurements and Unshielded Magnetocardiography." IEEE Transactions on Instrumentation and Measurement, 69: 2552-2560.
[JCGM100_08] JCGM 100:2008. "Evaluation of measurement data—Guide to the expression of uncertainty in measurement (GUM)." Joint Committee for Guides in Metrology.
[JCGM200_12] JCGM 200:2012. "International vocabulary of metrology—Basic and general concepts and associated terms (VIM)." Joint Committee for Guides in Metrology.
[KAAJAKARI] Kaajakari, V. "MEMS Tutorial: Mechanical noise in microelectromechanical systems." Technical Reference.
[KESHAVARZI19] Keshavarzi, M. and Yavand Hasani, J. (2019). "Design and optimization of fully differential capacitive MEMS accelerometer." Microsystem Technologies.
[KESTER05] Kester, W. (2005). The Data Conversion Handbook. Analog Devices / Newnes.
[MILES_MAGIC] Miles, D.M. et al. "NASA MAGIC Project: Magnetometer Materials Investigation." NASA Science Mission Directorate.
[NIELSEN95] Nielsen, O.V. et al. (1995). "Development, construction and analysis of the Ørsted fluxgate magnetometer." Measurement Science and Technology, 6: 1099-1115.
[NI_SENSOR] National Instruments. "Sensor Fundamentals: Definitions and Specifications." Technical Reference.
[NIST_ACCEL] NIST. "Acceleration Calibration Services." Calibration Program Documentation.
[NIST_SP250_81] Strouse, G.F. (2007). "Standard Platinum Resistance Thermometer Calibrations from the Ar TP to the Ag FP." NIST Special Publication 250-81.
[NIST_SP1065] Riley, W.J. (2008). Handbook of Frequency Stability Analysis. NIST Special Publication 1065.
[NPL_MAT65] NPL Report MAT 65. "Measurement of DC Magnetic Fields." UK National Physical Laboratory.
[NYQUIST28] Nyquist, H. (1928). "Thermal Agitation of Electric Charge in Conductors." Physical Review, 32: 110-113.
[PALLASARENY01] Pallàs-Areny, R. and Webster, J.G. (2001). Sensors and Signal Conditioning. Wiley-Interscience.
[PASIC_THESIS] Pasic, A. "Polynomial Interpolation Methods for Sensor Calibration." DCU Thesis.
[PMC_CALIB] "Calibration Method of Accelerometer Based on Rotation Principle Using Double Turntable Centrifuge." PMC, 2022.
[PMC_REVIEW20] "Micromachined Accelerometers with Sub-µg/√Hz Noise Floor: A Review." Sensors, 2020.
[PMC_TEMP23] "Combined Temperature Compensation Method for Closed-Loop MEMS Capacitive Accelerometer." PMC, 2023.
[PRIMDAHL79] Primdahl, F. (1979). "The fluxgate magnetometer." Journal of Physics E: Scientific Instruments, 12: 241-253.
[PRIMDAHL82] Primdahl, F. and Jensen, P.A. (1982). "Compact spherical coil for fluxgate magnetometer vector feedback." Journal of Physics E: Scientific Instruments, 15(2): 221.
[RILEY08] Riley, W.J. (2008). Handbook of Frequency Stability Analysis. NIST Special Publication 1065.
[RIPKA03] Ripka, P. (2003). "Advances in fluxgate sensors." Sensors and Actuators A, 106: 8-14.
[SASADA02] Sasada, I. (2002). "Orthogonal fluxgate mechanism operated with dc biased excitation." Journal of Applied Physics, 91: 7789-7791.
[SERCEL] Sercel. "High-Sensitivity MEMS Accelerometers: Noise Characterization." Technical Paper.
[SIWARE] Si-Ware Systems. "SWS1120 Capacitive Readout ASIC." Datasheet.
[SPRINGER23] "Temperature Bias Drift Phase-Based Compensation for MEMS Accelerometer." Nanomanufacturing and Metrology, Springer, 2023.
[TI_RTD] Texas Instruments. "A Basic Guide to RTD Measurements." Application Note SBAA275.
[TOFFNER16] Tøffner-Clausen, L. et al. (2016). "In-flight scalar calibration and characterisation of the Swarm magnetometry package." Earth, Planets and Space, 68: 129.
[UKAS_M3003] UKAS M3003 (2024). "The Expression of Uncertainty and Confidence in Measurement." Edition 6.
[VANBREE74] van Bree, J.L.M.J., Poulis, J.A., and Hooge, F.N. (1974). "Barkhausen noise in fluxgate magnetometers." Applied Scientific Research, 29: 59-68.
[VANDUSEN25] Van Dusen, M.S. (1925). "Platinum-Resistance Thermometry at Low Temperatures." Journal of the American Chemical Society.
[WANG21] Wang, S. et al. (2021). "Recent Progress of Fluxgate Magnetic Sensors: Basic Research and Application." Sensors, 21: 1500.
[WELCH67] Welch, P.D. (1967). "The use of fast Fourier transform for the estimation of power spectra." IEEE Transactions on Audio and Electroacoustics, 15(2): 70-73.
[WHITE96] White, D.R. et al. (1996). "The Status of Johnson Noise Thermometry." Metrologia, 33: 325-335.
Last updated