Sensor Calibration

Offset error is a constant bias added to the sensor's output, causing a zero-point shift where the sensor reads a non-zero value when the true input is zero. This is often caused by manufacturing tolerances, temperature effects, or inherent sensor drift.

• Correction: A calibration procedure measures the sensor's output at a known zero-input condition. This measured offset value is then subtracted from all subsequent readings.
• Example: An accelerometer at rest (0g) might output 0.05V. Calibration determines this 0.05V offset, and the firmware subtracts it, ensuring the true zero reading is reported.

Gain error is a multiplicative error where the sensor's sensitivity deviates from its ideal specification, causing the slope of the input-output relationship to be incorrect. This results in under-reporting or over-reporting of the measured magnitude.

• Correction: Calibration involves applying a known reference input (e.g., 1g for an accelerometer) and measuring the output. The gain correction factor is calculated as (Expected Output / Measured Output). All raw readings are then multiplied by this factor.
• Impact: Critical for quantitative measurements. An uncalibrated load cell could report 9.8N for a 10N weight, a 2% error directly affecting system accuracy.

Non-linearity describes the deviation of a sensor's response from a perfect straight-line (linear) relationship between its input and output across its operational range. The error is not constant and varies with the magnitude of the input.

• Correction: Advanced calibration maps multiple known reference points across the sensor's range. A piecewise linear or polynomial correction function (e.g., y_corrected = a*x^2 + b*x + c) is then fitted and applied to linearize the output.
• TinyML Consideration: Implementing higher-order polynomials on MCUs is computationally expensive. Engineers often use lookup tables (LUTs) for efficient, memory-light non-linearity correction.

Cross-axis sensitivity (or axis misalignment) occurs when a force or stimulus applied along one sensor axis produces an erroneous output on another, orthogonal axis. This is common in multi-axis sensors like IMUs (accelerometers, gyroscopes).

• Correction: Calibration involves applying known stimuli along each primary axis and measuring the output on all axes. A misalignment matrix is constructed. The raw sensor vector is multiplied by the inverse of this matrix to decouple the axes.
• Real-World Effect: In a 3-axis accelerometer, tilting the device along the X-axis might incorrectly register a small change on the Z-axis. Calibration removes this crosstalk.

Temperature drift refers to the variation in a sensor's offset, gain, and other parameters caused by changes in ambient temperature. This is a dominant source of error in uncompensated systems operating in non-climate-controlled environments.

• Correction: A temperature calibration characterizes the sensor's performance across its operational temperature range. Correction coefficients (for offset and gain vs. temperature) are stored. During operation, an on-board temperature sensor provides a reading, and the firmware dynamically applies the appropriate corrections.
• TinyML Imperative: Essential for outdoor IoT devices, industrial equipment, and automotive applications where temperature swings are significant.

Hysteresis is an error where the sensor's output depends not only on the current input but also on the history of previous inputs—specifically, the direction from which the input was approached (increasing or decreasing).

• Mechanism: Caused by internal friction, magnetic domains, or material stress in sensor elements. The output for an input of 5 units may differ if approached from 4 units vs. 6 units.
• Correction Challenge: Hysteresis is difficult to correct perfectly in software as it requires modeling the sensor's state history. Calibration often involves characterizing the hysteresis loop and applying a conservative error bound, or selecting sensors with minimal inherent hysteresis for precision applications.

Sensor fusion is the process of combining data from multiple, disparate sensors to produce a more accurate, complete, and reliable estimate of the environment or system state than is possible with any single sensor. It directly depends on well-calibrated sensor inputs.

• Purpose: To overcome the limitations of individual sensors (e.g., noise, drift, blind spots).
• Common Techniques: Include Kalman filters, particle filters, and Bayesian networks.
• TinyML Context: On microcontrollers, fusion algorithms must be extremely lightweight, often using complementary filters to merge, for example, accelerometer and gyroscope data for stable orientation estimation.

A Kalman filter is a recursive, optimal estimation algorithm that uses a series of measurements observed over time, containing statistical noise, to produce estimates of unknown variables. It is a cornerstone algorithm for dynamic sensor calibration and fusion.

• Core Function: It predicts the next state of a system, then corrects that prediction with a new, noisy measurement.
• Two-Step Process: Prediction (forecast) and Update (correction).
• Use in Calibration: Can continuously estimate and correct for sensor bias (offset) and scale factor (gain) errors in real-time, treating them as part of the system's state.

Signal-to-Noise Ratio (SNR) is a fundamental metric that compares the level of a desired signal to the level of background noise, quantifying the clarity and quality of a sensor measurement. Calibration aims to maximize the effective SNR.

• Calculation: Typically expressed in decibels (dB). A higher SNR indicates a cleaner signal.
• Impact of Calibration: Proper calibration removes systematic errors (like offset), which can be considered a form of structured noise, thereby improving the effective SNR.
• TinyML Implication: Low-power sensors on edge devices often have lower inherent SNR, making calibration and intelligent filtering even more critical for reliable feature extraction.

ADC sampling is the process of converting a continuous analog voltage from a sensor into a discrete digital number. The fidelity of this conversion is a prerequisite for all digital calibration and directly limits measurement accuracy.

• Key Parameters: Resolution (bits), sampling rate, and reference voltage stability.
• Calibration Link: Sensor calibration often includes correcting for ADC non-linearities and errors in the voltage reference. A known reference voltage is used to calibrate the ADC's transfer function.
• Nyquist Consideration: Must sample at least twice the signal's highest frequency (Nyquist rate) to avoid aliasing, which calibration cannot fix.

Feature extraction is the process of transforming raw, high-dimensional sensor data into a reduced set of informative, non-redundant values (features) that are suitable for machine learning model input. The quality of these features is wholly dependent on calibrated sensor data.

• Common Features for Sensor Data: Root Mean Square (RMS), Zero-Crossing Rate (ZCR), spectral features from FFT, or Mel-Frequency Cepstral Coefficients (MFCCs) for audio.

• Calibration's Role: Uncalibrated data leads to distorted features. For example, a DC offset in an accelerometer will corrupt the RMS calculation of vibration energy.

• TinyML Optimization: Extracting features on-device (e.g., calculating FFT on an MCU) reduces data transmission; these algorithms must be designed to work with fixed-point arithmetic from calibrated ADCs.

Digital Signal Processing (DSP) is the computational manipulation of discrete-time signals, such as digitized sensor data, using algorithms to filter, analyze, or transform them. Calibration is often implemented as a DSP algorithm.

• Core Techniques: Include filtering (FIR, IIR), transforms (FFT, DCT, Wavelet), and convolution.
• Calibration as DSP: Offset correction is a subtraction (a simple DSP operation). Gain correction is a multiplication. Non-linearity correction may involve a polynomial function or lookup table—all standard DSP constructs.
• Embedded Implementation: On microcontrollers, DSP routines for calibration and filtering are highly optimized, using integer math and circular buffers to manage real-time sensor streams efficiently.