Finite Impulse Response (FIR) Filter

An FIR filter's output depends only on a finite number of current and past input samples. Its defining equation is: y[n] = b₀x[n] + b₁x[n-1] + ... + bₙx[n-N], where b are the filter coefficients and N is the filter order. This structure guarantees that any impulse input will produce an output that settles to zero within N+1 samples, ensuring inherent stability and a bounded output for any bounded input—a critical safety feature for embedded systems.

A symmetric (or antisymmetric) coefficient arrangement produces a linear phase response, meaning all frequency components of a signal are delayed by the same amount of time. This is essential for applications where waveform shape must be preserved, such as:

• Digital communications (maintaining pulse shape)
• Audio processing (avoiding phase distortion)
• Image processing (preventing edge artifacts) Achieving linear phase with an Infinite Impulse Response (IIR) filter is generally impractical, making FIR the preferred choice for these tasks.

FIR filters have a non-recursive, feedforward-only architecture. The output is a weighted sum of inputs, with no feedback from past outputs. This eliminates issues with limit cycles and overflow oscillations that can plague IIR filters. The lack of feedback loops simplifies analysis, guarantees stability, and makes them highly amenable to pipelining and parallel implementation on digital signal processors (DSPs) and microcontrollers, enhancing throughput.

FIR filters offer precise control over frequency response through coefficient design. The window method is a common technique:

• 1. Specify an ideal frequency response (e.g., a low-pass 'brick wall' filter).
• 2. Compute its infinite-length impulse response via the Inverse Discrete Fourier Transform (IDFT).
• 3. Truncate this infinite sequence to a finite length using a window function (e.g., Hamming, Hanning, Blackman). Windows trade off between main lobe width (transition bandwidth) and side lobe attenuation (stopband ripple), allowing engineers to tailor the filter to specific passband/stopband requirements.

The primary drawback of FIR filters is their higher computational cost compared to IIR filters. To achieve a sharp frequency cutoff (narrow transition band), a high filter order (N) is required, leading to many multiply-accumulate (MAC) operations per output sample. This is a key consideration for TinyML and microcontroller deployment, where memory and CPU cycles are severely constrained. Optimization techniques like symmetric coefficient exploitation (reducing multiplications by ~50%) and polyphase structures for decimation/interpolation are critical for efficient implementation.

FIR filters are ubiquitous in microcontroller-based sensor processing pipelines:

• Noise Removal: A low-pass FIR filter can suppress high-frequency noise from an accelerometer or microphone signal.
• DC Offset Removal: A high-pass FIR filter (e.g., a moving average subtractor) can eliminate unwanted DC bias from sensor readings.
• Feature Preprocessing: Shaping signals before feature extraction for activity recognition or acoustic event detection.
• Anti-Aliasing: Implementing a digital low-pass filter after ADC sampling to prevent aliasing during downsampling operations. Their deterministic latency and stability make them ideal for these real-time, safety-conscious edge applications.

FIR filters are the primary tool for removing high-frequency noise from sensor streams while preserving the underlying signal shape. Their linear phase response prevents distortion of the signal's temporal features, which is critical for time-domain analysis.

• Example: A low-pass FIR filter applied to an accelerometer signal to isolate the low-frequency component of human gait for activity recognition, removing high-frequency vibration noise from motors or footsteps.
• Implementation: Coefficients are designed to attenuate frequencies above a cutoff (e.g., 10 Hz). The fixed, pre-calculated filter taps ensure predictable runtime and memory use on the MCU.

Raw sensor data is often transformed into a cleaner representation before feature calculation. FIR filters act as a preprocessing stage for algorithms extracting Mel-Frequency Cepstral Coefficients (MFCCs), Root Mean Square (RMS), or Zero-Crossing Rate (ZCR).

• Workflow: A microphone's ADC output is first passed through a band-pass FIR filter to isolate the human speech frequency range (300-3400 Hz) before computing MFCCs for a keyword spotting model.
• Benefit: By removing irrelevant frequency bands, the downstream features become more robust and the machine learning model requires less complexity to achieve high accuracy.

Before an analog signal is sampled by an Analog-to-Digital Converter (ADC), it must pass through an anti-aliasing filter. While this is often an analog circuit, a digital FIR filter can provide additional suppression of aliased frequencies in the digital domain.

• TinyML Constraint: The filter must be designed with a steep enough roll-off to meet the Nyquist criterion given the MCU's limited sampling rate.
• Design Trade-off: A sharper filter requires more taps, increasing computation. Engineers balance stop-band attenuation with the MCU's Multiply-Accumulate (MAC) budget.

The standard direct form implementation is straightforward but memory-intensive. TinyML employs optimized structures to reduce resource usage.

• Direct Form I: Stores all N past input samples in a buffer. Requires N multiplications and N-1 additions per output sample.
• Optimized Forms:
  • Symmetric FIR: Exploits coefficient symmetry in linear-phase filters to nearly halve the multiplications.
  • Polyphase Decomposition: Crucial for sample rate conversion, allowing efficient implementation of decimation or interpolation filters.
• Fixed-Point Arithmetic: Coefficients and data are represented as integers, replacing floating-point operations with faster integer MACs.

The choice between FIR and Infinite Impulse Response (IIR) filters is a key architectural decision.

TinyML Verdict: FIR is preferred when linear phase is critical or system stability is paramount, despite its higher computational cost per pole.

FIR filters enable core functionalities in commercial microcontroller deployments.

• Vibration Monitoring: A high-pass FIR filter removes the DC offset and low-frequency drift from a piezoelectric sensor, isolating the high-frequency vibration signatures of bearing faults in industrial motors.
• Heart Rate Monitoring (PPG): A band-pass FIR filter (0.5 Hz - 5 Hz) cleans the photoplethysmogram (PPG) signal from a wrist-worn device, removing motion artifact noise before peak detection calculates beats per minute.
• Wake-Word Detection: In always-on audio systems, a multi-band FIR filter bank decomposes the audio signal into frequency sub-bands, providing cleaner inputs for a small neural network to detect a trigger phrase with minimal false positives.
• Resource Profile: A typical 64-tap FIR low-pass filter running on a 80 MHz ARM Cortex-M4F core can process over 100,000 samples per second while consuming < 5% of the CPU budget, leaving ample headroom for the ML model.

An Infinite Impulse Response (IIR) filter is a type of digital filter whose output depends on both past inputs and past outputs, creating a feedback loop. This structure allows IIR filters to achieve a desired frequency response (e.g., a sharp cutoff) with far fewer coefficients than an equivalent FIR filter, making them computationally efficient.

• Key Trade-off: This efficiency comes with potential downsides: IIR filters can be unstable if their feedback coefficients are not carefully designed, and they generally have a non-linear phase response, which can distort the timing relationships within a signal.
• Common Use: They are often used in applications where phase linearity is not critical but computational resources are extremely limited, such as in audio equalization or simple sensor smoothing on microcontrollers.

A Moving Average (MA) filter is the simplest and most common type of FIR filter. It operates by averaging a fixed number of the most recent input samples to produce each output sample, effectively smoothing the signal by reducing high-frequency noise.

• Implementation: For a window of size N, each output y[n] = (x[n] + x[n-1] + ... + x[n-(N-1)]) / N.
• Characteristics: It has a low-pass frequency response. The width of the averaging window directly controls the cutoff frequency; a larger window provides more smoothing but also introduces greater latency.
• TinyML Relevance: Its computational simplicity (requiring only additions and a single division) makes it a ubiquitous first-step filter for denoising sensor streams (e.g., temperature, pressure) on microcontrollers before feature extraction.

Digital Signal Processing (DSP) is the overarching discipline concerned with the algorithmic manipulation of discrete-time signals. FIR filters are a core algorithmic tool within the DSP toolkit.

• Core Operations: DSP encompasses a wide range of operations beyond filtering, including transformations (FFT, DCT), convolution, correlation, and modulation/demodulation.
• Hardware Context: In TinyML, DSP algorithms are implemented on microcontrollers' CPU cores or, increasingly, on dedicated DSP accelerator units found in modern MCUs. These hardware blocks are optimized for the multiply-accumulate (MAC) operations that are fundamental to FIR filtering and other DSP routines, dramatically improving power efficiency and speed.

Convolution is the fundamental mathematical operation at the heart of an FIR filter. The filter's output signal is computed as the convolution of the input signal with the filter's coefficient sequence (its impulse response).

• Mathematical Definition: For an FIR filter with coefficients b[0]...b[M], the output y[n] = Σ (b[k] * x[n-k]) for k = 0 to M. This is the discrete-time convolution sum.
• Computational Load: This is a multiply-accumulate (MAC)-heavy operation. The efficiency of convolution implementation—often using optimized loops, unrolling, or leveraging SIMD instructions—is critical for real-time sensor processing on edge devices.
• Broader Importance: Convolution is also the core operation in convolutional neural networks (CNNs), creating a direct conceptual link between classical DSP and modern deep learning for sensor data.

A window function (or tapering function) is applied to an ideal, infinitely long filter impulse response to create a practical, finite-length FIR filter. Truncating the ideal response directly causes artifacts like ripple in the passband/stopband (Gibbs phenomenon).

• Purpose: Windowing reduces spectral leakage and controls the trade-off between main lobe width (transition bandwidth) and side lobe attenuation (stopband rejection).
• Common Windows: Different windows optimize for different properties:
  • Hamming/Hann: Good general-purpose balance.
  • Blackman: Higher side lobe attenuation, wider transition.
  • Kaiser: Parameterizable, offering precise control over ripple and attenuation.
• Design Step: Selecting and applying a window function is a standard step in the frequency-sampling method of FIR filter design.

A linear phase response is a key property of many FIR filters, where the phase shift introduced by the filter is a linear function of frequency. This means all frequency components of a signal are delayed by the same amount of time.

• Critical Implication: A constant group delay preserves the shape of waveforms passing through the filter. This is essential in applications where the temporal relationship between signal features must be maintained, such as in:
  • Digital communications (pulse shaping).
  • Biomedical signal processing (ECG, EEG analysis).
  • Audio crossover networks.
• FIR Advantage: Symmetric or antisymmetric FIR filter coefficient sets guarantee exact linear phase. This is a major advantage over IIR filters, which typically have a non-linear phase response that can cause smearing or distortion of complex signals.