Fundamentals Of Digital Signal Processing Using Matlab

Getting Started With DSP In Matlab

You open MATLAB and your first task is to generate a signal, apply a filter, and look at the result. That's how most courses begin, and it stays that way because it's straightforward until it isn't. Digital signal processing in MATLAB boils down to three operations you'll repeat constantly: creating or loading a time-domain signal, designing a filter, and applying that filter. Everything else is incremental complexity built on top of that loop. Start by understanding how MATLAB stores signals. A digital signal is just a vector. If you sample a sine wave at 1000 Hz for one second, your vector has 1000 elements. That simplicity is both the gift and the curse. There's no automatic sampling rate attached to the data. You carry that knowledge yourself, and if you lose track of it, every downstream calculation gets garbage results without any warning.

Fundamentals Of Digital Signal Processing Using Matlab

The core workflow begins with the sampling theorem. If your signal contains frequency content above half your sampling rate, aliasing occurs. MATLAB will not tell you this. It will happily process your aliased data and produce a clean-looking result that is completely wrong. I spent two days chasing a phantom noise source in a biomedical signal once, only to realize the analog anti-aliasing filter on the acquisition board had failed. The DSP side was fine. The data was garbage from the start. This happens more often than you want to admit. When you design filters, you choose between FIR and IIR structures. FIR filters are straightforward: you specify the order and the cutoff, call the appropriate design function, and MATLAB gives you coefficients. They are inherently stable and can achieve linear phase, which preserves the waveform shape. The tradeoff is that FIR filters require significantly higher orders to achieve sharp transitions. A 5th order IIR Butterworth filter might do what a 200th order FIR filter does, and that difference matters when you're processing long recordings or running on embedded hardware. The freqz function is your primary tool for inspecting filter behavior. It computes the frequency response of a digital filter given its numerator and denominator coefficients. Pass it your b and a vectors, and it returns the complex frequency response. Plot the magnitude and phase separately. Always check both. A filter can have a perfect magnitude response and terrible phase distortion, which will ruin transient signals like audio clicks or ECG spikes. Linear phase is the solution for those cases, and FIR filters are your route to it.

For practical implementation, designfilt is the most versatile function available. It creates a filter object that you can inspect, modify, and apply without manually tracking coefficient arrays. Here is what a typical design looks like in practice: Create a lowpass filter with a 500 Hz cutoff and a 2000 Hz sampling rate, allowing 100 Hz of transition bandwidth and 60 dB of stopband attenuation. The function returns a digital filter object. Apply it using filter or, better yet, filtfilt for zero-phase filtering when you are analyzing data rather than deploying in real time. filtfilt applies the filter forward and then backward, effectively squaring the magnitude response and eliminating phase distortion entirely. The cost is that it requires the entire signal to be present upfront, so it is useless for streaming or real-time applications. Spectral analysis is where beginners consistently trip up. The fft function is not magical. It computes the discrete Fourier transform, which assumes your signal is periodic and finite. If your signal does not complete an integer number of cycles within your analysis window, spectral leakage occurs. The energy from a single frequency spreads across adjacent bins, creating artifacts that look like real signal content. Window functions reduce this problem. Multiply your signal by a Hanning or Blackman-Harris window before taking the FFT. You lose some frequency resolution in exchange for dramatically reduced leakage. Choose your window based on what you care about: resolution or dynamic range.

One thing that is not obvious when you are learning: the FFT output is symmetric. For a real-valued input signal, the second half of the FFT result is the complex conjugate of the first half. You only need to look at the first N/2+1 points. The magnitude spectrum is what you plot against frequency. Convert the bin index to actual frequency by multiplying by the sampling rate and dividing by the FFT length. This conversion is simple but easy to skip, and skipping it means your frequency axis is wrong. Convolution and correlation are related but distinct operations. Convolution flips one signal before sliding it across the other. Correlation does not flip it. MATLAB provides conv for convolution and xcorr for cross-correlation. Use convolution when you want to apply a filter or compute the output of a linear system. Use correlation when you want to measure similarity between two signals or find the time delay between them. I once used conv instead of xcorr trying to align two audio recordings, and spent an hour wondering why the alignment was backwards before I caught the flip in my head. Quantization effects are another area that textbooks gloss over. When you represent analog signals with finite bit depths, you introduce quantization noise. A 16-bit ADC gives you approximately 96 dB of dynamic range. If your signal uses only 8 bits of that range, your effective dynamic range is 48 dB, and the quantization noise floor sits much higher than the theoretical limit. MATLAB's quantiz function or bit-shifting operations can model this. In practice, if you are processing already-quantized data, the noise floor is baked in. You cannot recover what was lost. The only mitigation is to ensure adequate headroom during acquisition or to use dithering before quantization.

The resample function handles sample rate conversion, which combines interpolation and decimation in a single call. This is convenient, but it is also easy to misuse. If you downsample without proper anti-aliasing filtering, high-frequency content folds back into the passband. resample applies an internal anti-aliasing filter by default, which is good, but the default parameters may not be optimal for your signal. Check the resulting frequency response with freqz after resampling to verify that nothing unexpected crept in. For adaptive filtering, MATLAB provides adaptfilt objects and functions like lms and nlms. These adjust filter coefficients in real time to track changing signal characteristics. They are useful for noise cancellation and system identification. The step size parameter controls convergence speed versus steady-state error. Larger step sizes converge faster but introduce more misadjustment noise. Smaller step sizes are more stable but may never track rapid changes. There is no universal optimum. You tune it for your specific application through trial and measurement. Real-world data is messy. You will encounter DC offsets, missing samples, non-stationary signals, and clipping. A DC offset in your signal appears as a large spike at zero frequency in the FFT. It is usually harmless for filtering but can dominate your spectral view and mask lower-amplitude features. Remove it with detrend before spectral analysis, or use a highpass filter if the offset is not of interest. Missing samples are harder. Most MATLAB functions assume uniformly spaced data. If you have gaps, you need to interpolate or use methods designed for irregular sampling, such as the Lomb-Scargle periodogram available through third-party toolboxes.

One specific edge case I encountered involved processing accelerometer data sampled at 10 kHz. I designed a 4th order Butterworth bandpass filter from 10 to 500 Hz using butter and filter. The results looked reasonable until I compared them against a hardware-filtered reference signal. The phase response of the IIR filter was introducing measurable time shifts in transient events. Switching to filtfilt eliminated the phase distortion, but I lost the ability to process the data in real time on the target embedded platform. The workaround was to design a linear-phase FIR filter with firls and use it for both offline analysis and the embedded implementation. The FIR filter was longer, but the phase linearity meant the transient timing was preserved exactly. MATLAB is excellent for prototyping and offline analysis. It is not ideal for production embedded deployment unless you export your filters carefully. The coder toolbox can generate C code from MATLAB DSP algorithms, but the generated code is often inefficient and difficult to maintain. For anything that runs on actual hardware, write the filter directly in C or use a dedicated DSP library. MATLAB's strength is in the design and verification stage, not in the final implementation. When debugging DSP code, the most useful technique is to trace the signal through each stage and inspect intermediate results. Plot the time domain signal, then the spectrum, then the filtered time domain, then the filtered spectrum. Mismatches between expected and actual behavior at any stage narrow the problem quickly. Do not skip the time domain plots. A filter that looks correct in the frequency domain can still produce ugly artifacts in time due to Gibbs phenomenon or poor stopband attenuation.

The pwelch function is generally preferable to a raw FFT for power spectral density estimation. It segments the signal, applies a window to each segment, computes the FFT of each segment, and averages the results. This reduces variance at the cost of frequency resolution. The parameters you control are segment length, overlap, and window type. Longer segments give better resolution but higher variance. More overlap smooths the estimate further. A Hanning window with 50% overlap is a reasonable default for most applications. Signal generation functions like sinc, cos, sawtooth, and square are building blocks you will use constantly. The sinc function deserves special attention because it is the ideal interpolation kernel. When you upsample a signal, the mathematically correct reconstruction involves convolving with a sinc function. MATLAB's upsample function does zero-insertion followed by a lowpass filter, which approximates sinc interpolation. For most practical purposes this is sufficient, but if you need high-quality resampling, use resample with carefully chosen parameters instead of manual upsampling. Memory management matters more than you might expect. Large FFTs on long signals can consume significant RAM. A 10 million sample float64 vector takes 80 MB. If you are running multiple analyses in a loop, do not let intermediate results accumulate. Clear variables you no longer need. Use realfft if your input is real-valued, since it returns only the positive frequency components and cuts memory usage roughly in half.

Finally, documentation in MATLAB is genuinely useful. Type doc fft or help filter and you get detailed explanations with examples. The Signal Processing Toolbox documentation includes design recommendations and common pitfalls for most functions. Read it before you assume a function behaves the way you expect. Different versions of MATLAB sometimes change default parameters or behavior, so what worked in one release may behave differently in another. Pin your MATLAB version if reproducibility matters for your work.