\subsection{Signal Normalization}
Each raw acceleration time series 
\(\mathbf{a}_{k}(n)\), \(n = 0,1,\dots,N-1\) with \(N=262144\) samples (collected at \(f_s=1024\) Hz over 256 s) :contentReference[oaicite:0]{index=0} is first standardized to zero mean and unit variance:
\[
\tilde a_{k}(n)
=\frac{a_{k}(n)-\mu_{k}}{\sigma_{k}},
\quad
\mu_{k}=\frac{1}{N}\sum_{n=0}^{N-1}a_{k}(n),
\quad
\sigma_{k}=\sqrt{\frac{1}{N}\sum_{n=0}^{N-1}\bigl(a_{k}(n)-\mu_{k}\bigr)^{2}}.
\]

\subsection{Framing and Windowing}
The normalized signal \(\tilde a_{k}(n)\) is chopped into overlapping frames of length \(W\) samples with hop size \(H\). The \(p\)-th frame is
\[
x_{k,p}[m]
=\tilde a_{k}(pH + m)\,w[m],
\quad
m=0,1,\dots,W-1,
\]
where \(w[m]\) is a chosen window function (e.g., Hamming).

\subsection{Short-Time Fourier Transform (STFT)}
For each frame \(x_{k,p}[m]\), compute its STFT:
\[
S_{k}(f,p)
=\sum_{m=0}^{W-1}x_{k,p}[m]\;e^{-j2\pi\,f\,m/W},
\]
where \(f=0,1,\dots,W-1\) indexes frequency bins :contentReference[oaicite:1]{index=1}.

\subsection{Spectrogram and Log-Magnitude}
Form the magnitude spectrogram
\[
M_{k}(f,p)
=\bigl|S_{k}(f,p)\bigr|,
\]
and apply log scaling for numerical stability:
\[
L_{k}(f,p)
=\log\bigl(1 + M_{k}(f,p)^{2}\bigr).
\]
This yields a time–frequency representation
\(\mathbf{L}_{k}\in\mathbb{R}^{F\times P}\), with \(F\) frequency bins and \(P\) frames.

\subsection{Feature Matrix Assembly}
For each column \(j\in\{1,\dots,5\}\), select only the two endpoint sensors:
\[
\mathbf{L}_{\text{bot},j} = \mathbf{L}_{(j)},\quad
\mathbf{L}_{\text{top},j} = \mathbf{L}_{(25+j)},
\]
and stack them:
\[
\mathbf{F}_{j}
=
\begin{bmatrix}
\mathbf{L}_{\text{bot},j} \\[6pt]
\mathbf{L}_{\text{top},j}
\end{bmatrix}
\;\in\mathbb{R}^{2F\times P}.
\]
Finally, flatten into a feature vector:
\[
\mathbf{f}_{j}
=\operatorname{vec}\bigl(\mathbf{F}_{j}\bigr)
\;\in\mathbb{R}^{2FP}.
\]