Fix _fftcorrelate and tests

src/solartoolbox/signalproc.py

@@ -82,27 +82,77 @@ def correlation(baseline, estimation, scaling='coeff'):
 
 
 def _fftcorrelate(baseline, estimation, scaling='coeff'):
+    """
+    Compute the cross correlation between two signals, including the full range
+    of possible time lags, using cross correlation. Can handle multiple time
+    series.
+
+    Parameters
+    ----------
+    baseline: pd.DataFrame
+    estimation: pd.DataFrame
+    scaling: str (default 'coeff')
+        Type of scaling to use for cross correlation. Note that debiased
+        scaling options are not available here. Valid options are:
+        'coeff' - computes the correlation coefficient
+        'none' - no scaling is applied
+
+    Returns
+    -------
+    corr : numeric
+        A vector of the cross correlations for each pairing in the inputs
+
+    lag : numeric
+        A vector of the lag for each cross correlation
+    """
+
+    # Compute the lags
+    lags = signal.correlation_lags(len(baseline), len(estimation))
+
+    # Rename for brevity
+    ts_inm = baseline
+    ts_outm = estimation
+
+    # Condition pandas input types
+    if isinstance(ts_inm, pd.DataFrame):
+        ts_inm = np.array(ts_inm).T
+    if isinstance(ts_outm, pd.DataFrame):
+        ts_outm = np.array(ts_outm).T
+
+    # Condition vector inputs
+    if ts_inm.ndim == 1:
+        ts_inm = np.expand_dims(ts_inm, axis=0)
+    if ts_outm.ndim == 1:
+        ts_outm = np.expand_dims(ts_outm, axis=0)
+
+    # Perform the scaling calculations
     if scaling == 'coeff':
-        baseline -= np.mean(baseline)
-        estimation -= np.mean(estimation)
-        den = len(baseline) * (np.std(baseline) * np.std(estimation))
+        # Subtract means
+        ts_inm = ts_inm - np.expand_dims(np.mean(ts_inm, axis=1), axis=1)
+        ts_outm = ts_outm - np.expand_dims(np.mean(ts_outm, axis=1), axis=1)
+        # Compute the scaling factor (normalizing by stdev)
+        corr_scale = (np.size(ts_inm, axis=1)
+                      * np.std(ts_inm, axis=1)
+                      * np.std(ts_outm, axis=1)) ** -1
     elif scaling == 'none':
-        den = 1
+        corr_scale = 1
     else:
         raise ValueError(f"Illegal scaling specified: {scaling}.")
 
-    # zero pad by nA+nB-1
-    nl = len(baseline) + len(estimation) - 1
-    baseline = np.pad(baseline, (0, nl-len(baseline)), 'constant')
-    estimation = np.pad(estimation, (0, nl-len(estimation)), 'constant')
-    ffta = np.fft.fft(baseline)
-    fftb = np.fft.fft(estimation)
-    corr = np.real(np.fft.ifft(ffta * np.conj(fftb)))
-    corr = np.roll(corr, nl//2)
+    na = np.size(ts_inm, axis=1)
+    nl = 2 * na - 1
+    addon = np.zeros((ts_inm.shape[0], nl - na))
+    ts_inm = np.concatenate([ts_inm, addon], axis=1)
+    ts_outm = np.concatenate([ts_outm, addon], axis=1)
+
+    # Compute correlation via fft and re-center to match lags
+    ffta = scipy.fft.fft(ts_inm, axis=1)
+    fftb = scipy.fft.fft(ts_outm, axis=1)
+    corrxy = np.real(np.fft.ifft(ffta * np.conj(fftb), axis=1)) * corr_scale
 
-    corr /= den
-    return corr
+    corrxy = np.roll(corrxy, nl // 2, axis=1)
 
+    return corrxy, lags
 
 
 def averaged_psd(input_tsig, navgs, overlap=0.5,

tests/test_signalproc.py

@@ -56,17 +56,19 @@ def test_correlation_illegal(corr_data):
 @pytest.mark.parametrize("scaling", ['coeff', 'none'])
 def test_fftcorrelate_identity(corr_data, scaling):
     dt, t, x1, x2, dly = corr_data
-    c = _fftcorrelate(x1, x1, scaling)
+    c, l = _fftcorrelate(x1, x1, scaling)
     cr, lag = correlation(x1, x1, scaling=scaling)
     assert np.allclose(c, cr)
+    assert np.allclose(l, lag)
 
 
 @pytest.mark.parametrize("scaling", ['coeff', 'none'])
 def test_fftcorrelate_shift(corr_data, scaling):
     d, t, x1, x2, dly = corr_data
-    c = _fftcorrelate(x1, x2, scaling)
+    c, l = _fftcorrelate(x1, x2, scaling)
     cr, lag = correlation(x1, x2, scaling=scaling)
     assert np.allclose(c, cr)
+    assert np.allclose(l, lag)
 
 
 def test_averaged_psd():