From 93e74a45df46532b9cc71b3fa50ebcd881228edc Mon Sep 17 00:00:00 2001 From: Alexandr Velikiy <39257464+VargBurz@users.noreply.github.com> Date: Tue, 25 Sep 2018 17:45:31 +0300 Subject: [PATCH] Anti-segments in drops model #142 (#173) --- analytics/models/drop_model.py | 45 ++++++++++++++++++++++++++++++++-- 1 file changed, 43 insertions(+), 2 deletions(-) diff --git a/analytics/models/drop_model.py b/analytics/models/drop_model.py index e508ec7..06a7aa5 100644 --- a/analytics/models/drop_model.py +++ b/analytics/models/drop_model.py @@ -23,6 +23,8 @@ class DropModel(Model): 'DROP_HEIGHT': 1, 'DROP_LENGTH': 1, 'WINDOW_SIZE': 240, + 'conv_del_min': 54000, + 'conv_del_max': 55000, } def do_fit(self, dataframe: pd.DataFrame, segments: list) -> None: @@ -75,9 +77,36 @@ class DropModel(Model): labeled_drop = data[self.idrops[n] - self.state['WINDOW_SIZE']: self.idrops[n] + self.state['WINDOW_SIZE'] + 1] labeled_drop = labeled_drop - min(labeled_drop) auto_convolve = scipy.signal.fftconvolve(labeled_drop, labeled_drop) - convolve_jump = scipy.signal.fftconvolve(labeled_drop, self.model_drop) + convolve_drop = scipy.signal.fftconvolve(labeled_drop, self.model_drop) convolve_list.append(max(auto_convolve)) - convolve_list.append(max(convolve_jump)) + convolve_list.append(max(convolve_drop)) + + del_conv_list = [] + for segment in segments: + if segment['deleted']: + segment_from_index = utils.timestamp_to_index(dataframe, pd.to_datetime(segment['from'], unit='ms')) + segment_to_index = utils.timestamp_to_index(dataframe, pd.to_datetime(segment['to'], unit='ms')) + segment_data = data[segment_from_index: segment_to_index + 1] + if len(segment_data) == 0: + continue + flat_segment = segment_data.rolling(window = 5).mean() + flat_segment_dropna = flat_segment.dropna() + pdf = gaussian_kde(flat_segment_dropna) + x = np.linspace(flat_segment_dropna.min() - 1, flat_segment_dropna.max() + 1, len(flat_segment_dropna)) + y = pdf(x) + ax_list = [] + for i in range(len(x)): + ax_list.append([x[i], y[i]]) + ax_list = np.array(ax_list, np.float32) + antipeaks_kde = argrelextrema(np.array(ax_list), np.less)[0] + segment_median = ax_list[antipeaks_kde[0], 0] + cen_ind = utils.intersection_segment(flat_segment.tolist(), segment_median) #finds all interseprions with median + drop_center = cen_ind[0] # or -1? test + segment_cent_index = drop_center - 5 + segment_from_index + deleted_drop = data[segment_cent_index - self.state['WINDOW_SIZE'] : segment_cent_index + self.state['WINDOW_SIZE'] + 1] + deleted_drop = deleted_drop - min(labeled_drop) + del_conv_drop = scipy.signal.fftconvolve(deleted_drop, self.model_drop) + del_conv_list.append(max(del_conv_drop)) if len(confidences) > 0: self.state['confidence'] = float(min(confidences)) @@ -103,6 +132,16 @@ class DropModel(Model): self.state['DROP_LENGTH'] = int(max(drop_length_list)) else: self.state['DROP_LENGTH'] = 1 + + if len(del_conv_list) > 0: + self.state['conv_del_min'] = float(min(del_conv_list)) + else: + self.state['conv_del_min'] = self.state['WINDOW_SIZE'] + + if len(del_conv_list) > 0: + self.state['conv_del_max'] = float(max(del_conv_list)) + else: + self.state['conv_del_max'] = self.state['WINDOW_SIZE'] def do_predict(self, dataframe: pd.DataFrame) -> list: data = dataframe['value'] @@ -133,6 +172,8 @@ class DropModel(Model): conv = scipy.signal.fftconvolve(convol_data, pattern_data) if conv[self.state['WINDOW_SIZE']*2] > self.state['convolve_max'] * 1.2 or conv[self.state['WINDOW_SIZE']*2] < self.state['convolve_min'] * 0.8: delete_list.append(segment) + if max(conv) < self.state['conv_del_max'] * 1.02 and max(conv) > self.state['conv_del_min'] * 0.98: + delete_list.append(segment) else: delete_list.append(segment) # TODO: implement filtering