spatial

spatial analysis and backup material

spatial latency

Query: resample_1sec. Code: etl_spikes

We analize the latency data depending on the position of the vehicle. We create a geohash per coordinate pair and calculate the average latency

latency_geo latency per geohash

We see a similar pattern per modem upload

latency_geo upload per geohash

The most interesting correlations with camera latency are on the spatial level

geo_correlation spatial correlation

We clearly see that incidents (camera_latency > 400ms) are clustered in space

geo_incident spatial distribution of incidents

We check the cell handover

geo_handover handover cases spatially distributed

Handover is strangely highly correlated but not with camera latency

geo_corr correlation between handover

backup material

Ongoing analysis

long short term memory

Code: stat_reample

We want to asset the performances using a LSTM starting with a baseline of a single layer

We first train a model with 16 fold cross validation and we than substitute each time some random value per feature. The performance of the model with a synthetic random feature should significantly drop for the most important predictors.

lstm_importance performance drop depending on the feature

dictionary learning

We create rolling windows of the time series to see how we can cluster these windows into an essential dictionary of elements.

We start first with 18 clusters composed by series of 16 data points and we create some fundamental clusters

series_dictionay dictionary of time series windows

We make sure that the dimension of the cluster is pretty much orthogonal

dictionary_ortho orthogonality of dictionary