This paper presents an efficient approach for subsequence search in datastreams. The problem consists in identifying coherent repetitions of a givenreference time-series, eventually multi-variate, within a longer data stream.Dynamic Time Warping (DTW) is the metric most widely used to implement patternquery, but its computational complexity is a well-known issue. In this paper wepresent an approach aimed at learning a kernel able to approximate DTW to beused for efficiently analyse streaming data collected from wearable sensors,reducing the burden of computation. Contrary to kernel, DTW allows forcomparing time series with different length. Thus, to use a kernel, a featureembedding is used to represent a time-series as a fixed length vector. Eachvector component is the DTW between the given time-series and a set of 'basis'series, usually randomly chosen. The vector size is the number of basis seriesused for the feature embedding. Searching for the portion of the data streamminimizing the DTW with the reference subsequence leads to a globaloptimization problem. The proposed approach has been validated on a benchmarkdataset related to the identification of users depending on their walkingactivity. A comparison with a traditional DTW implementation is also provided.