2 results about "Primary clustering" patented technology

This invention discloses a sequence recommendation method based on multi-scale encoding and capsule intent purification. The method acquires and preprocesses user dynamic interaction sequences, and extracts logically related sequence pairs based on the principle of prediction target consistency. Using a multi-scale segment mapping method, the preprocessed user dynamic interaction sequences are deconstructed into behavioral segments of different granularities, and features are extracted and weighted to form a fused sequence representation. The generated fused sequence representation is then subjected to primary clustering, and the dynamic routing mechanism of the capsule network is used to perform nonlinear mapping and semantic enhancement on the cluster centers, generating a purified high-order intent capsule set. A joint loss function is constructed to uniformly optimize the main recommendation task and the multi-dimensional self-supervised auxiliary task. The parameters of the multi-scale encoder, capsule network, and item embedding matrix are synchronously updated using the backpropagation algorithm until the recommendation model converges, resulting in a trained recommendation model. Recommended sequences are then obtained based on the trained recommendation model.

The application discloses a partial discharge diagnosis method based on regional clustering, which firstly extracts pulse repetitive discharge spectrum data and corresponding labels in a partial discharge training set and stores the data as a mapping matrix. Through a primary clustering function, corresponding selected region coordinate matrices are obtained. The selected region coordinate matrices are classified according to the labels, and the result is stored in a clustering data dictionary. The labels and the selected region coordinate matrices in the clustering data dictionary are iterated, and are sent into a secondary clustering function to obtain final diagnosis region coordinate matrices. Whether the collected PRPD spectrum data to be diagnosed is faulty is judged in combination with a discharge frequency threshold, a corresponding fault occurrence probability is calculated, and finally a partial discharge fault diagnosis result containing the fault occurrence probability is generated.