6 results about "Group frequency" patented technology

Various aspects of the present disclosure relate to group frequency hopping for ambient Internet of Things (AIoT) devices. For example, a non-orthogonal multiple access mechanism for AIoT devices enabling a group of devices to simultaneously and non-orthogonally transmit different transport blocks (TBs) to a receiver device (a base station, UE, and so on). The group of UEs may perform group frequency hopping to transmit data blocks at different frequencies and time slots. A receiver device may then detect the transmissions (e.g., superimposed signals) using multi-user detection techniques.

A method may include receiving a first symbol, and encoding the first symbol based on an initial codebook including a symbol frequency and a ranking for each symbol, and a group frequency for each group of symbols. The method may include incrementing a symbol frequency of the first symbol based on receiving the first symbol, swapping a rank of a lowest ranking symbol from among symbols in the initial codebook that have a symbol frequency lower than the incremented symbol frequency of the first symbol with a rank of the first symbol, in response to the symbol frequency of the lowest ranking symbol being lower than the incremented symbol frequency of the first symbol, and storing an updated codebook in a memory accessible by an encoder and / or a decoder.

The invention discloses a method for measuring interaction strength between two Rydberg atom ensembles. The method comprises the following steps: S1, preparing a first Rydberg atom ensemble A and a second Rydberg atom ensemble B; s2, respectively measuring to obtain a single group energy level frequency shift of the first Rydberg atom ensemble A and a single group energy level frequency shift of the second Rydberg atom ensemble B; s3, measuring to obtain a combined energy level frequency shift in a coupling state; s4, the interaction intensity VAB between the first Rydberg atom ensemble A and the second Rydberg atom ensemble B is obtained through calculation. According to the method, the single-group frequency shift of the two ensembles in the isolated state and the combined frequency shift of the two ensembles in the coupled state are measured respectively, and the interaction intensity between the ensembles is indirectly calculated according to the single-group frequency shift and the combined frequency shift, so that non-destructive measurement which can be dynamically monitored in real time and has high system universality is realized.

The invention discloses a frequency response data change feature generation method and device for transformer winding fault category identification and a storage medium. The method comprises the following steps: respectively converting normal frequency response amplitude-frequency data and frequency response amplitude-frequency data of each fault winding into respective corresponding frequency response amplitude-frequency polar coordinate graphs; calculating multiple types of geometric features of all the frequency response amplitude-frequency polar coordinate graphs, and calculating a difference value between the radius of each point in the frequency response amplitude-frequency polar coordinate graph of each fault winding and the radius of each point in the normal frequency response amplitude-frequency polar coordinate graph; the feature vector for the machine learning model is formed by the multiple types of geometric features and the radius difference values. The equipment and the storage medium are used for implementing the method. The method can provide important reference and guidance for the feature extraction problem in transformer winding fault diagnosis based on a frequency response analysis method and a machine learning method.

The invention discloses a packet aggregation approximate query processing method under local differential privacy, and belongs to the technical field of approximate query processing and privacy protection. The method is divided into two stages of sampling rate optimization and packet aggregation approximate query processing: firstly dividing user subsets, locally disturbing part of user data through a packet distribution preserving disturbance mechanism (GDP), estimating the frequency, the mean value and the variance of each packet in an unbiased manner by a server according to the local disturbance, calculating an error comprehensive coefficient and deducing the optimal sampling rate of each packet; after remaining users are disturbed by a PCKV-GRR mechanism, adaptive sampling is carried out according to the optimal sampling rate, and the server estimates and outputs an aggregation query result based on sampling data. According to the method, grouping statistics joint unbiased estimation is realized, grouping estimation errors are balanced, the privacy leakage is avoided by firstly disturbing and then sampling, the accuracy and efficiency of grouping aggregation query are improved while the local differential privacy is met, and the method is suitable for a large-scale grouping aggregation analysis scene of sensitive data.

This invention provides a method for identifying long-context attention heads based on static parameters of a large language model. The method includes: performing rotational position encoding on the static parameters of positions within a multi-head attention layer of the large language model; dividing each attention head within the multi-head attention layer into multiple frequency components to determine the frequency kernel of the attention head at the relative distance between positions; determining the effective rank of the frequency kernel based on its energy distribution across different frequency bands; determining the frequency energy distribution score and effective rank score corresponding to the energy distribution; and determining the long-context potential score of each attention head at the relative distance based on the frequency energy distribution score and effective rank score. This allows for the selection of attention heads with the potential to identify long-context modeling before deploying the large language model. This invention can improve the inference performance of the model in long-context tasks without changing the model structure or retraining the model.