CN-121997260-A - Bioelectric signal recognition model and reconfigurable hardware accelerator thereof

CN121997260ACN 121997260 ACN121997260 ACN 121997260ACN-121997260-A

Abstract

The application discloses a bioelectric signal identification model and a reconfigurable hardware accelerator thereof, which adopt a multi-view learning method to comprehensively extract characteristic information from a plurality of characteristic views of bioelectric signal data, can more comprehensively capture various characteristics of signals, firstly adopt a deep neural network to learn initial characteristics of three views to extract depth characteristics, secondly fuse the depth characteristics of all the views to form a unified multi-view characteristic representation, finally input the fused multi-view characteristics into a multi-layer perceptron to further learn the characteristics to obtain classification results, and finally finish classification decision of bioelectric signals. The reconfigurable hardware accelerator integrates a reconfigurable computing array, can realize the dynamic multiplexing of hardware resources through a time division multiplexing mechanism, is dynamically configured into a fast Fourier transform computing mode or a neural network reasoning computing mode according to the needs in different computing stages, and does not need to introduce additional computing resources. The application reduces the whole area of the accelerator and improves the utilization ratio of computing resources by the reconfigurable design, and the modules interact data through the data interface and complete instruction transmission through the instruction interface, thereby realizing flexible reconfiguration and ensuring that different computing stages can be efficiently executed in bioelectric signal processing tasks.

Inventors

LI LI
FU YUXIANG
HE SHUZHUAN
GU ZHENGLIN
WANG DAN
ZHANG HENG
Ben chi
LUO YOUBIN
WANG CHENGZHI
PAN HAOCHUAN
HE GUOQIANG

Assignees

南京大学

Dates

Publication Date: 20260508
Application Date: 20260122
Priority Date: 20251121

Claims (6)

1. A bioelectrical signal recognition model, the model comprising: the device comprises a depth time domain feature extraction module, a depth time domain feature extraction module and a depth time domain feature extraction module, wherein the depth time domain feature extraction module is used for extracting depth time domain features from time domain features extracted from original bioelectric signals, and the original bioelectric signals comprise electroencephalogram signals and/or electromyogram signals; The depth frequency domain feature extraction module is used for extracting depth frequency domain features from frequency domain features extracted from the original bioelectricity signals; the depth time-frequency domain feature extraction module is used for extracting depth time-frequency domain features from time-frequency domain features extracted from the original bioelectricity signals; The feature fusion module is used for carrying out self-adaptive weighted fusion on the depth time domain features, the depth frequency domain features and the depth time frequency domain features to obtain multi-view depth features; And the multi-layer perception decision module is used for classifying the multi-view depth features to obtain an identification result.
2. The bioelectrical signal recognition model of claim 1, wherein the feature fusion module is further configured to: And carrying out self-adaptive weighted fusion on the depth time domain features, the depth frequency domain features and the depth time frequency domain features based on the importance weights of the time domain features, the frequency domain features and the time frequency domain features to obtain multi-view depth features.
3. The bioelectrical signal recognition model of claim 1, wherein the model further comprises: The model starting module is used for acquiring the identification result of the single-view identification model on the original bioelectric signal; And calling the depth time domain feature extraction module, the depth frequency domain feature extraction module and the depth time frequency domain feature extraction module to acquire the depth time domain feature, the depth frequency domain feature and the depth time frequency domain feature under the condition that the identification result is the first result.
4. A reconfigurable hardware accelerator deployed with the bioelectrical signal recognition model of any one of claims 1 to 3, the reconfigurable hardware accelerator comprising: The calculation module is used for performing calculation on the input data of the bioelectric signal identification model to obtain a calculation result; The storage module is used for storing input data, intermediate result data, result data and weight data of the bioelectrical signal identification model; The control module comprises a main controller, a data transmission controller and a plurality of algorithm controllers, wherein the algorithm controllers comprise a convolution controller, a full connection controller, a pooling controller, a vector controller and a fast Fourier transformation controller; The main controller is used for generating a control instruction for controlling the calculation module and/or the storage module according to the input instruction configuration information; the data transmission controller is used for storing input data into the storage module and transmitting the data stored in the storage module in different storage areas; The algorithm controller is used for responding to the control instruction of the main controller, executing corresponding calculation and outputting a calculation result to the storage module.
5. The reconfigurable hardware accelerator of claim 4, wherein the computing module comprises: The reconfigurable computing array is used for realizing corresponding computing functions under a specified computing mode and configuration information, wherein the computing mode comprises a real multiplication and addition mode, a fast Fourier transform mode and a multiplication and accumulation tree computing mode; A computing resource controller, comprising: the decoder is used for receiving and decoding the instruction information and extracting the appointed calculation mode and configuration information; A reconstruction controller for reconstructing a data path of the reconfigurable compute array based on a compute mode and configuration information; The finite state machine is used for coordinating and controlling the reconstruction controller to execute reconstruction; An input buffer for receiving the input data; and the output buffer area is used for outputting the calculation result.
6. The reconfigurable hardware accelerator of claim 5, wherein the reconfigurable compute array comprises a comparator array and four compute units, each compute unit comprising 8 multipliers and 8 adders, the comparator array comprising 16 comparators; the reconfiguration controller is further configured to: Reconstructing the calculation unit into 2-path 4-parallel real number multiply-accumulate tree operation or 1-path 8-parallel real number multiply-accumulate tree operation for convolution calculation; and/or interconnecting the two calculation units to reconstruct a 1-way 16-parallel real multiplication accumulation tree operation for full-connection calculation; and/or reconstructing the computing unit into a base 2 fast fourier transform operation for a fast fourier transform computation.

Description

Bioelectric signal recognition model and reconfigurable hardware accelerator thereof The present application claims priority from the chinese patent office, application number 202511718543.1, entitled "bioelectric signal recognition model and reconfigurable hardware accelerator" filed on day 21, 11, 2025, the entire contents of which are incorporated herein by reference. Technical Field The application relates to the technical field of bioelectric signal detection, in particular to a bioelectric signal identification model and a reconfigurable hardware accelerator thereof. Background In the field of bioelectric signal detection, classification and identification of bioelectric signals mainly relate to four key processing stages, namely a data acquisition stage, a signal preprocessing stage, a characteristic extraction stage and a classification stage. In the feature extraction stage and the classification stage, the extracted features are analyzed and modeled by adopting a machine learning or deep learning method, so that the accuracy of bioelectric signal identification can be improved. For bioelectricity signals, the feature extraction mainly comprises three main types, namely time domain feature, frequency domain feature and time-frequency domain feature extraction, and the three features respectively reflect the statistical property, the frequency spectrum feature and the dynamic feature of the bioelectricity signals changing along with time from different angles. When only a single type of feature is extracted for use in constructing the recognition model, important information in other feature types may be ignored, thereby limiting the discriminatory power of the model. Therefore, how to effectively fuse different types of features is a key challenge to improve model performance. In the related scheme, a multi-view fusion method can be adopted to construct a model, wherein one is a pixel-level fusion method, initial features of different views are directly spliced or weighted and summed in a data input stage to form a new fusion feature, although the method can keep original information to the greatest extent, the computational complexity is greatly improved, redundant data can be possibly introduced to cause the great increase of the cost of mode training and reasoning, the other is a decision-level fusion method, decision combination is carried out in the output stage of a plurality of independent models or classifiers, the initial features of each view are firstly subjected to feature extraction and classification decision through independent sub-networks, and finally classification results are obtained through a voting method, a weighted average method and the like, but the mode has certain flexibility, the final decision result is limited by the independent decision capability of each sub-network, and if the decision result of a certain view is unreliable, the overall recognition performance is influenced. Therefore, it is necessary to construct a bioelectrical signal recognition model based on a new fusion strategy. Disclosure of Invention The invention provides a bioelectric signal identification model and a reconfigurable hardware accelerator thereof, which are used for solving the problem of lower identification performance in the prior art. In a first aspect, the present application provides a bioelectrical signal recognition model comprising: the device comprises a depth time domain feature extraction module, a depth time domain feature extraction module and a depth time domain feature extraction module, wherein the depth time domain feature extraction module is used for extracting depth time domain features from time domain features extracted from original bioelectric signals, and the original bioelectric signals comprise electroencephalogram signals and/or electromyogram signals; The depth frequency domain feature extraction module is used for extracting depth frequency domain features from frequency domain features extracted from the original bioelectricity signals; the depth time-frequency domain feature extraction module is used for extracting depth time-frequency domain features from time-frequency domain features extracted from the original bioelectricity signals; The feature fusion module is used for carrying out self-adaptive weighted fusion on the depth time domain features, the depth frequency domain features and the depth time frequency domain features to obtain multi-view depth features; And the multi-layer perception decision module is used for classifying the multi-view depth features to obtain an identification result. In some embodiments, the feature fusion module is further configured to: And carrying out self-adaptive weighted fusion on the depth time domain features, the depth frequency domain features and the depth time frequency domain features based on the importance weights of the time domain features, the frequency domain features and the time frequency domain featu