Research On CNN-based ECG Classification Algorithm And Energy-Efficient Acceleration Architecture

Heart disease is a serious threat to human health.As most heart diseases are characterized by intermittent attacks with low frequency,long-term electrocardiogram(ECG)analysis has become an important means of detecting heart disease clinically.With the popularity of smart mobile devices and the development of artificial intelligence,it has become possible to use mobile devices to automatically classify ECG in real-time.With the smart mobile devices as the carrier,high-performance inter-patient and patient-specific real-time ECG classification methods based on Convolutional Neural Network(CNN)are proposed,which focus on solving the problems of high complexity of data preprocessing and low detection accuracy of arrhythmia.Furthermore,considering that the calculation amount and computational energy consumption of CNN are generally large,this dissertation studies and proposes a CNN accelerator architecture,which effectively improves the operation performance of the algorithms and reduces the power consumption.Besides,to fully fit the multi-application scenario of smart mobile devices,the proposed accelerator architecture also supports the acceleration of a variety of common CNN models.The main contents and characteristics of this dissertation are as follows:1.Research of ECG Classification method based on feature-synthesis inputs and multi-resolution neural network.For inter-patient ECG classification,a CNNbased method exploiting multi-resolution using atrous convolution is proposed.ECG morphological features,RR-interval and historical classified information are combined to form the informative network inputs,and then filters with different receptive fields are used to extract multi-resolution features automatically,thus helping the network to recognize different classes of heartbeats.Evaluated with MIT-BIH arrhythmia database,this method achieves a detection accuracy of 92.9% in arrhythmia classes.2.Research of ECG Classification method based on folding morphological features and channel-wise attentional neural network.For patient-specific ECG classification,a method exploiting CNN with channel-wise attention is proposed.Twochannel ECG segments are folded into six-channel segments as network inputs,and parallel convolution layers are used to discretize and combine multi-channel features.Furthermore,the channel-wise attention model automatically emphasize the informative features that help distinguish different classes of heartbeats and thus significantly enhance the classification performance.Validated on the MIT-BIH arrhythmia database,this method demonstrates a classification accuracy of 98.0%.3.Research of energy-efficient CNN accelerator architecture suitable for smart mobile devices.TruthSeek,an energy-efficient instruction-driven accelerator architecture,is proposed,which can effectively support the acceleration of various CNN models including the ECG classification algorithms.TruthSeek realizes the merging operation of convolution and pooling layers by customizing the novel instruction set and register,thus effectively reducing the power consumption of memory access.At the same time,network sparsity is exploited to further reduce operation cycles and improve performance.Validated on 5 different types of CNN models with calculations ranging from 0.4M to 3.8G,TruthSeek accelerator only requires an average of 0.057 to 65.6ms for every inference,and obtains a task efficiency of 0.06 to 0.42 GOPS/mW,which indicates its application potential in the field of mobile devices.The key methods and technologies proposed in this dissertation have positive research significance and application value for solving the real-time ECG classification problem required by heart disease monitoring in mobile scenes.And the proposed CNN acceleration architecture has certain reference significance for accelerators in other fields.