| 摘要: | 基於卷積神經網路，將孕婦腹部心電訊號進行胎兒心電訊號與孕婦心電訊辨別，使得以非侵入式的方式偵測也能達到高準確度之偵測辨識結果。首先資料前處理需以250毫秒作為樣本分割與10毫秒的重疊區，，並將每個樣本標記，分為4類別，再運算短時距傅立葉轉換，資料轉為時頻二維的特徵圖，以及移除基準線漂移的步驟。而本文之卷積神經網路架構包含兩層卷積層、兩層池化層，與全連接層，最終輸出結果為4類別判別之機率值，其中卷積層包含了激活函數，本文使用Sigmoid函數，輸出層前也透過Softmax函數將資料轉為機率形式。反向傳播中運用損失函數計算誤差值，並用偏導數之運算搭配學習率的設定，將參數進行更新，而迭代次數的設定讓卷積神經網路在次數內進行訓練學習與測試，再加上融合之方式，4條導線的結果得以統一。本文將兩組不同心電訊號進行測試，測試結果最高可達98%之偵測率，偵測率明顯高於KNN演算法與隨機森林演算法。硬體設計方面，本文對其做複雜度評估與量化分析後，決定以Doubling Algorithm設計FFT架構減少乘法器數量並搭配Radix2^2，而卷積神經網路架構採記憶體型態方式做設計，卷積層之運算排程設計和加入區域暫存器，可以讓記憶體讀取寫入次數分別下降約7.2%與3.3%，兩層卷積層與激活函數之查表硬體皆共用，並將查表範圍由2^14下降為2^11，以減少硬體面積，最後並以FPFA進行驗證。;Based on convolutional neural network(CNN), we introduce a recognition and monitoring system for distinguish fetal electrocardiogram(fECG) signals and maternal electrocardiogram(mECG) signals from pregnant abdominal ECG recording, so that non-invasive electrodes detection can also achieve high detection rate. First, the data pre-processing make the abdominal ECG waveform is partitioned into 250ms as a segment with 10ms overlap area. Marking each segment into four different labels. In order to reduce the noise effect of the ECG signals in time domain, short-time Fourier transform is convert the ECG signals into two-dimensional time-frequency feature map. Then, we removing the baseline wandering. In this thesis, CNN architecture consists of two convolutional layers, two pooling layers, and fully connected layer, the final output result are probability value of 4 classes. The convolutional layer contains the activation function, using sigmoid function. Before the output layer, the data are also converted into the probability through the softmax function. In back propagation, calculating the error of the loss function, and the partial derivative with learning rate to update the parameters. In addition to the fusion, four leads record can generate one final result. We use two set of different ECG signal to test, the detection rates of pregnant a02, and a05 are 98.62% and 98.51%, respectively. Compared with the conventional K-nearest neighbor algorithm and random forest, a higher detection accuracy is achieved. For hardware design, after doing the complexity evaluation and quantization, we decided the FFT architecture with doubling algorithm to reduce the number of multipliers and use 〖radix2〗^2. The CNN architecture is designed by memory based, the schedule of the convolutional layer and local buffer design decreases the number of memory accesses to 7.2%, and 3.3%, respectively. The two hardware of convolutional layers are shared. And make the size of LUT reduced from 2^14 to 2^11 to reduce the area. Finally, the overall hardware design verified by FPGA. |
