Hybrid neural network

About: Hybrid neural network is a research topic. Over the lifetime, 1305 publications have been published within this topic receiving 18223 citations.

Estimation of Bearing Remaining Useful Life Based on Hybrid Deep Learning Framework

Abstract: Estimation of bearing remaining useful life (RUL) aims to accurately estimate the RUL using sensor data, it is essential for maintaining the safety of the equipment. However, most of the existing prognostic methods based on data-driven have the following two drawbacks: (1) Knowledge of experts are required to select multiple features. (2) The complex spatial-temporal sequence relationship between the sensor data and RUL is difficult to captured. To overcome these drawbacks, a hybrid deep learning framework for construct the health indicator (HI) and predict the RUL is proposed. In the data processing phase, a new feature extraction method based on complex morlet wavelet transform (CMWT) is presented to convert the sensor data to image, and the image is named as the power-spectrum-energy image. In the estimation of HI phase, a new hybrid neural network model based on two conventional neural network (CNN) models that combined with bidirectional gated recurrent unit (BiGRU) network is proposed to capture the spatial-temporal features of the power-spectrum-energy image. Compared to the other methods, the proposed framework can efficiently extract the spatial-temporal features of sensor data and require no prior expert knowledge on prognostic and signal processing. The bearing dataset provided by PRONOSTIA verifies the usability of the framework. The results confirm that the framework can effectively reduce the error value of RUL estimation.

A Hybrid Neural Network/Analog Model for Climate Downscaling

Abstract: Synoptic downscaling models are used in climatology to model local-scale climate variables from synoptic-scale atmospheric circulation variables. This paper presents a hybrid method for multi-site downscaling that combines an artificial neural network and an analog, i.e., k-nearest neighbor, model. The method can resolve complicated synoptic-to local-scale relationships while preserving spatial relationships between sites. Performance on both synthetic and real-world datasets indicates that the hybrid model is capable of outperforming other forms of analog models used in synoptic downscaling.

Adaptive-sized hybrid neural network for segmentation of breast cancer cells in pathology images

Abstract: Akira Hasegawa', Kevin J. Cullen2, Seong K. Mu&1 Georgetown University Medical Center, Department of Radiology2115 Wisconsin Ave., NW, Suite 603, Washington, DC 200072 Georgetown University Medical Center, Division of Medical OncologyABSTRACTIn this report, we describe a novel method to automatically segment several kinds of cells in breast cancerpathologyimages. The information on the number of cells is expected to assist pathologists in consistent diagnosis of breast cancer.Currently, most pathologists make a diagnosis based on a rough estimation of the number of cells on an image. Because of therough estimation, the diagnosis is not objective. To assist pathologists to make a consistent, objective and fast diagnosis, it isnecessary to develop a computer system to automatically recognize and count several kinds of cells. As first step of this purpose,we proposed a novel neural network model, called an adaptive-sized hybrid neural network (ASH-NN), and developed a method,based on this network model, to segment cells from breast cancer pathology images. The proposed neural network consists ofthree layers and the connection weights between the first and second layers are updated by self-organization, and the weightsbetween the second and third layers are determined based on supervised learning. The ASH-NN has the capability of (1) automaticadjustment of the number of hidden units and (2) quick learning.Keywords: neural networks, self-organization, color image segmentation, breast cancer, pathology image1. INTRODUCTIONDiagnosis of breast cancer based on pathology images is commonly done to identify malignancy of suspicious tumors ormicrocalcifications. Fig. 3 (a) shows an example of a breast cancer pathology image. Tissue sections of breast cancer are stained,and magnified by a microscope. In Fig. 3 (a), the rounded cell indicated by the red arrow is an epithelial cell, which is a cancercell, a long cell indicated by the blue arrow is a stromal cell, and the small black dots overall the image are insulin-like growthfactor-Il (IGF-II) mRNAs. Generally, IGF-II mRNA is a potent mitogen for a variety of cell types and is considered an importantregulator of breast cancer growth [1]. This means that cells overlapped with clustered IGF-II mRNAs are active and growing. Inthe treatment of breast cancer, it is important to recognize and count active cancer cells. However, without the help ofcomputers,counting cells is time-consuming work. Currently, most pathologists make a diagnosis based on a rough estimation of thenumber of cells on an image. Because of the rough estimation, the diagnosis is not objective. To assist pathologists to make aconsistent diagnosis, it is necessary to develop a computer system to automatically recognize and count several kinds of cells.The system will provide pathologists with not only a consistent but also fast diagnosis.A study on automatic segmentation of breast tissues by a back propagation neural network [2] has been reported by Okiiet al. [3]. However, they used input images with a notable difference only in brightness, but not color. In addition, the numberof the training examples used for training of a back propagation neural network was so small that the trained neural network'sperformance was not expected to be generalized. Generally, it is said that to get a generalized performance of a back propagationneural network, the necessary number of training examples is more than ten times of the number of connection weights in theneural network [4, 5]; nevertheless, Okii et al. used only 48 training examples for training a neural network with more than 360connection weights.For the image segmentation of general color images, several techniques have been reported. Methods based on thethresholding of image histograms have been proposed [6-8]. In these methods, on each color axis, color space is divided by meansof histogram thresholding, similar to a technique used for the segmentation on gray-scale images. Other techniques based on theK-means algorithm and the least sum of squares criterion have been reported [9, 10]. Recently, a method using neural network hasalso been proposed [1 1]. The basic idea of this method is a vector quantization, which is done by a competitive learning. Imagesegmentation is done by clustering color space by the vector quantization. In these methods, however, the number of segments isusually not specified and depends on the pathology image. On the other hand, in segmentation on pathology images, the numberof segments is assigned.

Wavelets Hybrid Neural Network for Harmonic Sources Recognition with the V-I Characteristics

Abstract: This paper presents a wavelets hybrid neural network (WHNN) for harmonic source recognition with voltage-current (V-I) characteristics. With time-domain technique, WHNN is applied in the application of harmonic sources classification. The proposed hybrid network is a two-subnetwork architecture, consisting of self-organizing feature lattice network (SOFL) and wavelet layer connected in cascade. The first Layer Morlet wavelets, as an extractor, are used to extract the features from voltages and currents, and SOFL is employed to classify the various feature patterns including electronic devices, DC/AC motor, and electric arc furnaces (EAFs) in second layer. The computational results have show that With the time-domain technique, the effectiveness of the proposed model is demonstrated by tests. The recognition results of different harmonic sources show the computational efficiency and accurate recognition.