It is difficult to diagnose the faults of large-scale digital-analog hybrid circuit because it has numerous fault modes, the circuit failure status is complex and can be propagated easily. To solve these problems, a new failure diagnosis method, namely Modularized Back Propagation (BP) neural network based on Fault Propagation (MBPFP), was proposed. Firstly, fault propagation between subcircuits was analyzed on the basis of circuit module division, and failure source and transmission source were modularized. Secondly, the set of fault causes was narrowed and the fault module was determined by the anomaly detection model of subcircuit in 1-order positioning. Finally, the fault location was realized and the fault mode was identified by the BP neural network of target module in 2-order positioning. The experimental results show that compared with the traditional BP neural network method, the proposed MBPFP method has a high fault coverage and the accuracy is improved by at least 8 percentage points, which is outperforms the traditional method based on BP neural network.

Focusing on the issues of poor real-time, low classification recognition rate and less features of the crowd abnormal detection, an abnormal behavior detection algorithm of small and medium crowd based on intelligent video surveillance was proposed. Firstly, the rapid population density detection algorithm was employed to extract the change information of crowd amount. Secondly, the improved Lucas-Kanade optical flow method was utilized to extract the average kinetic energy, the direction entropy and the distance potential energy of the crowd. Finally, the crowd behaviors were classified by using the Extreme Learning Machine (ELM) algorithm. UMN common data set was used for test, compared to abnormal crowd behavior detection algorithm in high and medium density and abnormal behavior detection algorithm based on Kinetic Orientation Distance (KOD) energy feature, the recognition rate of ELM algorithm in abnormal behavior detection of small and medium crowd increased by 7.13 percentage points and 5.89 percentage points respectively. On the part of the crowd density estimation, compared to the high and medium crowd density detection algorithm, the processing time for each frame of ELM algorithm reduced 106 ms almost 1/3, approximately. The experiments show that the proposed abnormal behavior detection of small and medium crowd based on intelligent video surveillance can effectively improve recognition rate and real-time performance of the abnormal behavior detection.

The Synthetic Aperture Radar (SAR) image uses Support Vector Machine (SVM) and Markov Random Field (MRF) or Conditional Random Field (CRF) to classify based on feature extraction of coarsely segmented pixel blocks. The traditional method exists the deviation issue of different type pixels inside the same pixel block and it only considers the adjacent area without using global information and structure information. Fully Convolutional Network (FCN) was introduced to solve the deviation problem, and the original classification probability of pixel was gotten by constructing convolutional layers based on pixel level for sample training and using ESAR images as samples. Then CRF-Recurrent Neural Network (CRF-RNN) was introduced as post layer to combine the original classification probability obtained by FCN with full image information transfer and structure information, which was produced by CRF structure. Finally, the RNN iteration was used to further optimize the experimental results. By taking advantages of global information and structure information, the proposed method based on pixel level solved some disadvantages of the traditional classification. The classification accuracy rate of the proposed method was improved by average 6.5 percentage points compared with SVM or CRF. The distance weight of CRF-RNN is fitted by Gaussian kernel, which can not be changed or determined according to the training data, thus it remains some deviation. So a convolutional network based on trainable full image distance weight was proposed to improve CRF-RNN. The experiment results show that the classification accuracy rate of the improved CRF-RNN is further improved by 1.04 percentage points.

The sound can be accurately located by human auditory system in noisy environment. The main element to realize the location is interaural time difference. But the effects are unsatisfactory when using interaural time difference to locate in noisy environment. In order to resolve this problem, this thesis put forward a sound source locating system based on cochlear nucleus model, and cochlear nucleus model simulated the process of how cochlear deals with auditory information. The process could draw the synchronization information and firing rate from the reaction of auditory nerve fibers to sound, thus realizing the inhibitory of noise, and locating the sound source in noisy environment. The location error of system in noisy environment was 1.297 degrees. The experimental results show that the improved sound locating system can complete locating in noisy environment.

The traditional Web services selection algorithms were analyzed and the problems existing in dynamic environment were pointed out. A personalized Web services selection method based on collaborative filtering was proposed to address these problems. And a personalized Web service selection framework was designed, which used the collaborative filtering to predict the Quality of Service (QoS) and selected the best service that met users' requirements. About 1.5 million real world QoS data were employed to evaluate the proposal with other four methods and the experimental results demonstrate that the proposed method is a feasible manner and it provides better prediction results.

In order to save the energy in using the parallel computer cluster, a scheme was presented in which Open Portable Batch System （OpenPBS） was used to control the parallel computers' boot and shutdown automatically. Homeostasis can be achieved between task computing and energy saving.