Aero-engine damage is an important factor affecting flight safety. There are two main problems in the current computer vision-based damage detection of engine borescope image:one is that the complex background of borescope image makes the model detect the damage with low accuracy; the other one is that the data source of borescope image is limited, which leads to fewer detectable classes for the model. In order to solve these two problems, a Mask R-CNN (Mask Region-based Convolutional Neural Network) based progressive detection network from binary classification to multiple classification was proposed for aero-engine damage images. By adding a binary classification detection branch to the Mask R-CNN, firstly, the damage in the image was detected in binary way and regression optimization was performed to the localization coordinates. Secondly, the original detection branch was used to progressively perform multiple classification detection, so as to further optimize the damage detection results by regression and determine the damage class. Finally, instance segmentation was performed to the damage through the Mask branch according to the results of multiple classification detection. In order to increase the detection classes of the model and verify the effectiveness of the method, a dataset of 1 315 borescope images with 8 damage classes was constructed. The training and testing results on this set show that the Average Precision (AP) and AP75 (Average Precision under IoU (Intersection over Union) of 75%) of multiple classification detection are improved by 3.34% and 9.71% respectively, compared with those of Mask R-CNN. It can be seen that the proposed method can effectively improve the multiple classification detection accuracy for damages in borescope images.

Most of the traditional network representation learning methods cannot consider the rich structure information and attribute information in the network at the same time, resulting in poor performance of subsequent tasks such as classification and clustering. In order to solve this problem, an Auto-Encoder based Multi-View Attributed Network Representation learning model(AE-MVANR) was proposed. Firstly, the topological structure information of the network was transformed into the Topological Structure View(TSV), and the co-occurrence frequencies of the same attributes between nodes were calculated to construct the Attributed Structure View(ASV). Then, the random walk algorithm was used to obtain a series of node sequences on two views separately. At last, by inputting all the generated sequences into an auto-encoder model for training, the node representation vectors that integrate structure information and attribute information were obtained. Extensive experiments of classification and clustering tasks on several real-world datasets were carried out. The results demonstrate that AE-MVANR outperforms the widely used network representation learning method based solely on structure information and the one based on both network structure information and node attribute information. In specific, for classification results of the proposed model, the maximum increase of accuracy is 43.75%, and for clustering results of the proposed model, the maximum increase of Normalized Mutual Information(NMI) is 137.95%, the maximum increase of Silhouette Coefficient is 1 314.63% and the maximum decrease of Davies Bouldin Index(DBI) is 45.99%.

The number of defect region samples is far less than that of the normal region samples in aircraft engine borescope inspection image defect detection task, and the defect samples cannot cover the whole sample space, which result in poor generalization of the detection algorithms. In order to solve the problems, a new interactive data augmentation method based on style transfer technology was proposed. Firstly, background image and defect targets were selected according to the interactive interface, and the informations such as size, angle and position of the target needed to be pasted were specified according to the background image. Then, the style of background image was transferred to the target image through style transfer technology, so that the background image and the target to be detected had the same style. Finally, the boundary of the fusion region was modified by Poisson fusion algorithm to achieve the effect of natural transition of the connected region. Two-class classification and defect detection were conducted to verify the effectiveness of the proposed method. The testers achieve 44.0% classification error rate for the two-class classification on the dataset with real images and augmented images averagely. In the detection task based on Mask Region-based Convolutional Neural Network (Mask R-CNN) model, the proposed method has the Average Precision (AP) of classification and segmentation improved by 99.5% and 91.9% respectively compared to those of the traditional methods.

The traditional category-driven approaches only consider the association between categories or organize them into flat or hierarchical structure, but the relationships between items and categories are complex, making other information be ignored. Aiming at this problem, a random walk recommendation algorithm based on combinational category space was proposed to better organize the category information of items and alleviate data sparsity. Firstly, a combinational category space of items represented by Hasse diagrams was constructed to map the one-to-many relationship between items and categories into one-to-one simple relationships, and represent the user's jumps between items in higher and lower levels, the same level and the cross-levels. Then the semantic relationships and two types of semantic distances - the links and the preferences were defined to better describe the changes of the user's dynamic preferences qualitatively and quantitatively. Afterwards,the user personalized category preference model was constructed based on random walking and combination of the semantic relationship, semantic distance, user behavior jumping, jumping times, time sequence and scores of the user's browsing graph in the combinatorial category space. Finally, the items were recommended to users by collaborative filtering based on the user's personalized category preference. Experimental results on MovieLens dataset show that compared with User-based Collaborative Filtering (UCF) model and category-based recommendation models (UBGC and GENC), the recommended F1-score was improved by 6 to 9 percentage points, the Mean Absolute Error (MAE) was reduced by 20% to 30%; compared with Category Hierarchy Latent Factor (CHLF) model, the recommended F1-score was improved by 10%. Therefore, the proposed algorithm has advantage in ranking recommendation and is superior to other category-based recommendation algorithms.

Spy technology in Positive and Unlabeled (PU) learning is easily susceptible to noise and outliers, which leads to the impurity of reliable positive instances, and the mechanism of selecting spy instances in the initial positive instances randomly tends to cause inefficiency in dividing reliable negative instances. To solve these problems, a PU learning framework combining new spy technology and semi-supervised self-training was proposed. Firstly, the initial labeled instances were clustered and the instances closer to the cluster center were selected to replace the spy instances. These instances were able to map the distribution structure of unlabeled instances effectively, so as to better assist to the selection of reliable negative instances. Then, the reliable positive instances divided by spy technology were purified by self-training, and the reliable negative instances which were divided as positive instances mistakenly were corrected by secondary training. The proposed framework can solve the problem of PU learning that the classification efficiency of traditional spy technology is susceptible to data distribution and random spy instances. The experiments on nine standard data sets show that the average classification accuracy and F-measure of the proposed framework are higher than those of Basic PU-learning algorithm (Basic_PU), PU-learning algorithm based on spy technology (SPY), Self-Training PU learning algorithm based on Naive Bayes (NBST) and Iterative pruning based PU learning (Pruning) algorithm.

Shuffle performance is an important indicator of affecting cluster performance for big data frameworks. The Shuffle memory allocation algorithm of Spark itself tries to allocate memory evenly for every Task in the memory pool, but it is found in experiments that the memory was wasted and the efficiency was low due to the imbalance of memory requirements for each Task. In order to solve the problem, an improved Spark Shuffle memory allocation algorithm was proposed. According to the amount of memory applications and historical operating data, the Task was divided into two categories based on memory requirements. The "split"processing was carried out for the Task of small memory requirements, while the memory was allocated based on the number of Task overflows and the waiting time after overflow for the Task of large memory requirements. By taking full advantage of the free memory of memory pool, the adaptive adjustment of Task memory allocation could be realized under the condition of unbalanced Task memory requirements caused by the data skew. The experimental results show that, compared with the original algorithm, the improved algorithm can reduce the overflow rate of the Task, decrease the turnaround time of the Task, and improve the running performance of the cluster.

Agent organization can be used for multi-agent problem solving so as to potentially decrease the difficulty of problem solving and complexity of interaction, and the organization structure, organization formation rule and object planning are the key problems among multi-agent cooperation. In this paper, a multi-agent organization model and a set of rules based on the model were presented and the method about multi-agent object planning and the coherence and perfection of the planning were described. Based on the above study, the formation procedure and the release procedure of multi-agent organization were presented, and all the procedures were formally described by extended temporal logic and pi-calculus. All these improved the study of multi-agent theory and programming practice which was not closely combined.