Patent price evaluation is an important part of intellectual property right transactions. When evaluating patent prices, the impact of the market, law, and technical dimensions on patent prices was not considered effectively by the existing methods. And the market factor of patent plays an important role in the evaluation of patent prices. Aiming at the above problem, an automatic patent price evaluation method based on recurrent neural network was proposed. In this method, based on the market approach, various other factors were considered comprehensively, and the Gated Recurrent Unit (GRU) neural network method was used to realize the automatic evaluation of patent prices. Example tests show that, with the qualitative evaluation results of experts as the benchmark, the average relative accuracy of the proposed method is 0.85. And this average relative accuracy of the proposed method is increased by 3.66%, 4.94% and 2.41% of the average relative accuracies of Analytic Hierarchy Process (AHP), rough set theory method and Back Propagation (BP) neural network method respectively.

In the capacity prediction process of telecom operation and maintenance, there are problems of too many capacity indicators and deployed business classes. Most of the existing researches do not consider the difference of indicator data types, and use the same prediction method for all types of data, which results in both good and bad prediction effects. In order to improve the efficiency of indicator prediction, a classification method of data type was proposed, and the data types were divided into trend type, periodic type and irregular type. Aiming at the prediction of periodical data, a periodic capacity indicator prediction model based on Bi-directional Recurrent Neural Network (BiRNN), called BiRNN-BiLSTM-BI, was proposed. Firstly, In order to analyze the periodic characteristics of capacity data, a busy and idle distribution analysis algorithm was proposed. Secondly, a Recurrent Neural Network (RNN) model was built, which included a layer of BiRNN and a layer of Bi-directional Long Short-Term Memory network (BiLSTM). Finally, the output of BiRNN was optimized by the system's busy and idle distribution information. Experimental results compared with the best one among Holt-Winters, AutoRregressive Integrated Moving Average (ARIMA) model and Back Propagation (BP) neural network model show that, the proposed BiRNN-BiLSTM-BI model has the Mean Square Error (MSE) reduced by 15.16% and 45.67% on the unified log dataset and the distributed cache service dataset respectively, showing that the prediction accuracy is greatly improved.

The abundant structure and association information contained in Knowledge Graph (KG) can not only alleviate the data sparseness and cold-start in the recommender systems, but also make personalized recommendation more accurately. Therefore, a knowledge graph driven end-to-end recommendation model of graph neural network, named KGLN, was proposed. First, a signal-layer neural network framework was used to fuse the features of individual nodes in the graph, then the aggregation weights of different neighbor entities were changed by adding influence factors. Second, the single-layer was extended to multi-layer by iteration, so that the entities were able to obtain abundant multi-order associated entity information. Finally, the obtained features of entities and users were integrated to generate the prediction score for recommendation. The effects of different aggregation methods and influence factors on the recommendation results were analyzed. Experimental results show that on the datasets MovieLen-1M and Book-Crossing, compared with the benchmark methods such as Factorization Machine Library (LibFM), Deep Factorization Machine (DeepFM), Wide&Deep and RippleNet, KGLN obtains an AUC (Area Under ROC (Receiver Operating Characteristic) curve) improvement of 0.3%-5.9% and 1.1%-8.2%, respectively.

In image caption methods, image feature information is not completely extracted and the vanishing gradient is generated by the Recurrent Neural Network (RNN). In order to solve the problems, a new image captioning algorithm based on multi-feature extraction was proposed. The constructed model was consisted of three parts:Convolutional Neural Network (CNN) was used for image feature extraction, ATTribute extraction model (ATT) was used for image attribute extraction, and Bidirectional Long Short-Term Memory (Bi-LSTM) network was used for word prediction. In the constructed model, image representation was enhanced by extracting image attribute information, so as to accurately describe the things in the image, and Bi-LSTM was used to capture bidirectional semantic dependency, so that the long-term visual language interaction learning was carried out. Firstly, CNN and ATT were used to extract the global image features and image attribute features respectively. Then, the two kinds of feature information were input into Bi-LSTM to generate sentences that were able to reflect the image content. Finally, the effectiveness of the proposed method was validated on Microsoft COCO Caption, Flickr8k, and Flickr30k datasets. Experimental results show that, compared with the multimodal Recurrent Neural Network (m-RNN) method, the proposed algorithm has improved the description performance by 6.8-11.6 percentage points. The proposed algorithm can effectively improve the semantic description performance of the constructed model for images.

In Fixed Word-Aligned (FWA) inverted index compression algorithms of Web search engines, due to the "greedy" block partition strategy of the inverted list and the interleaved storage of the codeword information, it is difficult for the algorithm to achieve the optimal compression performance. To solve the above problem, an FWA partition compression algorithm based on Directed Acyclic Graph (DAG) was proposed. Firstly, considering the small integer information in the inverted list brought by the clustering characteristics of Web pages, a novel FWA compression format with data area of 64-bit blocks was designed. In this compression format, the data area was divided into 16 storage modes suitable for continuous small integer compression through 4-bit selector area, and the selector area and data area in each block of the inverted list were stored separately, so as to ensure good batch decompression performance. Secondly, a DAG described Word-Aligned Partition (WAP) algorithm was proposed on the basis of the new compression format. In the algorithm, the inverted list block partitioning problem was regarded as a Single-Source Shortest-Path (SSSP) problem by DAG, and by considering the constraints of various storage modes of data area in FWA compression format, the structure and recursive definition of the SSSP problem were determined. Thirdly, the dynamic programming technique was used to solve the problem of SSSP and generate the pseudo-code and algorithm complexity of the optimal partition vector. The original storage modes of traditional FWA algorithms such as S9, S16 and S8b were partitioned and optimized based on DAG, and the computational complexities of the algorithms before and after optimization were compared and analyzed. Finally, the compression performance experiments were conducted with simulation integer sequence data and Text REtrieval Conference (TREC) GOV2 Web page index data. Experimental results show that, compared with the traditional FWA algorithms, the DAG based FWA partition algorithm can improve the compression ratio and decompression speed by batch decompression and partition optimization technology. At the same time, it can obtain a higher compression ratio than the traditional Frame Of Reference (FOR) algorithms for the compression of continuous small integer sequence.

Pedestrian re-identification tasks lack the consideration of the pedestrian feature scale variation during feature extraction, so that they are easily affected by environment and have low accuracy of pedestrian re-identification. In order to solve the problem, a pedestrian re-identification method based on multi-scale feature fusion was proposed. Firstly, in the shallow layer of the network, multi-scale pedestrian features were extracted through mixed pooling operation, which was helpful to improve the feature extraction capability of the network. Then, strip pooling operation was added to the residual block to extract the remote context information in horizontal and vertical directions respectively, which avoided the interference of irrelevant regions. Finally, after the residual network, the dilated convolutions with different scales were used to further preserve the multi-scale features, so as to help the model to analyze the scene structure flexibly and effectively. Experimental results show that, on Market-1501 dataset, the proposed method has the Rank1 of 95.9%, and the mean Average Precision (mAP) of 88.5%; on DukeMTMC-reID dataset, the proposed method has the Rank1 of 90.1%, and the mAP of 80.3%. It can be seen that the proposed method can retain the pedestrian feature information better, thereby improving the accuracy of pedestrian re-identification tasks.

Aiming at the limitations of traditional noise reduction methods for acoustic signals in low Signal-to-Noise Ratio (SNR) environment, a real-time noise reduction algorithm was proposed by combining adaptive threshold Voice Activity Detection (VAD) algorithm and Minimum Mean-Square Error Log-Spectral Amplitude estimation (MMSE-LSA). Firstly, the background noise was estimated in VAD algorithm by probability statistics based on the maximum value of the energy probability, and the obtained background noise was updated in real time and saved. Then, the background noise updated in real time was used as the reference noise of MMSE-LSA, and the noise amplitude spectrum was updated adaptively. Finally, the noise reduction processing was performed. The experimental results on four kinds of acoustic signals in real scenes show that the proposed algorithm can guarantee the real-time processing of low SNR acoustic signals; and compared with the traditional MMSE-LSA algorithm, it has the SNR of the noise reduction signal increased by 10-15 dB without over-subtraction. It can be applied to practical engineering.

In practical application,the patent quality evaluation is usually adopted by experts scoring or the quality evaluation index designed by the experts,so that the evaluation results are subjective and cannot be agreed by the both sides of the evaluation at the same time. In order to solve these problems,a deep learning patent quality evaluation method based on paper similarity calculation was proposed. Firstly,the papers were selected as the objective evaluation data,and the papers were used to calculate the similarity with the patent for augmented data. Then,a deep neural network was introduced to train the quality evaluation model,which was able to realize the map between the similarity of the paper and the quality of the patent to be evaluated. Finally,the quality evaluation model was used to access the patent quality. With perfect score of 100,the simulation results show that in different fields,compared to the corresponding expert evaluation result,the deviation of patent quality evaluation scores obtained by the proposed method is lower than 4,indicating that the proposed method has an effective patent quality evaluation ability.

Most of the existing automatic emotion annotation methods only extract the single recognition feature from acoustic layer or language layer. While Yi language is affected by the factors such as too many branch dialects and high complexity, so the accuracy of automatic annotation of Yi language with single-layer emotion feature is low. Based on the features such as rich emotional affixes in Yi language, a double-layer feature fusion method was proposed. In the method, the emotional features from acoustic layer and language layer were extracted respectively, the methods of generating sequence and adding units as needed were applied to complete the feature sequence alignment, and the process of automatic emotion annotation was realized through the corresponding feature fusion and automatic annotation algorithm. Taking the speech and text data of Yi language in a poverty alleviation log database as samples, three different classifiers were used for comparative experiments. The results show that the classifier has no obvious effect on the automatic annotation results, and the accuracy of automatic annotation after the fusion of double-layer features is significantly improved, the accuracy is increased from 48.1% of acoustic layer and 34.4% of language layer to 64.2% of double-layer fusion.

Aiming at the problems of insufficient generalization ability, poor stability and difficulty in meeting the real-time requirement of facial expression recognition, a real-time facial expression recognition method based on multi-scale kernel feature convolutional neural network was proposed. Firstly, an improved MSSD (MobileNet+Single Shot multiBox Detector) lightweight face detection network was proposed, and the detected face coordinates information was tracked by Kernel Correlation Filter (KCF) model to improve the detection speed and stability. Then, three linear bottlenecks of three different scale convolution kernels were used to form three branches. The multi-scale kernel convolution unit was formed by the feature fusion of channel combination, and the diversity feature was used to improve the accuracy of expression recognition. Finally, in order to improve the generalization ability of the model and prevent over-fitting, different linear transformation methods were used for data enhancement to augment the dataset, and the model trained on the FER-2013 facial expression dataset was migrated to the small sample CK+ dataset for retraining. The experimental results show that the recognition rate of the proposed method on the FER-2013 dataset reaches 73.0%, which is 1.8% higher than that of the Kaggle Expression Recognition Challenge champion, and the recognition rate of the proposed method on the CK+ dataset reaches 99.5%. For 640×480 video, the face detection speed of the proposed method reaches 158 frames per second, which is 6.3 times of that of the mainstream face detection network MTCNN (MultiTask Cascaded Convolutional Neural Network). At the same time, the overall speed of face detection and expression recognition of the proposed method reaches 78 frames per second. It can be seen that the proposed method can achieve fast and accurate facial expression recognition.

When the smart grid phasor measurement equipment competes for limited network communication resources, the data packets will be delayed or lost due to uneven resource allocation, which will affect the accuracy of power system state estimation. To solve this problem, a Sampling Awareness Weighted Round Robin (SAWRR) scheduling algorithm was proposed. Firstly, according to the characteristics of Phasor Measurement Unit (PMU) sampling frequency and packet size, a weight definition method based on mean square deviation of PMU traffic flow was proposed. Secondly, the corresponding iterative loop scheduling algorithm was designed for PMU sampling awareness. Finally, the algorithm was applied to the PMU sampling transmission model. The proposed algorithm was able to adaptively sense the sampling changes of PMU and adjust the transmission of data packets in time. The simulation results show that compared with original weighted round robin scheduling algorithm, SAWRR algorithm reduces the scheduling delay of PMU sampling data packet by 95%, halves the packet loss rate and increases the throughput by two times. Applying SAWRR algorithm to PMU data transmission is beneficial to ensure the stability of smart grid.

Aiming at temporal uncertain information fusion problem, to fully reflect the dynamic characteristic and the influence of time factor on temporal information fusion, a temporal evidence fusion method was proposed with considering decision maker's preference for time sequence based on evidence theory. Firstly, time sequence preference of decision maker was fused to temporal evidence fusion, through the analysis of characteristics of temporal evidence sequence, decision maker's preference for time sequence was measured based on the definition of temporal memory factor. Then, the evidence source was revised by time sequence weight vector obtained by constructing the optimal model and evidence credibility idea. Finally, the revised evidences were fused by Dempster combination rule. Numerical examples show that compared with other fusion methods without considering time factor, the proposed method can deal with conflicting information in temporal information sequence effectively and obtain a reasonable fusion effect; meanwhile, with the consideration of the credibility of temporal evidence sequence and the subjective preference of decision maker, the proposed method can reflect the influence of subjective factors of decision maker on temporal evidence fusion, giving a good expression to the dynamic characteristic of temporal evidence fusion.

In massive Multiple-Input-Multiple-Output (MIMO) system, as the increse of antenna number, traditional detection algorithms have lower performance, higher complexity, and they are not suitable for high order modulation. To solve the problem, based on the idea of shadow domain, a search tree detection algorithm combining Quadratic Programming (QP) and Branch and Bound (BB) algorithm was proposed. Firstly, with QP model constructed, the unreliable symbols from solution vector of first-order QP algorithm were extracted; then, BB search tree algorithm was applied to the unreliable symbols for the optimal solution; meanwhile three pruning strategies were proposed to reach a compromise between complexity and performance. The simulation results show that the proposed algorithm increases 20 dB performance gain compared with the traditional QP algorithm in 64 Quadrature Amplitude Modulation (QAM) and increases 21 dB performance gain compared with QP algorithm in 256 QAM. Meanwhile, applying the same pruning strategies, the complexity of the proposed algorithm is reduced by about 50 percentage points compared with the traditional search tree algorithm.

With the development of sharing economy, there is an urgent need for highly trusted distributed transaction management; however, traditional centralized information systems are difficult to meet it. Blockchain technology provides a shared ledger mechanism, which laid foundation for building credible distributed transaction management service. As blockchain 2.0 platform supporting smart contract, Ethereum platform was used as the basic framework to deeply study the operation mechanism and implementation technology of the decentralized shared goods transaction service system based on blockchain technology. Decentralized item sharing transaction service system framework based on Ethereum was designed, and a transaction management process based on intelligent contract mechanism was proposed. The system implementation technology including user interface was described in detail, and the performance of the system in transaction processing was tested. The experimental results indicate that the Ethereum-based transaction management system can ensure the creditability of the data and has a high operational efficiency, with average transaction processing speed of 21.7 items/s, and indexed average query speed of 117.6 items/s.

To overcome the randomness of RRT-Connect and slow convergence of B-RRT ^*(asymptotically-optimal Bidirectional Rapidly-exploring Random Tree) in path generation, an efficient path planning algorithm based on B-RRT ^*, abbreviated as EB-RRT ^*, was proposed. Firstly, an intelligent sampling function was intriduced to achieve more directional expansion of random tree, which could improve the smoothness of path and reduce the seek time. A rapidly-exploring strategy was also added in EB-RRT ^* by which RRT-Connect exploration mode was adopted to ensure rapidly expanding in the free space and improved asymptotically-optimal Rapidly-exploring Random Tree (RRT ^*) algorithm was adopted to prevent trapped in local optimum in the obstacle space. Finally, EB-RRT ^* algorithm was compared with Rapidly-exploring Random Tree (RRT), RRT-Connect, RRT ^* and B-RRT ^* algorithms. The simulation results show that the improved algorithm is superior to other algorithms in the efficiency and smoothness of path planning. It reduced the path planning time by 68.3% and the number of iterations by 48.6% compared with B-RRT ^* algorithm.

In order to solve the problem that traditional intrusion detection methods can not effectively solve instrusion data in high-dimensional networks, an intrusion detection method based on Stacked Sparse Denosing Autoencoder (SSDA) network was proposed. Firstly, SSDA was used to perform dimensionality reduction on the intrusion data. Then, the highly abstracted low-dimensional data was used as input data of softmax classifier to realize intrusion detection. Finally, in order to improve original intrusion data decoding ability of the network and intrusion detection ability of the model, an Improved model based on SSDA (ISSDA) was proposed, with new constraints added to the autoencoder. The experimental results show that compared with SSDA, ISSAD's detection accuracy of four types of attacks was improved by about 5%, and the false positive rate of ISSAD was also effectively reduced.

In the analysis process of high-dimensional data such as image data, genetic data and text data, when samples have redundant features, the complexity of the problem is greatly increased, so it is important to reduce redundant features before data analysis. The feature selection based on Mutual Information (MI) can reduce the data dimension and improve the accuracy of the analysis results, but the existing feature selection methods cannot reasonably eliminate the redundant features because of the single standard. To solve the problem, a feature selection method based on Maximum Conditional and Joint Mutual Information (MCJMI) was proposed. Joint mutual information and conditional mutual information were both considered when selecting features with MCJMI, improving the feature selection constraint. Exerimental results show that the detection accuracy is improved by 6% compared with Information Gain (IG) and minimum Redundancy Maximum Relevance (mRMR) feature selection; 2% compared with Joint Mutual Information (JMI) and Joint Mutual Information Maximisation (JMIM); and 1% compared with LW index with Sequence Forward Search algorithm (SFS-LW). And the stability of MCJMI reaches 0.92, which is better than JMI, JMIM and SFS-LW. In summary the proposed method can effectively improve the accuracy and stability of feature selection.

As a new research task, suggestion mining has important application value. Since traditional suggestion sentence classification methods have problems like complex rules, large labeling workload, high feature dimension and data sparsity, a PU (Positive and Unlabeled)-based suggestion sentence classification method was proposed. Firstly, some suggestion sentences were selected from an unlabeled review set by using a simple rule to form a positive example set; then a reliable negative example set was constructed by Spy technique in the feature space of autoencoder neural network to reduce the feature dimension and alleviate data sparsity; finally, Multi-Layer Perceptron (MLP) was trained by the positive example set and the reliable negative example set to classify the remaining unlabeled samples. On a Chinese dataset, the F1 value and the accuracy of the proposed method, reached 81.98% and 82.67% respectively. The experimental results show that the proposed method can classify suggestion sentences effectively without manually labelling the data.

In order to evaluate the reliability and fault-tolerant ability of multi-processor system which takes hypercubes as underlying networks, combining the fact that structural faults often occur when the system is invaded by computer viruses, three-length-path structure connectivity and substructure connectivity of the n-cube network were investigated. Firstly, by using the three-length-path structure-cut of the n-cube network, an upper bound of three-length-path structure connectivity of the network was obtained. Secondly, by using an equivalent transformation or a reductive transformation of the three-length-path substructure-set of the n-cube network, a lower bound of three-length-path substructure connectivity of the network was obtained. Finally, combining with the property that three-length-path structure connectivity of a network is not less than its three-length-path substructure connectivity, it was proved that both three-length-path structure connectivity and substructure connectivity of a n-cube network were half of n. The results show that to destroy the enemy's multi-processor system which take the n-cubes as underlying networks under three-length-path structure fault model, at least half of n three-length-path structures or substructures of the system should be attacked.

The multi-source heterogeneity and complicated relationships of spatio-temporal data of vehicle refueling bring great challenges to existing anomaly detection approaches. Aiming at the problem, a multi-view deep anomaly detection framework for vehicle refueling based on spatio-temporal data fusion was proposed. Firstly, the static information and dynamic activity data were correlated, fused and managed based on Unified Conceptual Model (UCM). Secondly, the spatio-temporal data were encoded and converted according to spatial view, temporal view and semantic view. Finally, a deep anomaly detection framework was constructed based on the above multi-views. The experimental results on vehicle refueling spatio-temporal dataset show that all anomaly detection approaches tested can achieve an average decrease in the Root Mean Square Error (RMSE) by 10.73%, and the proposed multi-view spatio-temporal anomaly detection framework can obtain a decrease in the RMSE by 19.36% compared to LSTM (Long Short-Term Memory), which gets the best results in the-state-of-the-art methods. And the Matthews Correlation Coefficient (MCC) of the proposed method on the credit card fraud dataset is increased by 32.78% compared with that of Logistic Regression model. All experimental results demonstrate the effectiveness of the proposed anomaly detection framework.

Android uses the permission-based access control method to protect the system resources, which has the problem of rough management. At the same time, some malicious applications can secretly access resources in a privacy scenario without the user's permission, bringing certain threats to user privacy and system resources. Based on the original permission management and control and with the introduction of Access Control List (ACL) mechanism, an Android fine-grained permission management and control system based on ACL mechanism was designed and implemented. The proposed system can dynamically set the access rights of the applications according to the user's policy, avoiding the access of malicious codes to protect system resources. Tests of compatibility and effectiveness show that the system provides a stable environment for applications.

Aspect sentiment analysis is a fine-grained task in sentiment classification. Concerning the problem that traditional neural network model can not accurately construct sentiment features of aspects, a Long Short-Term Memory with Multi-ATTention and Aspect Context (LSTM-MATT-AC) neural network model was proposed. Different types of attention mechanisms were added in different positions of bidirectional Long Short-Term Memory (LSTM), and the advantage of multi-attention mechanism was fully utilized to allow the model to focus on sentiment information of specific aspects in sentence from different perspectives, which could compensate the deficiency of single attention mechanism. At the same time, combining aspect context information of bidirectional LSTM independent coding, the model could capture deeper level sentiment information and effectively distinguish sentiment polarity of different aspects. Experiments on SemEval2014 Task4 and Twitter datasets were carried out to verify the effectiveness of different attention mechanisms and independent context processing on aspect sentiment analysis. The experimental results show that the accuracy of the proposed model reaches 80.6%, 75.1% and 71.1% respectively for datasets in domain Restaurant, Laptop and Twitter. Compared with previous neural network-based sentiment analysis models, the accuracy has been further improved.

Most existing forensic methods for digital speech aim at detecting a specific operation, which means that these methods can not identify various operations at a time. To solve the problem, a universal forensic algorithm for simultaneously detecting various operations, such as pitch modification, low-pass filtering, high-pass filtering, and noise adding was proposed. Firstly, the statistical moments of Mel-Frequency Cepstral Coefficients (MFCC) were calculated, and cepstrum mean and variance normalization were applied to the moments. Then, a multi-class classifier based on multiple two-class classifiers was constructed. Finally, the classifier was used to identify various types of speech operations. The experimental results on TIMIT and UME speech datasets show that the proposed universal features achieve detection accuracy over 97% for various speech operations. And the detection accuracy in the test of MP3 compression robustness is still above 96%.

In order to quickly and stably detect the faces with large change of angle and serious occlusion in image sequence, a new automatic Detection-Tracking-Detection (DTD) model was proposed by combining the fast and accurate target detection model MobileNet-SSD (MS) and the fast tracking model Kernel Correlation Filtering (KCF), namely MS-KCF face detection model. Firstly, the face was detected quickly and accurately by using MS model, and the tracking model was updated. Secondly, the detected face coordinate information was input into the KCF tracking model to track steadily, and the overall detection speed was accelerated. Finally, to prevent tracking loss, the detection model was updated again after tracking several frames, then the face was detected again. The recall of MS-KCF model in the FDDB face detection benchmark was 93.60%; the recall in Easy, Medium and Hard data sets of WIDER FACE benchmark were 93.11%, 92.18% and 82.97%, respectively; the average speed was 193 frames per second. Experimental results show that the MS-KCF model is stable and fast, which has a good detection effect on the faces with serious shadows and large angle changes.

Focused on the issue that lexicalized reordering models are faced with context independence and sparsity problems in machine translation, a reordering table reconstruction model based on semantic content for reordering orientation and probability prediction was proposed. Firstly, continuous distributed representation approach was employed to acquire the feature vectors of reordering rules. Secondly, Recurrent Neural Networks (RNN) were utilized to predict the reordering orientation and probability of each reordering rule that represented with dense vectors. Finally, the original reordering table was filtered and reconstructed with more reasonable reordering probability distribution for the purpose of improving the reordering information accuracy in reordering model as well as reducing the size of the reordering table to speed up subsequent decoding process. The experimental results show that the reordering table reconstruction model can provide BLEU point gains (+0.39) for Chinese to Uyghur machine translation task.

In multicast routing comunication of smart grid, concerning the long communication delay of multicast tree when transmitting control messages to high-power load device, which caused by only considering delay constraint without considering the demand of smart grid, a new multicast tree construction method that considered load and comunication delay at the same time was proposed, namely multicast routing algorithm based on Demand Response (DR) capability constraint. Firstly, a complete graph satisfying the constraint was generated according to the grid network topology. Then, a lower-cost multicast tree was constructed by using the Prim algorithm. Finally, the multicast tree was restored to the original network. The simulation results show that the proposed algorithm can effectively reduce the demand response delay of high-power load devices, and can significantly reduce the power frequency deviation compared with the multicast routing algorithm only considering delay constraint. This algorithm can actually improve the real-time demand response in the smart grid and stabilize the grid frequency.

For most spatial co-location pattern mining methods, distance threshold is used as a standard to measure the neighboring relation among instances of different objects, then to mine frequent co-location patterns, but the interation between instances with neighboring relations and the gain rate of patterns are not considered. In the spatial co-location patterns mining process, by introducing the interation rate between instances and the seasonal average income of objects, the concepts of object effect rate, suite total income and gain rate were defined, and a basic algorithm named NAGA and an efficient pruning algorithm named NAGA_JZ for mining high gain rate co-location patterns were put forward. Finally, a large number of experiments were carried out to verify the correctness and practicability of the basic algorithm, and the mining efficiency of the basic algorithm and the pruning algorithm were compared. The experimental results prove the high efficiency of the pruning algorithm.

Plant recognition includes two kinds of tasks:specimen recognition and real-environment recognition. Due to the existence of background noise, real-environment plant image recognition is more difficult. To reduce the weight of Convolutional Neural Networks (CNN), to improve over-fitting, to improve the recognition rate and generalization ability, a method of plant identification with Family Priority (FP) was proposed. Combined with the lightweight CNN MobileNet model, a plant recognition model Family Priority MobileNet (FP-MobileNet) was established by means of migration learning. On the single background plant dataset flavia, the MobileNet model achieved 99.8% of accuracy. For the more challenging real-environment flower dataset flower102, when the number of samples in the training set was greater than that in the test set FP-MobileNet achieved 99.56% of accuracy. When the number of samples in the training set was smaller than that in the test set, FP-MobileNet still obtained 95.56% of accuracy. The experimental results show that the accuracies of FP-MobileNet under two different data set partitioning schemes are both higher than those of the pure MobileNet model. In addition, FP-MobileNet weighs only occupy 13.7 MB with high recognition rate. It takes into account both accuracy and delay, and is suitable for promotion to mobile devices that require a lightweight model.

Aiming at the problem that the existing signal control logic cannot respond to the bus cumulative number and the sensitivity of control parameters is poor, a bus priority strategy trigger probability model was constructed to detect and analyze the methods for improving trigger accuracy. Based on the Siemens 2070 signal controller, the triggering theory of Transit Signal Priority (TSP) was analyzed, and the trigger probability models were constructed for green-extension strategy and early-green strategy. Taking the actual intersection as an example, the trigger probability results of different signal timing plans were calculated and compared by simulation, the trigger characteristic of TSP strategies was studied and the improvement was discussed. The research shows that the trigger probability of green-extension strategy is so far below the early-green strategy; the trigger probability of green-extension is inversely proportional to the mini-green and the max-green time, while the trigger probability of early-green strategy is mainly related to the number of buses which applying for priority in the non-favored signal phase; the trigger probability of green-extension strategy can be improved by optimizing the mini-green, max-green time and increasing the bus number applying for priority; the trigger probability of early-green strategy can be improved by optimizing the original signal timing scheme then adding TSP subsequently.

A large amount of alarms considerably complicate the root-cause analysis in telecom networks, thus a new alarm filtering algorithm was proposed to minimize the interference on the analysis. Firstly, a quantitative analysis for the alarm data, e.g., the quantity distribution and the average duration, was conducted, and the concepts of alarm impact and high-frequency transient alarm were defined. Subsequently, the importance of each alarm instance was evaluated from four perspectives:the amount of the alarms, the average duration of the alarms, the alarm impact, and the average duration of the alarm instance. Accordingly, an alarm filtering algorithm with O ( n) computation complexity in principle was proposed, where n is the number of alarms under analysis. Single-factor experimental analysis show that the compression ratio of the alarm data has a positive correlation with the alarm amount of a specific alarm element, the average duration of the alarms, the alarm impact, and the duration of the alarm instance; further, the accuracy of the proposed algorithm is improved by 18 percentage points at most compared with Flexible Transient Flapping Determination (FTD) algorithm. The proposed algorithm can be used both for off-line analysis of historical alarm data and for on-line alarm filtering.