The inherent “black-box” nature of neural network models leads to opaque internal logic and behavioral decision， restricting their application in critical areas. To address this issue， based on the Layer-wise Relevance Propagation （LRP） method， the interpretability of image classification models such as VGG16 and ResNet50 was researched， and an interpretation method for image classification models named SSGLRP （Selective Softmax Gradient for Layer-wise Relevance Propagation） was proposed. By introducing activation values for the positive gradients of output neurons and modifying the initial relevance values for non-target classes， the problem of LRP heatmaps containing noise and lacking class discrimination was effectively solved. The SSGLRP method was quantitatively evaluated using maximum patch masking and fixed-point game experiments， with classic LRP， Selective Layer-wise Relevance Propagation （SLRP）， and Softmax Gradient Layer-wise Relevance Propagation （SGLRP） as baseline methods. The results of the maximum patch masking experiments show that on VGG16， the prediction change obtained by SSGLRP is on average 91.8%， 60.3%， and 34.2% higher than those obtained by LRP， SLRP， and SGLRP， respectively； on ResNet50， the improvements are 61.9%， 64.9%， and 33.2%， respectively. The SSGLRP method has higher class discrimination ability and less noise with superior performance in interpreting VGG16 and ResNet50 models.

In response to the resource allocation problem in multi-beam scenarios of Low Earth Orbit （LEO） satellite， as the factors such as interference and noise between wave beams in actual satellite communication environments are complex and variable， conventional subcarrier dynamic allocation algorithms cannot adjust parameters dynamically to adapt to changes in the communication environment. By combining traditional communication scheduling algorithms with reinforcement learning techniques， with the goal of minimizing user packet loss rate， user’s scheduling situations were adjusted dynamically and resources of the entire satellite communication system were allocated dynamically to adapt to environmental changes. The dynamic characteristic model of LEO satellite was discretized by time slot division， and a Deep Reinforcement Learning （DRL）-based resource allocation strategy was proposed on the basis of the modeling of LEO satellite resource allocation scenarios. In this strategy， the scheduling opportunities for users with high latency were increased by adjusting the satellite scheduling queue situation， that is， adjusting the resource blocks in each beam of a single LEO satellite to correspond to qualifications of users， thereby ensuring a certain level of fairness and reducing the user packet loss rate at the same time. Simulation results show that under the condition meeting total power constraints， the user transmission fairness and system throughput are stable in the proposed Deep Reinforcement Learning based Resource Allocation algorithm （DRL-RA）， and users with large latency obtain more scheduling opportunities in DRL-RA due to priority improvement. Compared with Proportional Fairness （PF） algorithm and Maximum Carrier/Interference （Max C/I） algorithm， DRL-RA has the data packet loss rate reduced by 13.9% and 15.6% respectively. It can be seen that the proposed algorithm solves the problem of packet loss effectively during data transmission.

Running a large-scale Spiking Neural Network （SNN） on a distributed computing platform is one of the basic means to improve the level of brain-like computing intelligence. The difficulty lies in how to deploy the SNN to the corresponding number of computing nodes in order to make the overall system run with the best energy efficiency. To solve this problem， on the basis of NEural Simulation Tool-based （NEST-based） Workload Automatic Mapper for SNN （SWAM） proposed by others before， a workload automatic mapper for SNN， named SWAM2， based on precise communication modeling was proposed. In SWAM2， based on the NEST simulator， the communication part of the SNN workload was further accurately modeled； the quantization method of the parameters in the workload model was improved； the maximum network scale prediction method was designed. Experimental results on typical cases of SNN show that， the average prediction errors of SWAM2 were reduced by about 12.62 and 5.15 percentage points respectively compared with those of SWAM in workload communication and computing time prediction. When predicting the optimal mapping of the workload， the average accuracy of SWAM2 reached 97.55%， which was 13.13 percentage points higher than that of SWAM. SWAM2 avoids the process of manual trial and error by automatically predicting the optimal deployment/mapping of SNN workload on computing platform.

The detection precision is low when the diatom training sample size is small， so a Multi-scale Multi-head Self-attention （MMS） and Online Hard Example Mining （OHEM） based few-shot diatom detection model， namely MMSOFDD was proposed based on the few-shot object detection model Two-stage Fine-tuning Approach （TFA）. Firstly， a Transformer-based feature extraction network Bottleneck Transformer Network-101 （BoTNet-101） was constructed by combining ResNet-101 with a multi-head self-attention mechanism to make full use of the local and global information of diatom images. Then， multi-head self-attention was improved to MMS， which eliminated the limitation of processing single object scale of the original multi-head self-attention. Finally， OHEM was introduced to the model predictor， and the diatoms were identified and localized. Ablation and comparison experiments between the proposed model and other few-shot object detection models were conducted on a self-constructed diatom dataset. Experiment results show that the mean Average Precision （mAP） of MMSOFDD is 69.60%， which is improved by 5.89 percentage points compared with 63.71% of TFA； and compared with 61.60% and 60.90% the few-shot object detection models Meta R-CNN and Few-Shot In Wild （FSIW）， the proposed model has the mAP improved by 8.00 percentage points and 8.70 percentage points respectively. Moreover， MMSOFDD can effectively improve the detection precision of the detection model for diatoms with small size of diatom training samples.

At present， most deep learning models are difficult to deal with the classification of bird sound under complex background noise. Because bird sound has the continuity characteristic in time domain and high-low characteristic in frequency domain， a fusion model of homologous spectrogram features was proposed for bird sound classification under complex background noise. Firstly， Convolutional Neural Network （CNN） was used to extract Mel-spectrogram features of bird sound. Then， the time domain and frequency domain dimensions of the same Mel-spectrogram feature were compressed to 1 by specific convolution and down-sampling operations， so that frequency domain feature with only high-low characteristics and the time domain feature with only continuous characteristics were obtained. Based on the above operation to extract frequency domain and time domain features， the features of Mel-spectrogram were extracted both in time domain and frequency domain， the time-frequency domain features with continuity and high-low characteristics were obtained. Then the self-attention mechanism was applied to the obtained time domain， frequency domain and time-frequency domain features， strengthening their own characteristics. Finally， the results of these three homologous spectrogram features after decision fusion were used for bird sound classification. The proposed model was used for audio classification of 8 bird species on Xeno-canto website， achieved the better result in the comparison experiment with the Mean Average Precision （MAP） of 0.939. The experimental results show that the proposed model can deal with the problem of the poor classification effect of bird sound under complex background noise.

Designing a unified solution to the combinational optimization problems of undesigned heuristic algorithms has become a research hotspot in the field of machine learning. At present， mature technologies are mainly aiming at static combinatorial optimization problems， but the combinational optimization problems with dynamic changes are not fully solved. In order to solve above problems， a lightweight model called Dy4TSP （Dynamic model for Traveling Salesman Problems） was proposed， which combined multi-head-attention mechanism with distributed reinforcement learning to solve the traveling salesman problem on a dynamic graph. Firstly， the node representation vector from graph convolution neural network was processed by the prediction network based on multi-head-attention mechanism. Then， the distributed reinforcement learning algorithm was used to quickly predict the possibility that each node in the graph was output as the optimal solution， and the optimal solution space of the problems in different possibilities were comprehensively explored. Finally， the action decision sequence which could meet the specific reward function in real time was generated by the trained model. The model was evaluated on three typical combinatorial optimization problems， and the experimental results showed that the solution qualities of the proposed model are 0.15 to 0.37 units higher than those of the open source solver LKH3 （Lin-Kernighan-Helsgaun 3）， and are significantly better than those of the latest algorithms such as Graph Attention Network with Edge Embedding （EGATE）. The proposed model can reach an optimal path gap of 0.1 to 1.05 in other dynamic traveling salesman problems， and the results are slightly better.

Due to the multi-hierarchy, multi-attribute process of product design, the key technology of knowledge reuse in product design based on axiomatic design was studied. With the axiomatic design framework, the design process was decomposed, and the Z map of the requirement domain, functional domain, structure domain, process domain of the product design was put forward. Then, a case-based knowledge reuse method with overall consideration of inter-domanin and intra-domain design information was presented, which adopted the similarity-based retrieving in knowledge reuse algorithm, and got the best reuse objects. Finally, a case was provided to validate and prove the efficiency and feasibility of the method.

With regard to the thrashing problem of load-balancing algorithm in heterogeneous environments, a new scheduling algorithm called Dynamic Predetermination of Slow Task (DPST) was designed to reduce the probability in slow task scheduling and improve load-balancing. Through defining capability measure of heterogeneous task in heterogeneous nodes, the capacity of nodes which performed heterogeneous tasks was normalized. With the introduction of predetermination, thrashing result from heterogeneous environments was reduced. By using double queues of slow task and slow node, the efficiency of scheduling was improved. The experimental results show that the thrashing times in heterogeneous environments fell by more than 40% compared with Hadoop. Because thrashing times have been reduced effectively, DPST algorithm has better performance in reducing average response time and increasing system throughput in heterogeneous environments.

The limited node energy, high node redundancy and other characteristics of Wireless Sensor Network (WSN) make working-in-round mechanism one of the basic policies in solving the network coverage problem. In this paper, the authors investigated the network characteristics of stochastic scheduling model, analyzed the relationship between the number of effective nodes and rotation cycles, and finally provided an adaptive algorithm to adjust the node work probability according to the effective node number in the network. The algorithm can solve the performance problem in the later periods of the network, which is caused by reduction of effective nodes and static work probability, and thus guaranteeing the network performance in each round. Simulation verifies the effectiveness and correctness of the novel algorithm.

User requirement is not considered in existing spectrum allocation algorithms for the cognitive radio, which results in excessive allocation of resources to users with less demand. To solve the above problem, two spectrum allocation algorithms, which included spectrum allocation algorithm based on user requirement and joint proportional fairness algorithm, were proposed. They both consider user requirement as the needed factor of spectrum allocation. Numerical results indicate that these two algorithms are better satisfied with users requirement than the previous algorithms.

Based on thorough analysis on the existing PSO algorithms, an improved PSO algorithm with adaptive neighborhood was proposed, which included two special improvements: the mechanism of scheduled neighborhood adaptation, and a modified mechanism of scheduled interia weight adaptation. The proposed PSO algorithm is proved to be high-performing and high accurate by 7 standard banchmark functions.

A demand-based spectrum allocation algorithm between cognitive radio cells was proposed. Based on the combination of graph coloring model and the concept of spectrum quality grading, the proposed algorithm allocated the available spectrum to cognitive radio cells to meet demand maximally under QoS guarantee. Simulation results show that, compared with the original algorithm, the proposed algorithm can better satisfy demand in the network.

Different smoothing regions of SAR image can not be represented accurately in single wavelet domain. In order to solve this problem, a Besov ball projections algorithm in multiple wavelet domains was proposed. Different Besov balls radius were defined by the universal hidden Markov tree model and Besov norm, then an image was projected onto alternate Besov balls of proper radius corresponding to the projection onto convex sets algorithm based on preserving edge and strong scatterers (SPOCS) for image denoising. Experimental results demonstrate that this method can not only provide significant improvement over conventional algorithms based on single wavelet domain, but also preserve edge structure better.

How to improve the retrieval efficiency by means of term relationships was discussed. In order to integrate the specialties of term similarity and term correlativity, a new method to mine the term relationships was proposed. In this method, a correction factor from term correlativity to term similarity was used to adjust the level of relations with the term, which is a more accurate solution to the concept of the terms of semantic matching problem. Applying the new method to two retrieval models, experimental results show that the retrieval efficiency can be improved more obviously.

Word embedding models can map words to low-dimensional vector space for analyzing word semantics， which provides an effective way for computer understanding and text processing. Traditional Chinese word embedding models learn semantic information through the internal compositional information of Chinese words， however， for the utilization degree of Chinese characters and information of their different levels of components， different models have insufficient or excessive utilization problems. Thus， in order to utilize the information of different levels of components of Chinese characters better to generate high-quality word embeddings， a Multilevel-component Joint Chinese Word Embedding （MJWE） model was proposed to integrate the characteristics of words， Chinese characters， and multilevel-components， combine word embeddings with positional information， and construct multilevel-component embeddings composed of radicals and finer-grained components to capture the internal composition information of Chinese words more comprehensively. Meanwhile， a non-compositional word list was constructed to prevent the over utilization of the internal information of Chinese words. Experimental results show that MJWE model has the accuracy improved by 2.11% compared to JWE （Joint learning Word Embeddings） model on the word similarity task “WS-295”， by 2.52% compared to Skip-Gram （SG） model on the word analogy task “state”， by 6.58% compared to CBOW （Continuous Bag Of Words） model on the word analogy task “family”， by 0.71% compared to JWE model on the emotion classification task （two classes）， and by 8.60% compared to SG model on the emotion classification task （seven classes）. Meanwhile， MJWE model was applied to analyse literature on traditional Chinese medicine， for core drug identification in traditional Chinese medicine formulae， and MJWE model was able to identify the core drugs for treating different symptoms of chronic glomerulonephritis. It can be seen that MJWE can generate Chinese word embeddings with good quality， and combined with community detection algorithm， it can identify core drugs for treating different syndromes of chronic glomerulonephritis， which is conducive to assisting traditional Chinese medicine doctors in clinical decision-making.