在工业检测场景下,按照是否引入正常样本中不存在的异常,可以将异常检测问题分为结构异常检测和逻辑异常检测两类,逻辑异常检测对网络的全局理解能力提出了更高的要求。针对现有无监督异常检测模型在结构异常上已有较好的检测精度,但无...在工业检测场景下,按照是否引入正常样本中不存在的异常,可以将异常检测问题分为结构异常检测和逻辑异常检测两类,逻辑异常检测对网络的全局理解能力提出了更高的要求。针对现有无监督异常检测模型在结构异常上已有较好的检测精度,但无法适应逻辑异常检测需求的问题,提出一种包含空间聚合模块和通道聚合模块的双自编码器结构,主要由3部分组成。首先设计了并行空空间通道双自编码器架构,从空间和通道两个方向得到包含全局信息的特征向量,提升网络的长程依赖关系;其次设计一个选择性融合模块,融合双自编码器信息,放大包含重要信息的特征,以进一步提高对逻辑异常的表达能力;最后提出在自编码器与学生网络的损失函数中加入余弦损失,避免网络对单个像素差异过于敏感,从而关注于全局差异。在MVTec LOCO AD数据集上进行实验,逻辑异常检测精度达到89.4%,结构异常检测精度达到94.9%,平均检测精度92.1%,超越了基线方法和其他无监督缺陷检测方法,验证了方法的有效性和优越性。展开更多
This study presents analysis, control and comparison of three hybrid approaches for the direct torque control (DTC) of the dual star induction motor (DSIM) drive. Its objective consists of combining three different he...This study presents analysis, control and comparison of three hybrid approaches for the direct torque control (DTC) of the dual star induction motor (DSIM) drive. Its objective consists of combining three different heuristic optimization techniques including PID-PSO, Fuzzy-PSO and GA-PSO to improve the DSIM speed controlled loop behavior. The GA and PSO algorithms are developed and implemented into MATLAB. As a result, fuzzy-PSO is the most appropriate scheme. The main performance of fuzzy-PSO is reducing high torque ripples, improving rise time and avoiding disturbances that affect the drive performance.展开更多
A 512-bit EEPROM IP was designed by using just logic process based devices.To limit the voltages of the devices within 5.5 V,EEPROM core circuits,control gate(CG) and tunnel gate(TG) driving circuits,DC-DC converters:...A 512-bit EEPROM IP was designed by using just logic process based devices.To limit the voltages of the devices within 5.5 V,EEPROM core circuits,control gate(CG) and tunnel gate(TG) driving circuits,DC-DC converters:positive pumping voltage(VPP=4.75 V) ,negative pumping voltage(VNN=4.75 V) ,and VNNL(=VNN/2) generation circuit were proposed.In addition,switching powers CG high voltage(CG_HV) ,CG low voltage(CG_LV) ,TG high voltage(TG_HV) ,TG low voltage(TG_LV) ,VNNL_CG and VNNL_TG switching circuit were supplied for the CG and TG driving circuit.Furthermore,a sequential pumping scheme and a new ring oscillator with a dual oscillation period were proposed.To reduce a power consumption of EEPROM in the write mode,the reference voltages VREF_VPP for VPP and VREE_VNN for VNN were used by dividing VDD(1.2 V) supply voltage supplied from the analog block in stead of removing the reference voltage generators.A voltage level detector using a capacitive divider as a low-power DC-DC converter design technique was proposed.The result shows that the power dissipation is 0.34μW in the read mode,13.76μW in the program mode,and 13.66μW in the erase mode.展开更多
云时代,云API作为服务交付、数据交换和能力复制的最佳载体,已成长为当今面向服务软件开发和企业数字化转型不可或缺的核心要素.然而动态开放网络中持续增长的云API在给开发者提供了更多选择的同时,也将其淹没在海量的云API选择之中,设...云时代,云API作为服务交付、数据交换和能力复制的最佳载体,已成长为当今面向服务软件开发和企业数字化转型不可或缺的核心要素.然而动态开放网络中持续增长的云API在给开发者提供了更多选择的同时,也将其淹没在海量的云API选择之中,设计有效的云API推荐方法就此成为API经济健康发展中迫切要解决的现实问题.但是,现有研究主要利用搜索关键词、服务质量和调用偏好进行建模,生成质量高功能单一的云API推荐列表,没有考虑服务化软件实际开发中开发者对多元化高阶互补云API的客观需要.高阶互补云API推荐旨在为多个查询云API生成多元互补云API列表,要求推荐结果与查询云API均互补,以满足开发者的联合需求.针对此问题,本文提出基于概率逻辑推理的高阶互补云API推荐方法(Probabilistic Logic Reasoning for High-order Complementary Cloud API Recom⁃mendation,PLR4HCCR).首先,通过云API生态真实数据分析论证云API互补推荐需求的必要性和互补关系建模中替补噪声的客观存在,为云API互补推荐问题研究提供动机和数据支持.其次,采用Beta概率嵌入对云API及其之间的关系约束进行编码,以刻画云API间互补关系的不确定性和支持互补逻辑推理.接着,设计由投影、取反和交并三个基本逻辑算子构建的互补关系逻辑推理网络,使查询集中的每个云API获得非对称互补关系感知和替补噪声消解约束下的互补云API表示.然后,引入注意力机制为查询云API的互补云API分配不同权重,增强高阶互补云API基向量的表征能力.在此基础上,采用KL散度度量高阶互补云API基向量与候选云API之间的距离,并根据KL散度排序生成高阶互补性可感知下的云API推荐结果.最后,我们利用两个真实云API数据集在不同阶互补推荐场景下进行实验,实验表明,与传统启发式推荐方法和深度学习推荐方法相比,PLR4HCCR在互补关系感知推理和替补噪声消解方面均具有较大的优势,继而使其在低阶、高阶和混合阶互补云API推荐中均展示出更优的推荐效果和更强的泛化能力.进一步,超参数敏感性实验、实例分析和用户调查验证了方法的有效性、实用性和可行性,这使结合高阶互补关系的云API推荐方法PLR4HCCR不仅更有可能生成开发者满意的结果,而且可有效提升云API服务提供者的收益.展开更多
文摘在工业检测场景下,按照是否引入正常样本中不存在的异常,可以将异常检测问题分为结构异常检测和逻辑异常检测两类,逻辑异常检测对网络的全局理解能力提出了更高的要求。针对现有无监督异常检测模型在结构异常上已有较好的检测精度,但无法适应逻辑异常检测需求的问题,提出一种包含空间聚合模块和通道聚合模块的双自编码器结构,主要由3部分组成。首先设计了并行空空间通道双自编码器架构,从空间和通道两个方向得到包含全局信息的特征向量,提升网络的长程依赖关系;其次设计一个选择性融合模块,融合双自编码器信息,放大包含重要信息的特征,以进一步提高对逻辑异常的表达能力;最后提出在自编码器与学生网络的损失函数中加入余弦损失,避免网络对单个像素差异过于敏感,从而关注于全局差异。在MVTec LOCO AD数据集上进行实验,逻辑异常检测精度达到89.4%,结构异常检测精度达到94.9%,平均检测精度92.1%,超越了基线方法和其他无监督缺陷检测方法,验证了方法的有效性和优越性。
基金Project supported by Faculty of Technology,Department of Electrical Engineering,University of Batna,Algeria
文摘This study presents analysis, control and comparison of three hybrid approaches for the direct torque control (DTC) of the dual star induction motor (DSIM) drive. Its objective consists of combining three different heuristic optimization techniques including PID-PSO, Fuzzy-PSO and GA-PSO to improve the DSIM speed controlled loop behavior. The GA and PSO algorithms are developed and implemented into MATLAB. As a result, fuzzy-PSO is the most appropriate scheme. The main performance of fuzzy-PSO is reducing high torque ripples, improving rise time and avoiding disturbances that affect the drive performance.
基金Project supported by the Second Stage of Brain Korea 21
文摘A 512-bit EEPROM IP was designed by using just logic process based devices.To limit the voltages of the devices within 5.5 V,EEPROM core circuits,control gate(CG) and tunnel gate(TG) driving circuits,DC-DC converters:positive pumping voltage(VPP=4.75 V) ,negative pumping voltage(VNN=4.75 V) ,and VNNL(=VNN/2) generation circuit were proposed.In addition,switching powers CG high voltage(CG_HV) ,CG low voltage(CG_LV) ,TG high voltage(TG_HV) ,TG low voltage(TG_LV) ,VNNL_CG and VNNL_TG switching circuit were supplied for the CG and TG driving circuit.Furthermore,a sequential pumping scheme and a new ring oscillator with a dual oscillation period were proposed.To reduce a power consumption of EEPROM in the write mode,the reference voltages VREF_VPP for VPP and VREE_VNN for VNN were used by dividing VDD(1.2 V) supply voltage supplied from the analog block in stead of removing the reference voltage generators.A voltage level detector using a capacitive divider as a low-power DC-DC converter design technique was proposed.The result shows that the power dissipation is 0.34μW in the read mode,13.76μW in the program mode,and 13.66μW in the erase mode.
文摘云时代,云API作为服务交付、数据交换和能力复制的最佳载体,已成长为当今面向服务软件开发和企业数字化转型不可或缺的核心要素.然而动态开放网络中持续增长的云API在给开发者提供了更多选择的同时,也将其淹没在海量的云API选择之中,设计有效的云API推荐方法就此成为API经济健康发展中迫切要解决的现实问题.但是,现有研究主要利用搜索关键词、服务质量和调用偏好进行建模,生成质量高功能单一的云API推荐列表,没有考虑服务化软件实际开发中开发者对多元化高阶互补云API的客观需要.高阶互补云API推荐旨在为多个查询云API生成多元互补云API列表,要求推荐结果与查询云API均互补,以满足开发者的联合需求.针对此问题,本文提出基于概率逻辑推理的高阶互补云API推荐方法(Probabilistic Logic Reasoning for High-order Complementary Cloud API Recom⁃mendation,PLR4HCCR).首先,通过云API生态真实数据分析论证云API互补推荐需求的必要性和互补关系建模中替补噪声的客观存在,为云API互补推荐问题研究提供动机和数据支持.其次,采用Beta概率嵌入对云API及其之间的关系约束进行编码,以刻画云API间互补关系的不确定性和支持互补逻辑推理.接着,设计由投影、取反和交并三个基本逻辑算子构建的互补关系逻辑推理网络,使查询集中的每个云API获得非对称互补关系感知和替补噪声消解约束下的互补云API表示.然后,引入注意力机制为查询云API的互补云API分配不同权重,增强高阶互补云API基向量的表征能力.在此基础上,采用KL散度度量高阶互补云API基向量与候选云API之间的距离,并根据KL散度排序生成高阶互补性可感知下的云API推荐结果.最后,我们利用两个真实云API数据集在不同阶互补推荐场景下进行实验,实验表明,与传统启发式推荐方法和深度学习推荐方法相比,PLR4HCCR在互补关系感知推理和替补噪声消解方面均具有较大的优势,继而使其在低阶、高阶和混合阶互补云API推荐中均展示出更优的推荐效果和更强的泛化能力.进一步,超参数敏感性实验、实例分析和用户调查验证了方法的有效性、实用性和可行性,这使结合高阶互补关系的云API推荐方法PLR4HCCR不仅更有可能生成开发者满意的结果,而且可有效提升云API服务提供者的收益.
