2.845

2023影响因子

(CJCR)

  • 中文核心
  • EI
  • 中国科技核心
  • Scopus
  • CSCD
  • 英国科学文摘

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

一种同伴知识互增强下的序列推荐方法

胡开喜 李琳 吴小华 解庆 袁景凌

杨飞生, 汪璟, 潘泉, 康沛沛. 网络攻击下信息物理融合电力系统的弹性事件触发控制. 自动化学报, 2019, 45(1): 110-119. doi: 10.16383/j.aas.c180388
引用本文: 胡开喜, 李琳, 吴小华, 解庆, 袁景凌. 一种同伴知识互增强下的序列推荐方法. 自动化学报, 2023, 49(7): 1456−1470 doi: 10.16383/j.aas.c220347
YANG Fei-Sheng, WANG Jing, PAN Quan, KANG Pei-Pei. Resilient Event-triggered Control of Grid Cyber-physical Systems Against Cyber Attack. ACTA AUTOMATICA SINICA, 2019, 45(1): 110-119. doi: 10.16383/j.aas.c180388
Citation: Hu Kai-Xi, Li Lin, Wu Xiao-Hua, Xie Qing, Yuan Jing-Ling. A sequential recommendation method enhanced by peer knowledge. Acta Automatica Sinica, 2023, 49(7): 1456−1470 doi: 10.16383/j.aas.c220347

一种同伴知识互增强下的序列推荐方法

doi: 10.16383/j.aas.c220347
基金项目: 国家自然科学基金(62276196, 61602353), 湖北省重点研发计划项目(2021BAA030), 国家留学基金委基金(202106950041, 留金美[2020] 1509), 安康市科学技术研究发展计划(AK2020-GY-08)资助
详细信息
    作者简介:

    胡开喜:武汉理工大学计算机与人工智能学院博士研究生. 2018年获得重庆大学控制工程硕士学位. 主要研究方向为序列预测. E-mail: issac_hkx@whut.edu.cn

    李琳:武汉理工大学计算机与人工智能学院教授. 2009年获得日本东京大学博士学位. 主要研究方向为信息检索, 推荐系统. 本文通信作者. E-mail: cathylilin@whut.edu.cn

    吴小华:武汉理工大学计算机与人工智能学院博士研究生. 2019年获得西北大学计算机科学与技术硕士学位. 主要研究方向为可解释机器学习. E-mail: xhwu@whut.edu.cn

    解庆:武汉理工大学计算机与人工智能学院副教授. 2013年获得澳大利亚昆士兰大学博士学位. 主要研究方向为流数据挖掘与模式分析. E-mail: felixxq@whut.edu.cn

    袁景凌:武汉理工大学计算机与人工智能学院教授. 2004年获得武汉理工大学博士学位. 主要研究方向为分布式并行计算. E-mail: yjl@whut.edu.cn

A Sequential Recommendation Method Enhanced by Peer Knowledge

Funds: Supported by National Natural Science Foundation of China (62276196, 61602353), Key Research and Development Program of Hubei Province (2021BAA030), Foundation of China Scholarship Council (202106950041, LiuJinMei [2020] 1509), and Ankang Municipal Science and Technology Bureau (AK2020-GY-08)
More Information
    Author Bio:

    HU Kai-Xi Ph.D. candidate at the School of Computer Science and Artificial Intelligence, Wuhan University of Technology. He received his master degree in control engineering from Chongqing University in 2018. His main research interest is sequential prediction

    LI Lin Professor at the School of Computer Science and Artificial Intelligence, Wuhan University of Technology. She received her Ph.D. degree from University of Tokyo, Japan, in 2009. Her research interest covers information retrieval and recommender systems. Corresponding author of this paper

    WU Xiao-Hua Ph.D. candidate at the School of Computer Science and Artificial Intelligence, Wuhan University of Technology. He received his master degree in computer science and technology from Northwest University in 2019. His main research interest is explainable machine learning

    XIE Qing Associate professor at the School of Computer Science and Artificial Intelligence, Wuhan University of Technology. He received his Ph.D. degree from University of Queensland, Australia, in 2013. His research interest covers streaming data mining and pattern analysis

    YUAN Jing-Ling Professor at the School of Computer Science and Artificial Intelligence, Wuhan University of Technology. She received her Ph.D. degree from Wuhan University of Technology, in 2004. Her main research interest is parallel distributed computing

  • 摘要: 序列推荐(Sequential recommendation, SR)旨在建模用户序列中的动态兴趣, 预测下一个行为. 现有基于知识蒸馏(Knowledge distillation, KD)的多模型集成方法通常将教师模型预测的概率分布作为学生模型样本学习的软标签, 不利于关注低置信度序列样本中的动态兴趣. 为此, 提出一种同伴知识互增强下的序列推荐方法(Sequential recommendation enhanced by peer knowledge, PeerRec), 使多个具有差异的同伴网络按照人类由易到难的认知过程进行两阶段的互相学习. 在第1阶段知识蒸馏的基础上, 第2阶段的刻意训练通过动态最小组策略协调多个同伴从低置信度样本中挖掘出可被加强训练的潜在样本. 然后, 受训的网络利用同伴对潜在样本预测的概率分布调节自身对该样本学习的权重, 从解空间中探索更优的兴趣表示. 3个公开数据集上的实验结果表明, 提出的PeerRec方法相比于最新的基线方法在基于Top-k的指标上不仅获得了更佳的推荐精度, 且具有良好的在线推荐效率.
  • 信息物理融合系统(Cyber-physical systems, CPS)作为一种新型智能系统应运而生, 它是一类集成计算、网络和物理实体的复杂系统, 将三者进行有机融合与深度协作, 从而达到对大型物理系统与信息系统的实时感知, 动态控制和信息服务等[1], 典型CPS包括工控系统[2]、供水网络[3]等.智能电网从总体上可以视为由信息网和电力网这两个相互依存的网络构成的一个复合网络, 也是一个典型的CPS[4].同步相量测量装置(Phasor measurement units, PMUs)、广域测量系统(Wide-area measurement systems, WAMS)、变电站自动化等技术为智能电网的实现提供了坚实的基础, 但同时也增加了智能电网对信息资源的依赖.一旦信息网出错或崩溃, 电力网一般很难保持正常运行.这也为电力系统安全稳定运行带来了新的问题[5]:首先, 随着PMUs布点的增多, 调度数据网中传送的PMUs数据的比例将会越来越大, PMUs长期不间断且高刷新频率的传送导致海量的状态和控制信息等在通信网络上传送, 可能会产生网络拥塞, 影响数据实时传送; 其次, 与电力的物理系统相比, 信息系统对恶意攻击具有更明显的脆弱性.由于通信网络的开放性, 会导致电力系统面临各种类型网络攻击, 造成失稳甚至毁坏.如2015年12月23日乌克兰电网遭受协同攻击导致近8万用户家庭突发停电事故, 这次事故被认为是第一起由于网络攻击直接导致停电事故的案例.

    针对上述问题, 已有一些学者进行了相关的研究[5-10].文献[6]首次将事件触发机制引入多域电力系统负荷频率控制(Load frequency control, LFC)中, 有效减少了数据传输量; 文献[8]提出了一种弹性事件触发机制, 应用于多域电力系统LFC当中, 在考虑网络攻击的情况下保证电力系统的稳定性, 并减少了传输的数据量; 文献[9]提出一种针对电力系统状态估计的错误数据注入攻击防御与检测机制, 从保护和检测两方面入手; 基于保护的防御, 主要是识别和保护关键的传感器, 使系统更能抵御攻击; 基于检测的防御, 设计了基于空间和基于时间的检测方案, 以准确识别数据注入攻击.文献[10]介绍了DoS攻击下电网的LFC方法, 通过将电力系统建模为切换系统, 检测DoS (Denial-of-service, DoS)攻击的存在, 以双域电力系统为例, 分析了系统性能问题.由于电动汽车(Electric vehicles, EVs)具有良好的环境特征, 如温室气体排放量少, 噪声污染低等, 并且可用于提高电力系统的可靠性和灵活性[11].本文将EVs引入智能电网中, 与负荷频率控制相结合, 快速抑制系统扰动所引发的频率变化.在考虑DoS攻击的情况下, 对电力CPS进行稳定性分析, 并对事件触发机制和控制器进行联合设计, 从而达到理想的控制效果.

    在本文中, 假设存在大量可用的EVs, 即存在足够的电能储备以达到协助传统电力单元实现负荷频率调节的目的, 控制中心通过聚合器对EVs进行集中管理, 聚合器将分散的各EVs的信息和状态收集起来, 发送给控制中心.

    由于EVs的接入, LFC系统中产生了新的时变时延.本文假设所有的同步电机都有再热热涡轮机, 为了便于说明, 对于电力CPS, 我们将每个域的$M$个EVs等效为一个, 控制器的输出按比例分配给EVs和再热电机, 其中表示分配比例.如图 1所示, 不考虑弹性事件触发机制和网络环境下, 包含EVs的电力CPS动态模型可以描述为[12]

    $ \begin{align} \begin{cases} \dot{{\pmb x}}(t)=A{\pmb x}(t)+B_{0}{\pmb u}(t)+\\ \quad\qquad\sum\limits_{i=1}^{n}\hat{B}_i{\pmb u} (t-\tau_i(t))+ F{\pmb\Delta P_{d}}(t)\\ {\pmb y}(t)=C{\pmb x}(t) \end{cases} \end{align} $

    (1)

    $ \begin{align*} &{\pmb x}(t)=[{\pmb x_1}^{\rm T}(t) \ {\pmb x_2}^{\rm T}(t) \ \cdots \ {\pmb x_n}^{\rm T}(t) ]^{\rm T}\\ &{\pmb y}(t)=[{\pmb y_1}^{\rm T}(t) \ {\pmb y_2}^{\rm T}(t) \ \cdots \ {\pmb y_n}^{\rm T}(t) ]^{\rm T}\\ &{\pmb u}(t)=[{\pmb u_1}^{\rm T}(t) \ {\pmb u_2}^{\rm T}(t) \ \cdots \ {\pmb u_n}^{\rm T}(t) ]^{\rm T}\\ &{\pmb \Delta P_{d}}(t)=[\Delta P_{d1} \ \Delta P_{d2} \ \cdots \ \Delta P_{dn}]^{\rm T}\\ &{\pmb y_i}(t)=[ACE_i \ \int ACE_i{\rm d}t]^{\rm T}\\ &A=\left[\begin{array}{cccc} A_{11}&A_{12}&\cdots&A_{1n}\\ A_{21}&A_{22}&\cdots&A_{2n}\\ \vdots&\vdots&\ddots&\vdots\\ A_{n1}&A_{n2}&\cdots&A_{nn}\\ \end{array}\right] \end{align*} $

    图 1  基于弹性事件触发机制的电力CPS负载频率控制模型
    Fig. 1  Grid CPS LFC model with a resilient event-triggered

    $ \begin{align*} &A_{ii}= \left[\begin{array}{ccccccc} -\dfrac{D_i}{M_i}&\dfrac{1}{M_i}&0&0&\dfrac{1}{M_i}&-\dfrac{1}{M_i}&0\\[3mm] 0&-\dfrac{1}{T_{ci}}&\dfrac{1}{T_{ci}}&0&0&0&0\\[3mm] -\dfrac{F_{pi}}{R_iT_{gi}}&0&-\dfrac{1}{T_{ri}}&\dfrac{T_{gi}-F_{pi}T_{ri}} {T_{ri}T_{gi}}&0&0&0\\[3mm] -\dfrac{1}{R_iT_{gi}}&0&0&-\dfrac{1}{T_{gi}}&0&0&0\\[3mm] 0&0&0&0&-\dfrac{1}{T_{EVi}}&0&0\\ 2\pi\sum\limits_{j=1, j\ne i}^{n}T_{ij}&0&0&0&0&0&0\\ \beta_i&0&0&0&0&1&0 \end{array}\right]\\ &{\pmb x_i}(t)=[\Delta f_i \ \Delta P_{gi} \ \Delta P_{mi} \ \Delta X_{gi} \ \Delta P_{EVi} \ \Delta P_{tie-i} \ \int ACE_i{\rm d}t ]^{\rm T} \end{align*} $

    $ \begin{align*} &B_0=\left[\begin{array}{cccc} B_{011}&0&\cdots&0\\ 0&B_{022}&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ 0&0&\cdots&B_{0nn}\\ \end{array}\right]\\ &C={\rm diag}\{C_1, \ C_2, \ \cdots, \ C_n\}\\ &B_{0ii}=\left[0 \ 0 \ \frac{F_{pi}\alpha_{i0}}{T_{gi}} \ \frac{\alpha_{i0}}{T_{gi}} \ 0 \ 0 \ 0\right]^{\rm T}\\ &C_i=\left[\begin{array}{ccccccc} \beta_i&0&0&0&0&1&0\\ 0&0&0&0&0&0&1 \end{array}\right]\\ &F={\rm diag}\{F_1, \ F_2, \ \cdots, \ F_n\}\\ &F_i=\left[-\frac{1}{M_i} \ 0 \ 0 \ 0 \ 0 \ 0 \ 0\right]^{\rm T}\\ &\hat{B}_{i}=\left[\begin{array}{ccccc} 0&\cdots&0&\cdots&0\\ \vdots&\ddots&\vdots&\ddots&\vdots\\ 0&\cdots&B_{iii}&\cdots&0\\ \vdots&\ddots&\vdots&\ddots&\vdots\\ 0&\cdots&0&\cdots&0 \end{array}\right]\\ &B_{iii}=\left[0 \ 0 \ 0 \ 0 \ \frac{K_{EVi}\alpha_{i1}}{T_{EVi}} \ 0 \ 0\right]^{\rm T}\\ &A_{ij}=\left[\begin{array}{cccc} 0&0&\cdots&0\\ 0&0&\cdots&0\\ \vdots&\vdots&\ddots&\vdots\\ -2\pi T_{ij}&0&\cdots&0 \end{array}\right] T_{ij}=T_{ji} \end{align*} $

    表 1给出了系统中具体的符号意义.每个区域$i$的ACE (Area control error)信号定义为频率偏差与区域之间联络线电力交换之和:

    表 1  带EVs电力CPS负载频率控制模型参数($i=1, 2, \cdots, n$)
    Table 1  Parameters of power CPS LFC model including EV aggregators ($i=1, 2, \cdots, n$)
    参数符号Parameter notations 物理含义Physical meanings
    $M_i$ 惯性常数
    Inertia constant
    $D_i$ 负载阻尼常数
    Load-damping constant
    $T_{gi}$ 调速器时间常数
    Time constant of governor
    $T_{ci}$ 涡轮机时间常数
    Time constant of turbine
    $T_{ri}$ 再热时间常数
    Time constant of reheat
    $F_{pi}$ 总涡轮功率分数
    Fraction of total turbine power
    $R_i$ 转速
    Speed droop
    $\beta_i$ 频率偏差系数
    Frequency bias factor
    $K_{EVi}$ 电池系数
    Battery coefficient
    $T_{EVi}$ 电池时间常数
    Time constant of battery
    $T_{ij}$ 联络线同步系数
    Synchronizing coefficient of tie-line
    $\Delta f_i$ 频率偏差
    Deviation of frequency
    $\Delta P_{tie-i}$ 联络线的功率交换
    Power transfer of tie-line
    下载: 导出CSV 
    | 显示表格

    $ \begin{equation*} ACE_i=\beta_i \Delta f_i+\Delta P_{tie-i} \end{equation*} $

    假设传输线路为无损传输, 电力CPS控制各区域之间的联络线功率交换满足$\sum_{i=1}^n \Delta P_{tie-i}=0$.

    PI型LFC可以写成

    $ \begin{align} {\pmb u}(t)&=-K{\pmb y}(t) \end{align} $

    (2)

    其中, $K={\rm diag}\{{\pmb K_1}$, ${\pmb K_2}$, $\cdots$, ${\pmb K_n}\}$, ${\pmb K_i}=[K_{Pi}~K_{Ii}]$, $K_{Pi}$, $K_{Ii}$分别为比例增益和积分增益.

    注1[13].EVs参与负载频率控制可分为两种模式, SOC (State of charge)可控模式和SOC空闲模式. SOC空闲模式, 即EV从电网中消耗电能或释放电能不考虑EV电池的充电状态, 此时第$m$辆EV的增益$K_{Em}=\bar{K_{e}}$, 其中$\bar{K_{e}}=1$.而SOC可控模式中由于EV使用者的需要, 第$m$辆EV参与负载频率控制时需要考虑SOC, 因此通过SOC计算和当前SOC的值可以获得EV的增益$K_{Em}=\bar{K_e}-\bar{K_e}g_m(t)$, 其中$g_m(t)=(\frac{{\rm SOC}_m-{\rm SOC}_{{\rm low(high)}, m}}{{\rm SOC}_{{\rm max(min)}, m}-{\rm SOC}_{{\rm low(high)}, m}})^{v_m}$, $v_m$表示电池的设计规格, ${\rm SOC_{low(high)}}$为低(高)电池SOC, ${\rm SOC_{max(min)}}$为最大(最小)电池SOC.假设在$t$时刻, 区域$i$中有$Me$辆EVs参与LFC, 其中$Me_1$表示处于SOC空闲模式的EVs, $Me_2=Me-Me_1$表示处于SOC可控模式的EVs, 则区域$i$总的EVs增益$K_{EVi}=\sum_{m=1}^{Me}\frac{K_{Em}}{Me}=\frac{Me_1}{Me}\bar{K_e}+\frac{1}{Me}\bar{K_e}(Me_2-\sum^{Me}_{m=Me_1+1}g_m(t))$.

    早期事件触发机制是所谓的连续事件触发, 需要特殊硬件对状态进行连续监测.此外, 在触发机制设计中, 必须确保任意两个事件触发时刻之间的最小时间间隔严格大于零, 如果最小事件间隔时间为零, 就会出现无限事件发生在有限时间内的奇诺(Zeno)现象[14].为了解决这两个问题, 文献[15]提出一种基于采样数据的离散型事件触发机制:

    $ \begin{eqnarray} [{\pmb x}(t_k+jh)-{\pmb x}(t_k)]^{\rm T}\Omega[{\pmb x}(t_k+jh)-{\pmb x}(t_k)]\leq\nonumber\\ \sigma {\pmb x}^{\rm T}(t_k)\Omega {\pmb x}(t_k), j\in {\bf N} \end{eqnarray} $

    (3)

    当上述条件违背时, 传感器将采样数据传输给控制器.其中$t_k$为事件触发时刻, $\Omega>0$为触发矩阵, $\sigma\in(0, 1)$为触发参数, $h$为采样周期.该触发机制可以保证最小事件间隔$T_{etc}\geq h$, 避免出现Zeno现象, 并且不需要特殊硬件对状态进行连续监测.

    本文假设DoS攻击的能量是有限的, 即DoS攻击的持续时间是有限的, 当DoS攻击发生时通信中断. DoS攻击的发生将直接导致通信信道上正在传输的数据丢失, 因此并不是所有事件触发时刻的状态都能成功传输到控制器侧.假设DoS攻击发生时, 连续丢包量为$\tau_M$, 那么DoS攻击的持续时间$\tau_{\rm dos}\leq \tau_M T_{etc}$, 为了简单起见, 取$\tau_{\rm dos}\leq \tau_M h$.此时, 传统事件触发机制(3)将不能直接用于判断采样数据传输与否.为了消除DoS攻击所产生的影响, 提出下列弹性事件触发机制:

    $ \begin{eqnarray} [{\pmb x}(r_k+jh)-{\pmb x}(r_k)]^{\rm T}\Omega[{\pmb x}(r_k+jh)-{\pmb x}(r_k)]\leq\nonumber\\ \sigma_r {\pmb x}^{\rm T}(r_k)\Omega {\pmb x}(r_k), j\in {\bf N} \end{eqnarray} $

    (4)

    $\sigma_r$为事件触发参数.通常, $\sigma_r < \sigma$, 也就是说弹性事件触发机制将会产生更多的触发状态用以消除DoS攻击对系统所造成的影响.此时, 将存在三个时刻, 采样时刻$kh$, 触发时刻$r_k$和成功传输到控制器侧的时刻$t_k$.令$S_0=\{0, h, 2h, 3h, \cdots, jh\}, ~ j\in {\bf N}$, $S_1=\{0, r_1, r_2, r_3, \cdots, r_k\}, r_k/h\in {\bf N}$, $S_2=\{0, t_1, t_2, t_3, \cdots, t_k\}, t_k/h\in {\bf N}$, 则$S_2\subseteq S_1\subseteq S_0$.当$\sigma_r=\sigma$时, 意味着没有DoS攻击发生.下文将推导出系统所能容忍的DoS攻击最大持续时间.

    此时控制器输入${\pmb u}(t)={\pmb u}(t_k), t\in[t_k+d_{\tau_k}, t_{k+1}+d_{\tau_{k+1}})$, $d_{\tau_k}$表示传感器到控制器的传输时延, 最大值为$\bar{d_{\tau}}$, 此时系统(1)可以表示为

    $ \begin{eqnarray} \dot{{\pmb x}}(t)=A{\pmb x}(t)+B{\pmb u}(t_k)+F{\pmb\Delta P_{d}}(t) \end{eqnarray} $

    (5)

    其中, $B=B_{0}+\sum_{i=1}^{n}\hat{B}_i$.利用文献[15]中提出的时滞模型方法对系统进行稳定性分析和控制器设计.

    定义

    $ \begin{eqnarray*} d(t)= \left\{\begin{aligned} &t-t_{k}, t\in I_{1}\\ &t-t_{k}-lh, t\in I^l_{2}, \quad l=1, 2, \cdots, m_k-1\\ &t-t_{k}-m_kh, t\in I_{3}\\ \end{aligned}\right.\ \end{eqnarray*} $

    其中, $I_{1}=[t_{k}+d_{\tau_{k}}, t_{k}+h+\bar{d_\tau}), I^{l}_{2}=[t_{k}+lh+\bar{d_\tau}, t_{k}+lh+h+\bar{d_\tau}), I_{2}=\cup^{m_k-1}_{l=1}I^{l}_{2}$, 和$I_{3}=[t_{k}+m_kh+\bar{d_\tau}, t_{k+1}+d_{\tau_{k+1}})$. $[t_{k}+d_{\tau_{k}}, t_{k+1}+d_{\tau_{k+1}})=I_{1}\cup I_{2}\cup I_{3}$.

    因此, 可以得到$0\leq d(t) < \bar{d}$, 其中$\bar{d}=h+\bar{d_\tau}$.对于$\forall t\in[t_{k}+d_{\tau_{k}}, t_{k+1}+d_{\tau_{k+1}})$, 我们定义

    $ \begin{eqnarray*} {\pmb e}(t)= \begin{cases} 0, &t\in I_{1}\\ {\pmb x}(t_{k})-{\pmb x}(t_{k}+lh), &t\in I^l_{2}\\ {\pmb x}(t_{k})-{\pmb x}(t_{k}+m_kh), &t\in I_{3}\\ \end{cases} \end{eqnarray*} $

    结合$d(t)$和${\pmb e}(t)$定义, 当$t\in[t_{k}+d_{\tau_{k}}, t_{k+1}+d_{\tau_{k+1}})$时, 系统(5)可以变形为

    $ \begin{align} \dot{{\pmb x}}(t)=\, &A{\pmb x}(t)-BKC({\pmb e}(t)+\nonumber\\ &{\pmb x}(t-d(t)))+F{\pmb\Delta P_{d}}(t) \end{align} $

    (6)

    本文的主要目的是在DoS攻击的情况下研究电力系统的稳定性、设计弹性事件触发机制和控制器联合求解方案, 同时系统还满足如下条件:

    1) 当负载干扰为零时(即${\pmb\Delta P_d}=0$), 电力系统在DoS攻击下是渐近稳定的;

    2) 当系统的初值为零时, 对于任意非零${\pmb\Delta P_d}\in L_2[0, \infty)$, 有$\Vert {\pmb y}(t)\Vert _2 \leq \gamma \Vert {\pmb\Delta P_d}\Vert_2$, 其中$\gamma$是给定的$H_{\infty}$性能.

    引理1[16].令矩阵$R_1, R_2$为正定矩阵, 标量$a\in(0, 1)$, 以及向量${\pmb \omega_1}, {\pmb\omega_2} \in {\bf R}^m$, 那么对于任意矩阵$Y_1, Y_2 \in {\bf R}^{m\times m}$, 下列不等式成立:

    $ \begin{align*} &\frac{1}{a}{\pmb\omega_1}^{\rm T}R_1{\pmb\omega_1}+\frac{1}{1-a}{\pmb\omega_2}^{\rm T}R_2{\pmb\omega_2}\geq \nonumber\\ &~~~~~~~{\pmb\omega_1}^{\rm T}[R_1+(1-a)(R_1-Y_1R_2^{-1}Y_1^{\rm T})]{\pmb\omega_1}+ \nonumber \\ &~~~~~~~{\pmb\omega_2}^{\rm T}[R_2+a(R_2-Y_2^{\rm T}R_1^{-1}Y_2)]{\pmb\omega_2}+ \nonumber\\ &~~~~~~~2{\pmb\omega_1}^{\rm T}[aY_1+(1-a)Y_2]{\pmb\omega_2} \end{align*} $

    本节对基于事件触发机制的电力CPS进行稳定性分析, 选择Lyapunov-Krasovskii泛函为

    $ \begin{align*} V(t)=\, &V_{1}(t)+V_{2}(t)+V_{3}(t)\\ V_{1}(t)=\, &\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\int^{t}_{t-\bar{d}} {\pmb x}^{\rm T}(s){\rm d}s\end{array}\right]\times\\& U\left[\begin{array}{c}{\pmb x}(t)\\\int^{t}_{t-\bar{d}}{\pmb x}(s){\rm d}s\end{array}\right]\\ V_{2}(t)=\, &\int_{t-\bar{d}}^{t}\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\bar{d}{\pmb x}^{\rm T}(s)\end{array}\right]Q\left[\begin{array}{c}{\pmb x}(t)\\\bar{d}{\pmb x}(s)\end{array}\right]{\rm d}s\\ V_{3}(t)=\, &\bar{d}\int_{t-\bar{d}}^{t}\int_{v}^{t}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s{\rm d}v \end{align*} $

    其中, $U$, $Q$对称, $R>0$.应用Jensen不等式,

    $ \begin{align*} V_{2}(t)\geq\, &\frac{1}{\bar{d}}\int_{t-\bar{d}}^{t}\left[\begin{array}{c}{\pmb x}(t)\\ \bar{d}{\pmb x}(s)\end{array}\right]^{\rm T}{\rm d}s\times\\ &Q\int_{t-\bar{d}}^{t}\left[\begin{array}{c}{\pmb x}(t)\\ \bar{d}{\pmb x}(s)\end{array}\right]{\rm d}s=\\ &\bar{d}\left[\begin{array}{c}{\pmb x}(t)\\ \int^{t}_{t-\bar{d}}{\pmb x}(s){\rm d}s\end{array}\right]^{\rm T}\times\\&Q\left[\begin{array}{c}{\pmb x}(t)\\ \int^{t}_{t-\bar{d}}{\pmb x}(s){\rm d}s\end{array}\right] \end{align*} $

    从而得到:

    $ \begin{align*} &V(t)\geq\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\int^{t}_{t-\bar{d}}{\pmb x}^{\rm T}(s){\rm d}s\end{array}\right](U+\bar{d}Q)\times\\ &~~\left[\begin{array}{c}{\pmb x}(t)\\\int^{t}_{t-\bar{d}}{\pmb x}(s){\rm d}s\end{array}\right] +\bar{d}\int_{t-\bar{d}}^{t}\int_{v}^{t}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s{\rm d}v\\ \end{align*} $

    由上式, 当矩阵$U$和$Q$满足$U+\bar{d}Q>0$, $V(t)>0$, 此时并不需要$U, Q$和$R$都正定.对$V(t)$进行求导,

    $ \begin{align*} \dot{V}_{1}(t)=\, &2\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\int^{t}_{t-\bar{d}}{\pmb x}^{\rm T}(s){\rm d}s\end{array}\right]U \times\\ &\left[\begin{array}{c}\dot{{\pmb x}}(t)\\{\pmb x}(t)-{\pmb x}(t-\bar{d})\end{array}\right]\\ \dot{V}_{2}(t)=\, &\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\bar{d}{\pmb x}^{\rm T}(t)\end{array}\right]Q\left[\begin{array}{c}{\pmb x}(t)\\\bar{d}{\pmb x}(t)\end{array}\right]-\\ &\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\bar{d}{\pmb x}^{\rm T}(t-\bar{d})\end{array}\right]Q\left[\begin{array}{c}{\pmb x}(t)\\\bar{d}{\pmb x}(t-\bar{d})\end{array}\right]+\\ &2\int_{t-\bar{d}}^{t}\left[\begin{array}{cc}{\pmb x}^{\rm T}(t)&\bar{d}{\pmb x}^{\rm T}(s)\end{array}\right]Q\left[\begin{array}{c}\dot{{\pmb x}}(t)\\0\end{array}\right]{\rm d}s \end{align*} $

    $ \begin{align*} \dot{V}_{3}(t)%=\bar{d}^2\dot{x}^{T}(t)R\dot{x}(t)-\bar{d}\int_{t-\bar{d}}^{t}\dot{x}^T(s)R\dot{x}(s)ds\\ =\, &\bar{d}^2\dot{{\pmb x}}^{\rm T}(t)R\dot{{\pmb x}}(t)-\bar{d}\int_{t-d(t)}^{t}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s-\\ &\bar{d}\int_{t-\bar{d}}^{t-d(t)}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s \end{align*} $

    为了便于描述, 定义${\pmb\xi_1}:=\text{col}\{{\pmb x}(t)$, ${\pmb x}(t-d(t))$, ${\pmb x}(t-\bar{d})$, $\frac{1}{d(t)}\int_{t-d(t)}^{t}{\pmb x}(s){\rm d}s$, $\frac{1}{\bar{d}-d(t)}\int_{t-\bar{d}}^{t-d(t)}{\pmb x}(s){\rm d}s$, $\frac{1}{d^2(t)}\int_{t-d(t)}^{t}(t-s){\pmb x}(s){\rm d}s$, $\frac{1}{(\bar{d}-d(t))^2}\int_{t-\bar{d}}^{t-d(t)}(t-d(t)-s){\pmb x}(s){\rm d}s$, ${\pmb e}(t)$, ${\pmb\Delta P_d}(t)\}$.

    应用文献[17]中的Lemma 1,

    $ \begin{align} & \bar{d}\int_{t-\bar{d}}^{t-d(t)}{{{{\mathit{\boldsymbol{\dot{x}}}}}^{\text{T}}}}(s)R\mathit{\boldsymbol{\dot{x}}}(s)\text{d}s\ge \frac{1}{a}{{({{\Gamma }_{1}}{{\mathit{\boldsymbol{ }}\!\!\xi\!\!\text{ }}_{1}}(t))}^{\text{T}}}\hat{R}({{\Gamma }_{1}}{{\mathit{\boldsymbol{ }}\!\!\xi\!\!\text{ }}_{1}}(t)) \\ & \bar{d}\int_{t-d(t)}^{t}{{{{\mathit{\boldsymbol{\dot{x}}}}}^{\text{T}}}}(s)R\mathit{\boldsymbol{\dot{x}}}(s)\text{d}s\ge \frac{1}{1-a}{{({{\Gamma }_{2}}{{\mathit{\boldsymbol{ }}\!\!\xi\!\!\text{ }}_{1}}(t))}^{\text{T}}}\hat{R}({{\Gamma }_{2}}{{\mathit{\boldsymbol{ }}\!\!\xi\!\!\text{ }}_{1}}(t)) \\ & {{\Gamma }_{1}}:=\text{col}\{{{e}_{2}}-{{e}_{3}}, {{e}_{2}}+{{e}_{3}}-2{{e}_{5}}, {{e}_{2}}-{{e}_{3}}-6{{e}_{5}}+12{{e}_{7}}\} \\ & {{\Gamma }_{2}}:=\text{col}\{{{e}_{1}}-{{e}_{2}}, {{e}_{1}}+{{e}_{2}}-2{{e}_{4}}, {{e}_{1}}-{{e}_{2}}-6{{e}_{4}}+12{{e}_{6}}\text{ }\!\!\}\!\!\text{ } \\ & \hat{R}=\text{diag}\{R, 3R, 5R\} \\ \end{align} $

    其中, $a=\frac{\bar{d}-d(t)}{\bar{d}}$.进一步应用引理1, 可以得到:

    $ \begin{align*} &\bar{d}\int_{t-d(t)}^{t}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s+\\ &~~~~~\bar{d}\int_{t-\bar{d}}^{t-d(t)}\dot{{\pmb x}}^{\rm T}(s)R\dot{{\pmb x}}(s){\rm d}s\geq\\ %&~~~~\geq\frac{1}{a}(\Gamma_1\xi_1(t))^T\hat{R}(\Gamma_1\xi_1(t))\\ %&~~~~~+\frac{1}{1-a}(\Gamma_2\xi_1(t))^T\hat{R}(\Gamma_2\xi_1(t))\\ &~~~~~{\pmb\xi_1}^{\rm T}(t)[(2-a)\Gamma_1^{\rm T}\hat{R}\Gamma_1+(1+a)\Gamma_2^{\rm T}\hat{R}\Gamma_2+\\ &~~~~~{\rm Sym}\{\Gamma_1^{\rm T}[aY_1+(1-a)Y_2]\Gamma_2\}-\Pi_2]{\pmb\xi_1}(t)\\ &\Pi_2=a\Gamma_2^{\rm T}Y_2^{\rm T}\hat{R}^{-1}Y_2\Gamma_2+(1-a)\Gamma_1^{\rm T}Y_1\hat{R}^{-1}Y_1^{\rm T}\Gamma_1 \end{align*} $

    因此

    $ \begin{align*} \dot{V}(t)\leq\,&{\pmb \xi}_1^{\rm T}(t)(\Pi_1+\Pi_2){\pmb \xi_1}(t)\\ \Pi_1:=\, &\ell_1^{\rm T} Q\ell_1-\ell_2^{\rm T} Q\ell_2+\bar{d}^2\ell_0^{\rm T} R\ell_0-\\ &(2-a)\Gamma_1^{\rm T}\hat{R}\Gamma_1-(1+a)\Gamma_2^{\rm T}\hat{R}\Gamma_2+\\ &{\rm Sym}\{\ell_3^{\rm T}U\ell_4+ \bar{d}\ell_4^{\rm T}Q\ell_5-\\ &\Gamma_1^{\rm T}[aY_1+(1-a)Y_2]\Gamma_2\}\\ \ell_0:=\, &Ae_1-BKC(e_2+e_8)+Fe_9\\ \ell_1:=\, &{\rm col}\{e_1, \bar{d}e_1\}, \ell_2={\rm col}\{e_1, \bar{d}e_3\}\\ \ell_3:=\, &{\rm col}\{\ell_0, e_1-e_3\}\\ \ell_4:=\, &{\rm col}\{e_1, d(t)e_4+(\bar{d}-d(t))e_5\}\\ \ell_5:=\, &{\rm col}\{\ell_0, 0\} \end{align*} $

    结合传统事件触发机制(3), 对于$t\in[t_k+d_{\tau_k}$, $t_{k+1}+$ $d_{\tau_{k+1}})$, 我们可以得到:

    $ \begin{align*} \dot{V}(t)\leq\,&{\pmb\xi_1}^{\rm T}(t)(\hat{\Pi}_1+\Pi_2){\pmb\xi_1}(t)-\\ &{\pmb y}^{\rm T}(t){\pmb y}(t)+\gamma^2{\pmb \Delta P}^{\rm T}_d(t){\pmb \Delta P_d}(t)\\ \hat{\Pi}_1=\, &\Pi_1+\sigma (e_2+e_8)^{\rm T}\Omega (e_2+e_8)-e_8^{\rm T}\Omega e_8+\\ &e_1^{\rm T}C^{\rm T}Ce_1-\gamma^2 e_9^{\rm T}e_9 \end{align*} $

    使用Schur补, 可以得到当存在$R>0$, 对称矩阵$U, Q$, 实矩阵$Y_1$, $Y_2$以及标量$\bar{d}$, 满足$U+\bar{d}Q>0$, 并且LMIs (7)和(8)成立时, 系统(6)渐近稳定.

    $ \begin{eqnarray} \left[\begin{array}{cc} \hat{\Pi}_1&\Gamma_2^{\rm T}Y_2^{\rm T}\\ *&-\hat{R}\\ \end{array}\right]<0 \end{eqnarray} $

    (7)

    $ \begin{eqnarray} \left[\begin{array}{cc} \hat{\Pi}_1&\Gamma_1^{\rm T}Y_1\\ *&-\hat{R}\\ \end{array}\right]<0 \end{eqnarray} $

    (8)

    同时, 当LMIs (7)和(8)成立时, 我们可以得到:

    $ \begin{align*} \dot{V}(t)\leq -{\pmb y}^{\rm T}(t){\pmb y}(t)+\gamma^2{\pmb \Delta P}^{\rm T}_d(t){\pmb \Delta P_d}(t) \end{align*} $

    因为${\pmb x}(t)$在$t$上连续, 所以$\dot{V}(t)$在$t$上也连续.因此对不等式两边同时从$0$到$\infty$对$t$进行积分, 可得:

    $ \begin{align*} V(\infty)-V(0)\leq &\int^{\infty}_{0}[-{\pmb y}^{\rm T}(t){\pmb y}(t)+\\ &\gamma^2{\pmb \Delta P}^{\rm T}_d(t){\pmb \Delta P_d}(t)]{\rm d}s \end{align*} $

    从而在零初始条件下,

    $ \begin{align*} \int^{\infty}_{0}[-{\pmb y}^{\rm T}(t){\pmb y}(t)+\gamma^2{\pmb\Delta P}^{\rm T}_d(t){\pmb\Delta P_d}(t)]{\rm d}s \geq 0 \end{align*} $

    即, 对任意非零${\pmb\Delta P_d}(t) \in L_2[0, +\infty)$以及给定的$\gamma$, $\Arrowvert {\pmb y}(t)\Arrowvert_2 \leq \gamma\Arrowvert {\pmb\Delta P_d}\Arrowvert_2 $.在${\pmb\Delta P_d}(t)=0$的情况下, 存在$\varepsilon>0$使得${\pmb x}(t)\ne 0$时, $\dot{V}(t)\leq -\varepsilon \Arrowvert {\pmb x}(t)\Arrowvert_2$.因此, 系统(6)渐近稳定并具$H_{\infty}$范数界.

    本节将在DoS攻击发生时, 基于弹性事件触发机制(4)对系统进行稳定性分析, 找出弹性触发参数$\sigma_r$以及所能容忍的最大DoS攻击持续时间.

    定理1.对于给定的标量$\sigma>0$, $\bar{d}>0$, $\gamma>0$, 当存在实矩阵$R>0$, $\Omega>0$, 对称矩阵$Q, U$满足$U+\bar{d}Q>0$, 实矩阵$Y_1$, $Y_2$以及给定的控制器增益矩阵$K$使得LMIs (7)和(8)成立, 那么在弹性事件触发机制(4)下, 系统(1)渐近稳定, 并且$\sigma_r$满足:

    $ \begin{equation} \sigma_r \leq (\sqrt[\tau_M+1]{\sqrt{\sigma}+1}-1)^2 \end{equation} $

    (9)

    同时, 当$\sigma_r$已知的情况下, 由式(10)可以得到系统所能容忍的最大DoS攻击持续时间

    $ \begin{equation} \tau_{\rm dos}\leq \tau_{M}h \leq \lfloor[{\rm log}_{1+\sqrt{\sigma_r}}(1+\sqrt{\sigma})]\rfloor h \end{equation} $

    (10)

    其中, $\lfloor*\rfloor$表示向下取整.

    证明.为了符号表示方便, 我们假设在两次成功传输时刻的区间$[t_k, t_{k+1})$存在$\tau_M$个由于DoS攻击所造成的未成功传输但是触发的状态${\pmb x}(r_j)$, 其中$t_k=r_0 < r_1 < r_2 < \cdots < r_{\tau_M} < r_{\tau_M+1}=t_{k+1}$.

    因此, 区间$[t_k, t_{k+1})$可以分为多个小区间$[r_j, r_{j+1}), j\in\{0, 1, 2, \cdots, \tau_M\}$.

    $ \begin{align*} &\vert {\pmb e}(t)\vert=\vert {\pmb x}(t)-{\pmb x}(t_k)\vert\leq\\ &~~~~~~~~~~~~\sum^{j-1}_{p=0}\vert {\pmb x}(r_{p+1})-{\pmb x}(r_p) \vert+\vert {\pmb x}(t)-{\pmb x}(r_j)\vert \end{align*} $

    其中, $t\in[r_j, r_{j+1}), t=r_j+lh, l\in {\bf N}$.应用弹性事件触发机制(4),

    $ \begin{align*} &\vert {\pmb x}(t)-{\pmb x}(r_j)\vert \leq \sqrt{\sigma_r}\vert {\pmb x}(r_j)\vert \\ &\vert {\pmb x}(r_{p+1})-{\pmb x}(r_p)\vert \leq \sqrt{\sigma_r}\vert {\pmb x}(r_{p})\vert \end{align*} $

    因此

    $ %\begin{align*} %&\vert e(t)\vert=\vert x(t)-x(t_k)\vert\\ %&~~~~~~~~~~~~\leq \sum^{j-1}_{p=0}\sqrt{\sigma_r}\vert x(r_p) \vert+\sqrt{\sigma_r}\vert x(r_j)\vert\\ %&~~~~~~~~~~~~=\sum^{j}_{p=0}\sqrt{\sigma_r}\vert x(r_p) \vert %\end{align*} \begin{align*} &\vert {\pmb e}(t)\vert=\vert {\pmb x}(t)-{\pmb x}(t_k)\vert=\sum^{j}_{p=0}\sqrt{\sigma_r}\vert {\pmb x}(r_p) \vert \end{align*} $

    又由$\vert {\pmb x}(r_{p+1})-{\pmb x}(r_p)\vert \leq \sqrt{\sigma_r}\vert {\pmb x}(r_{p})\vert$可以得到$\vert {\pmb x}(r_{p+1})\vert \leq (1+\sqrt{\sigma_r})\vert {\pmb x}(r_{p})\vert$, 进一步可得,

    $ \begin{align*} &\vert {\pmb x}(r_{p+1}) \vert \leq (1+\sqrt{\sigma_r})^{p+1}\vert {\pmb x}(t_k)\vert, \\ &~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~p=0, 1, \cdots, j-1 \end{align*} $

    所以

    $ \begin{align} &\vert {\pmb e}(t)\vert=\vert {\pmb x}(t)-{\pmb x}(t_k)\vert\leq\nonumber \\ &~~~~~~~~~~~\sum^{j}_{p=0}\sqrt{\sigma_r}(1+\sqrt{\sigma_r})^{p}\vert {\pmb x}(t_k) \vert \leq\nonumber \\ &~~~~~~~~~~~((1+\sqrt{\sigma_r})^{\tau_M+1}-1)\vert {\pmb x}(t_k)\vert \end{align} $

    (11)

    由于$t\in[r_j, r_{j+1})$, 数据包没有成功传输, 因此,

    $ \begin{align} \vert {\pmb x}(t)-{\pmb x}(t_k)\vert \leq \sqrt{\sigma} \vert {\pmb x}(t_k)\vert \end{align} $

    (12)

    结合式(11)和(12)即可以得到定理1.

    本节将在第3.1节和第3.2节稳定性分析的基础上, 对弹性事件触发矩阵和控制器进行联合设计, 从而达到理想的控制效果.

    定理2.对于给定的参数$\gamma>0$, $\bar{d}>0$, 在弹性事件触发通信机制(4)情况下, 存在正定矩阵$R$, $\Omega$, 对称矩阵$Q$, $U$满足$U+\bar{d}Q>0$, 以及实矩阵$P$, $Y_1$和$Y_2$使得下列不等式成立, 此时系统(1)渐近稳定, $H_{\infty}$控制器增益$K=K_c(CP)^{+}$:

    $ \begin{eqnarray} \left[\begin{array}{cc} \widetilde{\widetilde{\Pi}}_1&\Gamma_2^{\rm T}Y_2^{\rm T}\\ *&-\hat{R}\\ \end{array}\right]<0 \end{eqnarray} $

    (13)

    $ \begin{eqnarray} \left[\begin{array}{cc} \widetilde{\widetilde{\Pi}}_1&\Gamma_1^{\rm T}Y_1\\ *&-\hat{R}\\ \end{array}\right]<0 \end{eqnarray} $

    (14)

    其中

    $ \begin{align*} &\widetilde{\widetilde{\Pi}}_1:=\ell_1^{\rm T} Q\ell_1-\ell_2^{\rm T} Q\ell_2+\bar{d}^2e_8^{\rm T}Re_8-\gamma^2e_{10}^{\rm T}e_{10}-\\ &~~~~~~~(1+a)\Gamma_2^{\rm T}\hat{R}\Gamma_2+{\rm Sym}\{\widetilde{\ell}_3^{\rm T}U\ell_4+\bar{d}\ell_4^{\rm T}Q\widetilde{\ell}_5-\\ &~~~~~~~\Gamma_1^{\rm T}[aY_1+(1-a)Y_2]\Gamma_2-(e_1+e_8)^{\rm T}[Pe_8-\\ &~~~~~~~APe_1+BK_c(e_9+e_2)-Fe_{10}]\}-e_9^{\rm T}\Omega e_9+\\ &~~~~~~~\sigma_r (e_2+e_9)^{\rm T}\Omega (e_2+e_9)+e_1^{\rm T}(CP)^{\rm T}CPe_1-\\ &~~~~~~~(2-a)\Gamma_1^{\rm T}\hat{R}\Gamma_1\\ &\widetilde{\ell}_3={\rm col}\{e_8, e_1-e_3\}, \widetilde{\ell}_5={\rm col}\{e_8, 0\} \end{align*} $

    证明.选择非奇异实矩阵$P$[18], 令${\pmb x}(t)=P{\pmb z}(t)$.则${\pmb u}(t)=-KC{\pmb x}(t)=-KCP{\pmb z}(t)=-K_c{\pmb z}(t)$, $K_c=KCP$, 当$t\in [t_k+d_{\tau_k}, t_{k+1}+d_{\tau_{k+1}})$系统(1)变形为

    $ \begin{align} \dot{{\pmb z}}(t)=\, &P^{-1}AP{\pmb z}(t)-P^{-1}BK_c{\pmb z}(t_k)+\nonumber\\ &P^{-1}F{\pmb\Delta P_d}(t) \end{align} $

    (15)

    与第2节中相似, 对$d(t)$和${\pmb e_z}(t)$进行定义, 其中$d(t)$的定义与第3节中相同.

    $ \begin{eqnarray*} {\pmb e_z}(t)= \begin{cases} 0, &t\in I_{1}\\ {\pmb z}(t_{k})-{\pmb z}(t_{k}+lh), &t\in I^l_{2}\\ {\pmb z}(t_{k})-{\pmb z}(t_{k}+m_kh), &t\in I_{3}\\ \end{cases} \end{eqnarray*} $

    则系统(1)可以进一步变形为

    $ \begin{align} &\dot{{\pmb z}}(t)=P^{-1}AP{\pmb z}(t)-P^{-1}BK_c({\pmb e_z}(t)+\nonumber\\ &~~~~~~~~~~~~~~~~~~~~~{\pmb z}(t-d(t)))+P^{-1}F{\pmb\Delta P_d}(t) \end{align} $

    (16)

    本节基于状态${\pmb z}$选择与之前相同的Lyapunov-Krasovskii泛函以及相同的证明方法, 不同的是在第2.1节中控制器增益$K$是预先给出的, 而定理2中可以同时求出控制器增益$K$以及事件触发矩阵.定义${\pmb\xi_2}:=\text{col}\{{\pmb z}(t)$, ${\pmb z}(t-d(t))$, ${\pmb z}(t-\bar{d})$, $\frac{1}{d(t)}\int_{t-d(t)}^{t}{\pmb z}(s){\rm d}s$, $\frac{1}{\bar{d}-d(t)}\int_{t-\bar{d}}^{t-d(t)}{\pmb z}(s){\rm d}s$, $\frac{1}{d^2(t)}\int_{t-d(t)}^{t}(t-s){\pmb z}(s){\rm d}s$, $\frac{1}{(\bar{d}-d(t))^2}\int_{t-\bar{d}}^{t-d(t)}(t-d(t)-s){\pmb z}(s){\rm d}s$, $\dot{{\pmb z}}(t)$, ${\pmb e_z}(t)$, ${\pmb\Delta P_d}(t)\}$.

    从而, 可得到

    $ \begin{align*} \dot{V}(t)\leq {\pmb\xi_2}^{\rm T}(\widetilde{\Pi}_1+\Pi_2){\pmb\xi_2} \end{align*} $

    $ \begin{align*} &\widetilde{\Pi}_1:=\ell_1^{\rm T} Q\ell_1-\ell_2^{\rm T} Q\ell_2+\bar{d}^2e_8^{\rm T}Re_8-(2-a)\Gamma_1^{\rm T}\hat{R}\Gamma_1-\\ &~~~~~~~(1+a)\Gamma_2^{\rm T}\hat{R}\Gamma_2+{\rm Sym}\{\widetilde{\ell}_3^{\rm T}U\ell_4+\bar{d}\ell_4^{\rm T}Q\widetilde{\ell}_5-\\ &~~~~~~~\Gamma_1^{\rm T}[aY_1+(1-a)Y_2]\Gamma_2\}-e_9^{\rm T}\Omega e_9-\gamma^2 e_{10}^{\rm T}e_{10}+\\ &~~~~~~~\sigma_r (e_2+e_9)^{\rm T}\Omega (e_2+e_9)+e_1^{\rm T}(CP)^{\rm T}CPe_1 \end{align*} $

    此外

    $ \begin{align*} &-2[{\pmb z}(t)+\dot{{\pmb z}}(t)]^{\rm T}[P\dot{{\pmb z}}(t)-AP{\pmb z}(t)+BK_c\times\\ &~~~~~~~~~~~~({\pmb z}(t-d(t))+{\pmb e_z}(t))-F{\pmb\Delta P_d}(t)]=0 \end{align*} $

    因此可以得到: $\dot{V}(t)\leq {\pmb\xi_2}^{\rm T}(\widetilde{\widetilde{\Pi}}_1+\Pi_2){\pmb\xi_2}$, 使用Schur补即可得到式(13)和(14).

    本节我们将第3节中的相关结论应用于三域LFC控制系统中, 验证所提出的弹性事件触发机制的有效性, 联合求出控制器增益$K$和事件触发矩阵.弹性事件触发通信机制与三域电力系统LFC的统一框架如图 1所示, 其中的相关参数[19]表 2.

    表 2  带EVs三域LFC模型参数($i=1, 2, 3$)
    Table 2  Parameters of three-area LFC model including EV aggregators ($i=1, 2, 3$)
    参数 取值
    $M_i$ 10
    $D_i$ 1.0
    $T_{gi}$ 0.1
    $T_{ci}$ 0.3
    $T_{ri}$ 10
    $F_{pi}$ 0.05
    $R_i$ 0.05
    $\beta_i$ 21
    $K_{EVi}$ 1
    $T_{EVi}$ 1
    $T_{ij}$ 0.026
    下载: 导出CSV 
    | 显示表格

    令$h=0.01$, 控制器增益$K_{Pi}=0.2$, $K_{Ii}=0.2$, $i=1, 2, 3$, 分配比例$\alpha_0=0.9$, $\alpha_1=0.1$.如表 3所示, 给定$\sigma$和$\sigma_r$, 基于定理1我们可以得到DoS攻击最大持续时间$\tau_{\rm dos}$的值.从中可以看出弹性事件触发机制可以容许DoS攻击所造成的数据包丢失, 当攻击持续时间小于$\tau_{\rm dos}$时, 系统保持稳定.此外, 当给定$\sigma$和$\tau_M$时, 也可以求出$\sigma_r$的值.从表 3可以看出, $\sigma_r$越大, $\tau_{\rm dos}$越小, 这是因为$\sigma_r$越大, 信道上传输的数据量越小, 从而在保证系统稳定的情况下, 所能允许的DoS攻击造成的数据丢失量越小, 因此系统所能承受的最大攻击持续时间越小.当$\sigma=\sigma_r=0.01$时, 系统触发次数为389次, 而根据文献[7$-$8]所提方法可以得到触发次数分别为398次和457次, 因此可以看出本文所提方法可以得到更少的触发次数, 有利于节约通信资源.

    表 3  给定不同的$\sigma$和$\sigma_r$, 最大连续丢包量$\tau_M$和攻击持续时间$\tau_{{\rm dos}}$的值
    Table 3  $\tau_M$ and $\tau_{{\rm dos}}$ for different $\sigma$ and $\sigma_r$
    $\sigma$ 0.1 0.1 0.3 0.3 0.5 0.5 0.5
    $\sigma_r$ 0.01 0.03 0.01 0.03 0.01 0.03 0.05
    $\tau_M$ 2 1 4 2 5 3 2
    $\tau_{{\rm dos}}$ 0.02 0.01 0.04 0.02 0.05 0.03 0.02
    下载: 导出CSV 
    | 显示表格

    基于定理2, 给定$\sigma_r=0.01$时, 使用Matlab/LMI工具箱, 可以同时得到控制器增益和弹性事件触发矩阵如下:

    $ \begin{align*} &K=\left[\begin{array}{cccccc} 2.38&-0.06&0&0&0&0\\ 0&0&2.38&-0.06&0&0\\ 0&0&0&0&2.38&-0.06\\ \end{array}\right]\\ &\Omega={\rm diag}\{\Omega_1, \Omega_2, \Omega_3\} \end{align*} $

    此时, 系统的响应曲线如图 2所示, 由图中可以看出系统处于稳定状态.

    $ \begin{align*} &\Omega_i=\left[\begin{array}{ccccccc} 3.64&-172.49&-11.01&85.51&20.83&-35.07&-2.18\\ -172.49&34\, 113.53&-477.46&-13\, 091.96&-15\, 595.51&-2\, 455.89&133.46\\ -11.01&-477.46&2\, 559.74&40\, 074.27&-171.10&339.19&-5.25\\ 85.51&-13\, 091.96&40\, 074.27&992\, 118.76&-1\, 436&-215.22&-11.12\\ 20.83&-15\, 595.51&-171.10&-1\, 436&8\, 599&2\, 037.21&-25.42\\ -35.07&-2\, 455.89&339.19&-215.22&2\, 037.21&1\, 125.41&-2.55\\ -2.18&133.46&-5.25&-11.12&-25.42&-2.55&7.96\\ \end{array}\right], \quad i=1, 2, 3 \end{align*} $

    图 2  系统变化频率曲线和功率交换曲线
    Fig. 2  The curve of frequency variation and power transfer

    本文将电动汽车引入电力CPS中, 提出了一种弹性事件触发机制, 能够容忍DoS攻击所造成的数据包丢失, 并给出了系统所能承受的最大DoS攻击持续时间.构建新型Lyapunov-Krasovskii泛函, 对系统进行稳定性分析, 并联合求出弹性控制器增益和事件触发矩阵.所得到的时滞依赖稳定性条件中, 矩阵$U$, $Q$只需要对称即可, 放松了对其正定性的要求.由于所提出的弹性事件触发通信机制, 在保证电力CPS稳定的情况下, LFC控制器输入只在需要的时候进行更新, 并可以消除DoS攻击对系统稳定性所造成的不利影响.最后, 通过三域电力系统仿真, 验证了所提出方法的有效性.


  • 11 一种基于最优传输理论衡量两个分布间距离的度量方式, 目前只在一维分布、高斯分布等少数几种分布上存在闭式解.
  • 图  1  用户动态兴趣在潜在空间中的表示与推断

    Fig.  1  The representation and inference of dynamic interests in latent representation spaces

    图  2  PeerRec模型的网络结构

    Fig.  2  The architecture of our proposed PeerRec

    图  3  基于刻意训练的互相学习

    Fig.  3  An illustration of deliberate practice based mutual learning

    图  4  PeerRec变体在HR@1指标上的对比

    Fig.  4  The comparison between the variants of PeerRec in terms of HR@1

    图  5  不同同伴网络数量的性能对比

    Fig.  5  The evaluation on different number of peers

    图  6  Batch大小设置为256时, 模型的迭代速率

    Fig.  6  The running speed of different models with batch size 256

    图  7  超参数敏感性分析

    Fig.  7  Sensitivity analysis of hyper-parameters

    表  1  实验集数据统计表

    Table  1  Statistics of dataset

    ML-1mLastFMToys
    用户数量6 0401 09019 412
    行为类别数量3 4163 64611 924
    最长序列的行为数量2 275897548
    最短序列的行为数量1633
    序列的平均行为数量163.5046.216.63
    序列行为数量的方差192.53 77.698.50
    下载: 导出CSV

    表  2  与基线模型在精度指标上的对比

    Table  2  The comparison with baselines in terms of accuracy based metrics

    数据集模型HR@1HR@5HR@10NDCG@5NDCG@10MRR
    ML-1mPOP0.04070.16030.27750.10080.13830.1233
    BERT4Rec[17]0.36950.68510.78230.53750.56900.5108
    S3-Rec[1]0.28970.65750.79110.45570.52660.4535
    HyperRec[9]0.31800.66310.77380.50140.53750.4731
    R-CE[11]0.39880.64780.74040.53270.56270.5179
    STOSA[14]0.32220.65460.78440.49670.53890.4716
    PeerRec (同伴1)0.42500.71970.81410.58430.61500.5600
    PeerRec (同伴2)0.42520.72250.81410.58600.61570.5610
    LastFMPOP0.02020.09080.17800.05440.08250.0771
    BERT4Rec[17]0.10910.32940.46140.22270.26480.2266
    S3-Rec[1]0.11560.28440.42290.20030.24520.2148
    HyperRec[9]0.11460.31470.46880.21500.26460.2241
    R-CE[11]0.06510.18350.28620.12430.15700.1397
    STOSA[14]0.07520.21650.34120.14580.18600.1556
    PeerRec (同伴1)0.12940.34950.47890.23390.27550.2341
    PeerRec (同伴2)0.12480.33580.48350.23180.27960.2378
    ToysPOP0.02600.10460.18480.06520.09090.0861
    BERT4Rec[17]0.13900.33790.45960.24090.28020.2444
    S3-Rec[1]0.09900.30230.43930.20210.24630.2081
    HyperRec[9]0.11470.28750.39090.20310.23650.2087
    R-CE[11]0.11300.31890.45290.21790.26110.2233
    STOSA[14]0.18380.35870.45500.27490.30590.2732
    PeerRec (同伴 1)0.17940.37030.47850.27850.31340.2810
    PeerRec (同伴 2)0.17820.37060.47780.27810.31270.2803
    下载: 导出CSV

    表  3  知识蒸馏与刻意训练对比

    Table  3  The comparison between knowledge distillation and deliberate practice

    数据集HR@1NDCG@5MRR
    知识蒸馏[39]ML-1m0.39520.56560.5386
    LastFM0.11190.23010.2314
    Toys0.16930.27610.2767
    刻意训练 PeerRecML-1m0.42510.58520.5605
    LastFM0.12710.23290.2360
    Toys0.17880.27830.2807
    下载: 导出CSV

    表  4  PeerRec模型采用不同初始化的性能对比

    Table  4  The performance comparison between different initializations of our PeerRec

    数据集初始化方式HR@1NDCG@5MRR
    ML-1mTND0.42510.58520.5605
    Xavier0.42630.58520.5600
    Kaiming0.42780.59110.5652
    LastFMTND0.12710.23290.2360
    Xavier0.12940.23970.2424
    Kaiming0.12470.22570.2342
    ToysTND0.17880.27830.2807
    Xavier0.17750.27940.2811
    Kaiming0.18060.27760.2804
    下载: 导出CSV
  • [1] Zhou K, Wang H, Zhao W X, Zhu Y T, Wang S R, Zhang F Z, et al. S3-Rec: Self-supervised learning for sequential recommendation with mutual information maximization. In: Proceedings of the 29th ACM International Conference on Information & Knowledge Management. New York, USA: ACM, 2020. 1893−1902
    [2] 饶子昀, 张毅, 刘俊涛, 曹万华. 应用知识图谱的推荐方法与系统. 自动化学报, 2021, 47(9): 2061-2077

    Rao Zi-Yun, Zhang Yi, Liu Jun-Tao, Cao Wan-Hua. Recommendation methods and systems using knowledge graph. Acta Automatica Sinica, 2021, 47(9): 2061-2077
    [3] Li X C, Liang J, Liu X L, Zhang Y. Adversarial filtering modeling on long-term user behavior sequences for click-through rate prediction. In: Proceedings of the 45th International ACM SIGIR Conference on Research and Development in Information Retrieval. Madrid, Spain: ACM, 2022. 1969−1973
    [4] 汤文兵, 任正云, 韩芳. 基于注意力机制的协同卷积动态推荐网络. 自动化学报, 2021, 47(10): 2438-2448

    Tang Wen-Bing, Ren Zheng-Yun, Han Fang. Attention-based collaborative convolutional dynamic network for recommendation. Acta Automatica Sinica, 2021, 47(10): 2438-2448
    [5] 郭磊, 李秋菊, 刘方爱, 王新华. 基于自注意力网络的共享账户跨域序列推荐. 计算机研究与发展, 2021, 58(11): 2524-2537

    Guo Lei, Li Qiu-Ju, Liu Fang-Ai, Wang Xin-Hua. Shared-account cross-domain sequential recommendation with self-attention network. Journal of Computer Research and Development, 2021, 58(11): 2524-2537
    [6] Rao X, Chen L S, Liu Y, Shang S, Yao B, Han P. Graph-flashback network for next location recommendation. In: Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining. Washington, USA: ACM, 2022. 1463−1471
    [7] 孟祥武, 梁弼, 杜雨露, 张玉洁. 基于位置的移动推荐系统效用评价研究. 计算机学报, 2019, 42(12): 2695-2721

    Meng Xiang-Wu, Liang Bi, Du Yu-Lu, Zhang Yu-Jie. A survey of evaluation for location-based mobile recommender systems. Chinese Journal of Computers, 2019, 42(12): 2695-2721
    [8] Hu K X, Li L, Liu J Q, Sun D. DuroNet: A dual-robust enhanced spatial-temporal learning network for urban crime prediction. ACM Transactions on Internet Technology, 2021, 21(1): Article No. 24
    [9] Wang J L, Ding K Z, Hong L J, Liu H, Caverlee J. Next-item recommendation with sequential hypergraphs. In: Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval. New York, USA: ACM, 2020. 1101−1110
    [10] 陈聪, 张伟, 王骏. 带有时间预测辅助任务的会话式序列推荐. 计算机学报, 2021, 44(9): 1841-1853

    Chen Cong, Zhang Wei, Wang Jun. Session-based sequential recommendation with auxiliary time prediction. Chinese Journal of Computers, 2021, 44(9): 1841-1853
    [11] Wang W J, Feng F L, He X N, Nie L Q, Chua T S. Denoising implicit feedback for recommendation. In: Proceedings of the 14th ACM International Conference on Web Search and Data Mining. New York, USA: ACM, 2021. 373−381
    [12] Neupane K P, Zheng E, Yu Q. MetaEDL: Meta evidential learning for uncertainty-aware cold-start recommendations. In: Proceedings of the IEEE International Conference on Data Mining (ICDM). Auckland, New Zealand: IEEE, 2021. 1258−1263
    [13] Fan Z W, Liu Z W, Wang S, Zheng L, Yu P S. Modeling sequences as distributions with uncertainty for sequential recommendation. In: Proceedings of the 30th ACM International Conference on Information and Knowledge Management. Queensland, Australia: ACM, 2021. 3019−3023
    [14] Fan Z W, Liu Z W, Wang Y, Wang A, Nazari Z, Zheng L, et al. Sequential recommendation via stochastic self-attention. In: Proceedings of the ACM Web Conference. Lyon, France: ACM, 2022. 2036−2047
    [15] Jiang J Y, Yang D Q, Xiao Y H, Shen C L. Convolutional Gaussian embeddings for personalized recommendation with uncertainty. In: Proceedings of the 28th International Joint Conference on Artificial Intelligence. Macao, China: AAAI Press, 2019. 2642−2648
    [16] Zhou X L, Liu H, Pourpanah F, Zeng T Y, Wang X Z. A survey on epistemic (model) uncertainty in supervised learning: Recent advances and applications. Neurocomputing, 2022, 489: 449-465 doi: 10.1016/j.neucom.2021.10.119
    [17] Sun F, Liu J, Wu J, Pei C H, Lin X, Ou W W, et al. BERT4Rec: Sequential recommendation with bidirectional encoder representations from transformer. In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management. Beijing, China: ACM, 2019. 1441−1450
    [18] Ovadia Y, Fertig E, Ren J, Nado Z, Sculley D, Nowozin S, et al. Can you trust your model's uncertainty? Evaluating predictive uncertainty under dataset shift. In: Proceedings of the 33rd International Conference on Neural Information Processing Systems. Vancouver, Canada: Curran Associates Inc., 2019. Article No. 1254
    [19] Fort S, Hu H Y, Lakshminarayanan B. Deep ensembles: A loss landscape perspective. arXiv preprint arXiv: 1912.02757, 2019.
    [20] Lakshminarayanan B, Pritzel A, Blundell C. Simple and scalable predictive uncertainty estimation using deep ensembles. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc., 2017. 6405−6416
    [21] Renda A, Barsacchi M, Bechini A, Marcelloni F. Comparing ensemble strategies for deep learning: An application to facial expression recognition. Expert Systems With Applications, 2019, 136: 1-11 doi: 10.1016/j.eswa.2019.06.025
    [22] Deng D D, Wu L, Shi B E. Iterative distillation for better uncertainty estimates in multitask emotion recognition. In: Proceedings of the IEEE/CVF International Conference on Computer Vision Workshops (ICCVW). Montreal, Canada: IEEE, 2021. 3550−3559
    [23] Reich S, Mueller D, Andrews N. Ensemble distillation for structured prediction: Calibrated, accurate, fast —— Choose three. In: Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP). Stroudsburg, USA: ACL, 2020. 5583−5595
    [24] Jiao X Q, Yin Y C, Shang L F, Jiang X, Chen X, Li L L, et al. TinyBERT: Distilling BERT for natural language understanding. In: Proceedings of the Findings of the Association for Computational Linguistics: EMNLP 2020. Stroudsburg, USA: ACL, 2020. 4163−4174
    [25] Zhu J M, Liu J Y, Li W Q, Lai J C, He X Q, Chen L, et al. Ensembled CTR prediction via knowledge distillation. In: Proceedings of the 29th ACM International Conference on Information and Knowledge Management. New York, USA: ACM, 2020. 2941−2958
    [26] Kang S K, Hwang J, Kweon W, Yu H. DE-RRD: A knowledge distillation framework for recommender system. In: Proceedings of the 29th ACM International Conference on Information and Knowledge Management. New York, USA: ACM, 2020. 605−614
    [27] Hinton G, Vinyals O, Dean J. Distilling the knowledge in a neural network. In: Proceedings of the 28th Conference on Neural Information Processing Systems. Montreal, Canada: Curran Associates, 2014. 1−9
    [28] Shen Z Q, Liu Z C, Xu D J, Chen Z T, Cheng K T, Savvides M. Is label smoothing truly incompatible with knowledge distillation: An empirical study. In: Proceedings of the 9th International Conference on Learning Representations. Virtual Event, Austria: ICLR, 2020. 1−17
    [29] Furlanello T, Lipton Z C, Tschannen M, Itti L, Anandkumar A. Born-again neural networks. In: Proceedings of the 35th International Conference on Machine Learning. Stockholm, Sweden: PMLR, 2018. 1602−1611
    [30] Romero A, Ballas N, Kahou S E, Chassang A, Gatta C, Bengio Y. FitNets: Hints for thin deep nets. In: Proceedings of the 3rd International Conference on Learning Representations. San Diego, USA: ICLR, 2015. 1−13
    [31] Lin T Y, Goyal P, Girshick R, He K M, Dollár P. Focal loss for dense object detection. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). Venice, Italy: IEEE, 2017. 2999−3007
    [32] Bengio Y, Louradour J, Collobert R, Weston J. Curriculum learning. In: Proceedings of the 26th Annual International Conference on Machine Learning. Montreal, Canada: ACM, 2009. 41−48
    [33] Ericsson K A. Deliberate practice and acquisition of expert performance: A general overview. Academic Emergency Medicine, 2008, 15(11): 988-994 doi: 10.1111/j.1553-2712.2008.00227.x
    [34] Song W P, Shi C C, Xiao Z P, Duan Z J, Xu Y W, Zhang M, et al. Autoint: Automatic feature interaction learning via self-attentive neural networks. In: Proceedings of the 28th ACM International Conference on Information and Knowledge Management. Beijing, China: ACM, 2019. 1161−1170
    [35] Qin Y Q, Wang P F, Li C L. The world is binary: Contrastive learning for denoising next basket recommendation. In: Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval. New York, USA: ACM, 2021. 859−868
    [36] 黄震华, 杨顺志, 林威, 倪娟, 孙圣力, 陈运文, 等. 知识蒸馏研究综述. 计算计学报, 2022, 45(3): 624-653

    Hung Zhen-Hua, Yang Shun-Zhi, Lin Wei, Ni Juan, Sun Sheng-Li, Chen Yun-Wen, et al. Knowledge distillation: A survey. Chinese Journal of Computers, 2022, 45(3): 624-653
    [37] 潘瑞东, 孔维健, 齐洁. 基于预训练模型与知识蒸馏的法律判决预测算法. 控制与决策, 2022, 37(1): 67-76

    Pan Rui-Dong, Kong Wei-Jian, Qi Jie. Legal judgment prediction based on pre-training model and knowledge distillation. Control and Decision, 2022, 37(1): 67-76
    [38] Zhao B R, Cui Q, Song R J, Qiu Y Y, Liang J J. Decoupled knowledge distillation. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). New Orleans, USA: IEEE, 2022. 11943−11952
    [39] Zhang Y, Xiang T, Hospedales T M, Lu H C. Deep mutual learning. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. Salt Lake City, USA: IEEE, 2018. 4320−4328
    [40] Zhao H J, Yang G, Wang D, Lu H C. Deep mutual learning for visual object tracking. Pattern Recognition, 2021, 112: Article No. 107796 doi: 10.1016/j.patcog.2020.107796
    [41] Chen D F, Mei J P, Wang C, Feng Y, Chen C. Online knowledge distillation with diverse peers. In: Proceedings of the 34th AAAI Conference on Artificial Intelligence. New York, USA: AAAI, 2020. 3430−3437
    [42] Vaswani A, Shazeer N, Parmar N, Uszkoreit J, Jones L, Gomez A N, et al. Attention is all you need. In: Proceedings of the 31st International Conference on Neural Information Processing Systems. Long Beach, USA: Curran Associates Inc., 2017. 6000−6010
    [43] Toneva M, Sordoni A, des Combes R T, Trischler A, Bengio Y, Gordon G J. An empirical study of example forgetting during deep neural network learning. In: Proceedings of the 7th International Conference on Learning Representations. New Orleans, USA: ICLR, 2019. 1−19
    [44] Glorot X, Bengio Y. Understanding the difficulty of training deep feedforward neural networks. In: Proceedings of the 13th International Conference on Artificial Intelligence and Statistics. Sardinia, Italy: JMLR, 2010. 249−256
    [45] He K M, Zhang X Y, Ren S Q, Sun J. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. In: Proceedings of the IEEE International Conference on Computer Vision (ICCV). Santiago, Chile: IEEE, 2015. 1026−1034
  • 期刊类型引用(1)

    1. 闫小如. 基于多标签集成学习的螺旋CT机故障诊断研究. 计算机测量与控制. 2024(11): 48-55 . 百度学术

    其他类型引用(3)

  • 加载中
  • 图(7) / 表(4)
    计量
    • 文章访问数:  759
    • HTML全文浏览量:  625
    • PDF下载量:  174
    • 被引次数: 4
    出版历程
    • 收稿日期:  2022-04-28
    • 录用日期:  2022-09-13
    • 网络出版日期:  2022-10-26
    • 刊出日期:  2023-07-20

    目录

    /

    返回文章
    返回