Page 96 - 《振动工程学报》2025年第11期
P. 96

2554                               振     动     工     程     学     报                     第 38 卷

                  Processing,2020,140:106725.                       face[EB/OL].  ( 2023-11-03) [2025-10-05].  https://arxiv.org/
              [4]  王冬,侯炳昌,王玉婷,等. 稀疏测度和复杂性测度及其                       abs/2303.17580.
                  在设备健康监测中的研究进展          [J]. 机械工程学报,2025,      [18]  HUANG Y C,YANG L F. Winning gold at IMO 2025 with a
                  61(1):123-139.                                    model-agnostic  verification-and-refinement  pipeline[EB/OL].
                  WANG  Dong, HOU  Bingchang, WANG  Yuting, et  al.  (2025-07-25)[2025-10-05]. http://arxiv.org/abs/2507.15855.
                  Advances  in  sparsity  measures  and  complexity  measures  for  [19]  WANG  L, MA  C, FENG  X  Y, et  al.  A  survey  on  large
                  machine  health  monitoring[J].  Journal  of  Mechanical  Engi-  language  model  based  autonomous  agents[J].  Frontiers  of
                  neering,2025,61(1):123-139.                       Computer Science,2024,18(6):186345.
              [5]  YAN R Q,SHANG Z G,XU H,et al. Wavelet transform  [20]  陈超,沈飞,严如强. 改进  LSSVM  迁移学习方法的轴承故
                  for  rotary  machine  fault  diagnosis: 10  years  revisited[J].  障诊断  [J]. 仪器仪表学报,2017,38(1):33-40.
                  Mechanical  Systems  and  Signal  Processing, 2023, 200:  CHEN  Chao, SHEN  Fei, YAN  Ruqiang.  Enhanced  least
                  110545.                                           squares  support  vector  machine-based  transfer  learning  strat-
              [6]  CHEN Y Y,ZHANG D,YAN R Q,et al. Applications of  egy  for  bearing  fault  diagnosis[J].  Chinese  Journal  of  Scien-
                  domain generalization to machine fault diagnosis:a survey[J].  tific Instrument,2017,38(1):33-40.
                  IEEE/CAA  Journal  of  Automatica  Sinica, 2025, 12( 10) :  [21]  QIN L C,ZHANG L F,FENG J G,et al. A hybrid tribo-
                  1963-1984.                                        electric-piezoelectric smart squirrel cage with self-sensing and
              [7]  严如强,商佐港,王志颖,等. 可解释人工智能在工业智                       self-powering  capabilities[J].  Nano  Energy, 2024, 124:
                  能诊断中的挑战和机遇:先验赋能             [J]. 机械工程学报,          109506.
                  2024,60(12):1-20.                             [22]  HUANG H,BADDOUR N. Bearing vibration data collected
                  YAN Ruqiang,SHANG Zuogang,WANG Zhiying,et al.     under  time-varying  rotational  speed  conditions[J].  Data  in
                  Challenges  and  opportunities  of  XAI  in  industrial  intelligent  Brief,2018,21:1745-1749.
                  diagnosis:priori-empowered[J]. Journal of Mechanical Engi-
                  neering,2024,60(12):1-20.
              [8]  林京,焦金阳. 可解释机械智能诊断技术的研究进展与挑                   第一作者:李 奇      (1997―),男,博士研究生。
                  战  [J]. 机械工程学报,2023,59(20):215-224.                   E-mail:liq22@tsinghua.org.cn
                  LIN Jing,JIAO Jinyang. Research progress and challenges of  通信作者:秦朝烨  (1980―),男,博士,副教授。
                  interpretable  mechanical  intelligent  diagnosis[J].  Journal  of  E-mail:qinzy@tsinghua.edu.cn

                  Mechanical Engineering,2023,59(20):215-224.
              [9]  LI  Q, LI  H, HU  W  Y, et  al.  Transparent  operator
                                                                附  录  A
                  network: a  fully  interpretable  network  incorporating  learn-
                  able  wavelet  operator  for  intelligent  fault  diagnosis[J].  IEEE

                  Transactions on Industrial Informatics,2024,20(6):8628-
                  8638.                                         A1    三个案例的反思内容
              [10]  LI Q,LIU Y K,SUN S L,et al. Deep expert network:a
                  unified method toward knowledge-informed fault diagnosis via  A1.1    案例一中审查者的反思迭代过程
                  fully interpretable neuro-symbolic AI[J]. Journal of Manufac-  (1)反思迭代  1
                  turing Systems,2024,77:652-661.                   当前       深度:2。智能体决策:需要重新规划
              [11]  LI  Q, QIN  L  C, XU  H  F, et  al.  Transparent  information  DAG
                                                                (推断)。决策理由: 当前           DAG  深度为    2,低于最低
                  fusion network:an explainable network for multi-source bear-
                  ing  fault  diagnosis  via  self-organized  neural-symbolic  要求  4。尽管已执行的步骤有效,但工作流过浅,需
                  nodes[J].  Advanced  Engineering  Informatics, 2025, 65:  要扩展以提取更多故障诊断相关特征。应当引入更
                  103156.
              [12]  LIAO X M,CHEN C,WANG Z W,et al. Large language  高级的信号处理技术,例如,可对已计算的                 RMS  值
                  model assisted fine-grained knowledge graph construction for  本身进行进一步的统计分析,或对原始信号(ch1, ch2)
                  robotic  fault  diagnosis[J].  Advanced  Engineering  Informat-  应用小波分解进行时频分析。此外,在  FFT  计算后,
                  ics,2025,65:103134.                           应执行如谱熵、谱峭度或主频识别等特征提取。
              [13]  SHI F C,CHEN L,ZHOU M S,et al. A stepwise intelli-
                                                                    (2)反思迭代
                  gence generative method for structured maintenance guidance   2
                  documents  based  on  knowledge  graph  augmented  LLM[J].  当前  DAG  深度:3。智能体决策: 需要重新规划
                  Advanced Engineering Informatics,2025,67:103523.
                                                                (推断)。决策理由: DAG          结构有效,但深度和特征
              [14]  LIN  L, ZHANG  S  H, FU  S, et  al.  FD-LLM: large
                                                                多样性有限。尽管时域特征(RMS、熵)和频域特征
                  language model for fault diagnosis of complex equipment[J].
                  Advanced Engineering Informatics,2025,65:103208.  (谱峭度)很有用,但整体工作流过浅,无法捕捉轴承
              [15]  BRITO L C,SUSTO G A,BRITO J N,et al. Fault diagno-  故障的复杂性。双通道的对称性处理很好,但需要
                  sis using explainable AI:a transfer learning-based approach
                                                                从  FFT  和希尔伯特变换结果中提取更多特征,例如
                  for rotating machinery exploiting augmented synthetic data[J].
                                                                从  FFT  输出中提取统计特征(均值、标准差等)、计
                  Expert Systems with Applications,2023,232:120860.
              [16]  ALUR R,DURRETT G,KRESS-GAZIT H,et al. Report  算不同频带能量或探索谱质心等。同时,应引入小
                  on NSF workshop on science of safe AI[EB/OL]. (2025-06-  波变换以分解信号。
                  24)[2025-10-05]. https://arxiv.org/abs/2506.22492.
              [17]  SHEN  Y  L, SONG  K  T, TAN  X, et  al.  HuggingGPT:  (3)反思迭代  3
                  solving  AI  tasks  with  ChatGPT  and  its  friends  in  hugging  当前  DAG  深度:3。智能体决策:需要重新规划
   91   92   93   94   95   96   97   98   99   100   101