构建神经网络分类器，使其能够依据热工参数的变化规律，精准识别包括复合故障在内的各类故障；同时，针对实际应用中存在的数据不平衡问题，提出切实可行的解决办法。首先基于双向长短期记忆网络（BiLSTM）构建故障诊断分类器，经仿真数据训练验证其基础性能达标；针对实际应用中故障样本较少的问题，补充构建条件生成对抗网络（CGAN）生成与原始数据规律一致的相似样本，提升分类器对少数类故障辨别能力；最终通过中船动力M320DM-PFI双燃料机实际试验数据验证性能。仿真数据验证显示，BiLSTM 分类器诊断准确率为100%；CGAN可保留原始特征生成新样本，生成数据的JS散度均在0.3以下，与原始数据分布近似；数据增强后，分类器的召回率从80%提升至100%，诊断性能显著优化。CGAN数据增强方法与BiLSTM分类器结合可有效提高少数类的诊断准确性，为实际的热工故障诊断提供参考。
To construct a neural network classifier capable of accurately identifying various faults, including compound faults, based on the variation patterns of thermal parameters. Meanwhile, a practical solution is proposed to address the data imbalance issue encountered in real-world applications. First, a fault diagnosis classifier was built based on the Bidirectional Long Short-Term Memory (BiLSTM) network, and its basic performance was verified to meet standards through training with simulation data. To tackle the problem of insufficient fault samples in practical applications, a Conditional Generative Adversarial Network (CGAN) was additionally constructed to generate similar samples that conform to the patterns of the original data, thereby enhancing the classifier’s ability to identify minority-class faults. Finally, the performance was validated using actual test data from the CSSC Power M320DM-PFI dual-fuel engine. Simulation data verification showed that the diagnostic accuracy of the BiLSTM classifier reached 100%. The CGAN could preserve original features while generating new samples, with the Jensen-Shannon (JS) divergence of the generated data all below 0.3, indicating a distribution similar to that of the original data. After data augmentation, the recall rate of the classifier increased from 80% to 100%, achieving a significant optimization in diagnostic performance. The combination of the CGAN data augmentation method and the BiLSTM classifier can effectively improve the diagnostic accuracy of minority-class faults, providing a reference for practical thermal fault diagnosis.
2026,48(7): 77-83 收稿日期：2025-7-31
DOI：10.3404/j.issn.1672-7649.2026.07.013
分类号：U664.12
作者简介：吴唐奕(2001-),男,硕士研究生,研究方向为内燃机仿真与故障诊断
参考文献：
[1] 刘长铖. 基于Modelica语言的柴油机建模仿真研究 [D]. 哈尔滨: 哈尔滨工程大学, 2017.
[2] 李方玉, 汪翔, 陆轶, 等. 大型动力装置的故障模拟及分析[J]. 机电工程技术, 2022, 51(2): 102-105 LI F Y, WANG X, LU Y, et al. Fault simulation and analysis of large-scale power equipment[J]. Mechanical & Electrical Engineering Technology, 2022, 51(2): 102-105
[3] 张海涛. 船舶双燃料发动机故障诊断系统优化研究[J]. 舰船科学技术, 2020, 42(4): 76-78 ZHANG H T. Research on the optimization of fault diagnosis system for marine dual fuel engine[J]. Ship Science and Technology, 2020, 42(4): 76-78
[4] 郭晓成. 甲醇/柴油双燃料发动机健康评估与故障诊断方法研究[D]. 镇江: 江苏科技大学, 2024.
[5] 张振京, 曹石, 窦全礼, 等. 深度学习与异常信息融合的内燃机燃油系统智能诊断方法[J]. 北京化工大学学报(自然科学版), 2025, 52(2): 76-87. ZHANG Z J, CAO S, DOU Q L, et al. A deep learning and anomaly information fusion intelligent diagnosis method for the fuel system in internal combustion engine[J]. Journal of Beijing University of Chemical Technology(Natural Science), 2025, 52(2): 76-87.
[6] 毛荣珍, 米洁, 甄真, 等. 基于数据增强的多输出分类旋转机械复合故障诊断[J]. 北京信息科技大学学报(自然科学版), 2025, 40(1): 94-101 MAO R Z, MI J, ZHEN Z, et al. Multi-output classification fault diagnosis of rotating machinery based on data augmentation[J]. Journal of Beijing Information Science and Technology University (Natural Science Edition), 2025, 40(1): 94-101
[7] 杨松翰. 基于多尺度孪生网络的旋转机械小样本故障诊断方法研究[D]. 广州: 华南理工大学, 2024.
[8] 李巍华, 蓝昊, 陈祝云, 等. 非完备数据驱动的装备复合故障智能解耦方法[J]. 机械工程学报, 2024, 60(24): 45–55. LI W H, LAN H, CHEN Z Y, et al. Incomplete data-driven intelligent decoupling method for equipment composite faults[J]. Journal of Mechanical Engineering, 2024, 60(24): 45–55.
[9] 田睿, 刘维可, 伍珣. 基于CGAN与LSTM的逆变器多类型故障诊断方法研究[J]. 湖南电力, 2023, 43(6): 151-158 TIAN R, LIU W K, WU X. Research on multi-type fault diagnosis method of inverter based on CGAN and LSTM[J]. Hunan Electric Power, 2023, 43(6): 151-158
[10] 张伟业, 缪维跑, 苏欢欢, 等. 基于不平衡样本数据的轴承故障诊断模型与方法研究[J]. 动力工程学报, 2025, 45(6): 904–912. ZHANG W Y, MIAO W P, SU H H, et al. Research on bearing fault diagnosis model and method based on imbalanced sample data[J]. Journal of Chinese Society of Power Engineering, 2025, 45(6): 904–912.
[11] 周建民, 夏晓枫, 李家辉. 改进生成对抗网络的不平衡数据下轴承故障诊断[J]. 控制工程, 2025, 32(6): 1039-1048 ZHOU J M, XIA X F, LI J H. Bearing fault diagnosis under imbalanced data using improved generative adversarial network[J]. Control Engineering of China, 2025, 32(6): 1039-1048
[12] CHAWLA N V, BOWYER K W, Hall L O, et al. SMOTE: Synthetic minority over-sampling technique[J]. Journal of Artificial Intelligence Research, 2002, 16(1): 321-357.
[13] 王琦, 邓林峰, 赵荣珍. 基于改进一维卷积神经网络的滚动轴承故障识别[J]. 振动与冲击, 2022, 41(3): 216-223 WANG Q, DENG L F, ZHAO R Z. Fault recognition of rolling bearings based on improved one-dimensional convolutional neural network[J]. Journal of Vibration and Shock, 2022, 41(3): 216-223
[14] 何汇汇. 基于生成对抗网络的滚动轴承故障诊断方法研究[D]. 秦皇岛: 燕山大学, 2024.
