Centrifugal Pump Fault Identification Technology Based on CEEMDAN-PCA-AC-CNN Model
-
摘要: 为确保离心泵的长期健康稳定运行，对其进行在线监测与故障识别显得尤为重要。本文提出了一种基于自适应噪声的集合经验模态分解（CEEMDAN）-主成分分析（PCA）-自相关（AC）-卷积神经网络（CNN）的设备故障识别模型。首先将采集到的振动信号进行CEEMDAN，对得到的内涵模态函数（IMF）分量进行判别，剔除噪声分量，重构第一轮去噪信号。再通过PCA对一轮去噪的信号进行二次降噪处理。然后将经历2次降噪处理后的信号进行AC处理，送入CNN作为输入数据，对模型进行训练。通过对某离心泵故障进行实验验证，结果表明：本文提出的方法相较于传统双层降噪结合CNN的算法、CEEMD-小波降噪-AC-CNN等算法具有更好的抗干扰性能与更快的模型收敛速度，具有更高的识别准确率与更好的鲁棒性，在同等量级下，识别准确率高达97.9%。
-
关键词:
- 自适应噪声的集合经验模态分解（CEEMDAN） /
- 主成分分析（PCA） /
- 信号降噪 /
- 卷积神经网络（CNN） /
- 故障识别
Abstract: In order to ensure the long-term healthy and stable operation of centrifugal pump, on-line monitoring and fault identification is particularly important. Therefore, this paper presents a fault-identification model based on adaptive noise, with the methods of Complete Ensemble Empirical Mode Decomposition with Adaptive Noise (CEEMDAN), Principal Component Analysis (PCA) and Convolution Neural Network (CNN) combined with Autocorrelation (AC). Firstly, vibration signals are collected and decomposed by CEEMDAN. Then, we discriminate the obtained Intrinsic Mode Functions (IMF) fraction and eliminate the noise fraction to reconstruct the first round of denoised signals. After that, PCA is adopted to remove noise in the denoised signals, and the signals which have been filtered twice are then processed by AC and input to CNN to train the model. Through the experimental verification of a centrifugal pump fault, the results show that compared to traditional methods like double-layer noise reduction with CNN and CEEMD-wavelet denoising-AC-CNN, the model presented in this paper is more resistant to interference and has faster convergence speed. Also, it has the advantages of higher precision and better robustness. At the same order of magnitude, its precision can reach 97.9%. -
图 4 经AC处理后3种分解算法降噪后的函数的时频域图像
Figure 4. Time frequency Domain Images of Autocorrelation Functions after Denoising Using Three Decomposition Algorithms
图 6 基于CEEMDAN-PCA-AC-CNN模型的离心泵故障识别算法流程图
Figure 6. Flow Chart of Centrifugal Pump Fault Identification Algorithm Based on CEEMDAN-PCA-AC-CNN
图 7 5种状态下的信号时域和频域图像
Figure 7. Time Domain and Frequency Domain of Signals under Five Conditions
表 1 数据集规模分布
Table 1. Data Set Size Distribution
故障特征 流量 训练集 测试集 标签 轻微磨损 额定 640 200 状态0 中度磨损 额定 640 200 状态1 轴套碎裂 额定 160 50 状态2 轴套碎裂 1.1倍额定 160 50 状态3 正常运行 额定 1280 400 状态4 
下载: 导出CSV
表 2 几类模型在测试集中的识别准确率
Table 2. Precision of Different Models in the Test Set
模型 准确率/% CEEMDAN-PCA-AC-CNN 97.9 CEEMDAN-小波强制降噪-AC-CNN 92.9 CEEMDAN-小波硬阈值降噪-AC-CNN 95.1 CEEMDAN-AC-CNN 96.5 CEEMDAN-PCA-CNN 85.8 EEMD-PCA-CNN 77.4
下载: 导出CSV
-
[1] 吴秀星,关佳亮. 旋转机械故障诊断与状态监测发展综述[J]. 设备管理与维修,2018(10): 170-173. [2] 赵磊,张永祥,朱丹宸. 复杂装备滚动轴承的故障诊断与预测方法研究综述[J]. 中国测试,2020, 46(3): 17-25. doi: 10.11857/j.issn.1674-5124.2019040083 [3] 傅露莹,齐慧君,李同春,等. 基于CEEMDAN和相关性分析的大坝位移预测[J]. 三峡大学学报: 自然科学版,2024, 46(1): 1-6. [4] 贾文哲,闻麒,李春,等. 改进CEEMDAN算法与CNN融合的深度学习系泊故障研究[J]. 热能动力工程,2023, 38(11): 141-150. [5] 张健,尤文斌,丁永红,等. 基于CEEMDAN-PCA的冲击波信号降噪研究[J]. 弹箭与制导学报,2022, 42(4): 24-28. [6] 刘廷宇,江泉,李晔,等. 旋转机械故障的信号处理与诊断方法综述[J]. 能源与环保,2018, 40(1): 163-166. [7] 李炳锐,李荣昌. 基于PCA-PSO-SVM的造纸污水处理过程故障诊断方法[J]. 自动化应用,2023, 64(21): 244-247. [8] 孙英淳,唐斌,蔡先阳. 基于倒谱-对称点图谱-卷积神经网络的内燃机增压器滚动轴承故障诊断[J]. 内燃机工程,2023, 44(6): 69-76. [9] 蒲小平,陈克兴,徐金梧,等. 基于神经网络的轴承故障诊断方法[J]. 北京科技大学学报,1993, 15(6): 610-613. [10] 吴昊臻,许燕,周建平,等. 基于卷积神经网络的抽油机故障诊断[J]. 燕山大学学报,2024, 48(1): 30-38. doi: 10.3969/j.issn.1007-791X.2024.01.004 -
计量
- 文章访问数: 16
- HTML全文浏览量: 7
- PDF下载量: 2
- 被引次数: 0
