基于注意力机制改进的PSO-BiLSTM刀具磨损预测

杨沛东; 黄华; 尉卫卫; 郭宝岛

doi:10.13700/j.bh.1001-5965.2023.0545

基于注意力机制改进的PSO-BiLSTM刀具磨损预测

doi: 10.13700/j.bh.1001-5965.2023.0545

兰州理工大学机电工程学院，兰州 730050

基金项目:

国家自然科学基金(52365057,51965037)

详细信息

通讯作者:
E-mail：hh318872@126.com

中图分类号: TH164；V262.3+3
计量
- 文章访问数: 260
- HTML全文浏览量: 56
- PDF下载量: 13
- 被引次数: 0
出版历程
- 收稿日期: 2023-08-25
- 录用日期: 2023-11-13
- 网络出版日期: 2023-12-13
- 整期出版日期: 2025-10-31

Tool wear prediction based on attention mechanism and PSO-BiLSTM

School of Mechanical and Electrical Engineering，Lanzhou University of Technology，Lanzhou 730050，China

Funds:

National Natural Science Foundation of China (52365057,51965037)

More Information

Corresponding author: E-mail：hh318872@126.com

摘要

摘要:
针对刀具磨损故障诊断中存在的监测数据单一和特征信号处理效果差的问题，提出了一种基于注意力机制(AM)改进的粒子群算法(PSO)优化双向长短时记忆(BiLSTM)神经网络来实现端到端的刀具磨损预测方法。根据传感器信号进行多域特征提取，构建优质的信号输入样本；利用卡尔曼滤波对输入样本进行多传感器数据融合，得到鲁棒性更高的融合数据样本，在此基础上，通过PSO对BiLSTM网络进行超参数寻优，根据优化的超参数建立神经网络模型；基于注意力机制赋予输入影响权重，改进PSO-BiLSTM以获得更好的刀具磨损预测效果。对比实验结果验证了所提模型在刀具磨损预测中的可行性，其精度相比传统深度学习方法有较大的提升。
- 刀具磨损 /
- 卡尔曼滤波 /
- 粒子群算法 /
- 注意力机制 /
- 双向长短时记忆神经网络
Abstract:
A bi-directional long short-term memory (BiLSTM) neural network optimization end-to-end tool wear prediction method based on attention mechanism (AM) and particle swarm optimization (PSO) algorithm is proposed to address the issues of single monitoring data and poor feature signal processing in tool wear fault diagnosis. Firstly, based on the sensor data, to construct high-quality signal input samples, extract the multi-domain feature. With multi-sensor data to obtain the fused data samples with higher robustness, the Kalman filter is used to fuse the input samples. On this basis, the hyperparameters of the BiLSTM are optimized by PSO, and the neural network model is built according to the optimized hyperparameters. Finally, based on the attention mechanism to give weights to the input influences, the PSO-BiLSTM is improved to obtain a better tool wear prediction. The suggested model's validity and feasibility in tool wear prediction are confirmed by comparative experimental findings, and its accuracy is significantly higher than that of classical deep learning. This approach offers a fresh viewpoint on tool wear prediction.
- tool wear /
- Kalman filter /
- particle swarm optimization /
- attention mechanism /
- bi-diretional long short-term memory neural network

HTML全文

图 1 模型框架

Figure 1. Model framework

下载: 全尺寸图片幻灯片

图 2 LSTM单元结构

Figure 2. LSTM cell structure

下载: 全尺寸图片幻灯片

图 3 BiLSTM单元结构

Figure 3. BiLSTM cell structure

下载: 全尺寸图片幻灯片

图 4 优化流程

Figure 4. Optimize the process

下载: 全尺寸图片幻灯片

图 5 注意力机制的核心思想

Figure 5. The core idea of attention mechanism

下载: 全尺寸图片幻灯片

图 6 刀具磨损系统示意

Figure 6. Schematic diagram of the tool wear system

下载: 全尺寸图片幻灯片

图 7 未优化的LSTM在C1、C4和C6上的预测效果

Figure 7. Predictive effects of unoptimized LSTM on C1, C4 and C6

下载: 全尺寸图片幻灯片

图 8 PSO-AM-BiLSTM在C1刀具上的预测结果

Figure 8. Prediction results of PSO-AM-BiLSTM on C1

下载: 全尺寸图片幻灯片

图 9 PSO-AM-BiLSTM在C4刀具上的预测结果

Figure 9. Prediction results of PSO-AM-BiLSTM on C4

下载: 全尺寸图片幻灯片

图 10 PSO-AM-BiLSTM在C6刀具上的预测结果

Figure 10. Prediction results of PSO-AM-BiLSTM on C6

下载: 全尺寸图片幻灯片

图 11 算法迭代过程

Figure 11. Process of optimization algorithms

下载: 全尺寸图片幻灯片

表 1 PSO-BiLSTM模型参数

Table 1. Parameters of the PSO-BiLSTM model

网络参数	初始值
初始化权重	[−0.5,0.5]
遗忘门偏置量	1
输入门偏置量	[0,0.8]
输出门偏置量	[0,0.8]
迭代次数	50
种群个数	5
学习因子c₁	2
学习因子c₂	2
隐藏层单元数m	[10,300]
学习率r	[0.01,0.15]

下载: 导出CSV

表 2 实验加工参数

Table 2. Experimental processing parameters

主轴速度/ (r·min⁻¹)	进给速度/ (mm·min⁻¹)	径向切深/ mm	轴向切深/ mm	铣削方式
10 400	1 555	0.125	0.2	顺铣

下载: 导出CSV

表 3 特征指标

Table 3. Characteristic index

序号	特征名称	序号	特征名称
1	峰峰值	11	峰值指标
2	方差	12	脉冲指标
3	均值	13	裕度指标
4	歪度	14	峭度指标
5	峭度	15	均值频率
6	均方值	16	频谱二阶矩
7	方根幅值	17	标准偏差频
8	均方根值	18	峭度频率
9	绝对均值	19	均方根频率
10	波形指标	20	中心频率

下载: 导出CSV

表 4 预测结果评价指标

Table 4. Evaluations of the prediction results

模型	RMSE			MAPE			R²
模型	C1	C4	C6	C1	C4	C6	C1	C4	C6
PSO-AM-BiLSTM	2.906	6.85	5.22	2.563	5.54	4.32	0.994	0.991	0.992
PSO-CNN	6.72	8.66	8.31	5.02	7.97	6.83	0.91	0.907	0.903
PSO-BPNN	11.384	13.027	17.51	10.34	12.76	11.85	0.87	0.81	0.83
PSO-LSTM	16.34	15.27	17.08	12.17	13.03	14.05	0.75	0.85	0.80

下载: 导出CSV

参考文献(17)

[1]	CHEN Y X, JIN Y, JIRI G. Predicting tool wear with multi-sensor data using deep belief networks[J]. The International Journal of Advanced Manufacturing Technology, 2018, 99(5): 1917-1926.
[2]	谢振龙, 岳彩旭, 刘献礼, 等. 基于EMD-SVM的钛合金铣削过程刀具磨损监测[J]. 振动、测试与诊断, 2022, 42(5): 988-996. XIE Z L, YUE C X, LIU X L, et al. Tool wear monitoring based on EMD-SVM in milling process of Ti-alloy[J]. Journal of Vibration, Measurement & Diagnosis, 2022, 42(5): 988-996(in Chinese).
[3]	ZHOU J J, YU J B. Chisel edge wear measurement of high-speed steel twist drills based on machine vision[J]. Computers in Industry, 2021, 128: 103436.
[4]	LI L H, AN Q B. An in-depth study of tool wear monitoring technique based on image segmentation and texture analysis[J]. Measurement, 2016, 79(2): 44-52.
[5]	MORGAN J, O’DONNELL G E. Cyber physical process monitoring systems[J]. Journal of Intelligent Manufacturing, 2018, 29(6): 1317-1328. doi: 10.1007/s10845-015-1180-z
[6]	戴稳, 张超勇, 孟磊磊, 等. 基于深度学习与特征后处理的支持向量机铣刀磨损预测模型[J]. 计算机集成制造系统, 2020, 26(9): 2331-2343. DAI W, ZHANG C Y, MENG L L, et al. Support vector machine milling wear prediction model based on deep learning and feature re-processing[J]. Computer Integrated Manufacturing Systems, 2020, 26(9): 2331-2343(in Chinese).
[7]	郑圆, 胡建中, 贾民平, 等. 一种基于参数优化变分模态分解的滚动轴承故障特征提取方法[J]. 振动与冲击, 2020, 39(21): 195-202. ZHENG Y, HU J Z, JIA M P, et al. A method for rolling bearing fault feature extraction based on parametric optimization VMD[J]. Journal of Vibration and Shock, 2020, 39(21): 195-202(in Chinese).
[8]	李聪波, 万腾, 陈行政, 等. 基于切削功率的数控车削批量加工刀具磨损在线监测[J]. 计算机集成制造系统, 2018, 24(8): 1910-1919. LI C B, WAN T, CHEN X Z, et al. On-line monitoring method of tool wear for NC turning in batch processing based on cutting power[J]. Computer Integrated Manufacturing Systems, 2018, 24(8): 1910-1919(in Chinese).
[9]	何彦, 凌俊杰, 王禹林, 等. 基于长短时记忆卷积神经网络的刀具磨损在线监测模型[J]. 中国机械工程, 2020, 31(16): 1959-1967. HE Y, LING J J, WANG Y L, et al. In-process tool wear monitoring model based on LSTM-CNN[J]. China Mechanical Engineering, 2020, 31(16): 1959-1967(in Chinese).
[10]	曹大理, 孙惠斌, 张纪铎, 等. 基于卷积神经网络的刀具磨损在线监测[J]. 计算机集成制造系统, 2020, 26(1): 74-80. CAO D L, SUN H B, ZHANG J D, et al. In-process tool condition monitoring based on convolution neural network[J]. Computer Integrated Manufacturing Systems, 2020, 26(1): 74-80(in Chinese).
[11]	王丽华, 杨家巍, 张永宏, 等. 基于堆叠降噪自编码的刀具磨损状态识别[J]. 中国机械工程, 2018, 29(17): 2038-2045. doi: 10.3969/j.issn.1004-132X.2018.17.004 WANG L H, YANG J W, ZHANG Y H, et al. Tool wear condition recognition based on SDAE[J]. China Mechanical Engineering, 2018, 29(17): 2038-2045(in Chinese). doi: 10.3969/j.issn.1004-132X.2018.17.004
[12]	LUO J H, ZHANG X. Convolutional neural network based on attention mechanism and Bi-LSTM for bearing remaining life prediction[J]. Applied Intelligence, 2022, 52(1): 1076-1091. doi: 10.1007/s10489-021-02503-2
[13]	ALJAMAL M A, ABDELGHAFFAR H M, RAKHA H A. Developing a neural-Kalman filtering approach for estimating traffic stream density using probe vehicle data[J]. Sensors, 2019, 19(19): 4325. doi: 10.3390/s19194325
[14]	沈涛, 李舜酩. 针对滚动轴承故障的批标准化CNN-LSTM诊断方法[J]. 计算机集成制造系统, 2022, 28(12): 3946-3955. SHEN T, LI S M. CNN-LSTM method with batch normalization for rolling bearing fault diagnosis[J]. Computer Integrated Manufacturing Systems, 2022, 28(12): 3946-3955(in Chinese).
[15]	吴飞, 农皓业, 马晨浩. 基于PSO-LSTM模型的刀具磨损预测方法[J]. 吉林大学学报(工学版), 2023, 53(4): 984-997. WU F, NONG H Y, MA C H. Tool wear predictionmethod based on PSO-LSTM model [J]. Journal of Jilin University(Engineering and Technology Edition), 2023, 53(4): 984-997(in Chinese).
[16]	ZHANG J S, JIANG Y C, WU S M, et al. Prediction of remaining useful life based on bidirectional gated recurrent unit with temporal self-attention mechanism[J]. Reliability Engineering & System Safety, 2022, 221: 108297.
[17]	PHM Society. PHM data challenge 2010[DB/OL]. (2023-07-25)[2010-05-18]. http://www.phmsociety. org/competition/phm/10.