A Sample-imbalance-oriented Method for Crack Damage Identification in Wing Spars

doi:10.3969/j.issn.1004-132X.2026.09.008

Abstract

Abstract:

Aiming at the problem of class sample imbalance in crack detection of wing spars， a crack classification model was proposed based on cross-scale feature fusion with an improved Swin Transformer V2（CSFF-ISTV2）， and a triplet dynamic weight balance loss function was designed. First， the global attention mechanism was introduced into the Swin Transformer V2 module to enhance the global long-range dependence modeling ability. At the same time， the multi-layer perceptron was optimized by combining the gated linear unit and the Swish activation function to improve the accuracy of feature expression. Second， a cross-scale attention fusion structure was constructed to make up for the deficiency of single-scale receptive field feature extraction. Finally， the recognition ability of a few crack samples was enhanced by the triplet dynamic weight balance loss. The experimental results show that the geometric mean criterion value of the proposed model is 86.27% and the accuracy is 98.29% under the condition of significant sample imbalance across multiple categories. Compared with the baseline Swin Transformer V2， the two indicators are improved by 33.5% and 4.44%， respectively.

Key words: wing spar, damage identification, Swin Transformer V2, feature fusion, weight balancing

摘要：

针对机翼大梁裂纹检测中普遍存在的类别样本不平衡问题，提出了一种基于跨尺度特征融合的改进Swin Transformer V2（CSFF-ISTV2）裂纹分类模型，并设计了三元组动态权重平衡损失函数。在Swin Transformer V2模块中引入全局注意力机制，增强全局长距离依赖建模能力，同时结合门控线性单元与Swish激活函数优化多层感知机，提高特征表达精度；构建跨尺度注意力融合结构，弥补单尺度感受野特征提取不足的缺陷；通过三元组动态权重平衡损失强化少数裂纹样本的识别能力。实验结果表明：在多类别显著样本不平衡工况下，所提模型几何平均准则值为86.27%，准确率为98.29%；相较于基线Swin Transformer V2，两项指标分别提高了33.5%和4.44%。

关键词: 机翼大梁, 损伤识别, Swin Transformer V2, 特征融合, 权重平衡

CLC Number:

V267

WANG Lei, FU Haipeng, QIN Zhongrui. A Sample-imbalance-oriented Method for Crack Damage Identification in Wing Spars[J]. China Mechanical Engineering, 2026, 37(9): 2153-2165.

王雷, 付海朋, 秦中瑞. 面向样本不平衡的机翼大梁裂纹损伤识别方法[J]. 中国机械工程, 2026, 37(9): 2153-2165.

Figures/Tables 30

Fig.1 Framework of the CSFF-ISTV2 model

Fig.2 ISTV2 module

Fig.3 SGMLP architecture

Fig.4 CSFF module

Fig.5 Flowchart of CSFF-ISTV2 damage identification algorithm

Fig.6 Layout of wing spar and sensor placement

Fig.7 Data waveform diagram

Fig.8 Schematic of sliding window sampling

Tab.1 Number and proportion of different experimental samples

实验名称	健康∶损伤1∶损伤2∶损伤3∶损伤4
轻度不平衡	1500∶750∶750∶750∶750（2∶1∶1∶1∶1）
中度不平衡	1500∶450∶450∶450∶450（10∶3∶3∶3∶3）
重度不平衡	1500∶150∶150∶150∶150（10∶1∶1∶1∶1）
类间不平衡	1500∶750∶450∶150∶75（20∶10∶6∶2∶1）

Tab.2 Model hyperparameters and runtime environment configuration

条件设置	参数名称	值
GAM超参数	注意力头数	4
	通道维度	128/256/512
	Q/K/V投影	Conv1d（1×1）
	输出投影	Conv1d（1×1）
SGMLP超参数	激活函数	SwiGLU
	扩展比率	4
	输入维度	128/256/512
	偏置设置	False
ISTV2超参数	各尺度维度	128/256/512
	各尺度深度	4/6/4
	各尺度注意力头数	8/16/32
	各尺度窗口大小	16/8/4
TDWB损失函数超参数	三元组边界	1
TDWB损失函数超参数	平衡系数	0.2
训练超参数	优化器	Adam
	学习率	0.0001
	批次大小	256
	轮次	300
运行环境	CPU	i9-14900
	GPU	RTX4060
	深度学习框架	Pytorch2.2.2

Tab.3 Model performance evaluation metrics

性能指标	数学表达式
准确率 $A c c$	$(T P + T N) / (T P + T N + F P + F N)$
召回率 $R e c a l l$	$T P / (T P + F N)$
精确率 $P r e c$	$T P / (T P + F P)$
F1分数 $F 1$	$2 T P / (2 T P + F P + F N)$
几何平均准则 $G m e a n s$	$(T P T N) [(T P + F N) (T N + F P)]$

Tab.3 Model performance evaluation metrics

性能指标	数学表达式
准确率 $A c c$	$(T P + T N) / (T P + T N + F P + F N)$
召回率 $R e c a l l$	$T P / (T P + F N)$
精确率 $P r e c$	$T P / (T P + F P)$
F1分数 $F 1$	$2 T P / (2 T P + F P + F N)$
几何平均准则 $G m e a n s$	$(T P T N) [(T P + F N) (T N + F P)]$

Tab.4 Per-class evaluation metrics on mildly imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	98.76	94.19	96.64	79.89	98.28	99	97.33	96	95.33	76	98.84	95.74	96.32	86.93	85.71
STV2	98	93.51	97.18	84.62	84.47	98	96	92	80.67	90.67	98	94.74	94.52	82.59	87.46
W-DenseNet	99.33	97.37	98.64	87.9	93.06	99.33	98.67	96.67	92	89.33	93.33	98.01	97.64	89.9	91.16
CSCNN	99.32	91.82	95.95	91.33	93.88	98	97.33	94.67	91.33	92	98.66	94.5	95.3	91.33	92.93
MSResNet	98	92.67	94.12	88.61	95.68	98	92.67	96	93.33	88.67	98	92.67	95.05	90.91	92.04
GAN	100	97.4	98.68	91.87	99.28	99	100	99.33	98	91.33	99.5	98.68	99	94.84	95.14
WGAN	100	94.23	97.3	96.6	95.42	98.67	98	96	94.67	97.33	99.33	96.08	96.64	95.62	96.37
CSFF-ISTV2	100	98.68	100	98.64	96.73	99.33	100	100	96.67	98.67	99.67	99.33	100	97.64	97.69

Tab.4 Per-class evaluation metrics on mildly imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	98.76	94.19	96.64	79.89	98.28	99	97.33	96	95.33	76	98.84	95.74	96.32	86.93	85.71
STV2	98	93.51	97.18	84.62	84.47	98	96	92	80.67	90.67	98	94.74	94.52	82.59	87.46
W-DenseNet	99.33	97.37	98.64	87.9	93.06	99.33	98.67	96.67	92	89.33	93.33	98.01	97.64	89.9	91.16
CSCNN	99.32	91.82	95.95	91.33	93.88	98	97.33	94.67	91.33	92	98.66	94.5	95.3	91.33	92.93
MSResNet	98	92.67	94.12	88.61	95.68	98	92.67	96	93.33	88.67	98	92.67	95.05	90.91	92.04
GAN	100	97.4	98.68	91.87	99.28	99	100	99.33	98	91.33	99.5	98.68	99	94.84	95.14
WGAN	100	94.23	97.3	96.6	95.42	98.67	98	96	94.67	97.33	99.33	96.08	96.64	95.62	96.37
CSFF-ISTV2	100	98.68	100	98.64	96.73	99.33	100	100	96.67	98.67	99.67	99.33	100	97.64	97.69

Tab.5 Overall evaluation metrics on mildly imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	93.78 ± 3.12	81.86
STV2	92.56 ± 2.91	79.57
W-DenseNet	95.89 ± 1.68	88.24
CSCNN	95.22 ± 1.17	86.84
MSResNet	94.44 ± 2.61	84.94
GAN	97.78 ± 1.35	93.82
WGAN	97.22 ± 1.58	92.48
CSFF-ISTV2	99±0.16	97.34

Tab.5 Overall evaluation metrics on mildly imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	93.78 ± 3.12	81.86
STV2	92.56 ± 2.91	79.57
W-DenseNet	95.89 ± 1.68	88.24
CSCNN	95.22 ± 1.17	86.84
MSResNet	94.44 ± 2.61	84.94
GAN	97.78 ± 1.35	93.82
WGAN	97.22 ± 1.58	92.48
CSFF-ISTV2	99±0.16	97.34

Fig.9 Confusion matrices of different models on mildly imbalance

Fig.10 The t-SNE visualization of different models on mild imbalance

Tab.6 Per-class evaluation metrics on moderately imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	100	87.06	82.18	88.73	77.88	96.67	82.22	92.22	70	97.78	98.31	84.57	86.91	78.26	86.7
STV2	94.9	94.87	94.38	87.1	73.5	99.33	82.22	93.33	60	95.56	97.07	88.1	93.85	71.05	83.09
W-DenseNet	96.13	93.9	92.55	90.24	89.13	99.33	85.56	96.67	82.22	91.11	97.7	89.53	94.57	86.05	90.11
CSCNN	99	91.58	92.39	89.16	90.11	98.67	96.67	94.44	82.22	91.11	98.83	94.05	93.41	85.55	90.61
MSResNet	100	94.62	96.59	85.71	82.65	99	97.78	94.44	80	90	99.5	96.17	95.51	82.76	86.17
GAN	99.66	95.7	98.82	88.37	86.73	99	98.89	93.33	84.44	94.44	99.33	97.27	96	86.36	90.43
WGAN	100	92.71	95.56	95.18	91.49	99	98.89	95.56	87.78	95.56	99.5	95.7	95.56	91.33	93.48
CSFF-ISTV2	100	100	96.74	97.65	94.68	100	98.89	98.89	92.22	98.89	100	99.44	97.8	94.86	96.74

Tab.6 Per-class evaluation metrics on moderately imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	100	87.06	82.18	88.73	77.88	96.67	82.22	92.22	70	97.78	98.31	84.57	86.91	78.26	86.7
STV2	94.9	94.87	94.38	87.1	73.5	99.33	82.22	93.33	60	95.56	97.07	88.1	93.85	71.05	83.09
W-DenseNet	96.13	93.9	92.55	90.24	89.13	99.33	85.56	96.67	82.22	91.11	97.7	89.53	94.57	86.05	90.11
CSCNN	99	91.58	92.39	89.16	90.11	98.67	96.67	94.44	82.22	91.11	98.83	94.05	93.41	85.55	90.61
MSResNet	100	94.62	96.59	85.71	82.65	99	97.78	94.44	80	90	99.5	96.17	95.51	82.76	86.17
GAN	99.66	95.7	98.82	88.37	86.73	99	98.89	93.33	84.44	94.44	99.33	97.27	96	86.36	90.43
WGAN	100	92.71	95.56	95.18	91.49	99	98.89	95.56	87.78	95.56	99.5	95.7	95.56	91.33	93.48
CSFF-ISTV2	100	100	96.74	97.65	94.68	100	98.89	98.89	92.22	98.89	100	99.44	97.8	94.86	96.74

Tab.7 Overall evaluation metrics on moderately imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	90.06±1.28	70.83
STV2	90.3±2.32	66.11
W-DenseNet	93.64±1.04	78.45
CSCNN	94.55±1.28	82.15
MSResNet	94.39±1.16	81.13
GAN	95.61±2.61	85.36
WGAN	96.52±1.65	88.57
CSFF-ISTV2	98.48±0.58	94.44

Tab.7 Overall evaluation metrics on moderately imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	90.06±1.28	70.83
STV2	90.3±2.32	66.11
W-DenseNet	93.64±1.04	78.45
CSCNN	94.55±1.28	82.15
MSResNet	94.39±1.16	81.13
GAN	95.61±2.61	85.36
WGAN	96.52±1.65	88.57
CSFF-ISTV2	98.48±0.58	94.44

Fig.11 Confusion matrices of different models on moderately imbalance

Fig.12 The t-SNE visualization of different models on moderately imbalance

Tab.8 Per-class evaluation metrics on severely imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	91.98	37.5	86.36	61.36	100	99.33	20	63.33	90	46.67	95.51	26.09	73.08	72.97	63.64
STV2	94.64	84.21	88	61.54	63.64	100	53.33	73.33	53.33	70	97.24	65.31	80	57.14	66.67
W-DenseNet	98.66	83.33	88.89	65.62	78.12	98.33	83.33	80	70	83.33	98.5	83.33	84.21	67.74	80.65
CSCNN	97.34	64.29	80	68.57	84.26	97.67	60	80	80	73.33	97.5	62.07	80	73.85	78.57
MSResNet	96.75	76.92	75.76	67.74	81.82	99.33	66.67	83.33	70	60	98.03	71.43	79.37	68.85	69.23
GAN	98.03	82.76	88.46	71.43	75	99.67	80	76.67	66.67	80	98.84	81.36	82.14	68.97	77.42
WGAN	98.35	82.14	89.29	79.41	92.59	99.33	76.67	83.33	90	83.33	98.84	79.31	86.21	84.38	87.72
CSFF-ISTV2	99.34	100	90.91	90.32	93.1	100	83.33	100	93.33	90	99.67	90.91	95.24	91.8	91.53

Tab.8 Per-class evaluation metrics on severely imbalanced test set

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	91.98	37.5	86.36	61.36	100	99.33	20	63.33	90	46.67	95.51	26.09	73.08	72.97	63.64
STV2	94.64	84.21	88	61.54	63.64	100	53.33	73.33	53.33	70	97.24	65.31	80	57.14	66.67
W-DenseNet	98.66	83.33	88.89	65.62	78.12	98.33	83.33	80	70	83.33	98.5	83.33	84.21	67.74	80.65
CSCNN	97.34	64.29	80	68.57	84.26	97.67	60	80	80	73.33	97.5	62.07	80	73.85	78.57
MSResNet	96.75	76.92	75.76	67.74	81.82	99.33	66.67	83.33	70	60	98.03	71.43	79.37	68.85	69.23
GAN	98.03	82.76	88.46	71.43	75	99.67	80	76.67	66.67	80	98.84	81.36	82.14	68.97	77.42
WGAN	98.35	82.14	89.29	79.41	92.59	99.33	76.67	83.33	90	83.33	98.84	79.31	86.21	84.38	87.72
CSFF-ISTV2	99.34	100	90.91	90.32	93.1	100	83.33	100	93.33	90	99.67	90.91	95.24	91.8	91.53

Tab.9 Overall evaluation metrics on severely imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	86.67±2.44	22.99
STV2	89.29±2.73	38.21
W-DenseNet	92.86±1.75	61.84
CSCNN	90.71±1.61	52.44
MSResNet	90.95±3.09	48.14
GAN	92.86±2.94	57.1
WGAN	94.76±1.6	68.99
CSFF-ISTV2	97.62±0.71	83.66

Tab.9 Overall evaluation metrics on severely imbalanced test set

模型	$A c c$ /%	$G m e a n s$ /%
CNN	86.67±2.44	22.99
STV2	89.29±2.73	38.21
W-DenseNet	92.86±1.75	61.84
CSCNN	90.71±1.61	52.44
MSResNet	90.95±3.09	48.14
GAN	92.86±2.94	57.1
WGAN	94.76±1.6	68.99
CSFF-ISTV2	97.62±0.71	83.66

Fig.13 Confusion matrices of different models on severely imbalance

Fig.14 The t-SNE visualization of different models on severely imbalance

Tab.10 Per-class evaluation metrics on test set with inter-class imbalance

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	99.65	80.83	67.14	67.65	66.67	94	80.83	78.33	76.67	80	96.74	80.83	72.31	71.88	72.73
STV2	98.36	91.19	91.57	70.97	87.5	99.67	96.67	84.44	73.33	46.67	99.01	93.85	87.86	72.13	60.87
W-DenseNet	99.67	96.1	91.86	64.52	61.54	100	98.67	87.78	66.67	53.33	99.83	97.37	89.77	65.57	57.14
CSCNN	91.41	88.52	78.31	50	62.5	99.33	72	72.22	76.67	33.33	95.21	79.41	75.14	60.53	43.48
MSResNet	96.09	90.54	88.04	65.71	100	98.33	89.33	90	76.67	20	97.2	89.93	89.01	70.77	33.33
GAN	99.01	94.84	96.47	78.79	88.89	100	98	91.11	86.67	53.33	99.5	96.39	93.71	82.54	66.67
WGAN	100	96.71	95.24	68.57	68.75	99.38	98	88.89	80	73.33	99.67	97.35	91.95	73.85	70.97
CSFF-ISTV2	100	96.15	100	88.24	100	99.67	100	93.33	100	80	99.83	98.04	96.55	93.75	88.89

Tab.10 Per-class evaluation metrics on test set with inter-class imbalance

模型	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
模型	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
CNN	99.65	80.83	67.14	67.65	66.67	94	80.83	78.33	76.67	80	96.74	80.83	72.31	71.88	72.73
STV2	98.36	91.19	91.57	70.97	87.5	99.67	96.67	84.44	73.33	46.67	99.01	93.85	87.86	72.13	60.87
W-DenseNet	99.67	96.1	91.86	64.52	61.54	100	98.67	87.78	66.67	53.33	99.83	97.37	89.77	65.57	57.14
CSCNN	91.41	88.52	78.31	50	62.5	99.33	72	72.22	76.67	33.33	95.21	79.41	75.14	60.53	43.48
MSResNet	96.09	90.54	88.04	65.71	100	98.33	89.33	90	76.67	20	97.2	89.93	89.01	70.77	33.33
GAN	99.01	94.84	96.47	78.79	88.89	100	98	91.11	86.67	53.33	99.5	96.39	93.71	82.54	66.67
WGAN	100	96.71	95.24	68.57	68.75	99.38	98	88.89	80	73.33	99.67	97.35	91.95	73.85	70.97
CSFF-ISTV2	100	96.15	100	88.24	100	99.67	100	93.33	100	80	99.83	98.04	96.55	93.75	88.89

Tab.11 Overall evaluation metrics on test set with inter-class imbalance

模型	$A c c$ /%	$G m e a n s$ /%
CNN	87.81±2.35	60.42
STV2	93.85±1.61	52.77
W-DenseNet	94.87±1.15	55.49
CSCNN	85.3±2.2	36.33
MSResNet	91.62±2.82	34.47
GAN	96.24±2.18	64.24
WGAN	95.73±1.74	71.25
CSFF-ISTV2	98.29±0.47	86.27

Tab.11 Overall evaluation metrics on test set with inter-class imbalance

模型	$A c c$ /%	$G m e a n s$ /%
CNN	87.81±2.35	60.42
STV2	93.85±1.61	52.77
W-DenseNet	94.87±1.15	55.49
CSCNN	85.3±2.2	36.33
MSResNet	91.62±2.82	34.47
GAN	96.24±2.18	64.24
WGAN	95.73±1.74	71.25
CSFF-ISTV2	98.29±0.47	86.27

Fig.15 Confusion matrices of different models under inter-class imbalance

Fig.16 The t-SNE visualization results of different models under inter-class imbalance

Fig.17 CD diagrams from Nemenyi test based on two metrics

Tab.12 Per-class evaluation metrics on test set for ablation study

M1	M2	M3	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
M1	M2	M3	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
			94.64	84.21	88	61.54	63.64	100	53.33	73.33	53.33	70	97.24	65.31	80	57.14	66.67
√			99	92	82.35	76	71.43	99.33	76.67	93.33	63.33	83.33	99.17	83.64	87.5	69.09	76.92
	√		97.37	82.61	84.85	86.96	72.97	98.67	63.33	93.33	66.67	90	98.01	71.7	88.89	75.47	80.6
		√	99.33	81.82	86.21	75	71.43	99	90	83.33	60	83.33	99.17	85.71	84.75	66.67	76.92
√	√		98.35	89.29	90.62	87.1	92.31	99.33	83.33	96.67	90	80	98.84	86.21	93.55	88.52	85.71
√		√	99.67	85.71	84.85	82.35	92.31	99.33	80	93.33	93.33	80	99.5	82.76	88.89	87.5	85.71
	√	√	99.67	93.33	93.33	87.5	77.14	100	93.33	93.33	70	90	99.83	93.33	93.33	77.78	83.08
√	√	√	99.34	100	90.91	90.32	93.1	100	83.33	100	93.33	90	99.67	90.91	95.24	91.8	91.53

Tab.12 Per-class evaluation metrics on test set for ablation study

M1	M2	M3	$P r e c$ /%					$R e c a l l$ /%					$F 1$ /%
M1	M2	M3	0	1	2	3	4	0	1	2	3	4	0	1	2	3	4
			94.64	84.21	88	61.54	63.64	100	53.33	73.33	53.33	70	97.24	65.31	80	57.14	66.67
√			99	92	82.35	76	71.43	99.33	76.67	93.33	63.33	83.33	99.17	83.64	87.5	69.09	76.92
	√		97.37	82.61	84.85	86.96	72.97	98.67	63.33	93.33	66.67	90	98.01	71.7	88.89	75.47	80.6
		√	99.33	81.82	86.21	75	71.43	99	90	83.33	60	83.33	99.17	85.71	84.75	66.67	76.92
√	√		98.35	89.29	90.62	87.1	92.31	99.33	83.33	96.67	90	80	98.84	86.21	93.55	88.52	85.71
√		√	99.67	85.71	84.85	82.35	92.31	99.33	80	93.33	93.33	80	99.5	82.76	88.89	87.5	85.71
	√	√	99.67	93.33	93.33	87.5	77.14	100	93.33	93.33	70	90	99.83	93.33	93.33	77.78	83.08
√	√	√	99.34	100	90.91	90.32	93.1	100	83.33	100	93.33	90	99.67	90.91	95.24	91.8	91.53

Tab.13 Overall evaluation metrics on test set for ablation study

M1	M2	M3	$A c c$ /%	$G m e a n s$ /%
			89.29	38.21
√			93.57	61.24
	√		92.86	59.16
		√	93.33	60.93
√	√		95.95	75
√		√	95.71	74.41
	√	√	96.19	74.08
√	√	√	97.62	83.66

Tab.13 Overall evaluation metrics on test set for ablation study

M1	M2	M3	$A c c$ /%	$G m e a n s$ /%
			89.29	38.21
√			93.57	61.24
	√		92.86	59.16
		√	93.33	60.93
√	√		95.95	75
√		√	95.71	74.41
	√	√	96.19	74.08
√	√	√	97.62	83.66

References 20

[1]	HE Changfu， HE Deqiang， WEI Zexian， et al. A Train Bearing Imbalanced Fault Diagnosis Method Based on Extended CCR and Multi-scale Feature Fusion Network［J］. Nonlinear Dynamics， 2024， 112（15）： 13147-13173.
[2]	ZHANG Yan， LIU Wenyi， WANG Xin， et al. A Novel Wind Turbine Fault Diagnosis Method Based on Compressed Sensing and DTL-CNN［J］. Renewable Energy， 2022， 194： 249-258.
[3]	陈昱，项薇，林文文，等. 基于半监督学习双模型结构的注塑产品异常检测［J］. 中国机械工程， 2025， 36（3）： 576-583.
	CHEN Yu， XIANG Wei， LIN Wenwen， et al. Abnormal Detection of Injection Molding Products Based on Semi-supervised Learning Dual-model Structure［J］. China Mechanical Engineering， 2025， 36（3）： 576-583.
[4]	JIAO Botao， GUO Yinan， GONG Dunwei， et al. Dynamic Ensemble Selection for Imbalanced Data Streams with Concept Drift［J］. IEEE Transactions on Neural Networks and Learning Systems， 2024， 35（1）： 1278-1291.
[5]	于越，丁磊，金朝阳. 考虑小样本不平衡的主动配电网预测辅助鲁棒状态估计［J］. 高电压技术， 2024， 50（10）： 4550-4560.
	YU Yue， DING Lei， JIN Zhaoyang. Robust Forecasting-aided State Estimation Method of Active Distribution Network Considering Small Sample Imbalance［J］. High Voltage Engineering， 2024， 50（10）： 4550-4560.
[6]	LI Lusi， HE Haibo， LI Jie. Entropy-based Sampling Approaches for Multi-class Imbalanced Problems［J］. IEEE Transactions on Knowledge and Data Engineering， 2020， 32（11）： 2159-2170.
[7]	ZHOU Ziyou. Fault Diagnosis Method for Imbalanced and Unlabeled Data Based on Bayesian Graph Balanced Learning［J］. Measurement Science and Technology， 2024， 35（9）： 096131.
[8]	GUO Zhaolin， ZHAO Yanming. An Intelligent Fault Diagnosis Method Based on Data Enhancement by Multi-information Driven Spiking Generative Adversarial Network［J］. Measurement Science and Technology， 2026， 37（4）： 046107.
[9]	LI Zhenxiang， ZHENG Taisheng， WANG Yang， et al. A Novel Method for Imbalanced Fault Diagnosis of Rotating Machinery Based on Generative Adversarial Networks［J］. IEEE Transactions on Instrumentation and Measurement， 2021， 70： 3500417.
[10]	王娜娜，栗文义，李小龙. 基于不均衡小样本DGA数据与改进CatBoost决策树的油浸式变压器故障诊断方法［J］. 电力系统保护与控制， 2024， 52（23）： 167-176.
	WANG Nana， LI Wenyi， LI Xiaolong. An Oil-immersed Transformer Fault Diagnosis Method Based on DGA Unbalanced Limited Sample Processing and Improved CatBoost［J］. Power System Protection and Control， 2024， 52（23）： 167-176.
[11]	LIU Ze， LIN Yutong， CAO Yue， et al. Swin Transformer： Hierarchical Vision Transformer Using Shifted Windows［C］∥2021 IEEE/CVF International Conference on Computer Vision （ICCV）. Montreal， 2021： 9992-10002.
[12]	ZENG Fei， REN Xiaotong， WU Qing. A Fault Diagnosis Method for Motor Vibration Signals Incorporating Swin Transformer with Locally Sensitive Hash Attention［J］. Measurement Science and Technology， 2024， 35（4）： 046121.
[13]	LIU Ze， HU Han， LIN Yutong， et al. Swin Transformer V2： Scaling up Capacity and Resolution［C］∥2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition （CVPR）. New Orleans， 2022： 11999-12009.
[14]	LI Zhenwei， HAN Yanqi， YANG Xiaoli. Multi-fundus Diseases Classification Using Retinal Optical Coherence Tomography Images with Swin Transformer V2［J］. Journal of Imaging， 2023， 9（10）： 203.
[15]	黄源峰，何剑锋，张峰源，等. 基于SwinV2-EfficientNetV2的铜矿石品位分类方法研究［J］. 矿业研究与开发， 2025， 45（4）： 220-228.
	HUANG Yuanfeng， HE Jianfeng， ZHANG Fengyuan， et al. Research on Copper Ore Grade Classification Method Based on SwinV2-EfficientNetV2［J］. Mining Research and Development， 2025， 45（4）： 220-228.
[16]	刘俊杰，谢俊，王虎，等. 基于全局注意力的Gam-EEGNet在SSVEP分类中的应用［J］. 电子测量技术， 2024， 47（22）： 76-83.
	LIU Junjie， XIE Jun， WANG Hu， et al. The Application of Gam-EEGNet with Global Attention Mechanism in SSVEP Classification［J］. Electronic Measurement Technology， 2024， 47（22）： 76-83.
[17]	张洪，盛永健，黄子龙，等. 基于W-DenseNet的减压阀不平衡样本故障诊断模型［J］. 控制与决策， 2022， 37（6）： 1513-1520.
	ZHANG Hong， SHENG Yongjian， HUANG Zilong， et al. W-DenseNet-based Fault Diagnosis Model of Pressure-reducing Valve with Unbalanced Samples［J］. Control and Decision， 2022， 37（6）： 1513-1520.
[18]	SOLEYMANPOUR S， SADR H， NAZARI SOLEI-MANDARABI M. CSCNN： Cost-sensitive Convolutional Neural Network for Encrypted Traffic Classification［J］. Neural Processing Letters， 2021， 53（5）： 3497-3523.
[19]	DU Linhao. Fault Diagnosis Method of Rotating Machinery Based on MSResNet Feature Fusion and CAM［J］. Journal of Vibroengineering， 2024， 26（7）： 1600-1615.
[20]	潘海洋，徐海锋，郑近德，等. 基于双加权不平衡矩阵分类器的机械故障诊断方法［J］. 机械工程学报， 2024， 60（3）： 170-180.
	PAN Haiyang， XU Haifeng， ZHENG Jinde， et al. Mechanical Fault Diagnosis Method Based on Twin Weighted Imbalanced Matrix Classifier［J］. Journal of Mechanical Engineering， 2024， 60（3）： 170-180.