Abstract:Most transfer learning-based fault diagnosis methods learn diagnostic information from the source domain to enhance performance in the target domain. However, in practical applications, usually there are multiple available source domains, and relying on diagnostic information from only a single source domain limits the transfer performance. To this end, a non-uniformly weighted multisource domain adaptation network is proposed to address the above challenge. In the proposed method, an intra-domain distribution alignment strategy is designed to eliminate multi-domain shifts and align each pair of source and target domains. Furthermore, a non-uniform weighting scheme is proposed for measuring the importance of different sources based on the similarity between the source and target domains. On this basis, a weighted multisource domain adversarial framework is designed to enhance multisource domain adaptation performance. Numerous experimental results on three datasets validate the effectiveness and superiority of the proposed method.
What problem does this paper attempt to address?
The problem that this paper attempts to solve is the application challenges of multi - source domain adaptation (MDA) networks in fault diagnosis under changing working conditions. Specifically, most fault diagnosis methods based on transfer learning learn diagnostic information from a single source domain to improve the performance of the target domain. However, in practical applications, there are usually multiple available source domains, and relying solely on the diagnostic information of a single source domain will limit the transfer performance. Therefore, the paper proposes a non - uniformly weighted multi - source domain adaptation network (NWMDAN), aiming to address the above challenges in the following ways:
1. **Cross - domain distribution alignment strategy**: A within - domain distribution alignment strategy is designed to eliminate multi - domain shift and align each source domain with the target domain.
2. **Non - uniform weighting scheme**: A non - uniform weighting scheme is proposed to measure the importance of different source domains according to the similarity between the source domain and the target domain.
3. **Weighted multi - source domain adversarial framework**: A weighted multi - source domain adversarial training framework is designed to enhance multi - source domain adaptation performance.
Through these methods, the paper aims to improve the fault diagnosis performance under changing working conditions, especially in the case of multi - source domains. The experimental results show that the proposed method is effective on three datasets and outperforms existing methods.
### Formula Summary
1. **Maximum Mean Discrepancy (MMD)**:
\[
D_H(X, Y) \triangleq \left\| \mathbb{E}_{X \sim p}[\phi_1(x)] - \mathbb{E}_{Y \sim q}[\phi_1(y)] \right\|_H
\]
\[
\text{MMD}^2(X, Y) = \frac{1}{n^2} \sum_{i,j = 1}^{n} K(x_i, x_j) - \frac{2}{mn} \sum_{i = 1}^{n} \sum_{j = 1}^{m} K(x_i, y_j) + \frac{1}{m^2} \sum_{i,j = 1}^{m} K(y_i, y_j)
\]
2. **Optimization objective of Domain - Adversarial Neural Network (DANN)**:
\[
L_{\text{DANN}} = \frac{1}{n_s} \sum_{x_i \in D_s} L_c(C(F(x_i)), y_i) - \lambda \frac{1}{n_s + n_t} \sum_{x_i \in (D_s \cup D_t)} L_d(D(F(x_i)), d_i)
\]
3. **Within - domain distribution alignment loss**:
\[
\text{MMD}^2(f_{s,k}, f_t) = \frac{1}{n_{s,k}^2} \sum_{i,j = 1}^{n_{s,k}} K(f_{s,k}^i, f_{s,k}^j) - \frac{2}{n_{s,k}n_t} \sum_{i = 1}^{n_{s,k}} \sum_{j = 1}^{n_t} K(f_{s,k}^i, f_t^j) + \frac{1}{n_t^2} \sum_{i,j = 1}^{n_t} K(f_t^i, f_t^j)
\]
\[
L_{\text{intra}} = \frac{1}{K} \sum_{k = 1}^{K} \text{MMD}^2(f_{s,k}, f_t)
\]
4. **Non - uniform weighting scheme**:
\[
\omega_{s,k} = \frac{\exp(-\eta \text{MMD}^2(f_{s,k}, f_t))}{\sum_{k = 1}^{K} \exp(-\eta \text{MMD}^2(f_{s,k}, f_t))}
\]
5. **Weighted domain adversarial loss**:
\[
L_d = \frac{1}{K} \sum_{k = 1}^{K} \om