Algorithm 1 GSG Pseudo-code

Input:

Training Set(D) = {Set(min), set(max)}

$\left|\mathrm{Set}(\min )\right|=\mathrm{N}$$,\left|\mathrm{Set}(\max )\right|=\mathrm{M}$ # N is the number of fault class samples.

# M is the number of normal class samples.

Output:

A training Set (D′) with more fault class samples than the original training Set(D).

1: $\mathrm{r}=\mathrm{N}/\mathrm{M}$ # Calculate the class imbalance rate of the original dataset.

2: b = random(r,1)

3: a = int(M*b-N) # Calculate the number of samples to be expanded in Set(min).

4: C = BIC(Set(min)) # Determine an optimal number of components.

5: Gi = GMM(Set(min), C)  # Use GMM to cluster Set(min) into C clusters.

6: for i ←1 to C do

7:      Set(G″i) = {} # Create C empty sample sets.

8:      if $\left|\mathrm{Set}{({\mathrm{G}}^{″}}_{\mathrm{i}})\right|<a\ast (\left|{\mathrm{G}}_{\mathrm{i}}\right|/\mathrm{N})$ then

9:         $\mathrm{Set}(\mathrm{x}\_\mathrm{new})=\mathrm{SMOTE}({\mathrm{G}}_{\mathrm{i}},\mathrm{int}(\mathrm{a}\ast (\left|{\mathrm{G}}_{\mathrm{i}}\right|/\mathrm{N})))$

# Use SMOTE to synthesize new samples based on the Gi cluster samples.

10:          G′I = GMM(Set(min) + Set(xnew_k), C)

# Use GMM again to cluster Set(min) and Set(xnew_k) into C clusters.

11:         for k ←1 to int(a* num(Gi)/N) do

12:               if xnew_k in G′i then               # The xnew_k is the k-th sample in Set(x_new).

13:                  Add the xnew_k to Set(G″i)                     # Add the xnew_k sample to Set(G″i).

14:               if xnew_k not in G′i then

15:                  Remove the xnew_k

16:               end if

17:               end if

18:         end for

19:      end if

20:      Set(N′) = concatenate (Set(G″i))

21: end for

22: Set(D′) = concatenate (Set(max), Set(min), Set(N′))

23: return Set(D′)