Introduction
Proposed technique
Principle of sparse-autoencoder for unsupervised feature extraction
Transformation of training and testing data by extracted features
Principle of the softmax regression classifier
Parameters used for the sparse autoencoders and softmax regression
Experimental setup and results
S. no. | Fault type | Number of data sets |
---|---|---|
1 | PGW (primary gear whining) | 64 |
2 | MRN (magneto rotor noise) | 65 |
3 | TAPPET | 59 |
4 | Healthy engine | 60 |
Ratio (%) | Pos1 | Pos2 | Pos3 | Pos4 | Majority voting |
---|---|---|---|---|---|
5–95 | 65.25 | 55.08 | 57.63 | 77.12 | 53.81 |
15–85 | 90.48 | 76.67 | 80.0 | 90.95 | 86.67 |
25–75 | 89.78 | 95.16 | 89.25 | 97.85 | 98.39 |
35–65 | 95.65 | 90.68 | 89.96 | 96.89 | 98.14 |
50–50 | 95.12 | 95.12 | 95.12 | 96.75 | 98.73 |
75–25 | 91.94 | 93.55 | 100 | 100 | 98.39 |
S. no. | Fault type | Pos 1 | Pos 2 | Pos 3 | Pos 4 | Majority voting |
---|---|---|---|---|---|---|
2 | PGW | 91.67 | 97.92 | 100 | 97.92 | 100 |
3 | MRN | 97.96 | 93.88 | 95.92 | 100 | 100 |
4 | TAPPET | 88.64 | 93.18 | 81.82 | 95.45 | 95.45 |
5 | Healthy engine | 84.44 | 95.56 | 77.78 | 100 | 97.78 |
Comparison of softmax regression with ANN-based classifier
Position | Classification performance |
---|---|
Position 1 | 90.86 |
Position 2 | 95.16 |
Position 3 | 89.25 |
Position 4 | 97.85 |
Ratio (%) | Pos1 | Pos2 | Pos3 | Pos4 | Majority voting |
---|---|---|---|---|---|
5–95 | 74.70 | 60.90 | 72.08 | 85.72 | 69.03 |
15–85 | 89.52 | 76.81 | 87.98 | 94.35 | 88.81 |
25–75 | 90.03 | 81.17 | 91.63 | 97.36 | 93.96 |
35–65 | 94.22 | 83.27 | 94.72 | 99.02 | 97.20 |
50–50 | 94.73 | 83.10 | 96.11 | 99.14 | 97.73 |
75–25 | 95.20 | 84.09 | 97.87 | 99.75 | 98.74 |
Comparison with existing techniques
S. no. | Technique | Classifier | No of faults | No of training data sets | Performance |
---|---|---|---|---|---|
1 | Spectrogram-based feature extraction [4] | ANN | 7 | 400 | Less than 93 % |
2 | FFT and correlation based [20] | Comparison with prototype engine correlation matrix | 4 | NA | Less than 93 % |
3 | WPT-based feature extraction [3] | ANN | 5 | 30 | Over 95 % in various engine operating conditions |
4 | Reducing the size of the matrix representation of the time–frequency image of the fault signal [22] | Fuzzy clustering | NA | NA | Very limited |
5 | Proposed technique | Softmax regression | 4 | 62 | 98 % |