A hybrid recognition model of microseismic signals for underground mining based on CNN and LSTM networks

Figures & data

Figure 1. Three-dimensional model of the Xiadian Gold Mine.

Figure 2. Three common types of signals recorded in MS monitoring system: (a) MS signal, (b) blasting signal, and (c) mechanical signal.

Figure 3. Histogram of characteristic parameters of three kinds of signals: (a) dominant frequency, (b) number of peaks and (c) maximum amplitude.

Figure 4. The typical network structure of the CNN.

Figure 5. Schematic diagram of the convolution operation. The left panel shows the feature map before convolution operation, and the right panel shows the feature map after convolution operation.

Figure 6. The implementation process of the Softmax layer.

Figure 7. LSTM units and connection mode of forge gate, input gate and output gate.

Figure 8. The structure of CNN-LSTM hybrid networks. MS, M, and BL stand for MS signal, mechanical signal, and blasting signal, respectively.

Figure 9. The flow chart of the SSA-CNN-LSTM hybrid networks: (a) data pre-processing, (b) extract RC of signals by SSA method, (c) build signal recognition hybrid networks, and (d) test the signal recognition performance.

Figure 10. SSA analysis process of the MS signal with severe noise: (a) original MS signal, (b) RC 1–RC 6, (c) energy ratio of RCs to the original MS signal, and (d) comparison of RC 1 and original MS signal.

Figure 11. Energy ratio of each RC of 30 MS signals. 1 to 8 in the horizontal axis represent the first to eighth RC of the MS signal.

Figure 12. Time-frequency amplitude diagrams of three typical monitoring signals before and after tht SSA processing: (a) MS signal, (b) blasting signal, and (c) mechanical signal.

Figure 13. The network structure of the proposed CNN-LSTM hybrid model.

Table 1. The network structure of the SSA-CNN-LSTM hybrid networks model.

Number of layers	Layer type	Activation function	Output size	Number and size of convolution kernels
1	Input layer		227 × 227 × 3
2	Folding layer		227 × 227 × 3
3	Convolution layer (Conv 1)	ReLU	55 × 55 × 96	96, 11 × 11
4	Maximum pool layer (Pool 1)		27 × 27 × 96	96, 3 × 3
5	Convolution layer (Conv 2)	ReLU	27 × 27 × 256	256, 5 × 5
6	Maximum pool layer (Pool 2)		13 × 13 × 256	256, 3 × 3
7	Convolution layer (Conv 3)	ReLU	13 × 13 × 384	384, 3 × 3
8	Convolution layer (Conv 4)	ReLU	13 × 13 × 384	384, 3 × 3
9	Convolution layer (Conv 5)	ReLU	13 × 13 × 256	256, 3 × 3
10	Maximum pool layer (Pool 5)		6 × 6 × 256	256, 3 × 3
11	Full connection layer (Fc 6)	ReLU	1 × 1 × 4096
12	Dropout layer (Drop 6)		1 × 1 × 4096
13	Full connection layer (Fc 7)	ReLU	1 × 1 × 4096
14	Dropout layer (Drop 7)		1 × 1 × 4096
15	Unfolding layer		1 × 1 × 4096
16	Flatten layer		4096
17	LSTM		4000
18	Dropout layer (Drop 8)		4000
19	Full connection layer (Fc 9)		3
20	Softmax layer	Softmax	3
21	Output layer

Figure 14. The rise time of three kinds of signals.

Figure 15. Confusion matrices of recognition results of four commonly used methods: (a) SVM, (b) KNN, (c) DT, (d) LDA. MS, M, and BL stand for MS signal, mechanical signal, and blasting signal.

Table 2. Signal recognition accuracy for four conventional methods using 5-fold cross-validation.

Models	Category	1st test	2nd test	3rd test	4th test	5th test	Mean
SVM	MS signal	92.1%	91.6%	92.2%	93.4%	92.4%	92.34%
	Blasting signal	88.2%	89.2%	89.3%	90.6%	88.5%	89.16%
	Mechanical signal	72.8%	74.4%	73.2%	75.0%	72.6%	73.6%
	Total	84.37%	85.07%	84.90%	86.33%	84.50%	85.03%
KNN	MS signal	90.1%	89.1%	91.2%	90.4%	89.7%	90.1%
	Blasting signal	85.2%	86.3%	87.1%	85.1%	86.6%	86.06%
	Mechanical signal	74.8%	74.6%	73.1%	73.2%	74.1%	73.96%
	Total	83.37%	83.33%	83.80%	82.90%	83.47%	83.37%
DT	MS signal	89.1%	89.4%	91.0%	89.5%	91.6%	90.12%
	Blasting signal	88.7%	87.6%	88.1%	90.1%	87.5%	88.4%
	Mechanical signal	71.5%	74.2%	70.6%	71.1%	72.3%	71.94%
	Total	83.10%	83.73%	83.23%	83.57%	83.80%	83.49%
LD	MS signal	89.6%	90.2%	88.4%	90.6%	89.9%	89.74%
	Blasting signal	80.5%	78.9%	81.1%	79.6%	79.4%	79.9%
	Mechanical signal	69.5%	70.6%	71.4%	71.8%	70.2%	70.7%
	Total	79.87%	79.90%	80.30%	80.67%	79.83%	80.11%

Table 3. Signal recognition accuracy for four neural network methods using 5-fold cross-validation.

Models	Category	1st test	2nd test	3rd test	4th test	5th test	Mean	Standard deviation
SSA-CNN-LSTM	MS signal	97.8%	96.3%	97.3%	97.4%	97.3%	97.2%	0.49
	Blasting signal	95.3%	94.4%	93.2%	93.3%	93.8%	94.0%	0.78
	Mechanical signal	92.8%	93.3%	92.5%	91.8%	92.0%	92.5%	0.54
	Total	95.3%	94.67%	94.3%	94.17%	94.37%	94.56%	0.40
CNN	MS signal	94%	93.8%	95.5%	94.8%	93.5%	94.32%	0.73
	Blasting signal	93.5%	93.0%	90.5%	91.0%	92.8%	92.16%	1.18
	Mechanical signal	89.5%	88.5%	90.3%	86.9%	90.8%	89.2%	1.39
	Total	92.33%	91.77%	92.10%	90.90%	92.37%	91.96%	0.54
LSTM	MS signal	92.5%	93%	94.3%	93.8%	94.1%	93.54%	0.68
	Blasting signal	90.8%	89.3%	87.8%	88.3%	89.5%	89.14%	1.04
	Mechanical signal	86.3%	86.1%	84.2%	88.8%	87.0%	86.48%	1.49
	Total	89.87%	89.47%	88.77%	90.3%	90.2%	89.72%	0.56
BP	MS signal	89.2%	87.4%	88.8%	87.2%	88.3%	88.18%	0.78
	Blasting signal	86.6%	83.3%	86.4%	83.3%	86.3%	85.18%	1.54
	Mechanical signal	81.9%	74.0%	77.9%	77.0%	75.9%	77.34%	2.62
	Total	85.90%	81.57%	84.37%	82.50%	83.50%	83.57%	1.50

Figure 16. Confusion matrices of recognition results of four neural network methods: (a) SSA-CNN-LSTM, (b) CNN, (c) LSTM, and (d) BP.

Figure 17. Comparison of signal recognition accuracy of eight methods.

Figure A1. The total performance of each distribution with and without over-sampling from Table A2 and Table A5. ID 1 shows no result after oversampling since it was already balanced.

Table A1. Nine distributions of 6000 signals for researching on the impact of imbalanced training data for the SSA-CNN-LSTM model.

ID	MS signals	Blasting signal	Mechanical signal
1	2000	2000	2000
2	1500	2500	2000
3	1000	3000	2000
4	2500	1750	1750
5	3000	1500	1500
6	1500	2250	2250
7	1000	2500	2500
8	2500	2500	1000
9	2250	2250	1500

Table A2. Nine distributions of 3000 signals for researching on the impact of imbalanced training data for the SSA-CNN-LSTM model.

ID	MS signals	Blasting signal	Mechanical signal
1	1000	1000	1000
2	750	1250	1000
3	500	1500	1000
4	1250	875	875
5	1500	750	750
6	750	1125	1125
7	500	1250	1250
8	1250	1250	500
9	1125	1125	750

Table A3. The total and the individual recognition accuracies (the total number of samples is 6000).

ID	MS signals	Blasting signal	Mechanical signal	Total
1	97.2%	94.0%	92.5%	94.57%
2	94.7%	92.9%	91.0%	92.87%
3	94.1%	95.1%	93.0%	94.07%
4	96.0%	89.0%	89.8%	91.60%
5	96.4%	84.4%	88.0%	89.60%
6	94.9%	92.2%	93.9%	93.67%
7	91.6%	94.3%	95.0%	93.63%
8	96.8%	94.5%	79.2%	90.17%
9	96.3%	92.3%	85.9%	91.50%

Table A4. The total and the individual recognition accuracies (the total number of samples is 3000).

ID	MS signals	Blasting signal	Mechanical signal	Total
1	92.1%	86.6%	87.5%	88.73%
2	91.8%	87.0%	87.1%	88.63%
3	85.0%	90.3%	86.5%	87.27%
4	92.0%	85.6%	87.2%	88.27%
5	93.3%	81.3%	86.8%	87.13%
6	90.3%	87.7%	86.2%	88.07%
7	87.0%	88.6%	88.1%	87.90%
8	91.2%	90.4%	73.0%	84.87%
9	91.8%	86.4%	78.0%	85.40%

Table A5. The total and the individual recognition accuracies with oversampling (the total number of samples is 3000).

ID	MS signals	Blasting signal	Mechanical signal	Total
2	92.1%	86.6%	87.5%	88.73%
3	91.2%	86.5%	87.1%	88.27%
4	90.5%	87.1%	86.5%	88.03%
5	91.5%	86.3%	88.1%	88.63%
6	91.9%	86.5%	87.1%	88.50%
7	90.4%	87.5%	88.4%	88.77%
8	91.0%	87.0%	90.1%	89.37%
9	91.1%	86.5%	84.8%	87.47%

ID 1 is not included since it was already balanced and thus was not oversampled.

Data availability statement

The data that support the finding of this study are available from the corresponding author upon reasonable request.