基于小波包多尺度模糊熵和加权KL散度的煤岩智能识别

Intelligent recognition of coal and rock based on wavelet packet multi-scale fuzzy entropy and weighted KL divergence

垮落煤岩智能识别是智能放煤的前提,通过垮落煤岩实时精准识别可避免人工放煤造成的顶煤“欠放”或“过放”问题。现有煤岩识别方法大多通过数据降维处理获得垮落煤岩特征向量,通过构建识别模型进行煤岩识别,但数据降维、模型建立和训练均需较长时间,一定程度上影响了连续综放开采效率。针对该问题,提出了一种基于小波包多尺度模糊熵和加权KL散度的煤岩智能识别方法。对不同工况(顶煤垮落、岩石垮落、大块顶煤垮落)下垮落煤岩冲击液压支架后尾梁的振动信号进行小波包分解,得到一系列频带;对各频带的序列进行粗粒化,计算各频带多个尺度粗粒化向量的模糊熵,即小波包多尺度模糊熵,将其作为特征向量;以小波包分解后各频带能量与振动信号总能量的比值作为加权KL散度的权重,比较待测未知样本与不同工况下样本特征向量的加权KL散度,实现垮落煤岩的实时精准识别。实验结果表明:基于小波包多尺度模糊熵和加权KL散度的方法可有效识别垮落煤岩类别,而基于多尺度模糊熵和KL散度的方法、基于小波包模糊熵和KL散度的方法识别效果不佳;将小波包多尺度模糊熵作为特征向量时,BP神经网络识别准确率达95%,进一步验证了小波包多尺度模糊熵可作为表征垮落煤岩的特征向量;整个煤岩识别过程耗时为1.063 9 s,基本满足垮落煤岩智能识别实时性需求,大大降低了对连续综放开采效率的影响,综合性能优于同类煤岩识别方法。

Intelligent recognition of collapsed coal and rock is a prerequisite for intelligent coal caving.Real-time and precise recognition of collapsed coal and rock can avoid the problem of"under caving"or"over caving"of top coal caused by manual coal caving.Most existing coal and rock recognition methods obtain collapsed coal and rock feature vectors through data dimensionality reduction processing,and construct recognition models for coal and rock recognition.However,data dimensionality reduction,model establishment,and training all require a long time.To some extent,these factors affect the efficiency of continuous fully mechanized caving mining.In order to solve the above problems,an intelligent coal and rock recognition method based on wavelet packet multiscale fuzzy entropy and weighted KL divergence is proposed. Wavelet packet decomposition is performed on the vibration signals of the tail beam after the hydraulic support is impacted by collapsed coal and rock under different working conditions (top coal collapse, rock collapse, and large top coal collapse) to obtain a series of frequency bands. The sequences of each frequency band are coarse-grained. The method calculates the fuzzy entropy under multiple scales of coarse-grained sequences in each frequency band, that is, wavelet packet multi-scale fuzzy entropy. The method uses it as a feature vector. The method uses the ratio of the energy of each frequency band after wavelet packet decomposition to the total energy of the vibration signal as the weight of the weighted KL divergence. The weighted KL divergence of the unknown samples to be tested and the sample feature vectors under different working conditions are compared. The real-time and precise recognition of collapsed coal and rock is achieved. The experimental results show that the method based on wavelet packet multi-scale fuzzy entropy and weighted KL divergence can effectively recognize the category of collapsed coal and rock. The method based on multi-scale fuzzy entropy and KL divergence and the method based on wavelet packet fuzzy entropy and KL divergence have poor recognition performance. When wavelet packet multi-scale fuzzy entropy is used as the feature vector, the recognition accuracy of the BP neural network reaches 95%. It further verifies that wavelet packet multi-scale fuzzy entropy can be used as the feature vector to characterize collapsed coal and rock. The entire coal and rock identification process takes 1.063 9 seconds, which basically meets the real-time requirements of intelligent recognition of collapsed coal and rock. At the same time, it greatly reduces the impact on the efficiency of continuous fully mechanized caving mining. Its comprehensive performance is superior to similar coal and rock recognition methods.