CN108896996A - A kind of Pb-Zn deposits absorbing well, absorption well water sludge interface ultrasonic echo signal classification method based on random forest - Google Patents

Abstract

The invention discloses a random forest-based method for classifying ultrasonic echo signals at the mud-water interface of underground suction wells in lead-zinc mines. Wavelet signal; decompose the collected signal by wavelet, use the Hesusure threshold selection method to calculate the threshold, use the soft threshold function for coefficient processing, and then reconstruct the signal to complete denoising; extract the modulus maximum feature from the denoised signal; Randomly extract modulus maxima features and partial samples with replacement to build a decision tree-based learner, and a random forest classifier composed of multiple decision trees is used for signal classification. The invention has high accuracy rate of classification of echo signals, low operation cost and great value for estimation of echo parameters under different mathematical models.

Description

A random forest-based ultrasonic echo signal of the mud-water interface in a lead-zinc mine suction well Classification

technical field

The invention belongs to the field of ultrasonic detection, and in particular relates to an ultrasonic echo classification method for sludge in an underground water-absorbing well of a lead-zinc mine.

Background technique

The safety of mine drainage is an important factor affecting the safe production of mining areas. Underground drainage is very important to the safety of underground workers and equipment, and is the most basic condition for underground operations. Due to underground mining and other reasons, the water in the well is relatively turbid. After the water accumulates in the water absorption well, it is easy to generate silt in the water absorption well. At present, the water level of the suction well is measured by an ultrasonic liquid level gauge, which gives the liquid level height by measuring the distance from the liquid level surface. This measurement method ignores the thickness of the silt deposited at the bottom of the well. As the water accumulates longer, the error becomes more and more serious. big. The color of the water in the suction well is dark, and it is impossible to estimate the mud level of the deposited silt with the eyes. When the silt height reaches the water pump suction level, the water pump cannot pump water or the water pump pumps silt, causing damage to the water pump. With the further development of computer automation technology, it is a development trend that unattended places such as the pump room are developing. If the water level measurement is inaccurate, or the thickness of the silt cannot be accurately known, it may cause serious consequences. Therefore, obtaining the accurate thickness of the silt is a reliable guarantee for unattended underground water pump rooms and other places.

When the water level of the suction well changes slowly, the solid particles in the water are fully settled, forming a relatively clear mud-water interface at the bottom of the well. Because the acoustic reactance of water and mud is quite different, when the ultrasonic waves propagate to the mud-water interface, they are all reflected , the ultrasonic sensor receives the ultrasonic echo signal of a single echo, and it is relatively easy to estimate the TOF (ultrasonic propagation time). When the solid particles are not sufficiently settled, the ultrasonic signal may generate multiple echo signals, and the different echo numbers lead to different mud level estimation methods. Therefore, it is necessary to classify the types of echo signals corresponding to different working conditions.

Ultrasonic ranging is based on the reflection signal generated when the ultrasonic signal encounters the measured object, and the distance between the ultrasonic probe and the measured object can be estimated according to the echo signal. At present, most ultrasonic ranging uses the threshold method to estimate the ultrasonic propagation time. In lead In the water absorption wells of zinc mines, there are interferences of heavy metal particles and other impurities in the mines, and it is difficult to obtain accurate information by using the traditional threshold method. Therefore, designing an estimation method based on the Gaussian echo signal model to estimate the ultrasonic propagation time can greatly improve the measurement accuracy, but different echo signals correspond to different echo mathematical models, and the echo signals under different models are classified. It is a prerequisite for accurate measurement.

Contents of the invention

The purpose of the present invention is to provide a method for classifying ultrasonic echo signals generated at the silt interface of underground water absorption wells. First, the signal is denoised to reduce the interference of environmental noise and the sensor's own noise on the subsequent feature extraction, and then the modulus maximum feature is extracted. Based on the extracted feature, combined with the random forest algorithm, the collected signal is used for training, and the random based The classifier of the forest algorithm classifies the signal.

The specific steps of the technical solution of the present invention are as follows:

S1: Install the ultrasonic transducer probe at a fixed height from the bottom of the water absorption well, use the ultrasonic transducer and signal acquisition system to obtain the echo signals of the ultrasonic waves reflected by the mud-water interface at the bottom of the water absorption well, and collect them under different working conditions P groups of echo signals, p satisfies 50≤p≤500;

S2: Perform 9-layer wavelet decomposition on the ultrasonic echo signal to obtain the wavelet coefficients of each layer, use the Hesusure threshold selection method for noise reduction processing, use the soft threshold function for coefficient processing, use the processed wavelet coefficients for signal reconstruction, and improve the signal SNR;

S3: Perform Mallat's binary wavelet decomposition algorithm on the denoised signal, decompose it into 6 scales, select 2 ² , 2 ³ , 2 ⁴ , 2 ⁵ scale space as the feature extraction space, and extract 4 modules in each layer respectively The maximum value and the time axis position corresponding to the modulus maximum value form a 32-dimensional feature vector F=[V ₁ , T ₁ , V ₂ , T ₂ ..., V ₁₆ , T ₁₆ ];

S4: Use the feature vector obtained in S3 as the selection attribute, randomly extract 4-dimensional features from it as the feature input of a single base learner, and randomly extract p _x samples from the signal samples obtained in S1 as samples of a single decision tree Input, use Gini impurity to select partition attributes, and build a CART decision tree;

S5: Repeat step S4 n _c times to establish a RandomForest classifier composed of n _c decision trees;

S6: For the newly collected data, input it into the RandomForest classifier established in the above steps, perform prediction classification in each CART decision tree, and select the final classification result by voting.

In the above S1, the ultrasonic transducer probe is installed at a fixed height underwater, the selected probe emission angle is 6°, the height of the probe from the bottom of the pool is h, the highest mud level in history is h _m , and h satisfies hh _m ≥ 30cm , when installing the probe, it is necessary to remove obstacles that may cause interference echoes. With the probe as the center point, the probe surface is the horizontal plane, and down to the bottom of the pool, the cylindrical space with a radius of h*tan3°+h′ is the measurement space. Solid obstacles that may cause interference echoes should be excluded in the measurement space, h' is the redundant amount, satisfying

In the S2, the wavelet function is used to decompose the original signal with 9 layers of wavelets to obtain the wavelet coefficients of each layer, using the Hesusure threshold selection rule, combined with the fixed threshold rule and the unbiased likelihood estimation adaptive threshold rule, the specific steps are as follows:

S21: First calculate the threshold specified by the fixed threshold rule N is the signal length, σ is the standard deviation of the noise;

S22: Then calculate the unbiased likelihood estimation adaptive threshold Rigrsure threshold, the wavelet coefficient square vector X=[x ₁ , x ₂ ,...,x _n ], where x ₁ ≤ x ₂ ≤... ≤ x _n , n is the layer The number of wavelet coefficients, let a risk vector be: S=[s ₁ , s ₂ ,…, s _i ,…, s _n ], each element in the risk vector is: Find the minimum value s _min from the risk vector as the risk value, and find the x _min corresponding to the subscript coefficient of the minimum risk value, then the adaptive threshold is:

S23: Combining the above two threshold selection rules, let K be the sum of squares of wavelet coefficients in a certain layer after wavelet decomposition, and define parameters parameter The value rule of hybrid threshold λ _h is After the threshold is determined, the soft threshold function is selected to process the wavelet coefficients, and the soft threshold function is:

Among them, x _n is the new wavelet system processed by the soft threshold function;

S24: After threshold processing is performed on the wavelet coefficients, signal reconstruction is performed to obtain a noise-reduced signal.

In S3, Mallat’s binary wavelet decomposition algorithm is used to decompose the noise-reduced signal into 6-level scales, and 2 ² , 2 ³ , 2 ⁴ , and 2 ⁵ scale spaces are selected as feature extraction spaces, and in each feature extraction space Extract 4 modulus maxima as the features of this layer, extract the time axis position parameters corresponding to each feature extremum, and form a 4*4*2=32-dimensional feature vector F=[V ₁ , T ₁ , V ₂ , T ₂ . . . , V ₁₆ , T ₁₆ ] are used as the classification attributes of the classifier.

In said S4, Gini impurity is selected as the division attribute selection rule, and a CART decision tree is established, and p _x sample data are randomly extracted from the sample data as a single decision tree data input, and there is replacement from the 32-dimensional feature vector Randomly extract 4 dimensions as the decision tree partition attribute, and the Gini impurity is defined as:

Where y is the number of sample categories, p _k is the proportion of samples of the kth category, and the Gini impurity of attribute Z is defined as:

For the extracted 4-dimensional eigenvalues, each dimension contains p _x sample data, and there are p _x +1 points that can be divided, and the candidate division points are defined as definition Where [z ¹ , z ² ,..., z ⁱ ,..., z ^j ] are j different values on attribute Z, j is the number of samples of a single decision tree, according to the formula

Select the optimal partition attribute and two optimal partition points, so that the data Gini impurity after partitioning is the smallest.

The S5 divides the echo signals of the mud-water interface of the lead-zinc mine water absorption well into three categories, single echo signal, double echo signal and triple echo signal, corresponding to three different mud-water interface situations, which can better estimate The location of the mud-water interface and the composition of the sediment.

The invention proposes a random forest-based method for classifying ultrasonic echo signals at the mud-water interface of a lead-zinc mine water absorption well. Underwater suction wells of lead-zinc mines have continuous inflow and extraction of groundwater generated during the mining process. Since the inflowing groundwater contains a large amount of silt and heavy metal particle impurities, sedimentation will occur at the bottom of the suction well, and the thickness of the silt will continue to change over time. thick. The different components contained in the sediment will produce different echo signals. For different echo signals, the corresponding models are different, and the same parameter model cannot be used for parameter estimation, so the echo signal category can be identified. is the only way to estimate parameters. First, the echo signal is collected as the input signal. The input signal contains a large number of noise components, mainly from the environmental noise and the sensor itself. Therefore, the wavelet denoising method is adopted, which combines the unbiased likelihood estimation adaptive threshold rule and fixed threshold. According to the rules, the soft threshold function is selected to process the wavelet coefficients. Experiments show that the denoising effect is the best when the number of wavelet decomposition layers is 9. The wavelet transform modulus maximum contains a large amount of information such as the structure and shape of the reflection interface. The scale parameter, amplitude parameter and corresponding time axis parameter of the modulus maximum are used as the target feature quantity to classify the target echo. This paper will The target signal is decomposed into 6-level scales by Mallat's binary wavelet algorithm, and 2 ² , 2 ³ , 2 ⁴ , and 2 ⁵ scale spaces are selected as the feature extraction space, and 4 modulus maxima and the modulus maxima are extracted in each layer respectively. The time axis position corresponding to the maximum value forms a 32-dimensional feature vector. In this paper, echo signals are divided into three categories. Since the sample features are continuous variables, two dividing points need to be selected. In this paper, the division attribute rules are defined as:

According to this rule, the division attribute and division point are selected. When the division point contains one of g ⁰ or g ^j+1 , it means that there are only two signal categories in the extracted sample. When the selected division point is g ⁰ and g j+1 When ^j+1 , it means that there is only one signal category in the extracted samples. As an integrated learning method, random forest avoids the problem of overfitting or misclassification of a single decision tree. It uses random extraction of features and samples to train each decision tree. A forest is composed of decision trees of a certain scale. The classification results of a decision tree, and finally vote to determine the final result, which greatly improves the classification accuracy.

Description of drawings

Fig. 1 is a schematic flow chart of the present invention;

Figure 2 is a comparison diagram of the original signal after unbiased adaptive threshold denoising and heuristic threshold denoising respectively.

Detailed ways

In order to describe the technical solutions and advantages of the present invention in more detail, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments.

Fig. 1 is a schematic flow diagram of the present invention, showing the basic flow sequence of the present invention. The specific process includes the following steps:

S1: In order to collect the echo signal, the ultrasonic transducer probe needs to be installed at a fixed height from the bottom of the water absorption well. The ultrasonic transducer used in this embodiment has an emission angle of 6°. If there is no interference signal, select an area 0.5m away from the pool wall, with slow water flow and sufficient precipitation, and use a long iron rod to fix the transducer probe. The upper end of the long iron rod is fixed on the baffle above the water absorption well, and the lower end is connected to the transducer probe, which is fixed at a selected height. The water absorption well of the lead-zinc mine selected in this example is 3.5m deep, the highest mud level in history is 0.9m, and the water level is generally kept above 1.5m, so the probe is fixed at a height of 1.4m from the bottom of the pool. After installing the ultrasonic transducer, run the signal acquisition system to obtain the echo signals after the ultrasonic waves are reflected by the mud-water interface at the bottom of the water absorption well, and collect 50 sets of echo signals under different working conditions.

S2: Perform 9-layer wavelet decomposition on the ultrasonic echo signal to obtain the wavelet coefficients of each layer, use the Hesusure threshold selection method for noise reduction processing, use the soft threshold function for coefficient processing, use the processed wavelet coefficients for signal reconstruction, and improve the signal SNR, combined with the fixed threshold rule and the unbiased likelihood estimation adaptive threshold rule, first calculates the threshold specified by the fixed threshold rule N is the signal length, σ is the standard deviation of the noise, and then calculate the unbiased likelihood estimation adaptive threshold Rigrsure threshold, vector X=[x ₁ , x ₂ ,…,x _n ], where x ₁ ≤x ₂ ≤…≤ x _n , n is the number of wavelet coefficients in this layer, let a risk vector be: S=[s ₁ , s ₂ ,…, s _i ,…, s _n ], each element in the risk vector is: Find the minimum value s _min from the risk vector as the risk value, and find the x _min corresponding to the subscript coefficient of the minimum risk value, then the adaptive threshold is: Combining the above two threshold selection rules, let K be the sum of squares of wavelet coefficients in a certain layer after wavelet decomposition, and define the parameter parameter The value rule of hybrid threshold λ _h is After the threshold is determined, the soft threshold function is selected to process the wavelet coefficients, and the soft threshold function is:

Wherein x _n is the new wavelet coefficient after the soft threshold function processing, and the present embodiment uses the hierarchical threshold Each layer of wavelet coefficients corresponds to a different threshold. Figure 2 shows the comparison after denoising using the heuristic threshold rule and the unbiased adaptive threshold rule respectively. It can be seen that the denoising effect under the heuristic threshold rule is significantly better than that without Partially adaptive threshold rules.

S3: Perform Mallat's binary wavelet decomposition algorithm on the denoised echo signal, decompose it into 6 scales, select 2 ² , 2 ³ , 2 ⁴ , 2 ⁵ scale space as the feature extraction space, and extract 4 A modulus maxima and the time axis position corresponding to the modulus maxima form a 32-dimensional feature vector F=[V ₁ , T ₁ , V ₂ , T ₂ ..., V ₁₆ , T ₁₆ ];

S4: Use the eigenvector obtained in S3 as the selection attribute, there are 32 dimensional features in total, use the formula d=log ₂ k to calculate the number of extracted features, which is 4 in this embodiment, and randomly extract 4 dimensional features from it as a single base learner 10 samples are randomly selected from the signal samples obtained in S1 as the sample input of a single decision tree, and the Gini impurity is used to select the partition attribute to establish a CART decision tree. The Gini impurity is defined as:

Where y is the number of sample categories, p _k is the proportion of samples of the kth category, and the Gini impurity of attribute Z is defined as:

For the extracted 4-dimensional eigenvalues, each dimension contains p _x sample data, and there are 11 dividing points. The candidate dividing points are defined as definition Where [z ¹ , z ² ,..., z ⁱ ,..., z ^j ] are j different values on attribute Z, j is the number of samples of a single decision tree, according to the formula

Select the optimal partition attribute and two optimal partition points, so that the data Gini impurity after partitioning is the smallest. The decision tree algorithm often falls into an over-fitting state, and the number of divisions is too many. It may take some of the characteristics of the training set data as the general characteristics of the input data and cause classification errors. Therefore, it is necessary to cut the tree when building a single decision tree. Branch handling and termination conditions are set to prevent overfitting. In this embodiment, each decision tree uses 10 sample data and randomly extracts 4-dimensional feature vectors. The data volume of a single decision tree is small, and the random forest algorithm itself can avoid the pruning operation of a single decision tree, so only It is enough to set the termination condition. In this embodiment, it is stipulated that each child node has only one type of signal or the Gini impurity after division of all division attributes is the same to stop splitting.

S5: Repeat step S4 100 times to establish a RandomForest classifier composed of 100 decision trees;

S6: For the newly collected data, input it into the RandomForest classifier established in the above steps, perform prediction classification in each CART decision tree, and select the final classification result by voting. This embodiment adopts the relative majority voting method, Define the prediction output of the random forest classifier described in this article as a 100-dimensional vector i represents the category corresponding to the prediction result, and the voting rule can be described as That is, the prediction result is the output category with the most votes.

The embodiments in the above description are only a part of the embodiments of the present invention, and the scope of protection claimed by the present invention is not limited to the above-mentioned specific implementation methods. Under the premise of not paying creative work, it is possible to obtain a substantially identical solution to the present invention. It belongs to the protection scope of the present invention.

Claims (5)

1. A kind of lead-zinc mine suction well mud-water interface ultrasonic echo signal classification method based on random forest, it is characterized in that comprising the following steps: S1: Install the ultrasonic transducer probe at a fixed height from the bottom of the water absorption well, use the ultrasonic transducer and signal acquisition system to obtain the echo signals of the ultrasonic waves reflected by the mud-water interface at the bottom of the water absorption well, and collect them under different working conditions P groups of echo signals, p satisfies 50≤p≤500; S2: Perform 9-layer wavelet decomposition on the ultrasonic echo signal to obtain the wavelet coefficients of each layer, use the Hesusure threshold selection method for noise reduction processing, use the soft threshold function for coefficient processing, use the processed wavelet coefficients for signal reconstruction, and improve the signal SNR; S3: Perform Mallat's binary wavelet decomposition on the noise-reduced signal, decompose it into 6-level scales, select 2 ² , 2 ³ , 2 ⁴ , and 2 ⁵ scale spaces as the feature extraction space, and extract 4 mode poles in each layer The maximum value and the time axis position corresponding to the modulus maximum value form a 32-dimensional feature vector F=[V ₁ , T ₁ , V ₂ , T ₂ ..., V ₁₆ , T ₁₆ ]; S4: Use the feature vector obtained in S3 as the selection attribute, randomly extract 4-dimensional features from it as the feature input of a single base learner, and randomly extract a sample from the signal samples obtained in S1 as the sample input of a single decision tree, Use Gini impurity to select partition attributes and build a CART decision tree; S5: Repeat step S4 n _c times to establish a RandomForest classifier composed of n _c decision trees; S6: For the newly collected data, input it into the RandomForest classifier established in the above steps, perform prediction classification in each CART decision tree, and select the final classification result by voting. 2. the method for classifying ultrasonic echo signals at the mud-water interface of lead-zinc mine water-absorbing wells based on random forest according to claim 1, characterized in that: S1 installs the ultrasonic transducer probe at a fixed height position underwater, and the selected probe emits The angle is 6°, the height of the probe from the bottom of the pool is h, the highest mud level in history is h _m , h satisfies hh _m ≥ 30cm, the probe is the center point, the probe surface is the horizontal plane, and the radius is h*tan3 down to the bottom of the pool The cylindrical space of °+h′ is the measurement space, h′ is the redundant quantity, satisfying 3. the ultrasonic echo signal classification method based on the random forest of lead-zinc mine water absorption well mud-water interface according to claim 1, is characterized in that: use wavelet function to carry out 9 layers of wavelet decomposition to original signal in S2, obtain the wavelet coefficient of each layer , using the Hesusure threshold selection rule, combined with the fixed threshold rule and the unbiased likelihood estimation adaptive threshold rule, first calculate the threshold specified by the fixed threshold rule σ is the standard deviation of the noise, N is the length of the signal, and then calculate the unbiased likelihood estimation adaptive threshold Rigrsure threshold, the wavelet coefficient square vector X=[x ₁ , x ₂ ,…, x _n ], where x ₁ ≤ x ₂ ≤...≤x _n , n is the number of wavelet coefficients in this layer, let a risk vector be: S=[s ₁ , s ₂ ,..., s _i ,..., s _n ], each element in the risk vector is: Find the minimum value s _min from the risk vector as the risk value, and find the x _min corresponding to the subscript coefficient of the minimum risk value, then the adaptive threshold is: Combining the above two threshold selection rules, let K be the sum of squares of wavelet coefficients in a certain layer after wavelet decomposition, and define the parameter parameter The value rule of hybrid threshold λ _h is After the threshold is determined, the soft threshold function is selected to process the wavelet coefficients, and the soft threshold function is: Among them, x _n is the new wavelet coefficient processed by the soft threshold function. 4. according to claim 1 based on the lead-zinc mine suction well mud-water interface ultrasonic echo signal classification method of random forest, it is characterized in that: in S4, Gini impurity is selected as the division attribute selection rule, and a CART decision tree is set up, from the sample Randomly extract p _x sample data that is replaced in the data as the data input of a single decision tree, and randomly extract 4 dimensions from the 32-dimensional feature vector with replacement as the division attribute of the decision tree. The Gini impurity is defined as: Where y is the number of sample categories, p _k is the proportion of samples of the kth category, and the Gini impurity of attribute Z is defined as: For the extracted 4-dimensional eigenvalues, each dimension contains p _x sample data, and there are p _x +1 points that can be divided, and the candidate division points are defined as definition Where [z ¹ , z ² ,..., z ⁱ ,..., z ^j ] are j different values on attribute Z, j is the number of samples of a single decision tree, according to the formula Select the optimal partition attribute and two optimal partition points, so that the data Gini impurity after partitioning is the smallest. 5. the method for classifying ultrasonic echo signals of the mud-water interface of the lead-zinc mine water absorption well based on random forest according to claim 1, characterized in that: the mud-water interface echo signals of the lead-zinc mine water absorption well are divided into three categories, corresponding to Three different mud-water interface situations, namely single-echo type, double-echo type and triple-echo type.