CN104655583A - Fourier-infrared-spectrum-based rapid coal quality recognition method - Google Patents

Abstract

The invention relates to a fourier-infrared-spectrum-based rapid coal quality recognition method which is suitable for rapid evaluation of coal quality. The method comprises the following steps: acquiring corresponding spectrum data of multiple coal quality samples, carrying out pretreatment on the acquired spectrum data, compressing spectroscopic data points through wavelet transform, establishing an expert database by use of neural network training, detecting a coal mine by use of an infrared spectrometer to acquire the characteristic data of the detected coal mine, comparing with data in the expert database and judging the coal quality of the detected coal mine. The method is simple, the recognition speed is high, and industrial production application is facilitated.

Description

A kind of ature of coal method for quickly identifying based on FTIR spectrum

Technical field

The present invention relates to a kind of ature of coal method for quickly identifying, be particularly useful for the ature of coal method for quickly identifying based on FTIR spectrum coal quality being evaluated and tested fast to use.

Background technology

Along with the development of science and technology, coal preparing plant automation degree improves gradually, has promoted the growth of coal washing efficiency and economic benefit.But, nowadays the automatic control of domestic coal preparation plant also rests on the such horizontality of automatic start-stop car, Media density bucket Liquid level, Automatic Dosing control, product ash content Real-Time Monitoring etc., substantially belong to scattered FEEDBACK CONTROL, Real-Time Monitoring and control can not be carried out to whole production link.

In actual production, on the one hand main or rely on the fast floating test findings feedback sorting information of fast ash, instruct the normal operation of screening installation and maintain the steady quality of product, but the method is quoted result need more than one hour from being sampled.On the other hand, although online ash measurer is vulnerable to the impact of high atomic number element content in ature of coal, especially for the product coal containing not de-dense medium to the greatest extent, the fluctuation of its measurement result is very large, unstable.

Summary of the invention

For the problems referred to above, the object of this invention is to provide a kind of method simple, identify rapidly accurately, the ature of coal method for quickly identifying based on Fourier spectrum of sample can be added simultaneously fast.

For achieving the above object, the method for quickly identifying of the ature of coal based on Fourier spectrum of the present invention, its step is as follows:

A. the ature of coal sample in multiple different mining area is chosen, choose many parts for often kind, every kind of an ature of coal sample is got and aly forms an ature of coal sample sets, obtain multiple ature of coal sample sets, use pressed disc method that all ature of coal samples are ground to form sample coal dust, sample coal dust compressing tablet is marked, uses infrared spectrometer to gather the spectrum line of all samples coal dust by ature of coal sample sets respectively, record spectrometer and export as spectroscopic data point after sampling digitizing;

B. because spectroscopic data point head end and end when gathering easily produce noise, chosen spectrum Mid-Section Data 340-1000nm; Utilize Savitzky-Golay wave filter when guaranteeing that the shape of spectroscopic data, width are constant respectively to the smoothing denoising of spectrum Mid-Section Data of all coal dust samples;

C. utilize wavelet transformation respectively all coal dust sample spectral data points after smoothing denoising to be carried out compression process, dimensionality reduction, reduce the number of spectroscopic data point, reduce the redundant information in spectroscopic data by compression, accelerate the training speed of neural network;

D. BP neural network is set up, all samples coal dust spectroscopic data point x after being compressed by wavelet transformation ₀, x ₁x _ibe normalized, afterwards as the neuron input BP neural network of BP neural network input layer, the node in hidden layer of BP neural network is M=log ₂n, wherein n is input layer number, and output layer neuron number depends on the classification of ature of coal; The hidden layer of BP neural network and output layer all adopt Sigmoid transport function: in formula, m is transfer function input, and e is natural constant, and S function will input from negative infinite interval to positive infinite mapping to (-1,1) and (0,1); Utilize formula: signal is forward propagated to hidden layer from input layer, and synchronous signal propagates into output layer from hidden layer and also adopts formula: transmission data, hidden layer exports and the neuronic sequence number y of classification corresponding in output layer output _jand y _n, wherein w _ijand w _jncorrespond to input layer and the neuronic connection weights of hidden layer, hidden layer and output layer respectively; x _ifor neuronic input signal, θ _jand θ _nthe weights of external bias;

E. above-mentioned steps is repeated, using each ature of coal sample sets all as the training sample of BP neural network, carry out repetitive exercise, probabilistic neural network is made to draw best weights, by the best weight value of the ature of coal sample in different mining area and colliery sample labeling corresponding record, thus the data craft storehouse of each sample obtained;

F. by the spectroscopic data point of sample coal dust to be identified input data craft storehouse, data craft storehouse can follow the spectroscopic data point of sample coal dust to be identified to carry out neural network test according to the sample best weight value stored, and judges the kind of the ature of coal of testing; When fruit experts database finds not comprise the spectroscopic data point of sample coal dust to be identified by contrast, then the study of the spectroscopic data point of this sample coal dust to be identified can be added experts database.

Spectroscopic data point after smoothing denoising is carried out compression method by described wavelet transformation, first wavelet transformation is carried out to the spectroscopic data point of smoothing denoising and obtain wavelet coefficient c, remove the gibberish being less than threshold values in wavelet coefficient c, the coefficient c after a spectral information specimens preserving _store, the storage space taken is far smaller than the storage space of raw data, when needs extract data by c _storeinverse transformation obtains reconstruct spectroscopic data point;

Described using ature of coal sample sets as in the training sample of BP neural network, utilize the error of sum square of training sample to evaluate the quality of estimation model: to establish total N number of sample, when to M sample training, the actual output of network after kth time repetitive exercise with desired output deviation for:

$\begin{array}{cc}{\mathrm{\δ}}_M^k=(y_M^k-c_M^k)& M=\mathrm{1,2}...N\end{array}$

The error of sum square of individual specimen is:

$E_k=\underset{M=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|{\mathrm{\δ}}_M^k\left|\right|}^2/2$

The error E obtained will be trained _kwith pre-set compared with error margin, if exceed maximum error, then calculate best weights according to error back propagation, then carry out the training error of calculation, until satisfy condition.

Beneficial effect: the present invention passes through Savitzky-Golay wave filter to the smoothing denoising of spectroscopic data, wavelet transformation is used for the compression to a large amount of spectroscopic data, thus reduction operand, simultaneously in conjunction with neural network ature of coal category authentication model, accelerate the training speed of neural network, by plurality of classes ature of coal Sample Establishing training pattern, thus the characteristic obtaining plurality of classes ature of coal sets up experts database, by infrared spectrometer by the characteristic comparison in the data of detection and experts database, thus the quick identification realized ature of coal, its method is simple, accuracy of identification is high, and speed is fast.

Accompanying drawing explanation

Fig. 1 is schematic flow sheet of the present invention;

Fig. 2 is BP network structure;

Fig. 3 is single neuronal structure schematic diagram;

Fig. 4 sets up neural network experts database block diagram.

Embodiment

Below in conjunction with accompanying drawing, embodiments of the present invention are further described:

As shown in Figure 1, the method for quickly identifying of the ature of coal based on Fourier spectrum of the present invention, its step is as follows:

A. the ature of coal sample in multiple different mining area is chosen, choose many parts for often kind, every kind of an ature of coal sample is got and aly forms an ature of coal sample sets, obtain multiple ature of coal sample sets, use pressed disc method that all ature of coal samples are ground to form sample coal dust, sample coal dust compressing tablet is marked, uses infrared spectrometer to gather the spectrum line of all samples coal dust by ature of coal sample sets respectively, record spectrometer and export as spectroscopic data point after sampling digitizing; Use the Handheld FieldSpec infrared spectrometer of ASD company of the U.S., resolution is 3.5nm, and measurement range is 325 ~ 1100nm, and analysis software adopts ASD View SpecPro, Matlab7.12;

B. because spectroscopic data point head end and end when gathering easily produce noise, chosen spectrum Mid-Section Data 340-1000nm; Utilizing Savitzky-Golay wave filter when guaranteeing that the shape of spectroscopic data, width are constant respectively to the smoothing denoising of spectrum Mid-Section Data of all coal dust samples, removing from high frequency random noise.The impacts such as baseline wander, sample are uneven, light scattering.

C. utilize wavelet transformation respectively all coal dust sample spectral data points after smoothing denoising to be carried out compression process, dimensionality reduction, reduce the number of spectroscopic data point, reduce the redundant information in spectroscopic data by compression, accelerate the training speed of neural network; Spectroscopic data point after smoothing denoising is carried out compression method by described wavelet transformation: first carry out wavelet transformation to the spectroscopic data point of smoothing denoising and obtain wavelet coefficient c, remove the gibberish being less than threshold values in wavelet coefficient c, the coefficient c after a spectral information specimens preserving _store, the storage space taken is far smaller than the storage space of raw data, when needs extract data by c _storeinverse transformation obtains reconstruct spectroscopic data point;

D. as shown in Figures 2 and 3, BP neural network is set up, all samples coal dust spectroscopic data point x after being compressed by wavelet transformation ₀, x ₁x _ibe normalized, afterwards as the neuron input BP neural network of BP neural network input layer, the node in hidden layer of BP neural network is M=log ₂n, wherein n is input layer number, and output layer neuron number depends on the classification of ature of coal; The hidden layer K of BP neural network ₁, K ₂... .K _jwith output layer f ₁... f _nall adopt Sigmoid transport function: in formula, m is transfer function input, and e is natural constant, and S function will input from negative infinite interval to positive infinite mapping to (-1,1) and (0,1); Utilize formula: signal is forward propagated to hidden layer from input layer, and synchronous signal propagates into output layer from hidden layer and also adopts formula: transmission data, hidden layer exports and the neuronic sequence number y of classification corresponding in output layer output _jand y _n, wherein w _ijand w _jncorrespond to input layer and the neuronic connection weights of hidden layer, hidden layer and output layer respectively; x _ifor neuronic input signal, θ _jand θ _nthe weights of external bias; Described using ature of coal sample sets as in the training sample of BP neural network, utilize the error of sum square of training sample to evaluate the quality of estimation model: to establish total N number of sample, when to M sample training, the actual output of network after kth time repetitive exercise with desired output deviation for:

$\begin{array}{cc}{\mathrm{\δ}}_M^k=(y_M^k-c_M^k)& M=\mathrm{1,2}...N\end{array}$

The error of sum square of individual specimen is:

$E_k=\underset{M=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|{\mathrm{\δ}}_M^k\left|\right|}^2/2$

The error E obtained will be trained _kwith pre-set compared with error margin, if exceed maximum error, then calculate best weights according to error back propagation, then carry out the training error of calculation, until satisfy condition;

E. above-mentioned steps is repeated, using each ature of coal sample sets all as the training sample of BP neural network, carry out repetitive exercise, probabilistic neural network is made to draw best weights, by the best weight value of the ature of coal sample in different mining area and colliery sample labeling corresponding record, thus the data craft storehouse of each sample obtained;

F. by the spectroscopic data point of sample coal dust to be identified input data craft storehouse, data craft storehouse can follow the spectroscopic data point of sample coal dust to be identified to carry out neural network test according to the sample best weight value stored, and judges the kind of the ature of coal of testing; When fruit experts database finds not comprise the spectroscopic data point of sample coal dust to be identified by contrast, then the study of the spectroscopic data point of this sample coal dust to be identified can be added experts database.

As Fig. 4 sets up neural network experts database method:

1. neural network initialization

Before neural network training, first will carry out arranging initial weight, weights rand function produces;

2. data normalization is inputted

First obtain the mean value of input amendment, deduct mean value with respective sample, by data mobile to coordinate axis center; Calculate sample standard deviation again, data, divided by standard deviation, make variance criterion

3. the training of neural network

Be input in network by the data after normalization, step-up error tolerance limit and maximum iteration time, when reaching error margin or reach maximum iteration time, training stops; If the error of calculation exceedes error margin when iteration, then calculate best weights according to error back propagation, then double counting error, until satisfy condition;

Using ature of coal sample sets all as in the training sample of BP neural network, utilize the error of sum square of training sample to evaluate the quality of estimation model, total total N number of sample, during to M sample training, the actual output of network after kth time repetitive exercise with desired output deviation for:

$\begin{array}{cc}{\mathrm{\δ}}_M^k=(y_M^k-c_M^k)& M=\mathrm{1,2}...N\end{array}$

The error of sum square of individual specimen is:

$E_k=\underset{M=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|{\mathrm{\δ}}_M^k\left|\right|}^2/2$

The error E obtained will be trained _kcompared with initial setting up error margin, if exceed maximum error, then calculate best weights according to error back propagation, then double counting error, until satisfy condition.

4. experts database test

Repeat above-mentioned steps, using each ature of coal sample sets all as the training sample of BP neural network, carry out repetitive exercise, probabilistic neural network is made to draw best weights, by the best weight value of the ature of coal sample in different mining area and colliery sample labeling corresponding record, thus the data craft storehouse of each sample obtained;

5) identification of unknown ature of coal

By the spectroscopic data point of sample coal dust to be identified input data craft storehouse, data craft storehouse can follow the spectroscopic data point of sample coal dust to be identified to carry out neural network test according to the sample best weight value stored, and judges the kind of the ature of coal of testing; If when experts database finds not comprise the spectroscopic data point of sample coal dust to be identified by contrast, then the study of the spectroscopic data point of this sample coal dust to be identified can be added experts database.

Claims (3)

1., based on an ature of coal method for quickly identifying for FTIR spectrum, it is characterized in that step is as follows:

A. the ature of coal sample in multiple different mining area is chosen, choose many parts for often kind, every kind of an ature of coal sample is got and aly forms an ature of coal sample sets, obtain multiple ature of coal sample sets, use pressed disc method that all ature of coal samples are ground to form sample coal dust, sample coal dust compressing tablet is marked, uses infrared spectrometer to gather the spectrum line of all samples coal dust by ature of coal sample sets respectively, record spectrometer and export as spectroscopic data point after sampling digitizing;

B. because spectroscopic data point head end and end when gathering easily produce noise, chosen spectrum Mid-Section Data 340-1000nm; Utilize Savitzky-Golay wave filter when guaranteeing that the shape of spectroscopic data, width are constant respectively to the smoothing denoising of spectrum Mid-Section Data of all coal dust samples;

C. utilize wavelet transformation respectively all coal dust sample spectral data points after smoothing denoising to be carried out compression process, dimensionality reduction, reduce the number of spectroscopic data point, reduce the redundant information in spectroscopic data by compression, accelerate the training speed of neural network;

D. BP neural network is set up, all samples coal dust spectroscopic data point x after being compressed by wavelet transformation ₀, x ₁x _ibe normalized, afterwards as the neuron input BP neural network of BP neural network input layer, the node in hidden layer of BP neural network is M=log ₂n, wherein n is input layer number, and output layer neuron number depends on the classification of ature of coal; The hidden layer of BP neural network and output layer all adopt Sigmoid transport function: in formula, m is transfer function input, and e is natural constant, and S function will input from negative infinite interval to positive infinite mapping to (-1,1) and (0,1); Utilize formula: signal is forward propagated to hidden layer from input layer, and synchronous signal propagates into output layer from hidden layer and also adopts formula: transmission data, hidden layer exports and the neuronic sequence number y of classification corresponding in output layer output _jand y _n, wherein w _ijand w _jncorrespond to input layer and the neuronic connection weights of hidden layer, hidden layer and output layer respectively; x _ifor neuronic input signal, θ _jand θ _nthe weights of external bias;

E. above-mentioned steps is repeated, using each ature of coal sample sets all as the training sample of BP neural network, carry out repetitive exercise, probabilistic neural network is made to draw best weights, by the best weight value of the ature of coal sample in different mining area and colliery sample labeling corresponding record, thus the data craft storehouse of each sample obtained;

F. by the spectroscopic data point of sample coal dust to be identified input data craft storehouse, data craft storehouse can follow the spectroscopic data point of sample coal dust to be identified to carry out neural network test according to the sample best weight value stored, and judges the kind of the ature of coal of testing; When fruit experts database finds not comprise the spectroscopic data point of sample coal dust to be identified by contrast, then the study of the spectroscopic data point of this sample coal dust to be identified can be added experts database.

2. the ature of coal method for quickly identifying based on FTIR spectrum according to claim 1, it is characterized in that: the spectroscopic data point after smoothing denoising is carried out compression method and is by described wavelet transformation, first wavelet transformation is carried out to the spectroscopic data point of smoothing denoising and obtain wavelet coefficient c, remove the gibberish being less than threshold values in wavelet coefficient c, the coefficient c after a spectral information specimens preserving _store, the storage space taken is far smaller than the storage space of raw data, when needs extract data by c _storeinverse transformation obtains reconstruct spectroscopic data point.

3. the ature of coal method for quickly identifying based on FTIR spectrum according to claim 1, it is characterized in that: described using ature of coal sample sets as in the training sample of BP neural network, utilize the error of sum square of training sample to evaluate the quality of estimation model: to establish total N number of sample, when to M sample training, the actual output of network after kth time repetitive exercise with desired output deviation for:

${\mathrm{\δ}}_M^k=(y_M^k-c_M^k),M=\mathrm{1,2}...N$

The error of sum square of individual specimen is:

$E_k=\underset{M=1}{\overset{N}{\mathrm{\Σ}}}{\left|\right|{\mathrm{\δ}}_M^k\left|\right|}^2/2$

The error E obtained will be trained _kwith pre-set compared with error margin, if exceed maximum error, then calculate best weights according to error back propagation, then carry out the training error of calculation, until satisfy condition.