CN101126680A - Thermal power plant ball mill load soft-sensing method - Google Patents

Abstract

The utility model discloses a load soft measurement method for a coal grinding machine in the thermal power plant, which aims to solve the load detecting difficulty existing long time ago of the coal grinding machine in the thermal power plant. The utility model acquires the information attributing the grinder load by the self-adaptive feature extraction from the grinder noise, the vibration and the differential pressure signal, creates a plurality of rules to identify the condition and the fitting polynomial under the rules after the fuzzy classification of the coal grindability factor, water content of the coal, steel balls loading and the load and finally integrates the output of all the rules to get the grinder load. The utility model has the advantages of overcoming the failure to adapt to the condition changes of the prior method, acquiring good accuracy and sensitivity under various conditions and providing guarantee for effective monitor of the grinder.

Description

A kind of thermal power plant ball mill load soft-sensing method

Technical field

The present invention relates to a kind of automatic measurement technical field, be particularly related to a kind of thermal power plant ball mill load soft-sensing method, this method is applicable to that the load of the barrel-shaped steel ball mill of thermal power plant's the ball type pulverizer system detects, adopt online human error rejecting, self-adaptive features information extraction, operating mode fuzzy recognition, rule-based technological means such as adaptive weighted, the variation according to operating mode that can be in good time provides the information on load of bowl mill.

Background technology

Pulverized coal preparation system is one of main backup system of fuel-burning power plant, because barrel-shaped steel ball mill is the strongest to the adaptability of coal, is domestic most widely used a kind of coal pulverizer.Coal pulverizer is the key equipment of pulverized coal preparation system, and can it normally move and whether operate in optimum condition, is directly connected to the work efficiency of pulverized coal preparation system.Mill load is accurately monitored, and is the condition precedent that pulverized coal preparation system is optimized control, thereby can realize saving energy and reduce the cost, keep the safety in production, protecting the purpose of environment.Therefore how accurately monitoring mill load becomes key and difficult point.

Because the grinding machine poor working environment, dust pollution is big, and internal environment is abominable, therefore can't directly measure, and can only detect by indirect method.At present, use more mill load detection method mill sound method, vibratory drilling method and power method are arranged.Existing mill sound method is to judge its load by the single sound transducer easy detection grinding machine noise sound intensity, shortcoming is that accuracy of detection is not high, failing effectively to remove ground unrest disturbs, particularly when many grinding machines moved in a workshop simultaneously, the noise that is sent when closing on the grinding machine operation can have a strong impact on the accuracy that load detects.Vibratory drilling method is when utilizing mill running, abrasive body and material are partial to a side of grinding machine, the rotating part of grinding machine is in serious non-equilibrium state, cause unbalanced centrifugal force, and make grinder system vibration, how many relevant these Characteristics Detection of grinding machine its oscillation intensity and grinding charge material amount when rotating speed is constant; The deficiency of vibratory drilling method is poor linearity, and accuracy is not high.The thinking of power method is that motor is consumed when measuring the grinding machine operation power is judged the load in the grinding machine.In actual applications, be the working current of measuring grinding machine.The weak point of this kind method is the working current of grinding machine in entire work process, and it is not very big changing, and measurement sensitivity is low.

In recent years, also there are some to take multiple input information comprehensively to obtain the method for mill load, as

" soft-sensing model that detects based on the thermal power plant mill load of ANFIS " (department is just complete, Cao Hui, Zhang Yanbin etc., Chinese journal of scientific instrument, the 4th phase supplementary issue II, 2007, vol.28) [1], " based on the soft measurement of loading of combined type neural network for thermal power plants barrel-shaped steel ball mill " (department is just complete, Cao Hui, Zhang Yanbin etc., thermal power generation, 2007, the 5 phases) [2]." the pulverized coal preparation system ball mill load soft based on neuroid is measured " (Wang Dongfeng, Song Zhiping) [3].But the problem that these methods exist is, it all is the soft-sensing model of setting up at single ball mill load operating mode, in case operating mode changes, the precision of prediction of soft-sensing model will variation, though document [3] has been set up the model of divided working status, it is distinguished operating mode and only adopts differential pressure to obtain to belong to the degree of empty mill and full mill, does not consider the working conditions change factor that raw coal, abrasive body cause, and the easier air quantity that is subjected to of differential pressure influences, and can not well discern operating mode; In addition, the input auxiliary variable is not rejected human error and rejected non-target information, because the characteristics of pulverized coal preparation system object, auxiliary variable wherein such as grinding machine noise, vibration, information such as pressure all contain a large amount of human errors, thereby do not weed out if these human errors are not tracked down, will influence the soft-sensing model precision of prediction of foundation, even make it draw error result, aspect the feature extraction of auxiliary variable, feature extraction is carried out to noise and vibration signal in document [1] and [2], but this feature extraction also reckons without working conditions change, the characteristics that its feature also changes thereupon, document [2] and [3] selection pressure class signal be as auxiliary variable, but do not consider that these pressure signals more are subject to the influence that throttle opening changes, and is not the reflection that this simple target of mill load changes.

Grinding machine noise and vibration information are to change closely-related auxiliary variable with mill load, and these two input quantities are easy to be subjected to the influence of coal, coal moisture content and steel ball loading capacity, variation along with operating mode, its feature that characterizes mill load also can change thereupon, pressure class signal also can reflect the variation of mill load to a certain extent, be subjected to the influence that air quantity changes but they are easier, so these influence factors must be rejected the variation that could reflect mill load.Therefore set up a kind of mill load soft-sensing model that can the self-adaptation operating mode, by coal, coal moisture content, steel ball loading capacity and the multiple factor of mill load mill working is carried out identification, and the feature of therefrom extracting the sign load enters soft-sensing model, thereby obtain mill load information accurately, will play important impetus the monitoring and the control of mill load.

Summary of the invention

The objective of the invention is to, the flexible measurement method that provides a kind of load of heat engine plant canister type steel ball coal pulverizer to detect is in order to solve the problem that thermal power plant's coal pulverizer load for a long time is difficult to detect.This flexible measurement method carries out identification by coal grindability factor, coal moisture content, steel ball loading capacity, mill load to operating mode, on the basis of operating mode identification, the characteristic information that extracts from grinding machine noise, vibration, gateway differential pressure signal is carried out rule-based adaptive weighted, thereby obtain mill load information accurately.

In order to realize above-mentioned task, the present invention takes following technical solution:

A kind of heat engine plant canister type steel ball mill load flexible measurement method, it is characterized in that, this method is set up hardware platform in the mill load measuring system, this hardware platform comprises sensor noise, vibration transducer, gateway differential pressure pick-up, hot blast family of power and influence position indicating device, cold-air flap valve position indicating device, reduced air family of power and influence position indicating device, mill exhauster inlet baffle valve position indicating device, Gravimetric Coal Feeders controller, data acquisition unit and computing machine, carries out soft measurement by operation Survey Software on computers; Survey Software and computing machine communication obtain real-time process data and human-machine interaction data by the data collecting card collection, specifically may further comprise the steps:

1) soft-sensing model input auxiliary variable is selected: a class is the load characteristic info class, is used for match mill load information, comprises the characteristic information F that extracts from noise signal _ZY, the characteristic information F that from vibration signal, extracts _ZDWith the characteristic information F that from the differential pressure signal of gateway, extracts _CYAnother kind of is operating mode identification information class, is used to form the rule of identification operating mode, comprises coal grindability factor λ _m, coal moisture content λ _s, steel ball loading capacity W _GQ, mill load L _GK

2) acquisition methods of input auxiliary variable

Noise and vibration signal characteristics information F _ZYAnd F _ZD: the noise and the vibration data sequence of gathering are carried out power spectrumanalysis, obtain characteristic spectra information, the energy in the characteristic spectra is carried out adaptive weighted acquisition characteristic information;

Gateway differential pressure characteristic information F _CY: to gateway differential pressure P _CY, hot blast door aperture μ _RF, cold-air flap aperture μ _LF, recycle air door aperture μ _ZXF, mill exhauster inlet baffle aperture μ _PRFFive information polynomial fittings obtain, and have eliminated the influence of throttle opening changing factor in the gateway differential pressure characteristic information of acquisition, and it changes the situation of change that only reflects the grinding machine internal loading.

F _CY＝p ₀+p ₁P _CY+p ₂μ _RF+p ₃μ _LF+p ₄μ _ZXF+p ₅μ _PRF

Parameter { p wherein ₀, p ₁, p ₂, p ₃, p ₄, p ₅Obtain by the least square training, training sample obtains in the following way: adopt control coal supply system automatically, the coal position is remained on the different levels, on every kind of coal position level, in the practical operation scope, change each throttle opening respectively, write down corresponding data, form training sample data collection.

Coal grindability factor λ _mWith coal moisture content λ _s: the human-machine interaction data by operating personnel's typing obtains; By thermal power plant's pulverized coal preparation system working specification, will regularly chemically examine these parameters of raw coal.

Steel ball loading capacity: by first loading amount W ₀, magnitude of recruitment W _B(k) and wear extent W _SHCalculate acquisition, the steel ball loading capacity computing formula is as follows:

$W_{\mathrm{GQ}}=[{\mathrm{<figure-callout\; id="0"\; label="W"\; filenames="A20071001863800211.png,A20071001863800221.png"\; state="\{\{state\}\}">W</figure-callout>}}_0+\underset{k}{\mathrm{\&Sigma;}}{W}_B\left(k\right)-W_{\mathrm{SH}}]=[{\mathrm{<figure-callout\; id="0"\; label="W"\; filenames="A20071001863800211.png,A20071001863800221.png"\; state="\{\{state\}\}">W</figure-callout>}}_0+\underset{k}{\mathrm{\&Sigma;}}{W}_B\left(k\right)-\mathrm{\&Sigma;}{w}_{\mathrm{FH}}\left(L_{\mathrm{GK}}\right)w_{\mathrm{MF}}]$

In this formula, first loading amount W ₀, magnitude of recruitment W _B(k), unit powder process dimension ball loss w under the different operating modes _FHObtain powder process amount w by human-machine interaction data _MFObtain by the Gravimetric Coal Feeders data on flows.

Steel ball is loading amount W just ₀Only, in normal course of operation, can't change the grinding machine steel ball being unloaded again, load the back change; Each steel ball magnitude of recruitment W _B(k) be in the grinding machine operational process, by the amount decision that operating personnel replenish steel ball, total magnitude of recruitment is each steel ball magnitude of recruitment W _B(k) result who adds up; Wear extent is according to unit powder process dimension ball loss W under the different load situation _FH(L _GK) to powder process amount w _MFThe result of accumulative total, because grinding machine is being under the different operating modes such as empty mill, full mill, the wearing and tearing of steel ball will great changes will take place, therefore it distinguished operating mode, to obtain wear extent more accurately.Unit powder process dimension ball loss w wherein _FH(L _GK) be in basic, normal, high three kinds of situations according to mill load, set three kinds of loss value w1, w2, w3, shown in following formula: $w_{\mathrm{FH}}=\left\{\begin{array}{cc}w1& L_{\mathrm{GK}}<0.4\\ \mathrm{<figure-callout\; id="2"\; label="w"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">w</figure-callout>}2& 0.4\&le;L_{\mathrm{GK}}\&le;0.7\\ w3& L_{\mathrm{GK}}>0.7\end{array}\right.$

Powder process amount w _MFDetermine by the feeder coal-supplying amount, accumulation calculating is carried out in the steel ball loss in the operational process, can obtain the total wear extent of steel ball.[] is rounding operation in the formula, and promptly steel ball loading capacity can't all change in each computation process, but only just can change when steel ball loading capacity increases or reduce to surpass under 1 ton the situation, can significantly reduce the calculated amount of soft-sensing model like this.

Mill load L _GKObtain by last periodic model output result;

3) noise, vibration, the employing of gateway differential pressure signal are carried out pre-service based on the gross error detection method of intermediate value distance, reject the human error influence;

4) adopt rule-based adaptive weighted soft-sensing model to calculate mill load information

Described noise and vibration signal characteristics frequency range and unconventional a certain continuous frequency range, but the set that Frequency point is formed is (because in the actual application, sample sequence is carried out frequency-domain analysis, that obtain is discrete point in frequency f, there is the fixing f=k * Δ f that concerns between frequency f and its frequency domain sequence subscript k, and adopt k to calculate in the actual computation process, so all frequency information is described in this instructions with k) its acquisition methods is as follows:

We provide the definition of characteristic spectra: the set K that characteristic spectra is made up of a series of Frequency points _Feature, wherein each Frequency point should meet the following conditions when any given two kinds of change working processes:

$\left|\frac{{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(i\right)}}^2-{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(j\right)}}^2}{\max ({{\left|X\left(k\right)\right|}_{\mathrm{state}\left(i\right)}}^2,{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(j\right)}}^2)}\right|{\mathrm{\&lambda;}}_{\mathrm{ij}},k=0~k_h$

k _h＝min{[f _h/Δf]，[f _s/2Δf]}

The frequency domain data sequence of X (k) wherein for data time series x (n) being carried out obtain after FFT calculates, skin || for taking absolute value computing, internal layer || for asking modular arithmetic, [] is rounding operation; λ _IjThe threshold of sensitivity for state state (i) and state (j) conversion; f _hBe the highest frequency of actual signal, determine by sensor hardware; f _sBe sample frequency; Δ f is a signal resolution, and is relevant with FFT computing points N.

Definite step of characteristic spectra is as follows:

1) selected three kinds of load conditions of grinding machine are in empty mill, normally, completely grind three kinds of state state (1), state (2), state (3) as selected grinding machine;

2) threshold of sensitivity of selection mode conversion is as λ ₁₂, λ ₂₃

3) determine from state (1) to state (2) transfer process the Frequency point set K that satisfies condition _F1

4) determine from state (2) to state (3) transfer process the Frequency point set K that satisfies condition _F1

5) determine set Kf _Eature, K _Feature=K _F1∩ K _F2

Because in noise and the vibration signal, except that containing the information that can characterize mill load, also contain background noise information, intrinsic mechanical noise information, adjacent mill and other interfere informations etc., in addition, grinding machine is under the different load operating mode, the feature power spectrum center of gravity of noise and vibration and all variations to some extent that distributes, therefore must be able to follow working conditions change carries out feature extraction to noise and vibration signal, and the characteristic information after the extraction just can enter into soft-sensing model and calculate.

Described noise and vibration signal self-adaptive features information extracting method, for each the Frequency point energy in the characteristic spectra is carried out adaptive weighted acquisition, adopt following formula to determine:

$E=\underset{k\&Element;K_{\mathrm{feature}}}{\mathrm{\&Sigma;}}\mathrm{\&mu;}\left(k\right)M{\left|X\left(k\right)\right|}^2=M\underset{k\&Element;K_{\mathrm{feature}}}{\mathrm{\&Sigma;}}\mathrm{\&mu;}\left(k\right){\left|X\left(k\right)\right|}^2$

Wherein μ (k) is adaptive weighted coefficient, adopts the double gauss type function to realize that formula is as follows

$\mathrm{\&mu;}\left(k\right)=\left\{\begin{array}{cc}\exp (-\frac{{(k-c_1)}^2}{{{2\mathrm{\&delta;}}_1}^2})& k<{\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="A20071001863800191.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\\ 1& {\mathrm{<figure-callout\; id="1"\; label="c"\; filenames="A20071001863800191.png"\; state="\{\{state\}\}">c</figure-callout>}}_1\&le;k\&le;{\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">c</figure-callout>}}_2\\ \exp (-\frac{{(k-c_2)}^2}{{{2\mathrm{\&delta;}}_2}^2})& {k>\mathrm{<figure-callout\; id="2"\; label="c"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">c</figure-callout>}}_2\end{array}\right.$

Parameter { c wherein ₁, δ ₁, c ₂, δ ₂Obtain by following formula:

$c_1:\underset{k_c}{\mathrm{<figure-callout\; id="1"\; label="max\; k\; c"\; filenames="A20071001863800191.png"\; state="\{\{state\}\}">max</figure-callout>}}\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_{c1}}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2{\&GreaterEqual;\mathrm{\&alpha;}}_1$

${\mathrm{\&delta;}}_1:\underset{\mathrm{\&Delta;<figure-callout\; id="1"\; label="k"\; filenames="A20071001863800191.png"\; state="\{\{state\}\}">k</figure-callout>}1}{\min }\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=c_1-{\mathrm{\&Delta;k}}_1}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2{\&GreaterEqual;\mathrm{\&alpha;}}_2$

$c_2:\underset{k_c}{\mathrm{<figure-callout\; id="2"\; label="min\; k\; c"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">min</figure-callout>}}\underset{\underset{k\&Element;K_{\mathrm{<figure-callout\; id="2"\; label="feature\; k\; c"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">feature</figure-callout>}}}{k=k_c}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2{\&GreaterEqual;\mathrm{\&alpha;}}_3$

${\mathrm{\&delta;}}_2:\underset{{\mathrm{\&Delta;<figure-callout\; id="2"\; label="k"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">k</figure-callout>}}_2}{\min }\underset{\underset{k\&Element;{\mathrm{<figure-callout\; id="2"\; label="K\; feature\; c"\; filenames="A20071001863800191.png,A20071001863800221.png"\; state="\{\{state\}\}">K</figure-callout>}}_{\mathrm{feature}}}{k=k_c}}{\overset{c_{}}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2{\&GreaterEqual;\mathrm{\&alpha;}}_4$

α ₁, α ₂, α ₃, α ₄, be threshold value, be chosen as 0.30,0.85,0.30 respectively, 0.85k _cBe the gravity frequency of feature power spectrum, by following formula decision

$k_c=\frac{f_c}{\mathrm{\&Delta;f}}=\frac{f_c}{f_s/N}=\frac{\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}k{\left|X\left(k\right)\right|}^2}{\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2}$

Described grinding machine noise, vibration and gateway differential pressure signal are carried out the data pre-service, the method for rejecting human error is the gross error detection method based on intermediate value, for an X of slip data queue _n, judge sample X _iThe criterion that whether is human error is:

|X _i-med(X _n)|≥δ _i*mmd(X _n)

δ wherein _tBe decision threshold, med (X _n) be X _nIntermediate value, mmd (X _n) be the intermediate value distance, calculate by following formula and obtain mmd (X _n)=med (med (d (X ₁)), med (d (X ₂)) ... med (d (X _n))), d (X _i) the expression sample Xi one-dimensional vector that the distance of other samples is formed in the formation.

Described rule-based adaptive weighted soft-sensing model, its regular citation form is:

ifλ _m is A _i1 and λ _s is B _i2 and W _GQ is C _i3 and L _GK is D _i4，

then L _m＝p _m F _ZY+q _mF _ZD+r _mC _YY+s _m

In the formula, i1, i2, i3, i4=1～3, m=1～81 have 81 rules.

Consisting of of soft-sensing model:

1) the operating mode identification class variable λ to importing _m, λ _s, W _GQ, L _GKCarry out obfuscation: get coal grindability factor λ _mLinguistic variable be { A ₁, A ₂, A ₃, expression grindability { good, general, poor }; Get coal moisture content λ _sLinguistic variable be { B ₁, B ₂, B ₃, expression coal water content { many, general, few }; The linguistic variable of getting steel ball loading capacity WGQ is { C ₁, C ₂, C ₃, expression steel ball loading capacity { on the high side, suitable, on the low side }; Get mill load L _GKLinguistic variable be { D ₁, D ₂, D ₃, expression mill load { full, suitable partially, empty partially }; Select Gauss's membership function that each input variable is carried out obfuscation;

$\left\{\begin{array}{c}{\mathrm{\&mu;}}_{A_{i1}}\left({\mathrm{\&lambda;}}_m\right)=\exp \left[\frac{-{({\mathrm{\&lambda;}}_m-d_{i1})}^2}{{{\mathrm{\&delta;}}^2}_{i1}}\right]\\ {\mathrm{\&mu;}}_{B_{i2}}\left({\mathrm{\&lambda;}}_s\right)=\exp \left[\frac{-{({\mathrm{\&lambda;}}_s-d_{i2})}^2}{{{\mathrm{\&delta;}}^2}_{i2}}\right]\\ {\mathrm{\&mu;}}_{C_{i3}}\left(W_{\mathrm{GQ}}\right)=\exp \left[\frac{-{(W_{\mathrm{GQ}}-d_{i3})}^2}{{{\mathrm{\&delta;}}^2}_{i3}}\right]\\ {\mathrm{\&mu;}}_{D_{I4}}\left(L_{\mathrm{GK}}\right)=\exp \left[\frac{-{(L_{\mathrm{GK}}-d_{i4})}^2}{{{\mathrm{\&delta;}}^2}_{i4}}\right]\end{array}\right.$

2) obtain the intensity of activation of sample to fuzzy rule:

${\mathrm{\&omega;}}_m={\mathrm{\&mu;}}_{A_{i1}}\left({\mathrm{\&lambda;}}_m\right)\&times;{\mathrm{\&mu;}}_{B_{i2}}\left({\mathrm{\&lambda;}}_s\right)\&times;{\mathrm{\&mu;}}_{C_{i3}}\left(W_{\mathrm{GQ}}\right)\&times;{\mathrm{\&mu;}}_{D_{i4}}\left(L_{\mathrm{GK}}\right)$

3) each bar rule intensity is carried out normalized:

4) output under every rule of calculating:

5) calculate finally output:

The parameter of the needs study in this soft-sensing model is former piece parameter set { d _I1, d _I2, d _I3, d _I4, δ _I1, δ _I2, δ _I3, δ _I4And consequent parameter set { p _m, q _m, r _m, s _m, adopt the hybrid algorithm training to obtain, by steepest gradient descent method training former piece parameter, by least square in training consequent parameter.

The present invention is by carrying out self-adaptive feature extraction to grinding machine noise, vibration and gateway differential pressure signal, obtain the information that can only characterize mill load, coal grindability factor, coal water content, steel ball loading capacity and load are blured division, foundation can the identification operating mode rule and each polynomial fitting under regular, the output to strictly all rules at last comprehensively obtains mill load.The coal pulverizer load soft-sensing method that the present invention proposes has overcome the shortcoming that detection method in the past can not adaptation condition changes, and all can obtain good accuracy and sensitivity under various operating modes, for the effective monitoring of grinding machine provides guarantee.

Description of drawings

Fig. 1 is the soft measurement hardware system of a mill load arrangement plan, label among the figure is represented respectively: 1, feeder controller, 2, cold-air flap valve position feedback device, 3, hot blast family of power and influence position feedback assembly, 4, reduced air family of power and influence position feedback assembly, 5, vibration transducer, 6, sensor noise, 7, the gateway differential pressure pick-up, 8, mill exhauster inlet baffle valve position feedback device, 9, data acquisition unit and computing machine, 10, soft-sensing model; The label that relates to pulverized coal preparation system equipment among the figure is represented respectively: 11, run coal bin, 12, feeder, 13, coal pulverizer, 14, mill separator, 15, pulverized-coal collector, 16, Pulverized Coal Bin, 17, mill exhauster.

Fig. 2 is the Survey Software FB(flow block);

Fig. 3 is that the sensitivity of grinding machine noise signal is with frequency distribution;

Fig. 4 is grinding machine noise gross error detection result;

Fig. 5 is grinding machine feature of noise power spectrum, gravity frequency and weighting coefficient; Wherein (a) is operating mode I, (b) is operating mode II, (c) is operating mode III;

Fig. 6 is the soft-sensing model figure that predicts the outcome;

The present invention is described in further detail below in conjunction with the drawings and specific embodiments.

Embodiment

With certain thermal power plant's the ball type pulverizer system is example, provides concrete an application of the present invention.This ball type pulverizer system is equipped with two coal pulverizers, and model is DTM350/700, and rotating speed is 17.57r/min, and design rating is 60t/h, adopts belt-type Gravimetric Coal Feeders coal supply.Its workflow is: feeder is sent into the raw coal in the run coal bin in the coal pulverizer, hot blast, cold wind, reduced air also enter coal pulverizer simultaneously, raw coal is through crushing grinding, the coal dust that grinds is transferred out by air-flow, what come out from coal pulverizer is primary mixture, and behind mill separator, thick excessively coal dust returns the coal pulverizer inlet again and grinds, qualified coal dust is brought into pulverized-coal collector and carries out the separation of gas powder, and qualified once more coal dust falls into Pulverized Coal Bin.According to the needs of boiler load, machine supplying powder is sent airduct of the input of the coal dust in the Pulverized Coal Bin into the stove internal combustion again.

Mill load measuring system structure has been equipped with following instrument as shown in Figure 1 in original system:

Feeder controller 1: in the control coal supply, (also i.e. equivalence is real-time powder process amount w to feed back to real-time coal supply flow _MF);

Hot blast family of power and influence position feedback assembly 3: hot blast door aperture indication μ _RF

Cold-air flap valve position feedback device 2: cold-air flap aperture indication μ _LF

Reduced air family of power and influence position feedback assembly 4: cold-air flap aperture indication μ _ZXF

Mill exhauster inlet baffle valve position feedback device 8: mill exhauster inlet baffle aperture μ _PRF

Gateway differential pressure pick-up 7: coal pulverizer gateway differential pressure signal P _CY

According to technical scheme of the present invention, increase following measurement instrument:

Sensor noise 6: be used to measure grinding machine operational process middle cylinder body noise, be installed on, point to steel ball whereabouts rum point apart from grinding machine 1/3 place that enters the mouth; Select the MPA206 microphone of popularity company, transducer sensitivity is 32mv/Pa, and response frequency is 20Hz～10kHz;

Vibration transducer 5: be used for measuring mill body Oscillation Amplitude, be installed on grinding machine inlet spring bearing place;

Select the 608A11 acceleration transducer of PCB company to detect vibratory output, transducer sensitivity 100mv/g, response frequency is 20Hz～10kHz;

Soft Survey Software operates on the independent computing machine, increases a high-speed data acquisition card PCI1714UL (grinding magnificent analog signal input card, 4 passages, maximum slew rate 30MHz, 12 AD) on this computing machine, gathers sensor noise and vibration sensor signal in real time; All the other signals obtain the correlated process data by the mode with the PLC communication from original system.To noise and vibration signals collecting cycle is 500ms, with the sampling rate of 51.2k 1024 point data of sampling, other process datas is upgraded with the cycle of 500ms at every turn.Soft Survey Software adopts the VC programming, by the mode and the PLC exchange process data of free mouthful communication.

The Survey Software FB(flow block) mainly is divided into two processes as shown in Figure 2, and one is the off-line calibration process, mainly finishes definite characteristic spectra, trains former piece and consequent parameter in inlet differential pressure feature extraction multinomial coefficient, the training soft-sensing model; Two is the on-line prediction process, the real-time online of realizing mill load calculates, and mainly comprises process data collection, the calculating of characteristic frequency spectrum gravity frequency, weighting coefficient calculating, each characteristics of variables information calculations, each regular normalization intensity of activation calculating, each rule module such as prediction output calculating and final prediction output calculating down.

In steel ball loading capacity computation process, according to this bowl mill ruuning situation, the unit powder process dimension ball loss of selecting mill load to be under basic, normal, high three kinds of situations is w1=240 gram/ton, w2=150 gram/ton, w3=110 gram/ton.

In gateway differential pressure characteristic information calculated, because this grinding machine is in operational process, cold-air flap and recycle air door kept the complete shut-down state, and the mill exhauster inlet baffle keeps full-gear, can be F with simplified formula therefore _CY=p ₀+ p ₁P _CY+ p ₂μ _RF, through the sample data training, it is as follows to obtain coefficient: p ₀=0, p ₁=1.52 * 10 ^-4, p ₂=0.14.

In definite process of characteristic spectra, be example with the grinding machine noise signal, we select empty mill state (1), normally move state (2), completely grind (3) three states of state and judge, select λ ₁₂=λ ₂₃=0.20, k _h=min{[10kHz/50Hz], [51.2kHz/2 * 50Hz] }=200, f _h=10kHz, sensitivity with the distribution of frequency as shown in Figure 3, as can be seen from the figure, the grinding machine noise is at sky mill and completely grind Duan Junyou sensitivity preferably, and determining to be arranged in the set that the Frequency point of top left region forms by the selected threshold of sensitivity is the characteristic spectra of grinding machine noise.For the grinding machine vibration signal, select λ ₁₂=0.05; λ ₂₃=0.20, k _h=min{[10kHz/50Hz], [51.2kHz/2 * 50Hz] }=200, f _h=10kHz.

In the process of gross error detection and rejecting, we select δ _t=1.3, length of data queue is n=15, is example with the grinding machine noise signal in the coal process, the result of gross error detection as shown in Figure 4, as seen because the grinding machine special objects has comprised a large amount of human errors in the collection result, will produce material impact to soft measurement.The gross error detection method that adopts the present invention to propose has been rejected the human error in the measurement data preferably, has guaranteed the accuracy that soft measurement predicts the outcome.

In the self-adaptive features information extraction process, with the grinding machine noise signal is example, be given in empty mill (operating mode I), normal operation (operating mode II) and full grind under (operating mode III) state feature power spectrum, gravity frequency and weighting coefficient curve as shown in Figure 5, its correlation parameter is as shown in the table; Therefrom as can be seen, grinding machine is under the different load conditions, and the center of gravity of feature power spectrum and distribution situation all can change, and along with increasing of mill load, gravity frequency moves to the low frequency direction.Through the characteristic information after adaptive weighted, can be good at adaptation condition and change, all can have the good sensitivity and the linearity in various operating modes.

Signal	Operating mode	k c	c 1	δ 1	c 2	δ 2
							Noise	I	80	62	31	102	41
II	73	58	38	86	56
						III		67	54	40	77	60

The soft-sensing model that adopts the present invention to propose at the various operating mode run durations of pulverized coal preparation system, can both self-adaptation be adjusted the calculating of characteristic information and be determined each rule intensity according to operating mode according to operating mode, thereby obtain output accurately.Fig. 6 is for experiencing the prediction output under step coal, disconnected coal and the linear coal operating mode continuously at grinding machine, as can be seen from the figure, the prediction of this soft-sensing model output can be good at reflecting grinding machine internal loading situation, and accuracy is better.

Claims (5)

1. heat engine plant canister type steel ball mill load flexible measurement method, it is characterized in that, this method is set up hardware platform in the mill load measuring system, this hardware platform comprises sensor noise, vibration transducer, gateway differential pressure pick-up, hot blast family of power and influence position indicating device, cold-air flap valve position indicating device, reduced air family of power and influence position indicating device, mill exhauster inlet baffle valve position indicating device, Gravimetric Coal Feeders controller, data acquisition unit and computing machine, carries out soft measurement by operation Survey Software on computers; Survey Software and computing machine communication obtain real-time process data and human-machine interaction data by the data collecting card collection, specifically may further comprise the steps:

1) soft-sensing model input auxiliary variable is selected: a class is the load characteristic info class, is used for match mill load information, comprises the characteristic information F that extracts from noise signal _ZY, the characteristic information F that from vibration signal, extracts _ZDWith the characteristic information F that from the differential pressure signal of gateway, extracts _CYAnother kind of is operating mode identification class, is used to form the rule of identification operating mode, comprises coal grindability factor λ _m, coal moisture content λ _s, steel ball loading capacity W _GQ, mill load L _GK

2) acquisition methods of input auxiliary variable

Noise and vibration signal characteristics information: the noise and the vibration data sequence of gathering are carried out power spectrumanalysis, obtain characteristic spectra information, the energy in the characteristic spectra is carried out adaptive weighted acquisition characteristic information;

Gateway differential pressure characteristic information obtains: to gateway differential pressure P _CY, hot blast door aperture μ _RF, cold-air flap aperture μ _LF, recycle air door aperture μ _ZXF, mill exhauster inlet baffle aperture μ _PRFThe influence of throttle opening changing factor has been eliminated in five information match acquisitions as follows in the gateway differential pressure characteristic information of acquisition, its variation only reflects that the situation of change of grinding machine internal loading is as follows:

F _CY＝p ₀+p ₁P _CY+p ₂μ _RF+p ₃μ _LF+p ₄μ _ZXF+p ₅μ _PRF

Wherein, F _CYFor characterizing the characteristic information that the grinding machine internal loading changes, parameter { p ₀, p ₁, p ₂, p ₃, p ₄, p ₅Obtain by the least square training, training sample obtains in the following way:

Adopt control coal supply system automatically, the coal position is remained on the different levels, on every kind of coal position level, in the practical operation scope, change each throttle opening respectively, write down corresponding data, form training sample data collection;

Coal grindability factor λ _mWith coal moisture content λ _s: the human-machine interaction data by operating personnel's typing obtains, and by thermal power plant's pulverized coal preparation system working specification, will regularly chemically examine these parameters of raw coal;

Steel ball loading capacity: by first loading amount W ₀, magnitude of recruitment W _B(k) and wear extent W _SHCalculate acquisition, wherein magnitude of recruitment W _B(k) result who adds up for each steel ball magnitude of recruitment, wear extent are according to unit powder process dimension ball loss w under the different load situation _FHTo powder process amount w _MFThe result of accumulative total; The steel ball loading capacity computing formula is as follows:

$W_{\mathrm{GQ}}=[W_0+\underset{k}{\mathrm{\&Sigma;}}{W}_B\left(k\right)-W_{\mathrm{SH}}]=[W_0+\underset{k}{\mathrm{\&Sigma;}}{W}_B\left(k\right)-\mathrm{\&Sigma;}{w}_{\mathrm{FH}}\left(L_{\mathrm{GK}}\right)w_{\mathrm{MF}}]$

In the formula: first loading amount W ₀, magnitude of recruitment W _B(k), unit powder process dimension ball loss w under the different operating modes _FHObtain powder process amount w by human-machine interaction data _MFObtain by the Gravimetric Coal Feeders data on flows;

Mill load L _GKObtain by last periodic model output result;

3) noise, vibration, the employing of gateway differential pressure characteristic information are carried out pre-service based on the gross error detection method of intermediate value distance, reject human error influence in the real-time measuring data;

4) set up rule-based adaptive weighted soft-sensing model, mill load is calculated.

2. the method for claim 1 is characterized in that described noise and vibration signal characteristics frequency range are not a certain continuous frequency range, but the set K that forms by a series of Frequency points that meet the following conditions _Fealure:

$\left|\frac{{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(i\right)}}^2-{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(j\right)}}^2}{\max ({{\left|X\left(k\right)\right|}_{\mathrm{state}\left(i\right)}}^2,{{\left|X\left(k\right)\right|}_{\mathrm{state}\left(j\right)}}^2)}\right|\&GreaterEqual;{\mathrm{\&lambda;}}_{\mathrm{ij}},$ k＝0～k _h

k _h＝min{[f _h/Δf]，[f _s/2Δf]}

The frequency domain data sequence of X (k) wherein for data time series x (n) being carried out obtain after FFT calculates, skin || for taking absolute value computing, internal layer || for asking modular arithmetic, [] is rounding operation; λ _IjThe threshold of sensitivity for state state (i) and state (j) conversion; f _hBe the highest frequency of actual signal, determine by sensor hardware; f _sBe sample frequency; Δ f is a signal resolution, and is relevant with FFT computing points N.

3. the method for claim 1 is characterized in that described noise and vibration signal self-adaptive features information extracting method, for each the Frequency point energy in the characteristic spectra is carried out adaptive weighted acquisition, adopts following formula to determine:

$E=\underset{k\&Element;K_{\mathrm{feature}}}{\mathrm{\&Sigma;}}\mathrm{\&mu;}\left(k\right)M{\left|X\left(k\right)\right|}^2=M\underset{k\&Element;K_{\mathrm{feature}}}{\mathrm{\&Sigma;}}\mathrm{\&mu;}\left(k\right){\left|X\left(k\right)\right|}^2$

Wherein μ (k) is adaptive weighted coefficient, adopts the double gauss type function to realize that formula is as follows

$\mathrm{\&mu;}\left(k\right)=\left\{\begin{array}{cc}\exp (-\frac{{(k-c_1)}^2}{2{{\mathrm{\&delta;}}_1}^2})& k<c_1\\ 1& c_1\&le;k\&le;c_2\\ \exp (-\frac{{(k-c_2)}^2}{2{{\mathrm{\&delta;}}_2}^2})& k>c_2\end{array}\right.$

Parameter { c wherein ₁, δ ₁, c ₂, δ ₂Obtain by following formula:

$c_1:\underset{k_{c1}}{\max }\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_{c1}}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2\&GreaterEqual;{\mathrm{\&alpha;}}_1$

${\mathrm{\&delta;}}_1:\underset{\mathrm{\&Delta;k}1}{\min }\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=c_1-\mathrm{\&Delta;}{k}_1}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{k_c}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2\&GreaterEqual;{\mathrm{\&alpha;}}_2$

$c_2:\underset{k_{c2}}{\min }\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{k_{c2}}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2\&GreaterEqual;{\mathrm{\&alpha;}}_3$

${\mathrm{\&delta;}}_2:\underset{\mathrm{\&Delta;}{k}_2}{\min }\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{c_2+\mathrm{\&Delta;}{k}_2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2/\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=k_c}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2\&GreaterEqual;{\mathrm{\&alpha;}}_4$

α ₁, α ₂, α ₃, α ₄, be threshold value, be chosen as 0.30,0.85,0.30,0.85 respectively;

k _cBe the gravity frequency of feature power spectrum, determine by following formula:

$k_c=\frac{f_c}{\mathrm{\&Delta;f}}=\frac{f_c}{f_s/N}=\frac{\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}k{\left|X\left(k\right)\right|}^2}{\underset{\underset{k\&Element;K_{\mathrm{feature}}}{k=0}}{\overset{(N-1)/2}{\mathrm{\&Sigma;}}}{\left|X\left(k\right)\right|}^2}.$

4. the method for claim 1 to the data preprocessing method that noise, vibration and gateway differential pressure characteristic information are taked, is characterized in that described gross error detection method based on intermediate value is for an X of slip data queue _n, judge sample X _iThe criterion that whether is human error is:

|X _i-med(X _n)|≥δ _i*mmd(X _n)

δ wherein _iBe decision threshold, med (X _n) be sample set X _nIntermediate value, mmd (X _n) be the intermediate value distance, calculate by following formula and obtain d (X _i) expression sample X _iThe one-dimensional vector that the distance of other samples is formed in the formation.

mmd(X _n)＝med(med(d(X ₁))，med(d(X ₂))，…med(d(X _n)))

5. the method for claim 1 is characterized in that described rule-based adaptive weighted soft-sensing model, and its regular citation form is:

if λ _m is A _i1 and λ _s is B _i2 and W _GQ is C _i3 and L _GK is D _i4，

then L _m＝p _m F _ZY+q _m F _ZD+r _m F _CY+s _m

In the formula, i1, i2, i3, i4=1～3, m=1～81 have 81 rules.

Consisting of of this soft-sensing model:

1) the operating mode identification class variable λ to importing _m, λ _s, W _GQ, L _GKCarry out obfuscation: get coal grindability factor λ _mLinguistic variable be { A ₁, A ₂, A ₃, expression grindability { good, general, poor }; Get coal moisture content λ _sLinguistic variable be { B ₁, B ₂, B ₃, expression coal water content { many, general, few }; Get steel ball loading capacity W _GQLinguistic variable be { C ₁, C ₂, C ₃, expression steel ball loading capacity { on the high side, suitable, on the low side }; Get mill load L _GKLinguistic variable be { D ₁, D ₂, D ₃, expression mill load { full, suitable partially, empty partially }; Select Gauss's membership function that each input variable is carried out obfuscation;

$\left\{\begin{array}{c}{\mathrm{\&mu;}}_{A_{i1}}\left({\mathrm{\&lambda;}}_m\right)=\exp \left[\frac{-{({\mathrm{\&lambda;}}_m-d_{i1})}^2}{{{\mathrm{\&delta;}}^2}_{i1}}\right]\\ {\mathrm{\&mu;}}_{B_{i2}}\left({\mathrm{\&lambda;}}_s\right)=\exp \left[\frac{-{({\mathrm{\&lambda;}}_s-d_{i2})}^2}{{{\mathrm{\&delta;}}^2}_{i2}}\right]\\ {\mathrm{\&mu;}}_{C_{i3}}=\left(W_{\mathrm{CQ}}\right)=\exp \left[\frac{-{(W_{\mathrm{GQ}}-d_{i3})}^2}{{{\mathrm{\&delta;}}^2}_{i3}}\right]\\ {\mathrm{\&mu;}}_{D_{i4}}\left(L_{\mathrm{GK}}\right)=\exp \left[\frac{-{(L_{\mathrm{GK}}-d_{i4})}^2}{{{\mathrm{\&delta;}}^2}_{i4}}\right]\end{array}\right.$

2) obtain the intensity of activation of sample to fuzzy rule:

${\mathrm{\&omega;}}_m={\mathrm{\&mu;}}_{A_{i1}}\left({\mathrm{\&lambda;}}_m\right)\&times;{\mathrm{\&mu;}}_{B_{i2}}\left({\mathrm{\&lambda;}}_s\right)\&times;{\mathrm{\&mu;}}_{C_{i3}}\left(W_{\mathrm{GQ}}\right)\&times;{\mathrm{\&mu;}}_{D_{i4}}\left(L_{\mathrm{GK}}\right)$

3) each bar rule intensity is carried out normalized:

4) output under every rule of calculating:

5) calculate finally output:

The parameter of the needs study in this soft-sensing model is former piece parameter set { d _I1, d _I2, d _I3, d _I4, δ _I1, δ _I2, δ _I3, δ _I4And consequent parameter set { p _m, q _m, r _m, s _m, adopt the hybrid algorithm training to obtain: by steepest gradient descent method training former piece parameter, by least square in training consequent parameter.

Images