CN104535646B - Method for detecting imperfection of food grains - Google Patents

Abstract

The invention relates to a method for detecting the imperfection of food grains. The method comprises the following steps: acquiring collision sound information of each food grain and a specific object, extracting set characteristic information from the collision sound information, and comparing and judging imperfect grains in the food grains according to the characteristic information. The method disclosed by the invention has the following advantages that an impact excitation sound signal of the food grains and a metal plate in the free falling body process is utilized, and the imperfect grains are detected by virtue of extraction and analysis of the characteristic parameters of the signal, so that the detection method has the non-contact characteristic. In the whole detection process, the physical structures of the food grains do not need to be damaged, so that the detection method has the nondestructive testing characteristic; and moreover, according to the detection method, special requirements and limitations on the types and falling postures of the food grains are avoided, and the universality is high.

Description

A kind of grain seed imperfection detection method

Technical field

The present invention relates to a kind of grain seed imperfection detection method, belong to grain quality online test method.

Background technology

Grain quality is to weigh the overall target of grain quality.It is diversified to influence the factor of grain quality, and its Middle imperfection seed too high levels are the major reasons for restricting grain quality.According to standard GB/T 1351-2008, by grain seed The embryo or endosperm of grain are subject to physical damnification or microorganism to encroach on, but the grain seed of also use value is defined as unsound grain. Unsound grain is mainly including broken kernel, injured kernel, sprouted kernel, mouldy grain etc..Due to the presence of grain unsound grain, grain is reduced The kind quality and edible quality of food, while the technological quality and edible quality (such as mouthfeel, smell) of grain are also reduced, separately On the one hand, unsound grain is highly prone to the infringement of insect and mould, is difficult safe storage, so to the accurate of grain unsound grain Inspection is particularly important.

Detection to grain unsound grain at present is mainly according to standard GB/T/T5494-2008 by artificial to sampling Seed is counted, but in the general treatment of grain and transportation, unsound grain is to mix with normal grain , and both ratios are very small, therefore in order to objective and accurate evaluation grain quality, it is necessary to the grain seed of sampling Quantity is a lot.But, up to the present, also without effective unsound grain detection method, it is cursory that commonly used approach includes Method, x-ray method, worm's ovum are creeped and the sound-detection method, carbon dioxide detection method and Immunological Method chewed, but these methods It is different degrees of to exist that detection speed is slow, workload is big, expensive equipment, be only capable of the defects such as quantitative determination.Due to unsound grain Its internal physics or architectural characteristic change compared with normal grain, therefore the present invention proposes to be produced by detecting that seed is collided Raw acoustic signal detects the imperfection of grain seed.

The content of the invention

It is an object of the invention to provide a kind of grain seed imperfection detection method, it is used to solve traditional to grain seed The detection method of grain imperfection has that detection speed is slow, workload big, expensive equipment, be only capable of asking for the defects such as quantitative determination Topic.

To achieve the above object, the solution of the present invention includes a kind of grain seed imperfection detection method, including following Step：The collision acoustic information of each grain seed and specific thing is obtained, the characteristic information of setting is extracted from collision acoustic information, Compare, judge the imperfection of grain seed according to characteristic information.

The detection method is comprised the following steps：

1), for same grain variety, S normal grain and T unsound grain are chosen respectively, by obtaining each seed respectively The time domain and/or frequency domain information of grain, extract at least one characteristic parameter；

2), using the characteristic parameter of each seed as three layers of input of BP neural network, using described S normal grain and T Grain unsound grain is trained to the neutral net, to realize that the imperfection of grain seed is detected and recognized.

Following parameter is obtained from the time domain and frequency domain information of each seed：Collide acoustic information amplitude maximum and its Corresponding maximum sampled point；Three parameters of Weibull functions：α、β、x₀；The 8 first normalization sides of window function in short-term DifferenceWith5 second peak signals in short-term in window function Amplitude M1, M2, M3, M4 and M5；Collide the maximum amplitude and its corresponding frequency of the discrete Fourier transform frequency spectrum of acoustic information Rate point；Spectral magnitude before and after frequency spectrum maximum amplitude corresponding to each 15 Frequency points；The maximum of first-order difference spectral function and its Corresponding Frequency point；At least one characteristic parameter is chosen from parameter.

Weibull functional forms are：

Also,

x₀=4 (M_1,0,3M_1,0,0-M_1,0,1 ²)/(4M_1,0,3+M_1,0,0-4M_1,0,1),

In formula, Г be Г functions, α (>0) it is form parameter, β (>0) it is scale parameter, x₀It is the starting of Weibull functions Point, x is Weibull argument of functions, x>x₀, Y is the corresponding amplitudes of sampled point x, and N is sampling number.

The width of the described first window function in short-term is 50 points, calculates the described first variances sigma in short-term in window function_i ²：

Wherein,

According toObtain 8 normalization variance yields With

5 width of selection are 10 points of the second window function, since the maximum sampled point, described 5 are calculated respectively Maximum amplitude in second window function, respectively M1, M2, M3, M4 and M5.

It is calculated as follows the first-order difference spectrum of frequency spectrum function：

F ' (u)=F (u)-F (u-1), F (u) are frequency spectrum function, F ' (u) first-order differences spectrum；Obtain first-order difference spectrum most Big value and its corresponding Frequency point.

At least one is chosen from the parameter using methods such as Stepwise Discriminatory Analysis, multiple linear regression and principal component analysis Individual characteristic parameter.

The invention has the advantages that：

(1) using voice signal is excited with the shock of metallic plate during grain seed freely falling body, by the signal The extraction of characteristic parameter and analysis detection imperfection, therefore this detection method has non-contacting feature.

(2) in whole detection process, the physical arrangement without destroying grain seed, therefore this detection method have it is lossless The characteristics of detection.

(3) this detection method does not have special requirement and limitation, highly versatile to the species and whereabouts attitude of grain seed.

(4) in detection process, produced without using any chemicals, "dead" material, collection and the place of data Reason computer is automatically performed, thus the method have it is pollution-free, excellent without manual intervention, the high, real-time online of detection efficiency etc. Point.

(5) the grain seed in the present invention can be the main agriculture such as paddy rice, brown rice, rice, wheat, corn and soybean, peanut Any one in product, therefore this detection method has wide applicability.

Brief description of the drawings

Fig. 1 is detecting system structure chart；

Fig. 2 is window function nonlinear filtering schematic diagram；

Fig. 3 is that window function variance calculates schematic diagram in short-term；

Fig. 4 is that short time-window function maxima calculates schematic diagram；

Fig. 5 is grain seed imperfection overhaul flow chart.

Specific embodiment

The present invention will be further described in detail below in conjunction with the accompanying drawings.

A kind of grain seed imperfection detection method, comprises the following steps：Obtain each grain seed and specific thing Collision acoustic information, the characteristic information of setting is extracted from collision acoustic information, is compared according to characteristic information, is judged grain seed Imperfection.

Based on above technical scheme, with reference to accompanying drawing, be given with next specific embodiment.

Implement detection method of the invention, using a grain seed detecting system based on acoustics, including light-emitting diodes Pipe 1, photodiode 2, computer 3, wave filter 4, amplifier 5, microphone 6 and 7, metallic plate 8, vibratory sieve 9 and charging aperture 10. Grain seed freely falling body clashes into metallic plate and produces voice signal, the signal that the input of amplifier is connected into after being detected through microphone End, the signal after amplification is connected into computer through high-pass filtering, and the collection signal of its Computer is triggered by photodiode.System System structure is as shown in Figure 1.

By function, the system can be divided into three parts：(1) signal is produced.The part is by vibratory sieve, charging aperture, metallic plate structure Into its function is mainly the generation acoustic signal relevant with grain seed sophistication；(2) signal acquisition.The part is main by two Individual microphone, amplifier, wave filter are constituted, and its function is mainly collects the voice signal that grain seed is produced with metallic plate collision And preliminary treatment is carried out to the signal；(3) signal transacting.The part is main to be completed by computer, and its major function is by carrying The time-frequency and frequency domain character parameter of voice signal are taken, by judgement and then detection and the imperfection for identifying grain seed.

System work process is as follows：

Grain seed is with charging aperture center by grain free-falling, Light-Emitting Diode and photodiode from charging aperture Benchmark, symmetry axis is placed, and makes the light-emitting area of Light-Emitting Diode horizontally aligned with the light receiving surface of photodiode. When there is no seed to pass through, because light path is unobstructed, so photodiode output stabilization photosignal, and work as seed via light path When, the effect of blocking due to seed to light, the signal for receiving photodiode diminishes, and exports pulse, and the pulse is defeated Enter to computer sound card as sound signal collecting starting point trigger signal.

When on grain seed freely falling body to metallic plate, will be collided with metallic plate, and then send voice signal.

Here metallic plate is one piece of stainless steel plate of polishing, and its size is 7.5 (length) × 5 (width) × 10 (thickness) cm.Compare For metallic plate, the quality of grain seed can be ignored, thus in knockout process metallic plate vibration relative to seed The voice signal that collision sends is very small, is ignored.In order to keep the uniformity of seed and metallic plate impingement position, enter The distance of material mouth to metallic plate is 30-50cm, and in order to prevent seed from producing secondary impact, metal plate level on a metal plate Incline 30 degree.

By two microphones while, be amplified for the signal by amplifier, then by wave filter by collected sound signal Filter low-frequency noise, and filtered signal is delivered into computer and processed and analyzed.

A kind of grain variety is selected, S normal grain and T unsound grain are chosen respectively, the value of S, T is 800- here Between 1000, C main characteristic parameters corresponding to each seed are obtained using following steps.

The randomness of attitude, causes sound when otherness and seed in view of grain seed kind are clashed into metallic plate The intensity difference of signal is larger, therefore voice signal is gathered simultaneously by two microphones with preposition amplification, wherein microphone 6 Multiplication factor be 1V/Pa, the multiplication factor of microphone 7 is 10V/Pa, two installation directions of microphone and rum point normal Direction is parallel, and apart from metal sheet surface 20-25mm, and two voice response frequencies of microphone are in the range of 0-100KHz. The output signal of microphone is further amplified by amplifier, then by high-pass filter filter below 60KHz low-frequency noises with And dc noise, filtered signal finally is delivered into computer sound card carries out A/D conversions, and the sample frequency of sound card is 200KHz, resolution ratio is 16.The trigger signal of sound card is provided by the output pulse of photodiode, once receive triggering letter Number, N number of data point (N ∈ 2000-3000) is gathered in the data signal that computer is exported from sound card, the number of data point is according to grain Food species is suitably chosen.

The selecting step of voice signal：

1. the voice signal for two microphones being collected is designated as f1 (x) and f2 (x), wherein x correspondence sampled points, note respectively For x1, x2 ..., xN.The absolute value of the number of winning the confidence f1 (x) and f2 (x), is designated as | f1 (x) |, | f2 (x) |；

2. the quantity of | f1 (x) | and the middle saturation points of | f2 (x) | is calculated respectively；

If 3. saturation point in signal | f2 (x) | (it is the signal that collects of microphone of 10V/Pa to correspond to multiplication factor) Number is less than or equal to 6, then | f1 (x) | otherwise, (is corresponded to multiplication factor by the voice signal for the signal being collided as seed For the signal that the microphone of 1V/Pa is collected) collide voice signal as pending seed.Selected voice signal is remembered It is f (x), f (x) is one of f1 (x), f2 (x).The signal after absolute value is asked f (x) to be designated as g (x), g (x) is | f1 (x) |, | f2 (x) | one of them.

Calculate voice signal Weibull parameters：

1. the maximum amplitude g of signal g (x) is searched for_max(x) and its corresponding sampled point x_max；

2. nonlinear filtering is carried out to signal g (x), filtering can be described as：Using the window function that width is 7 points With signal convolution, as shown in Fig. 2 maximum in window function replaces its central value, wherein 3 points of signal starting (correspondence x1, x2, X3) keep constant at 3 points with the signal value of neighbouring (corresponding xN-2, xN-1, xN) for terminating.

I.e.：J=i-3, i-2, i-1, i, i+1, i+2, i+3, i ∈ [4, N-3]

3. time domain fitting is carried out using three parameter Weibull function pairs signals.Three parameter Weibull functional forms are：

In formula, α (>0) it is form parameter, β (>0) it is scale parameter, x₀It is the starting point of Weibull functions, x is Weibull argument of functions x>x₀, Y is the corresponding amplitudes of sampled point x.It is the time-domain signal of N, each parameter for sampling number Be calculated as follows：

x₀=4 (M_1,0,3M_1,0,0-M_1,0,1 ²)/(4M_1,0,3+M_1,0,0-4M_1,0,1)

α=ln2/ln [(M_1,0,0-2M_1,0,1)/2(M_1,0,1-2M_1,0,3)]

Here, Г is Г functions, can obtain characterizing three parameters of Weibull functions by above formula：α, β and x₀。

Window function variance treatment in short-term：

Choose a width and be 50 points of window function, as shown in figure 3, calculating the variances sigma of signal in the window_i ²：

Here x_jIt is the signal amplitude of each sampled point in window, x is the average of all signals in window.Wherein first window letter Several window widths is 50 sampled points, the sampled point x corresponding to window initial point from voice signal maximum amplitude_maxPreceding 40 points Start, then on the basis of first initial point of window function, be incremented by with 30 point step sizes and obtain second window function, and calculate the window The variance of intraoral signal.Therefore 20 sampled data points are overlapped between two neighboring window, by that analogy, 8 windows is taken altogether, respectively Calculate in each window the variance of signal and 8 variances of window and, the normalization side of each window is then obtained using following formula Difference is：

It is hereby achieved that 8 normalization variance yieldsWith

Calculate short time-window function maxima：

Sampled point x corresponding to maximum from time-domain signal_maxStart, one width of selection is 10 points of window function, meter The maximum amplitude in window function is calculated, is then step-length moving window with 10 sampled points, calculate the maximum of signal in next window Amplitude, the sampled point in two neighboring window function is not overlapped, according to this maximum amplitude in 5 window functions of method Continuous plus, The maximum signal amplitude value of each window function is denoted as M1, M2, M3, M4 and M5 respectively.

Discrete Fourier transform (DFT) treatment of voice signal：

Sampled point x corresponding to maximum from time-domain signal f (x)_maxQ preceding sampled point starts to carry out signal Hamming function adding windows, adding window width should be able to cover the time-continuing process that grain seed collides metallic plate, and the value of Q is in 50- here Between 100, suitably chosen according to grain variety.Then 256 Fourier transformations are carried out to the function after adding window, signal is obtained Frequency spectrum function F (u), wherein u are frequency values, and u ∈ [0,100kHz] calculate the maximum F of frequency spectrum function_maxIt is (u) and its corresponding Frequency point u_max.Meanwhile, record u_maxThe corresponding spectral magnitude of front and rear each 15 Frequency points, 30 range values, are designated as F altogether_j (u), j=1,2 ... ..., 30.

Calculate the Difference Spectrum of signal discrete Fourier transformation (DFT)：

The first-order difference for being calculated as follows frequency spectrum function F (u) composes F ' (u):

F ' (u)=F (u)-F (u-1)

Search the maximum F ' of plain difference spectral function_max(u) and its corresponding Frequency point u_cfmax。

It should be noted that the acquisition to parameter can individually carry out time domain or frequency domain treatment, one is each obtained A little parameters, and selected, it is of course also possible to time domain and frequency domain are all processed, more parameters can be obtained, and selected Select.

To sum up, processed according to above-mentioned time domain and frequency domain, obtain 52 time-frequency characteristics parameters, including：Signal amplitude maximum g_max(x) and its corresponding sampled point x_max；Three parameters of Weibull functions：α、β、x₀；8 normalizings of window function in short-term Change variance yieldsWith5 peak signals in short-term in window function Amplitude M1, M2, M3, M4 and M5；The maximum amplitude F of discrete Fourier transform (DFT) frequency spectrum of signal_maxIt is (u) and its corresponding Frequency point u_max；Spectral magnitude F before and after frequency spectrum maximum amplitude corresponding to each 15 Frequency points_j(u), j=1,2 ... ..., 30；One The maximum F ' of order difference spectral function_max(u) and its corresponding Frequency point u_cfmax.Use Stepwise Discriminatory Analysis, multiple linear regression C (C >=1) individual main characteristic parameters are chosen from features above parameter with the method such as principal component analysis.

Using S of above-mentioned selection normal grain and T C characteristic parameter of each seed of unsound grain as three layers of BP god Through the input of network, the neutral net includes 1 input layer, 1 hidden layer and 1 output layer.It is normal using S of the selection Grain and T unsound grain are trained to the neutral net, obtain the detection model of grain seed imperfection.

So far, detection model has been completed.

When needing any seed to same grain variety to be detected and recognized, by these any seeds using above-mentioned The detection model of completion is detected and recognized, you can detection grain seed.Testing process is as shown in Figure 5.

Specific embodiment is presented above, but the present invention is not limited to described implementation method.Base of the invention This thinking is above-mentioned basic scheme, and for those of ordinary skill in the art, various changes are designed in teaching of the invention The model of shape, formula, parameter simultaneously need not spend creative work.It is right without departing from the principles and spirit of the present invention Change, modification, replacement and the modification that implementation method is carried out are still fallen within protection scope of the present invention.

Claims (8)

1. a kind of grain seed imperfection detection method, it is characterised in that the detection method is comprised the following steps：Obtain every Individual grain seed and the collision acoustic information of specific thing, the characteristic information of setting are extracted from the collision acoustic information, according to institute State the imperfection that characteristic information compares, judges the grain seed；

When the collision acoustic information is obtained, the microphone with two with different amplification gathers collision voice signal, The voice signal of the wherein less microphone collection of multiplication factor is designated as f1 (x), the sound of the larger microphone collection of multiplication factor Message number is designated as f2 (x), and x is sampled point, be designated as x1, x2 ..., xN, N is sampling number, and the number of winning the confidence f1 (x) and f2's (x) is exhausted To value, | f1 (x) |, | f2 (x) | are designated as；The quantity of | f1 (x) | and the middle saturation points of | f2 (x) | is calculated respectively；If signal | f2 (x) | middle saturation points are less than or equal to 6, the then voice signal that | f2 (x) | is collided as seed using signal；If signal | f2 (x) | middle saturation points are more than 6, then | f1 (x) | is collided into voice signal as seed.

2. grain seed imperfection detection method according to claim 1, it is characterised in that the detection method includes Following steps：

1), for same grain variety, S normal grain and T unsound grain are chosen respectively, by obtaining each seed respectively Time domain and/or frequency domain information, extract at least one characteristic parameter；

2), using the characteristic parameter of each seed as three layers of input of BP neural network, using described S normal grain and T not Improve grain to be trained the neutral net, to realize that the imperfection of grain seed is detected and recognized.

3. grain seed imperfection detection method according to claim 2, it is characterised in that from the time domain of each seed With following parameter is obtained in frequency domain information：Collide the amplitude maximum and its corresponding maximum sampled point of acoustic information； Three parameters of Weibull functions：α、β、x₀；The 8 first normalization variance yields of window function in short-term With5 second maximum signal amplitude value M1 in short-term in window function, M2, M3, M4 and M5；Collide the maximum amplitude and its corresponding Frequency point of the discrete Fourier transform frequency spectrum of acoustic information；Frequently Spectral magnitude before and after spectrum maximum amplitude corresponding to each 15 Frequency points；The maximum of first-order difference spectral function and its corresponding frequency Rate point；At least one characteristic parameter is chosen from parameter.

4. grain seed imperfection detection method according to claim 3, it is characterised in that the Weibull functions Form is：

Also,

$M_{1,0,1}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i(1-\frac{i-0.35}{N}),$

$M_{1,0,3}=\frac{1}{N}\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i{(1-\frac{i-0.35}{N})}^3,$

x₀=4 (M_1,0,3M_1,0,0-M_1,0,1 ²)/(4M_1,0,3+M_1,0,0-4M_1,0,1),

$\mathrm{\β}=(M_{1,0,0}-x_0)/\mathrm{\Γ}\[ln\left(\frac{M_{1,0,0}-2M_{1,0,1}}{M_{1,0,1}-2M_{1,0,3}}\right)/ln2\],$

In formula, Г is Г functions, and α (＞ 0) is form parameter, and β (＞ 0) is scale parameter, x₀It is the starting point of Weibull functions, X is Weibull argument of functions, x ＞ x₀, Y is the corresponding amplitudes of sampled point x, and N is sampling number.

5. grain seed imperfection detection method according to claim 4, it is characterised in that the first short time-window letter Several width is 50 points, calculates the described first variances sigma in short-term in window function_i ²：

${\mathrm{\σ}}_i^2=\frac{1}{49}\underset{j=1}{\overset{50}{\mathrm{\Σ}}}{(x_j-\stackrel{\‾}{x})}^2,$

Wherein,

According toI=1,2 ... ..., 8, obtain 8 normalization variance yields With

6. grain seed imperfection detection method according to claim 5, it is characterised in that it is 10 to choose 5 width Second window function of point, since the maximum sampled point, calculates the maximum amplitude in 5 second window functions respectively, Respectively M1, M2, M3, M4 and M5.

7. grain seed imperfection detection method according to claim 6, it is characterised in that be calculated as follows frequency spectrum letter Several first-order difference spectrums：

F ' (u)=F (u)-F (u-1), F (u) are frequency spectrum function, F ' (u) first-order differences spectrum；Obtain the maximum of first-order difference spectrum And its corresponding Frequency point.

8. grain seed imperfection detection method according to claim 7, it is characterised in that use successive Discrimination point The methods such as analysis, multiple linear regression and principal component analysis choose at least one characteristic parameter from the parameter.