CN105301099A - Food crispness detection method - Google Patents

Abstract

The invention discloses a food crispness detection method in order to solve the problems that in the prior art, manual interference is serious, time and labor are wasted and detection equipment is expensive. The method includes the steps that firstly, whether a food crispness detection model is built or not is judged, if not, the fourth step namely food crispness actual measurement starts, and if yes, the second step namely building of the food crispness detection model starts; secondly, the step of building the food crispness detection model includes the sub-steps that firstly, food crispness mechanical characteristics are obtained, secondly, food breaking sound signals are collected, thirdly, the food breaking sound signals are denoised, fourthly, a time domain eigenvalue and a frequency domain eigenvalue of the sound signals are extracted, and fifthly, the food crispness detection model is built; thirdly, whether the food crispness detection model is already built or not is judged, if no, the food crispness detection model continues to be built, and if yes, the fourth step namely food crispness actual measurement starts for use; fourthly, food crispness actual measurement is conducted, wherein the extracted time domain eigenvalue is substituted into the built food crispness detection model, a food crispness value is calculated, and a food crispness measurement result is output.

Description

Food brittleness detection method

Technical field

The present invention relates to a kind of detection method belonging to food technology field, more precisely, the present invention relates to food brittleness detection method.

Background technology

Food brittleness evaluates the important quality side of food aesthetic quality, and brittleness also reflects freshness and the degree of ripeness situation of food to a certain extent.

At present, the assay method of food brittleness mainly contains sense organ mensuration, dynamics measurement and acoustics determination three kinds of methods.Sense organ measure waste time and energy, and be easily subject to estimator's hobby, health, term understanding, cultural difference impact cause error; Dynamics measurement equipment costly and have certain difference with human sensory; Acoustic detection can well overcome the defect of above-mentioned two kinds of methods, and during subjective appreciation food brittleness, the sound of the fracture of food is often as an important indicator of subjective appreciation, and this also absolutely proves that the fragility of the voice signal that food ruptures and fruits and vegetables has good correlativity.The researchists such as Sanz (Sang Si) choose 10 28-60 year consumer and 11 evaluate group members, carry out subjective appreciation to 6 kinds of fragility fruits and vegetables, research shows, sound during fracture has very large correlativity to food brittleness.Drake (Drake) researcher finds, in the time-domain diagram of the sound sent when stinging toast bread, the peak value of voice signal increases along with the increase of bread baking time; The roasting time is longer, and toast bread is more crisp.The researchists such as Belie (Bu Lai) utilize Fourier transform to analyze apple and chew the amplitude of sound frequency domain figure, energy and frequency, analyze the brittleness of apple.The Acoustic detection of current domestic food fragility is still in conceptual phase, and good acoustic detection method can quantitatively not detect fruits and vegetables brittleness.

Summary of the invention

Technical matters to be solved by this invention overcomes prior art artificial interference seriously, wastes time and energy, and the problems such as checkout equipment is expensive, provide a kind of food brittleness detection method based on food fracture voice signal.

For solving the problems of the technologies described above, the present invention adopts following technical scheme to realize: the step of described food brittleness detection method is as follows:

1) judge whether structuring food prods brittleness detection model, or not do not enter the 4th) step instant food brittleness practical measurement step, be enter the 2nd) construction step of step instant food brittleness detection model;

2) structure of food brittleness detection model:

(1) acquisition of food brittleness mechanical characteristics;

(2) collection of food fracture voice signal;

(3) denoising of food fracture voice signal;

(4) time domain of voice signal and the extraction of frequency domain character value;

(5) structure of food brittleness detection model;

3) judge whether food brittleness detection model has built, or not do not continue structuring food prods brittleness detection model; Being, for entering the 4th) step instant food brittleness practical measurement step is for subsequent use;

4) food brittleness practical measurement.

The step of the acquisition of the food brittleness mechanical characteristics described in technical scheme is as follows:

1) sample pretreatment

By pane consistent sized by sample preparation, for arranging brittleness gradient, sample blocks being inserted in drying box and carrying out drying process;

2) acquisition of each optimum configurations of Texture instrument and sample breakage value

Test condition: single compresses, test speed: 1mm/s, measuring distance: 6mm, trigger point load: 0.1N, chooses wedge shaped pressure head, is put in by sample on self-control sample stage and carries out fragmentation, obtain first peak value in force diagram, be sample breakage value, unit N.

The step of the collection of the food fracture voice signal described in technical scheme is as follows:

1) utilize the broken food of food inspection device, adopt the sound transducer collected sound signal be connected with the computing machine being provided with AdobeAuditon3.0 software in food inspection device when crushed food;

2) select monophony when gathering food fracture voice signal, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sound transducer is 4cm.

The step of the denoising of the food fracture voice signal described in technical scheme is as follows:

Spectrum-subtraction is utilized to carry out denoising to the food fracture voice signal collected:

1) suppose that pure voice signal is S (t), noise is N (t), voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 1:

X(t)＝S(t)+N(t)(1)

2) X (t), S (t), N (t) is set through Fourier transform as X (w), S (w), N (w), therefore:

X(w)＝S(w)+N(w)(2)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(3)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](4)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(5)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(6)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{{\left|X\left(w\right)\right|}^2-{\left|N\left(w\right)\right|}^2\right\}}^{\frac{1}{2}}---\left(7\right).$

The time domain of the voice signal described in technical scheme and the extraction of frequency domain character value refer to:

1) waveform index Y ₂:

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The computing formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}{\left|x\left(i\right)\right|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(8\right)$

Wherein, i represents sampling number, and N represents sampled point number;

2) difference in magnitude D:

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(9)

Wherein, n represents sampling number;

3) marginal spectrum feature F:

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum feature is A _i, marginal spectrum feature formula (10) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(10\right)$

Wherein, f represents frequency values.

The structure of the food brittleness detection model described in technical scheme refers to and utilizes BP neural network food brittleness detection model, and concrete steps are:

1) tentatively determine hidden layer unit number span by three experimental formulas (11) ~ formula (13), then investigate and finally select hidden layer best-of-breed element number in the network convergence speed of just to determine each hidden layer unit number in scope, training error and test error situation;

H＝2I+1(11)

$H=\sqrt{I+O}+a---\left(12\right)$

H＝log ₂I(13)

Wherein, H is node in hidden layer, and I is input layer number, and O is output layer nodes, and α is constant, and value is 1 to 10;

2) determination of other parameters:

Learning rate establishes 0.05, and hidden layer uses sigmoid function, output layer to use linear function as activation function;

With the time-frequency characteristics value of sample breakage voice signal for input, with sample breakage index for exporting, build BP neural network food brittleness detection model for subsequent use.

The step of the food brittleness practical measurement described in technical scheme is as follows:

1) sample pretreatment:

By pane consistent sized by sample preparation, use wedge shaped pressure head, the sample be opposite on self-control sample stage carries out break process; Test condition is: single compresses, test speed 1mm/s, measuring distance 6mm, trigger point load 0.1N;

2) collection of food fracture voice signal:

Adopt food inspection device to adopt the sound transducer collected sound signal be connected with the computing machine being provided with AdobeAuditon3.0 software when crushed food, monophony is selected during collection, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sensor is 4cm;

3) denoising of food fracture voice signal:

Adopt spectrum-subtraction, step is as follows:

Suppose that pure voice signal is S (t), noise is N (t), and the voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 14:

X(t)＝S(t)+N(t)(14)

If X (t), S (t), N (t) they are X (w), S (w), N (w) through Fourier transform, therefore:

X(w)＝S(w)+N(w)(15)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(16)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](17)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so

E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(18)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(19)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{{\left|X\left(w\right)\right|}^2-{\left|N\left(w\right)\right|}^2\right\}}^{\frac{1}{2}}---\left(20\right)$

4) time domain of voice signal and the extraction of frequency domain character value

(1) waveform index Y ₂

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}{\left|x\left(i\right)\right|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(21\right)$

Wherein, i represents sampling number, and N represents sampled point number;

(2) difference in magnitude D

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(22)

Wherein, n represents sampling number.

(3) marginal spectrum feature F

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum is A _i, marginal spectrum feature formula (23) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(23\right)$

Wherein, f represents frequency values;

5) food brittleness measurement result:

In food brittleness detection model constructed by the time-frequency characteristics value extracted is substituted into, calculate food brittleness value, export food brittleness measurement result.

Compared with prior art the invention has the beneficial effects as follows:

1. the quantitative detection utilizing the food brittleness detection method of food fracture voice signal property can realize food brittleness (breaking property) of the present invention.

2. the food brittleness detection method of food fracture voice signal property that utilizes of the present invention does not rely on valuation officer's subjective judgement, only needs an operator can complete batch evaluation.

3. the food brittleness detection method detection speed of food fracture voice signal property that utilizes of the present invention is fast, can complete in 3 minutes even shorter time.

4. the equipment price adopted in the food brittleness detection method of food fracture voice signal property that utilizes of the present invention is cheap, economical.

Accompanying drawing explanation

Below in conjunction with accompanying drawing, the present invention is further illustrated:

The schematic diagram that Fig. 1 forms for the food inspection apparatus structure adopted in food brittleness detection method of the present invention;

Fig. 2 is the FB(flow block) of food brittleness detection method of the present invention;

Fig. 3-a is the time-domain diagram of sample breakage voice signal before the denoising that utilizes food brittleness detection method of the present invention and obtain;

Fig. 3-b is the time-domain diagram of sample breakage voice signal after the denoising that utilizes food brittleness detection method of the present invention and obtain;

In figure: 1. chopping block, 2. sample, 3. pressure head, 4. sound transducer, 5. computing machine, 6. pressure transducer.

Embodiment

Below in conjunction with accompanying drawing, the present invention is explained in detail:

The food brittleness detection method of food fracture voice signal property that utilizes of the present invention is for utilizing the food fracture voice signal characteristic in signal time domain and Hilbert frequency domain extraction, construct the BP neural network model according to sound characteristic value, prediction food brittleness.

Sound transducer 4, computing machine 5 that the drying box that the food inspection device adopted in food brittleness detection method, equipment include Texture instrument that model is CT3, model is PH070A, model are AT8033.

Model is the Texture instrument of CT3:

Mainly relevant with the mechanical characteristic texture of food characteristic that Texture instrument reflects, the physical property concept of sample is made to the statement of datumization, have compression, puncture, shear and the multiple test pattern that stretches, often kind of pattern has multiple probe selective, can carry out texture profile analysis (TPA), puncture, shears and tensile property test food.

Model is the drying box of PH070A:

Drying box is a kind of conventional instrument and equipment, is mainly used to dry sample, also can provide the temperature environment needed for experiment.

Model is the sound transducer of AT8033:

Sound transducer 4 is electric microphones with frequency response similar to human ear.It is used for receiving sound wave, the vibrational diagram of display sound.But can not measure the intensity of noise.

Consult Fig. 1, the position relationship in food inspection device between each parts and annexation:

When gathering food fracture voice signal, pressure head 3 is positioned at directly over chopping block 1, and sample 2 is placed under the seaming chuck 3 of chopping block 1, and pressure transducer 6 is positioned at below chopping block 1, and pressure transducer 6 is connected for contacting with testing table below with chopping block 1 above.Sound transducer 4 is positioned at the 4cm place, right side of sample 2, and meanwhile, sound transducer 4 is connected with computing machine 5 respectively by connecting line with pressure transducer 6.

Consult Fig. 2, the step of food brittleness detection method of the present invention is as follows:

Food of the present invention all refers to the obvious rhizome vegetable of brittleness feature, such as: carrot, ternip, potato, pachyrhizus.

1. judge whether structuring food prods brittleness detection model, or not do not enter the 4th step instant food brittleness practical measurement step; Enter the structure of the 2nd step instant food brittleness detection model;

2. the structure of food brittleness detection model:

1) acquisition of food brittleness mechanical characteristics

This method is in the sample force diagram drawn by Texture instrument, choose breaking property, length of curve, curve and coordinate axis surround area, top and all peak-to-averages ratio as brittleness mechanical characteristics, finally choose breaking property value and characterize food brittleness value, food brittleness value can adopt Texture instrument to obtain.

Experiment material: the obvious rhizome vegetables of brittleness feature such as carrot, ternip, potato, pachyrhizus;

The step measuring vegetables sample breaking property is as follows:

(1) sample pretreatment

By pane consistent sized by sample preparation, for arranging brittleness gradient, sample blocks being inserted in drying box and carrying out drying process.

(2) acquisition of each optimum configurations of Texture instrument and sample breakage

Test condition: single compresses, test speed: 1mm/s, measuring distance: 6mm, and trigger point load: 0.1N, chooses wedge shaped pressure head, is put in sample on self-control sample stage and carries out fragmentation, obtain first peak value in force diagram, be sample breakage, unit N.

2) collection of food fracture voice signal

Food fracture sound signal collecting step is as follows:

(1) utilize the broken food of food inspection device of the present invention, adopt sound transducer 4 collected sound signal be connected with the computing machine 5 being provided with AdobeAuditon3.0 software in food inspection device when crushed food.

(2) select monophony when gathering, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sound transducer 4 is 4cm.

3) denoising of food fracture voice signal

In sound admission process, due in environment and Texture instrument running time noise, noise may impact the extraction of eigenwert afterwards and modeling.Because the noise in bad border and when Texture instrument operates is comparatively steady, therefore spectrum-subtraction is utilized to carry out denoising to the food fracture voice signal collected.

Spectrum-subtraction is stable based on hypothesis noise or converts additivity noise slowly, and separate with short-term stationarity signal, thus utilize the power spectrum of the voice signal of band noise to deduct the power spectrum of noise, thus obtain pure useful voice signal.

Step is as follows:

(1) suppose that pure voice signal is S (t), noise is N (t), voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 1:

X(t)＝S(t)+N(t)(1)

(2) X (t), S (t), N (t) is set through Fourier transform as X (w), S (w), N (w), therefore:

X(w)＝S(w)+N(w)(2)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(3)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](4)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(5)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(6)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{{\left|X\left(w\right)\right|}^2-{\left|N\left(w\right)\right|}^2\right\}}^{\frac{1}{2}}---\left(7\right)$

4) time domain of voice signal and the extraction of frequency domain character value

This food brittleness detection method is extracted time domain and the frequency domain related features value of food fracture voice signal, and wherein, frequency domain adopts Hilbert-Huang transform (HHT) to process;

(1) waveform index Y ₂

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}{\left|x\left(i\right)\right|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(8\right)$

Wherein, i represents sampling number, and N represents sampled point number.

(2) difference in magnitude D

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(9)

Wherein, n represents sampling number.

(3) marginal spectrum feature F

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum is A _i, marginal spectrum feature formula (10) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(10\right)$

Wherein, f represents frequency values.

5) structure of food brittleness detection model

Utilize BP neural network food brittleness detection model, food brittleness is predicted.

The learning process of BP neural network is mainly divided into two stages: the forward-propagating of signal and the backpropagation of error.The process that forward-propagating outputs signal to output layer and at output layer from input layer to hidden layer by signal; Backpropagation and error signal, from output layer to hidden layer to the process that input layer transmits, are modified to the connection weights of every layer, to satisfy condition in process.BP neural network will through iteration repeatedly to reduce error.

(1) determination of node in hidden layer:

Specified by the scope of experimental formula to node in hidden layer, and then adopt method of trial and error determination implicit function nodes.

H＝2I+1(11)

$H=\sqrt{I+O}+a---\left(12\right)$

H＝log ₂I(13)

Wherein, H is node in hidden layer, and I is input layer number, and O is output layer nodes, and α is constant, and value is 1 to 10.

(2) determination of other parameters:

Learning rate establishes 0.05, and hidden layer uses sigmoid function, output layer to use linear function as activation function.

With the time-frequency characteristics value of sample breakage voice signal for input, with sample breakage index for exporting, build the BP neural network food brittleness detection model that input layer number is 3, output layer nodes is 1, by the waveform index Y of input food fracture voice signal _2,difference in magnitude D, marginal spectrum feature F tri-eigenwerts, can draw food brittleness testing result.

3. judge whether food brittleness detection model has built, or not do not continue structuring food prods brittleness detection model; Be, for subsequent use for entering the 4th step instant food brittleness practical measurement step;

4. food brittleness practical measurement

1) sample pretreatment

By pane consistent sized by sample preparation, use wedge shaped pressure head, the sample be opposite on self-control sample stage carries out break process.Test condition is: single compresses, test speed 1mm/s, measuring distance 6mm, trigger point load 0.1N.

2) collection of food fracture voice signal

When crushed food, the voice signal ruptured by sound transducer 4 collected specimens be connected with the computing machine 5 being provided with AdobeAuditon3.0 sound signal collecting software.During collection, select monophony, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sound transducer 4 is 4cm.

3) denoising of food fracture voice signal

Adopt spectrum-subtraction, step is as follows:

Suppose that pure voice signal is S (t), noise is N (t), and the voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 14:

X(t)＝S(t)+N(t)(14)

If X (t), S (t), N (t) they are X (w), S (w), N (w) through Fourier transform, therefore:

X(w)＝S(w)+N(w)(15)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(16)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](17)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(18)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(19)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{{\left|X\left(w\right)\right|}^2-{\left|N\left(w\right)\right|}^2\right\}}^{\frac{1}{2}}---\left(20\right)$

4) time domain of voice signal and the extraction of frequency domain character value

(1) waveform index Y ₂

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}{\left|x\left(i\right)\right|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(21\right)$

Wherein, i represents sampling number, and N represents sampled point number.

(2) difference in magnitude D

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(22)

Wherein, n represents sampling number.

(3) marginal spectrum feature F

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum is A _i, marginal spectrum feature formula (23) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(23\right)$

Wherein, f represents frequency values.

5) food brittleness measurement result:

In food brittleness detection model constructed by the time-frequency characteristics value extracted is substituted into, calculate food brittleness value, export food brittleness measurement result.

Embodiment

Based on food fracture voice signal food brittleness Forecasting Methodology step following (for potato):

1. the structure of food brittleness detection model

1) acquisition of sample brittleness mechanical characteristics

(1) sample preparation

Be the pane that length is respectively 2cm × 1cm × 1cm by sample preparation, for arranging brittleness gradient, sample blocks inserted in drying box and carrying out drying process.Bake out temperature is 60 DEG C, and drying time is spaced apart 5min, till enrolling less than voice signal, obtains 10 gradation of moisture samples altogether when sample breakage.

(2) test condition: single compresses, test speed: 1mm/s, measuring distance: 6mm, trigger point load: 0.1N, chooses wedge shaped pressure head and carries out fragmentation to the sample be put on self-control sample stage, obtain first peak value in force diagram, be sample breakage, unit N.Each gradation of moisture sample gets 6, totally 60 samples.

2) collection of sample breakage voice signal

While disrupted sample, the voice signal ruptured by sound transducer 4 collected specimens be connected with the computing machine 5 being provided with AdobeAuditon3.0 software.During collection, select monophony, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sound transducer 4 is 4cm, enrolls to obtain the voice signal of 60 samples altogether.

3) denoising of sample breakage voice signal

Consult Fig. 3-a, Fig. 3-b, the voice signal choosing front 20 frames, as " silent frames ", is namely designated as N (t), carry out estimating noise, utilize spectrum-subtraction, denoising is carried out to the voice signal of sample breakage, as shown in FIG., Fig. 3-a is effect before denoising to denoising effect, and Fig. 3-b is effect after denoising.

4) time-frequency characteristics value is extracted

Extract temporal signatures value and the Hilbert frequency domain character value of sample breakage voice signal respectively.

(1) waveform index: $Y_2=\underset{i=0}{\overset{N-1}{\Σ}}{\left|x\left(i\right)\right|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(8\right)$

Wherein, i represents sampling number, and N represents sampled point number;

(2) difference in magnitude: D=maxx (n)-minx (n) (9)

Wherein, n represents sampling number;

(3) marginal spectrum feature: $F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(10\right)$

5) structure of food brittleness detection model

Utilize BP neural network structuring food prods brittleness detection model.

(1) determination of node in hidden layer: by experimental formula (11) ~ formula (13), the scope of node in hidden layer is specified, and then with method of trial and error determination implicit function nodes.

Because input layer number is 3, output node number is 1, substitutes into experimental formula, formula (11) ~ formula (13), calculates node in hidden layer scope for [112], and then to choose node in hidden layer by method of trial and error be 5.

(2) arrange learning rate and establish 0.05, hidden layer uses sigmoid function, output layer to use linear function as activation function, to choose in MATLAB traingdm as training function, with the waveform index Y of sample breakage voice signal _2,difference in magnitude D and marginal spectrum feature F is input, with sample breakage index for exporting, builds the BP neural network food brittleness detection model that input layer number is 3, output layer nodes is 1, by the waveform index Y of input food fracture voice signal _2,difference in magnitude D, marginal spectrum feature F tri-eigenwerts, can draw sample brittleness testing result.

6) food brittleness practical measurement

After the sample process of the same terms, signals collecting, denoising and feature extraction, get 10 samples, using the difference in magnitude of acquisition, waveform index and marginal spectrum feature as input value, substitute in the food brittleness detection model based on BP neural network, brittleness testing result is as shown in table 1.

Table 1 testing result

Conclusion:

Based on time domain-Hilbert frequency domain, utilize the relative error of neural network prediction potato breaking property to be up to 1.71%, minimum is-0.03%, and average relative error is-0.05%, is less than perfect error.In sum, this food brittleness detection method based on food fracture voice signal provided by the invention, not only can realize the Fast Evaluation to the obvious rhizome vegetable brittleness of brittleness feature, and have very high accuracy.

Claims (7)

1. a food brittleness detection method, is characterized in that, the step of described food brittleness detection method is as follows:

1) judge whether structuring food prods brittleness detection model, or not do not enter the 4th) step instant food brittleness practical measurement step, be enter the 2nd) construction step of step instant food brittleness detection model;

2) structure of food brittleness detection model:

(1) acquisition of food brittleness mechanical characteristics;

(2) collection of food fracture voice signal;

(3) denoising of food fracture voice signal;

(4) time domain of voice signal and the extraction of frequency domain character value;

(5) structure of food brittleness detection model;

3) judge whether food brittleness detection model has built, or not do not continue structuring food prods brittleness detection model; Being, for entering the 4th) step instant food brittleness practical measurement step is for subsequent use;

4) food brittleness practical measurement.

2. according to food brittleness detection method according to claim 1, it is characterized in that, the step of the acquisition of described food brittleness mechanical characteristics is as follows:

1) sample pretreatment

By pane consistent sized by sample preparation, for arranging brittleness gradient, sample blocks being inserted in drying box and carrying out drying process;

2) acquisition of each optimum configurations of Texture instrument and sample breakage value

Test condition: single compresses, test speed: 1mm/s, measuring distance: 6mm, trigger point load: 0.1N, chooses wedge shaped pressure head, is put in by sample on self-control sample stage and carries out fragmentation, obtain first peak value in force diagram, be sample breakage value, unit N.

3. according to food brittleness detection method according to claim 1, it is characterized in that, the step of the collection of described food fracture voice signal is as follows:

1) utilize the broken food of food inspection device, adopt sound transducer (4) collected sound signal be connected with the computing machine (5) being provided with AdobeAuditon3.0 software in food inspection device when crushed food;

2) select monophony when gathering food fracture voice signal, sample frequency is 44100Hz, and sampling resolution is 16bit, and the distance of sound source and sound transducer (4) is 4cm.

4. according to food brittleness detection method according to claim 1, it is characterized in that, the step of the denoising of described food fracture voice signal is as follows:

Spectrum-subtraction is utilized to carry out denoising to the food fracture voice signal collected:

1) suppose that pure voice signal is S (t), noise is N (t), voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 1:

X(t)＝S(t)+N(t)(1)

2) X (t), S (t), N (t) is set through Fourier transform as X (w), S (w), N (w), therefore:

X(w)＝S(w)+N(w)(2)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(3)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](4)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(5)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(6)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{\right|X\left(w\right){|}^2-\left|N\left(w\right){|}^2\right\}}^{\frac{1}{2}}---\left(7\right).$

5. according to food brittleness detection method according to claim 1, it is characterized in that, the time domain of described voice signal and the extraction of frequency domain character value refer to:

1) waveform index Y ₂:

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The computing formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}|x\left(i\right){|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(8\right)$

Wherein, i represents sampling number, and N represents sampled point number;

2) difference in magnitude D:

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(9)

Wherein, n represents sampling number;

3) marginal spectrum feature F:

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum is A _i, marginal spectrum feature formula (10) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(10\right)$

Wherein, f represents frequency values.

6. according to food brittleness detection method according to claim 1, it is characterized in that, the structure of described food brittleness detection model refers to and utilizes BP neural network food brittleness detection model, and concrete steps are:

1) tentatively determine hidden layer unit number span by three experimental formulas (11) ~ formula (13), then investigate and finally select hidden layer best-of-breed element number in the network convergence speed of just to determine each hidden layer unit number in scope, training error and test error situation;

H＝2I+1(11)

$H=\sqrt{I+O}+a---\left(12\right)$

H＝log ₂I(13)

Wherein, H is node in hidden layer, and I is input layer number, and O is output layer nodes, and α is constant, and value is 1 to 10;

2) determination of other parameters:

Learning rate establishes 0.05, and hidden layer uses sigmoid function, output layer to use linear function as activation function;

With the time-frequency characteristics value of sample breakage voice signal for input, with sample breakage index for exporting, build BP neural network food brittleness detection model for subsequent use.

7. according to food brittleness detection method according to claim 1, it is characterized in that, the step of described food brittleness practical measurement is as follows:

1) sample pretreatment

By pane consistent sized by sample preparation, use wedge shaped pressure head, the sample be opposite on self-control sample stage carries out break process; Test condition is: single compresses, test speed 1mm/s, measuring distance 6mm, trigger point load 0.1N;

2) collection of food fracture voice signal

Adopt food inspection device to adopt sound transducer (4) collected sound signal be connected with the computing machine being provided with AdobeAuditon3.0 software (5) when crushed food, monophony is selected during collection, sample frequency is 44100Hz, sampling resolution is 16bit, and the distance of sound source and sound transducer (4) is 4cm;

3) denoising of food fracture voice signal:

Adopt spectrum-subtraction, step is as follows:

Suppose that pure voice signal is S (t), noise is N (t), and the voice signal containing noise is X (t), and S (t), N (t) are separate, then the relation obtaining three is represented by formula 14:

X(t)＝S(t)+N(t)(14)

If X (t), S (t), N (t) they are X (w), S (w), N (w) through Fourier transform, therefore:

X(w)＝S(w)+N(w)(15)

Both sides square obtain:

|X(w)| ²＝|S(w)| ²+|N(w)| ²+S ^*(w)N(w)+N ^*(w)S(w)(16)

Both sides are got respectively and are expected:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)+E[S(w) ^*N(w)]+E[N ^*(w)S(w)](17)

Owing to supposing S (t) before, N (t) is separate, therefore S (w), N (w) are independent, so

E [S (w) ^*n (w)]+E [N ^*(w) S (w)]=0, that is:

E(|X(w)| ²)＝E(|S(w)| ²)+E(|N(w)| ²)(18)

For signal being short-term stationarity, therefore desirable appropriate frame length obtains:

|X(w)| ²＝|S(w)| ²+|N(w)| ²(19)

The estimated value finally obtaining pure voice signal is:

$\left|S\left(w\right)\right|={\left\{\right|X\left(w\right){|}^2-\left|N\left(w\right){|}^2\right\}}^{\frac{1}{2}}---\left(20\right)$

4) time domain of voice signal and the extraction of frequency domain character value

(1) waveform index Y ₂

Waveform index is stable, a responsive characteristic parameter, i.e. the ratio of voice signal energy and amplitude root mean square.The formula of waveform index is as follows:

$Y_2=\underset{i=0}{\overset{N-1}{\Σ}}|x\left(i\right){|}^2/\sqrt{\frac{1}{N}\left(\underset{i=0}{\overset{N-1}{\Σ}}\right|x\left(i\right)\left|\right)}---\left(21\right)$

Wherein, i represents sampling number, and N represents sampled point number;

(2) difference in magnitude D

The i.e. difference of voice signal maximum amplitude and minimum amplitude in time domain scale, unit dB, if x (n) is discrete voice signal, difference in magnitude computing formula is as follows:

D＝maxx(n)-minx(n)(22)

Wherein, n represents sampling number.

(3) marginal spectrum feature F

Marginal spectrum feature unit dB, if the amplitude of marginal spectrum is A _i, marginal spectrum feature formula (23) calculates:

$F=\underset{i=0}{\overset{f-1}{\Σ}}{A}_i---\left(23\right)$

Wherein, f represents frequency values;

5) food brittleness measurement result:

In food brittleness detection model constructed by the time-frequency characteristics value extracted is substituted into, calculate food brittleness value, export food brittleness measurement result.