CN112185419A

CN112185419A - Glass bottle crack detection method based on machine learning

Info

Publication number: CN112185419A
Application number: CN202011055692.1A
Authority: CN
Inventors: 张涛; 丁碧云; 刘赣俊
Original assignee: Tianjin University
Current assignee: Tianjin University
Priority date: 2020-09-30
Filing date: 2020-09-30
Publication date: 2021-01-05

Abstract

A glass bottle crack detection method based on machine learning comprises the following steps: collecting a sound signal generated by knocking a glass bottle body to be detected by a metal rod by using a pickup; extracting the characteristics of the acquired sound signals, and extracting three major characteristics of the traditional characteristics, the short-term characteristics and the time-frequency characteristics of the sound signals to obtain an initial characteristic set; performing feature selection on the initial feature set by adopting a shuffled frog-leaping algorithm to obtain an optimal feature subset; performing BPNN training to obtain model parameters by taking the optimal characteristic subset as the input of the BPNN, and judging whether the glass bottle has cracks according to the output of the BPNN to finally obtain a crack detection result; and according to the crack detection result, removing the glass bottles with cracks by adopting removing equipment. The glass bottle crack detection provided by the invention has the characteristics of high detection precision and high efficiency.

Description

Glass bottle crack detection method based on machine learning

Technical Field

The invention relates to a nondestructive detection technology for defects of glass bottles in the beverage industry, in particular to a nondestructive crack detection method for empty glass bottles before the glass bottles are filled with beverage in the beverage industry.

Background

1. Glass bottle crack detection

Glass bottles in the beverage industry are easy to be damaged in the production or transportation process, and have the defects of bottle body damage, crack generation and the like. When the glass bottle with the cracks is used for containing beverage, the beverage leaks to cause unnecessary waste, and the reputation of a manufacturer is even affected, so that direct or indirect economic loss is caused. The problem is particularly obvious in a liquor factory, and the poor-quality glass bottles for filling the liquor not only can cause the waste of the liquor, but also can influence the reputation of the factory when flowing into the market and cause economic loss. Therefore, in order to prevent accidental loss, it is necessary to perform a crack check on the empty bottle before filling.

At present, aiming at the problem that the crack of the glass bottle can cause the leakage of the beverage, the beverage industry generally adopts the traditional manual detection method to detect the crack of the glass bottle, as shown in figure 1. This method is widely used in many beverage plants, such as wineries. The method comprises two parts of detection before filling and detection after filling.

The detection before filling is to identify the defects of the bottles through the perception of human ears (namely human hearing) and preliminarily remove the cracked bottles. This part aims at improving detection efficiency through the preliminary detection to bottle defect, and its detection accuracy is about 95%. And the detection after filling is to place the glass bottle filled with the beverage on a paperboard, and after 24 hours, whether the glass bottle at the corresponding position has cracks is identified by judging whether the watermark exists on the paperboard. The detection accuracy of the part is close to 100%.

Generally, the classification accuracy of the manual detection method can reach 100%, but the method completely depends on manual work, and has the defects of low efficiency, large occupied space, beverage waste and the like. In addition, as the beverage industry moves from traditional manual operation and manual tactics to mechanized, automated, informatization and intelligent production processes, the crack detection of the glass bottles provides new challenges of full-automatic online detection and elimination and the like. Therefore, the machine is adopted to replace manual labor, the automation of detection is realized, the labor intensity is reduced, the production efficiency is improved, and the product quality is ensured.

In order to overcome the challenges, an automatic detection method is provided to replace the detection before filling of manual detection and even the whole manual detection method by combining the production characteristics of wine bottles on the production line. Common automatic detection methods in the field of defect detection include computer vision-based methods, ultrasonic-based methods, vibration-based methods, and acoustic-based methods. The acoustic detection method has the advantages of being simple to operate, high in detection speed, small in environmental limitation, low in cost and the like. According to the method, firstly, the glass bottle is knocked, then acoustic response, namely sound, generated after knocking is collected, and the sound is analyzed and processed, so that whether cracks exist in the glass bottle is judged. However, the acoustic detection method has a problem of low classification performance in actual production.

2. Shuffled frog leaping algorithm

The Shuffled Frog Leaping Algorithm (SFLA) is a novel swarm intelligence optimization Algorithm proposed by Eusuff et al in 2003, which simulates the foraging and migration processes of Frog populations. The whole process can be described as follows: the frogs together form a large population, which is divided into initial sub-populations, each sub-population containing several frogs. Each frog has different adaptability to the environment. The adaptability of the frog is here expressed in terms of the distance from the food. In order to obtain food, each frog is close to the frog closest to the food in the sub-population, namely, each frog jumps to the frog closest to the food, so that the distance between the frog and the food is shortened, and the adaptability of the frog is improved. After the whole frog population finishes one jump, different frogs can exchange and share information, and the purpose is to improve the adaptability of the whole frog population. In order to avoid frogs accumulating in the same location in a sub-population, a new sub-population is re-created after each exchange of information. Each frog carries its own information to add to the new sub-population. The frog with the highest adaptability in the sub-populations is the jumping direction of the frog in the sub-populations, and the sub-populations are regenerated each time, so that information interaction among the sub-populations is guaranteed. By alternating the hopping of each frog to the optimal frog within a sub-population (local search) and the reshuffling between all frog sub-populations (global search), it is ensured that the entire frog population is moving in the optimal direction.

The mathematical description of the algorithm is as follows: first, an initial population of N individual frogs P ═ X is generated₁，X₂，…，X_NAnd after the initial population is generated, firstly, arranging the initial population in a descending order according to the fitness value of each frog, grouping the frog populations according to a modular factorial grouping method, and dividing the frog populations into p sub-populations, wherein each sub-population comprises q frogs and satisfies the relation of N being equal to p multiplied by q.

The rules of the modular factorization method are as follows: the 1 st frog is divided into the 1 st sub-population, the 2 nd frog is divided into the 2 nd sub-population, the p th frog is divided into the p th sub-population, then the p +1 th frog is divided into the 1 st sub-population, and so on until all the frogs are divided into a certain sub-population. The relationship between the frog sequencing sequence number y and the grouping number k in the whole distribution process is as follows:

marking the frog X with the maximum fitness value in each group as the optimal frog X in the group_bThe worst frog X in the group is the lowest fitness value_wAnd the global optimal frog X in the whole frog population with the maximum global fitness value_g。

Then for X in each sub-population_wJump, update rule as follows:

D＝r·(X_b-X_w)

X′_w＝X_w+D

where r is a random number between 0 and 1 and D represents the frog jump distance. If updated, new frog X'_wThe adaptability is stronger than the original adaptability, and the original adaptability is replaced; if not improved, use X_gIn place of X_bUpdating is carried out; if there is still no improvement, a new frog is randomly generated to replace X_w. After each sub-population completes the local search, all the frogs are mixed and sequenced again, then the sub-populations are divided again, and then the local search is carried out until a preset convergence condition is reached, such as the mixed iteration times are reached, the frogs do not jump within the appointed iteration times, and the like.

3. Feature extraction and selection for non-destructive defect detection

Feature extraction refers to deriving one or more parameters from a raw signal by way of mathematical transformation or statistical analysis, which parameters are capable of representing a characteristic of the signal in a certain respect. These parameters are called the features of the signal, and the process of obtaining these features is feature extraction. Feature extraction is used in the fields of data mining and pattern recognition, and is one of the key technologies. Through feature extraction, parameters capable of reflecting data characteristics in signals can be obtained, and the quality of the extracted features determines the performance of data mining and pattern recognition.

In the field of non-destructive defect detection, efficient feature extraction is crucial to the overall detection process, as a set of important and compact features will make correct detection easier. Feature extraction obtains relevant discrimination information (i.e., defect information) directly from the input signal. For an efficient system, features extracted from the same type of sample should be similar. Also, the features extracted from different classes of samples should be very different. Therefore, the features can effectively represent defect information, thereby greatly improving classification performance.

The traditional method for extracting the characteristics of the acoustic defect detection system comprises a time domain method, a frequency domain method and a wavelet transform method. The method comprises the characteristics of dimensional zero crossing rate and short-time energy, and the characteristics of dimensionless mean value, peak value, variance, root mean square value, peak value factor, impulse factor, margin factor, wave form factor, K factor, kurtosis, autocorrelation analysis parameters and the like. The frequency domain features include spectral peaks, spectral locations, spectral areas, spectral statistical parameters, and the like. The wavelet-based features include energy and variance of each node obtained after wavelet packet decomposition and reconstruction.

To better characterize the defect information in a nondestructive defect inspection task, a large number of features need to be extracted. However, not all features are significant, as many of them may be redundant, even independent of the classification task. Therefore, a significant and suitably dimensioned subset of features needs to be selected from the extracted large number of features. The feature selection is to select a feature subset from the original feature set, wherein the dimension of the feature subset is much smaller than the dimension of the original feature set. The feature selection selects a part of features from the original feature set on the premise of not changing the properties of the original feature set to form a new feature space. The result of feature selection directly affects the complexity, accuracy and stability of the classifier. Therefore, selecting the appropriate feature subset is also important for defect detection.

The feature selection aims to improve the classification performance of the classifier and reduce the computational complexity. Current research on feature selection methods focuses mainly on search strategies and evaluation criteria. According to the forming process of the feature subset, the basic search strategy of feature selection can be divided into the following 3 types: global search, random search, and heuristic search. Although the global search method such as the branch-and-bound method can obtain the optimal feature subset, the global search method has the problem of high computational complexity. Random search methods such as the Relief series algorithm have high uncertainty, more cycles are required, and setting of parameters has certain difficulty. Heuristic search such as individual optimal combination, a sequence forward selection method, a generalized sequence forward selection method (GSFS), a sequence backward selection method (SBS), a generalized sequence backward selection method (GSBS), an l-increasing and r-removing selection method, a generalized l-increasing and r-removing selection method, a floating search method and the like. Heuristic search strategies, while efficient, come at the cost of sacrificing global optimality. A particular search algorithm will employ two or more basic search strategies, e.g., genetic algorithms based on random searches and heuristic search algorithms.

Feature selection is very challenging due to the large search space and the correlation between features. At present, a large number of effective feature selection methods are proposed, but there are many disadvantages in practical application. Most feature selection methods have the problem of local optimization or high computational complexity.

4. Hilbert-yellow Transform (Hilbert-Huang Transform, HHT)

For nonlinear and non-stationary signals, due to their strong time-varying property, time-frequency analysis methods are generally adopted to extract the characteristics of the signals. The signal time-frequency analysis method can analyze the local part quantity on the basis of expanding the signal on a time-frequency space, and the method can better reflect the time-frequency characteristic of the audio signal. Common time-frequency analysis methods include short-time Fourier transform (STFT), Wigner-Viller distribution (WVD) and Wavelet Transform (WT), and although the methods can better analyze the time-frequency characteristics of an audio signal, the methods also have certain problems, such as the Heisenberg inaccurate measurement principle and the single resolution problem of the STFT, the serious cross interference problem of the WVD, and the limitation that the wavelet transform excessively depends on wavelet base selection, energy leakage and the like.

HHT is a time-frequency domain analysis method proposed by tsuba, NASA, usa in 1998. The method is completely based on the signal, has self-adaptability, has good processing effect on nonlinear non-stationary signals, is particularly suitable for analysis of the non-stationary and non-linear signals, is widely applied to the fields of marine signal analysis, seismic signal analysis, biomedicine, health monitoring and the like, and obtains good results. HHT is largely divided into two parts, first an Empirical Mode Decomposition (EMD) of the signal yields a series of eigenmode functions (IMF) with frequencies from high to low, and then a Hilbert transform of the IMF component yields meaningful instantaneous frequencies (instantaneous properties of the signal). Wherein each IMF component must satisfy the following 2 conditions: (1) in the whole data range, the number of the extreme points and the number of the zero-crossing points must be equal or differ by one at most; (2) at any point, the average value of the upper envelope formed by all maximum value points and the lower envelope formed by all minimum value points is always zero.

The steps of EMD decomposition are as follows:

1) for a signal x (t) to be processed, finding out all local maximum values and minimum value points of the signal;

2) carrying out spline interpolation on the extreme values to obtain an upper envelope line formed by all local maximum value points and a lower envelope line formed by all local minimum value points, which are respectively marked as u (t) and v (t);

3) the mean of the upper and lower envelope lines is calculated as:

4) let h (t) x (t) -m (t), verify whether h (t) satisfies the condition of IMF component, if yes, h (t) is the first IMF component; if not, taking h (t) as input, continuing the previous steps until the first IMF component is obtained and is marked as c₁(t)；

5) Will r is₁(t)＝x(t)-c₁(t) repeating steps (1) to (4) as a new analysis signal to obtain a second IMF component c₂(t), at this time, r is expressed₂(t)＝r₁(t)-c₂(t) repeating the above steps until a remainder r is obtained_m(t) is a monotone signal or its value is smaller than a certain predetermined threshold, the decomposition is finished, all IMF components are obtained;

if Hilbert transform is performed on each obtained IMF component, the instantaneous frequency and the instantaneous amplitude of the signal are obtained. Wherein the content of the first and second substances,

the Hilbert transform is used to find the instantaneous parameters of the signal. The Hilbert transform of IMF components c (t) is defined as follows:

i.e. the convolution of the signal and the inverse of time. For each IMF component c obtained after EMD decomposition_i(t) performing a Hilbert transform,obtaining a corresponding analytic signal:

wherein, a_i(t) is the amplitude of the analytic signal, θ_iAnd (t) is the phase of the analytic signal. The instantaneous frequency of the signal is defined as:

thereby c_i(t) can be expressed as:

(Re represents a real part)

A is to_i(t) is on the combined time-frequency plane, i.e. c is obtained_iHilbert spectrum of (t):

the residual component after EMD decomposition has large energy and can interfere with the analysis of other IMF components, and the information of the signal is usually located in high-frequency components, so the residual component is generally ignored when Hilbert transform is carried out.

After obtaining the Hilbert spectrum for each IMF component, the Hilbert spectrum of signal x (t) can be found by the following equation:

the margin spectrum of the signal can be defined by the Hilbert spectrum:

the marginal spectrum of the signal more accurately reflects the frequency of the signal than the frequency spectrum. From the marginal spectrum it can be seen at which frequencies the signal is mainly concentrated over time. The energy of a certain frequency in the marginal spectrum represents on the whole time axis, and a vibration wave with a frequency possibly occurs locally, and the larger the amplitude of the vibration wave is, the more the possibility of representing the frequency is.

Unlike other fourier transform-based time-frequency analysis methods, HHT does not suffer from the drawbacks of fourier spectrum analysis, as HHT can process any signal without a priori knowledge and is not limited by the Heisenberg uncertainty principle.

However, HHT methods have difficulty accurately resolving acoustic signals due to spurious IMF components and modal aliasing problems. In response to the above problems, an improved HHT method combining WPD and HHT is proposed. The method utilizes a wavelet packet decomposition method to extract a series of narrow-band signals from an acoustic signal so as to solve the problem of mode aliasing, and then utilizes an EMD method to extract a plurality of IMF components from the narrow-band signals. And then screening out a real IMF component based on the cross correlation between the IMF component and the original narrowband signal so as to solve the false IMF component. Finally, transient attributes are extracted from the actual IMF components through Hilbert transform, and relevant time-frequency characteristics are extracted from the transient attributes.

5. Mutual information

Mutual Information (MI) is a basic concept in Information theory, and represents the content of Mutual Information between two discrete variables, which can be used to evaluate the similarity and dependency between the two variables. For two discrete variables X and Y, assuming their edge probability distributions p (X) and p (Y), the mutual information I (X; Y) between them can be calculated by the following formula:

where p (x, y) is the joint probability distribution of x and y. In general, mutual information between two variables is calculated through information entropy, and the calculation formula is as follows:

I(X；Y)＝H(X)-H(X|Y)

wherein, H (X) represents the information entropy of the variable X, and the calculation formula is as follows:

h (X | Y) represents the conditional entropy of X with respect to Y, and is calculated as follows:

mutual information can be used to indicate the degree of dependency between two variables, where X and Y are completely independent, the mutual information value between them is 0, and the larger the degree of similarity between X and Y, the larger the mutual information value.

Disclosure of Invention

The invention aims to solve the technical problems of low detection efficiency, beverage waste and the like of the existing manual detection method, and provides a glass bottle crack detection method which is high in detection accuracy and efficiency and saves more manpower.

The technical scheme adopted by the invention is as follows: a glass bottle crack detection method based on machine learning comprises the following steps:

1) collecting a sound signal generated by knocking a glass bottle body to be detected by a metal rod by using a pickup, and transmitting the sound signal to an industrial personal computer;

2) in an industrial personal computer, extracting the characteristics of the acquired sound signals, and extracting three major characteristics of traditional characteristics, short-time characteristics and time-frequency characteristics of the sound signals to obtain an initial characteristic set;

3) performing feature selection on the initial feature set by adopting a shuffled frog-leaping algorithm to obtain an optimal feature subset;

4) performing BPNN training to obtain model parameters by taking the optimal characteristic subset as the input of the BPNN, and judging whether the glass bottle has cracks according to the output of the BPNN to finally obtain a crack detection result;

5) and according to the crack detection result, removing the glass bottles with cracks by adopting removing equipment.

The step 2) of performing feature extraction on the acquired sound signals comprises the following steps:

(2.1) extracting the traditional characteristics of the sound signal, and marking as { F1 }. The conventional features of sound signals include: time domain features, frequency domain features, and time-frequency domain features. Wherein the time domain features include: zero crossing rate, energy, mean, variance, root mean square, peak factor, impulse factor, margin factor, form factor, K factor, kurtosis, and root mean square, peak, and peak factors of the autocorrelation sequence. The frequency domain features include: spectral area, the first 5 spectral amplitude peaks and their frequency locations, center of gravity frequency, mean square frequency, and root mean square, standard deviation, and variance of frequency. The time-frequency domain characteristics comprise the energy and variance of all nodes of the third layer obtained after decomposition and reconstruction of the three layers of wavelet packets;

(2.2) extracting short-time characteristics of the sound signal, and recording the short-time characteristics as { F2 }. Firstly, framing the signals to obtain a plurality of subframe signals, then respectively extracting the traditional characteristics of each subframe signal, and finally combining the traditional characteristics of each subframe signal to obtain a short-time characteristic set { F2 };

and (2.3) extracting time-frequency characteristics of the sound signals based on the improved HHT method, and recording the time-frequency characteristics as { F3 }. The method comprises the following steps:

(2.3a) adopting Daubechies wavelets to carry out N-layer wavelet packet decomposition and reconstruction on the input sound signals x (t) to obtain 2^NNarrowband signals of different frequency bands;

(2.3b) performing empirical mode decomposition on all the obtained narrow-band signals respectively to obtain a plurality of IMF components;

(2.3c) calculating mutual information quantity of the narrow-band signal and each IMF component thereof, comparing the mutual information quantity with an IMF component screening threshold Mi, and screening out real IMF components capable of reflecting signal characteristics;

(2.3d) sorting the screened real IMF components according to the sequence of the frequency from high to low to obtain the final IMF component of the whole signal;

(2.3e) Hilbert transforming the IMF components of the final overall signal to obtain instantaneous properties of the signal, including: the instantaneous frequency and instantaneous amplitude of each IMF component, and the margin spectrum;

(2.3f) extracting time-frequency characteristics of the sound signal based on the instantaneous attributes of the signal, including: the spectral area, bandwidth, cut-off frequency and variance of the marginal spectrum, the peak value and peak position of the first 5 marginal spectra, the variance of the marginal spectrum, the maximum value and minimum value of the third and fourth IMF components, and the mean, variance and effective value of the instantaneous amplitude of the first 6 IMF components.

The step 3) of selecting the features of the initial feature set by using the shuffled frog leaping algorithm comprises the following steps:

(1) setting initial parameters of the shuffled frog-leaping algorithm, comprising: the frog population number, the number of each group, the maximum iteration times of the algorithm and the iteration times iter are equal to 0;

(2) setting dimension m of the optimal feature subset output in feature selection, wherein m corresponds to the position coding number of the frog individual in the shuffled frog-leaping algorithm, and the position of the frog corresponds to one dimension of the initial feature set with the dimension n to be the m feature subset;

(3) randomly generating individual information of the frog population according to the parameters set in the step (1) and the step (2), wherein the position code of each frog individual is represented by m unordered and non-repetitive sequences consisting of integers from 0 to n-1, each number in the sequences corresponds to a feature, and n is the dimension of an initial feature set;

(4) and carrying out fitness calculation on the position information of the individual frog by using a fitness function to obtain the fitness value of the individual frog and whether the fitness value meets the optimized constraint condition, namely calculating the evaluation function value of the characteristic subset corresponding to the scheme in the characteristic selection. The fitness function expression is:

where S is { S ═ S₁，S₂，…，S_MThe position information of the individual frogs is corresponding to the characteristic subset expressed by the individual frogs, wherein S belongs to F and S_iAnd S_jRespectively representing the ith and jth features in S. And L represents a target class label corresponding to the sample data. I (S)_iAnd L) represents the average mutual information quantity of the ith characteristic and the target class label. I (S)_i，S_j) And the average mutual information quantity of the ith characteristic and the jth characteristic is represented. Evaluating the position of each individual of the frog population to obtain the fitness value of each individual of the frog population;

(5) sorting the whole frog population in a descending order according to the fitness value of the frog individual, grouping the frog population according to a modular factorial grouping method, and determining the intra-group optimal frog, the intra-group worst frog and the global optimal frog in the whole frog population in each group;

(6) in each group, firstly, updating the worst frog in the group by using the information of the optimal frog in the group, and finishing an updating process if the fitness value of the updated worst frog in the group is found to be superior to that of the previous frog and meets the constraint condition of the feature subset; otherwise, updating by adopting the position information of the global optimal frog, and finishing the updating process if the fitness value of the worst frog in the updated group is superior to that of the previous frog and meets the constraint condition of the feature subset; if the frog still can not be updated successfully, updating the optimal frog in the group by adopting a random updating mode;

(7) adding 1 to the iteration number iter;

(8) the entire frog population is shuffled. If the iteration times are smaller than the maximum iteration times of the algorithm, turning to the step (5); otherwise, outputting the global optimal frog information;

(6) and obtaining the code of the global optimal solution, namely the optimal characteristic subset, according to the global optimal frog information.

Aiming at a specific glass bottle crack detection task, the invention adopts a characteristic extraction method of improving HHT to expand the traditional sound characteristic so as to improve the glass bottle crack detection precision, then adopts a shuffled frog-leaping algorithm to optimize a characteristic set so as to obtain a more compact and obvious characteristic subset, and finally adopts a BP neural network to make a decision on whether the glass bottle has cracks or not. The method can effectively solve the problems of excessive labor consumption, low detection efficiency, beverage waste and the like in the existing manual detection method in the production line of the beverage industry, thereby ensuring high-precision and automatic detection of the cracks of the glass bottles. Compared with the traditional method for detecting the cracks of the artificial glass bottle, the method for detecting the cracks of the glass bottle has the characteristics of high detection precision and high efficiency.

Drawings

FIG. 1 is a schematic diagram of a conventional glass bottle crack detection method in the beverage industry.

Fig. 2 is a flow chart of the improved HHT feature extraction algorithm of the present invention.

Fig. 3 is a flow chart of the shuffled frog-jump algorithm of the present invention.

FIG. 4 is a block diagram of the machine learning based crack detection for glass bottles in accordance with the present invention.

Detailed Description

The following describes a machine learning-based glass bottle crack detection method according to the present invention in detail with reference to the following embodiments and accompanying drawings.

As shown in fig. 4, the method for detecting cracks of glass bottles based on machine learning of the invention is characterized by comprising the following steps:

2) in an industrial personal computer, feature extraction is carried out on the acquired sound signals, and three major features of the traditional features, the short-time features and the time-frequency features of the sound signals are extracted to obtain an initial feature set. The method comprises the following steps:

3) And performing feature selection on the initial feature set by adopting a shuffled frog-leaping algorithm to obtain an optimal feature subset. The method comprises the following steps:

(7) adding 1 to the iteration number iter;

The optimal parameter settings are given below:

(1) the wavelet packet decomposition uses a db6 wavelet in the Daubechies wavelet family, and the optimal parameter of the number of layers N of the wavelet packet decomposition is set to 3.

(2) The threshold value of mutual information quantity adopted for screening the IMF components is generally 0.1, and when the mutual information quantity of the IMF components and corresponding narrow-band signals is less than 0.1, the IMF components are judged as false IMF components.

(3) The dimension m of the feature subset is typically between 5% and 20% of the dimension of the initial feature set, and the optimal parameter m in the present invention is set to 10.

(4) In the initial parameters of the shuffled frog-leaping algorithm, the optimal parameter of the frog population number is set to be 20, the optimal parameter of each group number is set to be 5, and the optimal parameter of the maximum iteration number of the algorithm is set to be 1000.

Claims

1. A glass bottle crack detection method based on machine learning is characterized by comprising the following steps:

2. The method for detecting cracks of glass bottles based on machine learning as claimed in claim 1, wherein the step 2) comprises:

3. The method for detecting cracks of glass bottles based on machine learning as claimed in claim 1, wherein step 3) comprises:

(7) adding 1 to the iteration number iter;