Vibration signal time-frequency analysis and feature fusion-based pear early-stage moldy heart disease nondestructive testing method

Abstract

The invention discloses a nondestructive testing method for an early moldy heart disease of pomes based on vibration signal time-frequency analysis and feature fusion, which comprises the following steps: acquiring vibration signals of moldy pomes and healthy pomes by using a vibration device, converting time-domain signals into time-frequency images by using a time-frequency analysis method, carrying out feature extraction on a time domain, a frequency domain and the time-frequency images, and calculating the early moldy heart disease of the pomes. And then, a machine learning algorithm is utilized to respectively establish single-domain and multi-domain fusion vibration multi-domain spectrum feature discrimination models, and accurate discrimination of early mildewed pears is realized. The method has excellent performance in pear fruit quality classification and prediction, improves the commodity rate and market competitiveness of the pear fruits, and has great significance in promoting rapid and healthy development of the pear fruit industry.

Description

Technical Field

[0001] The invention belongs to the technical field of rapid nondestructive detection of agricultural product quality, and in particular relates to a nondestructive detection method for early internal diseases of pear fruit based on vibration multi-domain spectrum. Background Art

[0002] Pear fruit core mold disease is an internal disease caused by infection by Alternaria fungi. The weakly parasitic fungal spores enter the ventricle from the stigma through the calyx tube during the flowering period of the pear fruit and then remain dormant. This state continues during the growth and ripening period of the fruit until the harvest and storage period, when the moldy symptoms of the diseased fruit begin to spread outward from the core chamber of the fruit and rot. If it can be identified as early as possible, it will help fruit merchants make wise decisions on storage or sales in a timely manner; otherwise, early core moldy fruit is very likely to be mixed with healthy fruit and flow into the circulation links such as storage and sales. The hidden mold will quietly develop into the flesh, making the fruit lose its edible value, seriously affecting the reputation of fruit merchants and the repurchase rate of pear fruit by consumers. It may even infect a large number of healthy fruits, causing fruit merchants and fruit preservation companies to suffer heavy losses. Therefore, studying the monitoring and rapid detection methods of the pathogenesis of pear fruit core mold disease is of great practical significance for its prevention and control.

[0003] In recent years, domestic and foreign researchers have conducted a lot of basic research on the non-destructive detection of internal diseases of pears. The methods used for non-destructive detection of internal diseases of pears mainly include near-infrared spectroscopy, hyperspectral technology, X-ray imaging, electronic nose technology, magnetic resonance technology and acoustic vibration. From the perspective of the technical means used, near-infrared spectroscopy has limitations in penetration depth and resolution when detecting deep tissues. Hyperspectral technology has many bands, large data volume and strong correlation between adjacent bands, which brings difficulties to data processing. X-ray imaging equipment is usually expensive and takes a long time to obtain high-quality 3D images. At the same time, X-ray detection involves radiation, which may affect the health of operators. Electronic nose technology has high requirements on sensor performance and gas collection is easily affected. Magnetic resonance technology has the disadvantages of high equipment cost and specialized and complex signal analysis. Acoustic vibration method has attracted more and more attention in the field of non-destructive detection of early diseases in fruits due to its advantages of simple equipment and low technical cost. Zhang et al. (2021) extracted statistical feature parameters from the time and frequency domain signals of the vibration response to identify the internal early browning pear fruit. The SVM model constructed by combining the statistical feature parameters in the time and frequency domains achieved a classification accuracy of 91.84% for the internal early browning pear fruit with a disease degree less than 30%, but the classification accuracy of the browning pear fruit with a disease degree less than 18% was still low, only 81.82%. Obviously, only extracting a limited number of statistical feature parameters from the two analysis domains of the time and frequency domains of the vibration signal failed to fully explore other more abundant and comprehensive potential information related to the internal diseases of the pear fruit, which is probably one of the reasons for the low detection accuracy of the acoustic vibration method for early internal diseased pear fruit. If the time and frequency domain statistical features extracted from the vibration signal and the time-frequency image texture features are fused to solve the problem of less information covered by the single domain features, we can have a more comprehensive understanding of the development of the internal disease of the pear fruit, which will help to significantly improve the detection accuracy of the early internal diseases of the pear fruit. This is a research idea worth exploring for the identification of early moldy pear fruit. Summary of the invention

[0004] The main technical problem solved by the present invention is to provide a nondestructive detection method for early internal diseases of pear fruit based on vibration multi-domain spectrum, and the extracted time domain and frequency domain statistical features and time-frequency image texture features are multi-domain fused to construct an early discrimination model for pear fruit core mold disease, laying a solid technical foundation for the research and development of early online detection devices for pear fruit core mold disease, and solving the current problem of low accuracy in early discrimination of pear fruit core mold disease.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is as follows: comprising the following steps: Step 1: Taking Korla pears as the research object, they were picked at Xinyang Family Farm. Fruit farmers selected apples suspected of molty core disease and healthy apples, and then stored them in a fruit preservation warehouse at -2~0℃ and relative humidity of 85%~95%; Step 2: Under the condition that the experimental device is stable and reliable, the time domain signal of the pear fruit is collected by using a laser Doppler vibrometer device, and the time domain signal is converted into a frequency domain signal by using a fast Fourier transform (FFT) algorithm; Step 3: Use time-frequency analysis methods, short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wagewiener distribution (WVD) to convert the acquired pear fruit time domain signal into the corresponding time-frequency image; Step 4: Cut the pear fruit in half along the equator and use a digital camera to take a photo of the cross section of the pear fruit. Use Matlab 2018b Image Processing Toolbox to extract the RGB components of the pear fruit color image, use the Canny edge detector to identify the edge of the B channel image, and use the extracted edge as the contour of the mask. S 1 The cross-sectional area of ​​the pear S 2 The percentage of pear fruit mold is defined as the disease severity of pear fruit mold S .

[0006] Step 5: Use statistical analysis to extract statistical features in the time and frequency domains, and use the gray-level gradient co-occurrence matrix (GLGCM) algorithm to extract time-frequency image texture features to establish a vibration single-domain / multi-domain spectrum feature set; Step 6: Based on the established vibration single-domain / multi-domain spectrum feature set, Pearson correlation analysis is used to evaluate the sensitivity of feature parameters to classification targets, principal component analysis (PCA) is used to reduce the dimension of vibration multi-domain feature parameters, and a machine learning algorithm is used to construct a pear fruit internal disease discrimination analysis model; Step 7: After determining the optimal model for early detection of pear fruit core mold disease, use the optimization algorithm to optimize the model.

[0007] The time-frequency analysis methods for converting the time domain into the time-frequency image in the step three are: short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wygvener distribution (WVD). STFT is an analysis method that combines the time domain and the frequency domain. The STFT external processing method for time series signals is to select a time-frequency localized window function. The Hann window can provide higher frequency resolution and reduce spectrum leakage when processing vibration signals. It also has smaller sidelobe gain and fast attenuation, as well as good time domain smoothness. The calculation formula is as follows: CWT has the ability to continuously adjust the time-frequency resolution as the frequency changes. When the frequency component of the analysis signal changes from low to high, the time domain resolution gradually increases, while the frequency domain resolution gradually decreases as the frequency increases. During the entire analysis process, the size of the time-frequency analysis window remains unchanged. HHT includes two links, EMD and Hilbert transform, and is widely used in the field of non-stationary and nonlinear signal processing. The calculation formula is as follows: HHT overcomes the defect of using meaningless harmonic components to represent non-stationary signals in traditional time-frequency analysis methods, and can provide extremely high time-frequency resolution and good time-frequency aggregation. The calculation formula is as follows: WVD is a bilinear time-frequency distribution, defined as the Fourier transform of the signal center covariance function. In WVD, the resolution of time and frequency are not mutually restricted, and it has excellent time-frequency aggregation capabilities. The calculation formula is as follows: In step five, the method for extracting multi-domain spectral feature parameters of vibration signals is as follows: 14 commonly used time-domain statistical features are extracted by statistical analysis, including maximum value, minimum value, mean value, standard deviation, root mean square value and square root of amplitude value, which are dimensioned; while skewness, kurtosis, absolute mean, peak factor, gap factor, shape factor, impulse factor and peak-to-peak value are dimensionless features. In addition, 7 frequency-domain statistical features are extracted, among which mean frequency, frequency standard deviation, kurtosis frequency, centroid frequency, mean square frequency and root mean square frequency are dimensioned features, while frequency variance is dimensionless feature. Grayscale gradient co-occurrence matrix (GLGCM) is used to extract 15 texture features of time-frequency images, including energy, grayscale mean, gradient mean, grayscale variance, gradient variance, correlation and inertia, which are used to quantify the statistical characteristics of image texture. Grayscale entropy, gradient entropy and mixed entropy are used to evaluate the uncertainty of grayscale and gradient distribution in images. Small gradient advantage and large gradient advantage are used to reflect the intensity of image grayscale changes, while gradient distribution unevenness and grayscale gradient distribution unevenness indicate the unevenness of image texture distribution. The inverse matrix reflects the similarity between a pixel in an image and its neighboring pixels.

[0008] The idea of ​​constructing the pear fruit internal disease discrimination analysis model based on the established vibration single-domain / multi-domain spectral feature set in step six is ​​as follows: ① Use the Kennard-Stone algorithm to randomly divide the 375 pear fruit vibration multi-domain data sets into training sets and test sets in an 8:2 ratio. ② Use Pearson correlation analysis to evaluate the correlation between the 14 time-domain statistical feature parameters, 7 frequency-domain statistical feature parameters, and 15 texture feature parameters obtained from the vibration signal and the three types of labels of pear fruit to determine the sensitivity of each feature parameter to the classification target. Use principal component analysis to reduce the dimension of the vibration multi-domain feature parameters. ③ Use the single-domain and multi-domain fusion feature parameters as input, respectively, to construct KNN, RF and SVM pear fruit early core mold disease detection models. ④ Determine the optimal input feature parameters through the accuracy of the training set, and use accuracy, recall rate, F 1 , and Kappa and other model performance evaluation indicators were used to select the most suitable machine learning model for the detection and classification of early core mold disease of pear fruit.

[0009] The optimization ideas based on the optimal early detection model of pear fruit rotten core disease in step 7 are as follows: ① Use common model optimization algorithms, particle swarm optimization algorithm (PSO), sparrow search algorithm (SSA) and gray wolf optimization algorithm (GWO) to further improve the performance of the pear fruit early rotten core disease detection classification model. ② Through the accuracy, recall rate, F 1 The classification model improved based on the optimization algorithm was evaluated by the Kappa model performance evaluation index to determine the final machine learning model for detection and grading of early core rot of pear fruit.

[0010] The beneficial effects of the present invention are: The present invention utilizes STFT, CWT, HHT and WVD algorithms to convert acoustic vibration signals into time-frequency images, uses statistical analysis to extract statistical features in time domain and frequency domain, and uses grayscale gradient co-occurrence matrix method to extract texture features of time-frequency images. Pearson correlation analysis and principal component analysis are used to process feature parameters, and a discriminant model with single-domain / multi-domain spectral features as input is established.

[0011] The present invention is used for non-destructive detection of early moldy core disease of pear fruit. Since the external symptoms of early moldy core pear fruit are not obvious and cannot be directly identified by naked eyes, mixing with healthy pear fruit will not only affect the overall quality of pear fruit and have an adverse impact on the reputation of fruit merchants, but also easily rot during storage, causing serious economic losses to fruit merchants. Therefore, using this method to perform non-destructive detection on pear fruit, the early moldy core pear fruit can be effectively removed.

[0012] The present invention applies machine learning algorithm in combination with time-frequency analysis and multi-domain fusion technology to the construction of a detection model for early moldy core disease of pear fruit, giving full play to the advantages of machine learning in processing raw data. At the same time, the machine learning algorithm is trained on a multi-domain spectrum data set to capture more comprehensive features and improve the accuracy of detection, providing a research idea worthy of exploration for the identification of early moldy core pear fruit. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] Figure 1 It is a flow chart of the time-frequency analysis and feature fusion method of the present invention. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present invention will be described in detail below in conjunction with the accompanying drawings so that those skilled in the art can more easily understand the advantages and features of the present invention and thus more clearly define the protection scope of the present invention.

[0015] Implementation steps: Figure 1 , embodiments of the present invention include: Taking pear fruit core mold disease as an example, the non-destructive detection of other internal diseases of fruits and vegetables can refer to this method for detection. By establishing a new knowledge base based on the evaluation criteria of the tested samples, non-destructive detection of such products can be carried out.

[0016] The early core mold disease detection system for pear fruit consists of a vibration signal system, a time-frequency analysis system, a feature fusion system and pattern recognition; the vibration signal acquisition device is a laser Doppler vibrometer detection device; the time-frequency analysis system uses short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wagewiener distribution (WVD) algorithm to convert the time domain signal into a time-frequency image; the feature fusion system uses Pearson correlation analysis and principal component analysis to process the extracted feature parameters; the early discriminant machine learning models KNN, RF and SVM for pear fruit core mold disease are constructed, and finally the disease degree of the tested pear fruit is determined by a computer.

[0017] The specific steps are as follows: First, experienced fruit farmers select pears suspected of having core mold disease and healthy pears, and then store them in a fruit preservation warehouse with a temperature of -2~0℃ and a relative humidity of 85%~95%.

[0018] The time domain signals of pear fruits were collected by laser Doppler vibrometer, and the time domain signals were converted into frequency domain signals by fast Fourier algorithm. The time domain signals were converted into time-frequency images corresponding to short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wagewiener distribution (WVD) by time-frequency analysis method. The pear fruit was cut into two halves along the equator, and the cross section of the pear fruit was photographed with a digital camera. The RGB components of the pear fruit color image were extracted using the Matlab2018b image processing toolbox, and the edges of the B channel image were identified using the Canny edge detector. The extracted edges were used as the contour of the mask. S 1 The cross-sectional area of ​​the pear S 2 The percentage of pear fruit mold is defined as the disease severity of pear fruit mold S .

[0019] The method for extracting multi-domain spectral feature parameters of vibration signals in step 5 is as follows: 14 most commonly used time domain statistical features, 7 frequency domain statistical features are extracted by statistical analysis, and 15 texture features of time-frequency images are extracted by gray gradient co-occurrence matrix (GLGCM). The time domain and frequency domain statistical features are shown in Tables 1 and 2; the time-frequency image texture features are shown in Table 3.

[0020]

[0021]

[0022] The Kennard-Stone algorithm was used to randomly divide the 375 pear fruit vibration multi-domain data sets into training sets and test sets in an 8:2 ratio. Pearson correlation analysis was used to evaluate the correlation between the 14 time-domain statistical feature parameters, 7 frequency-domain statistical feature parameters, and 15 texture feature parameters obtained from the vibration signal and the three types of pear fruit labels to determine the sensitivity of each feature parameter to the classification target. Principal component analysis was used to reduce the dimension of the vibration multi-domain feature parameters. The single-domain and multi-domain fusion feature parameters were used as input to construct the KNN, RF, and SVM pear fruit early core mold disease detection models. The optimal input feature parameters were determined by the accuracy of the training set, and the accuracy, recall, and F were used to determine the optimal input feature parameters. 1 , and Kappa and other model performance evaluation indicators were used to select the most suitable machine learning model for the detection and classification of early core mold disease of pear fruit.

[0023] Based on the optimal early detection model of pear fruit core mold disease, the performance of the pear fruit early core mold disease detection classification model was further improved using common model optimization algorithms, particle swarm optimization algorithm (PSO), sparrow search algorithm (SSA) and gray wolf optimization algorithm (GWO). 1 The classification model improved based on the optimization algorithm was evaluated by the Kappa model performance evaluation index to determine the final machine learning model for detection and grading of early core rot of pear fruit.

[0024] The above contents are only embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structure or process transformation based on this specification and drawings, or direct or indirect application to other related technical fields, shall be included in the patent protection scope of the present invention.

Claims

1. A nondestructive detection method for early core mold disease of pear fruit based on time-frequency analysis of vibration signals and feature fusion, characterized in that: Follow these steps: Step 1: Taking apples and pears as the research objects, experienced fruit farmers selected pears suspected of having molty core disease and healthy pears, and then stored them in a fruit preservation warehouse at -2~0℃ and relative humidity of 85%~95%; Step 2: Under the condition that the experimental device is stable and reliable, the time domain signal of the pear fruit is collected by using a laser Doppler vibrometer detection device, and the time domain signal is converted into a frequency domain signal by using a fast Fourier transform (FFT) algorithm; Step 3: Use time-frequency analysis methods, short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wagewiener distribution (WVD) to convert the acquired pear fruit time domain signal into the corresponding time-frequency image; Step 4: Cut the pear fruit in half along the equator and use a digital camera to take a photo of the cross section of the pear fruit. Use Matlab2018b image processing toolbox to extract the RGB components of the pear fruit color image, use the Canny edge detector to identify the edge of the B channel image, and use the extracted edge as the contour of the mask. S 1 accounts for the cross-sectional area of ​​the pear S The percentage of 2 is defined as the disease degree of pear fruit mold S .

2. Step 5: Use statistical analysis to extract statistical features in the time and frequency domains, and use the gray-level gradient co-occurrence matrix (GLGCM) algorithm to extract time-frequency image texture features to establish a vibration single-domain / multi-domain spectrum feature set; Step 6: Based on the established vibration single-domain / multi-domain spectrum feature set, Pearson correlation analysis is used to evaluate the sensitivity of feature parameters to classification targets, principal component analysis (PCA) is used to reduce the dimension of vibration multi-domain feature parameters, and a machine learning algorithm is used to construct a pear fruit internal disease discrimination analysis model; Step 7: After determining the optimal model for early detection of pear fruit core mold disease, use the optimization algorithm to optimize the model.

3. The nondestructive detection method for early core mold disease of pear fruit based on vibration signal time-frequency analysis and feature fusion according to claim 1, characterized in that: According to the requirement that the ratio of healthy pears, pears with mild core mold disease, and pears with moderate to severe core mold disease is 1.6:1.4:1, repeat step 2 to collect vibration signals of pears; The nondestructive detection method for early core mold disease of pear fruit based on vibration signal time-frequency analysis and feature fusion according to claim 1 is characterized in that: In the step 5, on the basis of completing the extraction of the above-mentioned vibration signal characteristic variables, Pearson correlation analysis is used to evaluate the correlation between the time domain statistical features, frequency domain statistical features and texture features extracted from the vibration signal and the degree of pear fruit mildew, so as to determine the sensitivity of each characteristic parameter to the classification target; the single-domain features are multi-domain fused, and if there is repeated expression of information in the vibration multi-domain spectrum characteristic parameters, the principal component analysis (PCA) method is used for dimensionality reduction, so that the selected principal components are uncorrelated with each other, and the cumulative variance contribution rate reaches more than 95%, and finally a multi-domain spectrum feature set that can effectively characterize internal diseases is constructed according to the number and score of the selected principal components; The nondestructive detection method for early core mold disease of pear fruit based on vibration signal time-frequency analysis and feature fusion according to claim 1 is characterized in that: The machine learning algorithms in step 6 include K-nearest neighbor algorithm (KNN), random forest (RF) and support vector machine (SVM); The nondestructive detection method for early core mold disease of pear fruit based on vibration signal time-frequency analysis and feature fusion according to claim 1 is characterized in that: The model optimization algorithms in step seven include particle swarm optimization algorithm (PSO), sparrow search algorithm (SSA) and grey wolf optimization algorithm (GWO).

4. A nondestructive detection method for early core mold disease of pear fruit based on vibration signal time-frequency analysis and feature fusion as described in claim 1, wherein the early core mold disease detection system of pear fruit consists of a vibration signal system, a time-frequency analysis system, a feature fusion system and pattern recognition; wherein the vibration signal acquisition device is a laser Doppler vibrometer detection device; the time-frequency analysis system converts the time domain signal into a time-frequency image using short-time Fourier transform (STFT), continuous wavelet transform (CWT), Hilbert-Huang transform (HHT) and Wagewiener distribution (WVD) algorithm; the feature fusion system uses Pearson correlation analysis and principal component analysis to process the extracted feature parameters; constructs machine learning models KNN, RF and SVM for early discrimination of pear fruit core mold disease, and finally determines the disease degree of the tested pear fruit by a computer.

Images