Method for identifying load state of drum of chili pepper harvester based on vibration signals

Abstract

A method for identifying a load state of a drum of a chili pepper harvester based on vibration signals is provided. The method includes: acquiring vibration signals on a drum wall plate 1, a seedling pressing wheel connecting plate 4, a drum cover plate 5, a drum frame 8, and a ground wheel connecting rod 10; inputting the vibration signals into a load state identification model of the drum of the chili pepper harvester; and obtaining the load state of the drum of the chili pepper harvester. A load state identification model can be made by acquiring vibration signals at the multiple points on the chili pepper harvester in no load, low load, normal load, and high load conditions, analysing the frequencies, amplitudes and other feature parameters of the vibration signals to determine a harvest load state of a drum of the chili pepper harvester in operation.

Description

TECHNICAL FIELD

[01] The present invention relates to the technical field of agricultural intelligent equipment, in particular to a method for identifying a load state of a drum of a chili pepper harvester based on vibration signals. BACKGROUND ART

[02] Chili pepper is one of the most widely cultivated crops in China, and is in the forefront of the world in terms of yield and planting area. With the gradual expansion of the chili pepper area, the mechanized harvesting of chili peppers has gradually become popular. An elastic-tooth drum type chili pepper harvester is one of the mainstream harvesting models at present, which can realize the harvesting of chili pepper fruits by rotating a drum to make elastic teeth continuously brush the chili pepper fruits. However, in actual work, a load state of the drum of the chili pepper harvester will be in a state of constant change due to the difference in the growth of chili pepper plants, different planting modes, complex field environments, etc. A high load state will cause the chili pepper harvester to work under excessive pressure and damage core components, while a low load state will cause the chili pepper harvester not to exert performances fully, resulting in low efficiency, both of which will affect the normal operation of the chili pepper harvester.

[03] At present, there are few studies on a load state of a drum of a chili pepper harvester in China, and it is still in a state of preliminary exploration. At present, the judgment for the load state of the chili pepper harvester basically depends on the manual judgment of a manipulator. This way of relying solely on the manual judgment for the load state of the drum of the chili pepper harvester depends excessively on the actual experience of the manipulator, resulting in poor accuracy, which easily causes the chili pepper harvester to enter an abnormal load state, thereby affecting the normal operation of the chili pepper harvester.

[04] In view of this, the present invention provides a method for identifying a load state of a drum of a chili pepper harvester based on vibration signals. According to the method, a multi-classification model composed of a plurality of support vector machine models is trained by acquiring the vibration signals, so that the chili pepper harvester can effectively identify four conditions of no load, low load, normal load and high load. This method will be able to get rid of the current situation that the chili pepper harvester only relies on manual load judgment, and achieve the automatic and real-time identification of the load state of the drum of the chili pepper harvester. SUMMARY

[05] In order to overcome the technical defects that exist in the harvesting process of an elastic-tooth drum type chili pepper harvester, the present invention provides a method for identifying a load state of a drum of a chili pepper harvester based on vibration signals.

[06] The present invention aims to achieve the above technical purpose by the following technical means.

[07] An embodiment of the present invention provides a method for identifying a load state of a drum of a chili pepper harvester based on vibration signals. The method mainly includes the following steps:

[08] acquiring vibration signals on a drum wall plate (1), a seedling pressing wheel connecting plate (4), a drum cover plate (5), a drum frame (8), and a ground wheel connecting rod (10);

[09] inputting the vibration signals into a load state identification model of the drum of the chili pepper harvester; and

[10] obtaining a load state of the drum of the chili pepper harvester.

[11] Furthermore, the load state identification model of the drum of the chili pepper harvester specifically includes:

[12] acquiring a large number of vibration signals on the drum wall plate (1), the seedling pressing wheel connecting plate (4), the drum cover plate (5), the drum frame (8), and the ground wheel connecting rod (10) when the drum (6) of the chili pepper harvester of a same model is in no load, low load, normal load and high load respectively;

[13] preprocessing four groups of vibration signals respectively;

[14] extracting features from the preprocessed vibration signals;

[15] performing dimensionality reduction on extracted feature data using kernel principal component analysis;

[16] dividing the feature data subjected to dimensionality reduction into a training set and a test set;

[17] inputting the data in the training set into a machine learning model for training, and adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error reaches a minimum; and

[18] obtaining the load state identification model of the drum the chili pepper harvester.

[19] Furthermore, the preprocessing the vibration signals specifically includes:

[20] decomposing the four groups of vibration signals using fully adaptive noise ensemble empirical mode decomposition to obtain several intrinsic mode components;

[21] calculating a correlation coefficient between each intrinsic mode component and an original signal by a formula p, = .                    where p: is the correlation coefficient between the ith intrinsic mode component and the original signal;

[22] calculating an intrinsic mode component selection threshold using p4 = where ph is the selection threshold; and

[23] selecting the intrinsic mode component which is greater than the threshold for reconstruction to obtain four groups of reconstructed vibration signals.

[24] Furthermore, the extracting the features from the preprocessed vibration signals specifically includes:

[25] performing frequency domain processing on the signals using a fast Fourier transform and a power spectral density output, and extracting a mean, a standard deviation, a maximum, a minimum, skewness, kurtosis, a peak, a crest factor, a wave factor, a pulse factor, a margin factor, and a root mean square of each of the four groups of vibration signals as features.

[26] Furthermore, the load state identification model of the drum of the chili pepper harvester according to claim 2 is characterized in that the performing dimensionality reduction on extracted feature data using kernel principal component analysis specifically includes:

[27] selecting a radial basis function as a kernel function, and calculating a kernel matrix using a formula = exp (—

[28] centralizing the kernel matrix to obtain a new matrix Kc;

[29] calculating eigenvalues of Kc and corresponding eigenvectors;

[30] determining a number p of principal components, selecting eigenvectors corresponding to largest p eigenvalues, and constructing a dimensionality reduction space; and

[31] projecting an original matrix into the dimensionality reduction space to obtain feature data subjected to dimensionality reduction.

[32] Furthermore, the load state identification model of the drum of the chili pepper harvester according to claim 2 is characterized in that the machine learning model specifically includes:

[33] the machine learning model is composed of a plurality of support vector machine models, and a Gaussian function is selected as a kernel function of the machine learning model, wherein a first-level support vector machine model is to distinguish between no load and other cases, a second-level support vector machine model is to distinguish between normal load and other cases, and a third-level support vector machine model is to distinguish between low load and high load.

[34] Furthermore, the load state identification model of the drum (6) of the chili pepper harvester according to claim 2 is characterized in that the adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error tested by the test set reaches a minimum specifically includes:

[35] initializing parameters of the particle swarm optimization algorithm;

[36] encoding the penalty coefficient c and the kernel function parameter y as a position of a particle;

[37] training the model using the training set;

[38] calculating a fitness value using a formula f = —             2 F, where f is the fitness value, y(i) is a predicted species, and yi is an actual species;

[39] updating the position and velocity of the particle; and

[40] determining whether a number of iterations is reached; if so, substituting the current penalty coefficient c and kernel function parameter y into the model, and verifying the model using the test set; and if not, continuing to iterative calculation.

[41] According to the method for identifying the load state of the drum of the chili pepper harvester based on the vibration signals as provided by the embodiment of the present invention, a multi-classification model composed of a plurality of support vector machines is established by acquiring the vibration signals, and a particle swarm optimization algorithm is used to optimize the model parameters, which effectively improves the accuracy of model classification. At the same time, the problem that the load state identification of the drum of the chili pepper harvester only relies on manual work, thereby achieving the real-time load identification of the chili pepper harvester. BRIEF DESCRIPTION OF THE DRAWINGS

[42] FIG. 1 is a specific schematic diagram of a method for identifying a load state of a drum of a chili pepper harvester based on vibration signals as provided by an embodiment of the present invention.

[43] FIG. 2 is a schematic diagram of an installation position of an acceleration sensor as provided by an embodiment of the present invention.

[44] FIG. 3 is a training flowchart of a load state identification model of chili peppers as provided by an embodiment of the present invention.

[45] FIG. 4 is a schematic diagram of a load state identification model of chili peppers as provided by an embodiment of the present invention.

[46] FIG. 5 is a flowchart of the adjustment of a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm as provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[47] In order to make the objects, technical solutions and advantages of the present invention clearer, the following will describe the present invention in detail in conjunction with accompany drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the present invention, but not used to limit the present invention.

[48] According to the embodiments of the present invention, when the chili pepper harvester is in operation, a load state of a drum 6 of a chili pepper harvester will be in a state of constant change due to the difference in the growth of chili pepper plants, different planting modes, complex field environment, etc. Identifying the load state of the drum 6 of the chili pepper harvester is of great significance for ensuring the normal operation of the chili pepper harvester. In order to enable the chili pepper harvester to identify the current load state of the drum 6 in real time, a method for identifying a load state of a drum 6 of the chili pepper harvester based on machine learning is proposed. As shown in FIG. 1, and the method includes but is not limited to the following steps:

[49] acquiring vibration signals on a dmm wall plate 1, a seedling pressing wheel connecting plate 4, a drum cover plate 5, a drum frame 8, and a ground wheel connecting rod 10;

[50] inputting the vibration signals into a load state identification model of the drum 6 of the chili pepper harvester; and

[51] obtaining a load state of the drum 6 of the chili pepper harvester.

[52] As shown in FIG. 2, the vibration signal on the drum wall plate 1 is acquired by an acceleration sensor 7, the vibration signal on the seedling pressing wheel connecting plate 4 is acquired by an acceleration sensor 2, the vibration signal on the drum cover plate 5 is acquired by an acceleration sensor 3, the vibration signal on the drum frame 8 is acquired by an acceleration sensor 9, and the vibration signal on the ground wheel connecting rod 10 is acquired by an acceleration sensor 11. The reason for the acquisition of vibration signals from a plurality of positions is that the vibration of the chili pepper harvester comes from a plurality of components and is affected by terrains during the harvesting process, and the vibration signals acquired by a single acceleration sensor cannot well represent actual load states. A plurality of acceleration sensors is used to acquire the vibration signals on the drum wall plate 1, the seedling pressing wheel connecting plate 4, the drum cover plate 5, the drum frame 8, and the ground wheel connecting rod 10 respectively, which can provide the vibration conditions of various components under different load states, thereby reducing the influences of vibration caused by terrains, material impact and the like.

[53] Furthermore, the load state identification model of the drum 6 of the chili pepper harvester specifically includes:

[54] acquiring a large number of vibration signals on the drum wall plate 1, the seedling pressing wheel connecting plate 4, the drum cover plate 5, the drum frame 8, and the ground wheel connecting rod 10 when the drum 6 of the chili pepper harvester of a same model is in no load, low load, normal load and high load respectively;

[55] preprocessing four groups of vibration signals respectively;

[56] extracting features from the preprocessed vibration signals;

[57] performing dimensionality reduction on extracted feature data using kernel principal component analysis;

[58] dividing the feature data subjected to dimensionality reduction into a training set and a test set;

[59] inputting the data in the training set into a machine learning model for training, and adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error tested by the test set reaches a minimum; and

[60] obtaining the load state identification model of the drum 6 of the chili pepper harvester.

[61] FIG. 3 shows a process of training this model.

[62] Furthermore, the preprocessing the vibration signals specifically includes:

[63] decomposing the four groups of vibration signals using fully adaptive noise ensemble empirical mode decomposition to obtain several intrinsic mode components;

[64] calculating a correlation coefficient between each intrinsic mode component and an original signal by a formula p, - ;                   where p: is the correlation coefficient between the ith intrinsic mode component and the original signal;

[65] calculating an intrinsic mode component selection threshold using p = where ph is the selection threshold; and

[66] selecting the intrinsic mode component which is greater than the threshold for reconstruction to obtain four groups of reconstructed vibration signals.

[67] The fully adaptive noise ensemble empirical mode decomposition is a novel signal decomposition algorithm, which can effectively solve the problem of mode aliasing. By decomposing the original signal into several intrinsic mode components, and then reconstructing the intrinsic mode components with correlation coefficients greater than a threshold, the noise in the original signal can be effectively removed.

[68] Furthermore, the extracting the features from the preprocessed vibration signals specifically includes:

[69] performing frequency domain processing on the signals using a fast Fourier transform and a power spectral density output, and extracting a mean, a standard deviation, a maximum, a minimum, skewness, kurtosis, a peak, a crest factor, a wave factor, a pulse factor, a margin factor, and a root mean square of each of the four groups of vibration signals as features.

[70] Furthermore, the load state identification model of the drum of the chili pepper harvester according to claim 2 is characterized in that the performing dimensionality reduction on extracted feature data using kernel principal component analysis specifically includes:

[71] selecting a radial basis function as a kernel function, and calculating a kernel matrix using a formula = exp (        );

[72] centralizing the kernel matrix to obtain a new matrix Kc;

[73] calculating eigenvalues of Kc and corresponding eigenvectors;

[74] determining a number p of principal components, selecting eigenvectors corresponding to largest p eigenvalues, and constructing a dimensionality reduction space; and

[75] projecting an original matrix into the dimensionality reduction space to obtain feature data subjected to dimensionality reduction.

[76] Kernel principal component analysis (PCA) is a data dimensionality reduction algorithm that can reduce the dimensionality of nonlinear data. By means of kernel PCA, data points in an input space are mapped to a feature space through a kernel function, so that the data points in the feature space can be more easily processed and analyzed.

[77] Furthermore, the load state identification model of the drum 6 of the chili pepper harvester according to claim 2 is characterized in that the machine learning model specifically includes:

[78] the machine learning model is composed of a plurality of support vector machine models, and a Gaussian function is selected as a kernel function of the machine learning model, wherein a first-level support vector machine model is to distinguish between no load and other cases, a second-level support vector machine model is to distinguish between normal load and other cases, and a third-level support vector machine model is to distinguish between low load and high load.

[79] The support vector machine is a machine learning model for binary classification problems, which separates samples of different classes by finding a division hyperplane in the sample space, and at the same time makes a minimum distance between two point sets to this plane the largest, and can solve the multi-classification problem when a plurality of support vector machine models are combined. The structure of the model is shown in FIG. 4.

[80] Furthermore, the load state identification model of the drum 6 of the chili pepper harvester according to claim 2 is characterized in that the adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error reaches the minimum specifically includes:

[81] initializing parameters of the particle swarm optimization algorithm;

[82] encoding the penalty coefficient c and the kernel function parameter y as a position of a particle;

[83] training the model using the training set;

[84] calculating a fitness value using a formula / - —         — y,J where f is the fitness value, y(i) is a predicted species, and yi is an actual species;

[85] updating the position and velocity of the particle; and

[86] determining whether the model error reaches a number of iterations; if so, substituting the current penalty coefficient c and kernel function parameter y into the model, and verifying the model using the test set; and if not, continuing to iterative calculation.

[87] The particle swarm optimization algorithm is an algorithm derived from the study on the foraging behavior of birds, in which each solution of the optimization problem is set to each particle, each particle has two parameters, positions and velocities, and the total velocity determines the direction of the particle flight. The particle swarm will be searched in a solution space following the current optimal particles. By randomly initializing the particle swarm in a given solution space, each particle updates its spatial position and flight speed in the solution space by tracking individual extrema and global extrema, and finally finds the global optimum. Here, the penalty coefficient c and the kernel function parameter y are encoded as a position of a particle, and in the process of training the model, these two parameters will be continuously updated with the training, so as to play the effect of optimization. FIG. 5 shows a process of this algorithm.

[88] The foregoing descriptions are merely preferred embodiments of the present invention, and are not intended to limit the present invention. Within the spirit and principles of the present invention, any modifications, equivalent substitutions, improvements, etc., are within the protection scope of the present invention.

Claims

1. A method for identifying a load state of a drum of a chili pepper harvester based on vibration signals, comprising the following steps:S1, acquiring vibration signals on a drum wall plate, a seedling pressing wheel connecting plate, a drum cover plate, a drum frame, and a ground wheel connecting rod;S2, inputting the vibration signals into a load state identification model of the drum of the chili pepper harvester; andS3, obtaining the load state of the drum of the chili pepper harvester.

2. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 1, wherein identifying the load state of the drum of the chili pepper harvester specifically comprises:SI, acquiring a large number of vibration signals on the drum wall plate, the seedling pressing wheel connecting plate, the drum cover plate, the drum frame, and the ground wheel connecting rod when the drum of the chili pepper harvester of a same model is in no load, low load, normal load and high load respectively;S2, preprocessing four groups of vibration signals respectively;S3, extracting features from the preprocessed vibration signals;S4, performing dimensionality reduction on extracted feature data using kernel principal component analysis;S5, dividing the feature data subjected to dimensionality reduction into a training set and a test set;S6, inputting the data in the training set into a machine learning model for training, and adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error tested by the test set reaches a minimum; andS7, obtaining the load state identification model of the drum of the chili pepper harvester.

3. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 2, wherein the preprocessing the vibration signals specifically comprises:SI, decomposing the four groups of vibration signals using fully adaptive noise ensemble empirical mode decomposition to obtain several intrinsic mode components;S2, calculating a correlation coefficient between each intrinsic mode component and an^)(y-y))original signal by a formula ln * _ 2      “T7, wherein p; is the correlation coefficienti£i=1(xi~x) l£i=1(y—y)between the ith intrinsic mode component and the original signal;S3, calculating an intrinsic mode component selection threshold using = ^^ax^s’ wherein ph is the selection threshold; andS4, selecting the intrinsic mode component which is greater than the threshold for reconstruction to obtain four groups of reconstructed vibration signals.

4. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 2, wherein the extracting features from the preprocessed vibration signals specifically comprises:performing frequency domain processing on the signals using a fast Fourier transform and a power spectral density output, and extracting a mean, a standard deviation, a maximum, a minimum, skewness, kurtosis, a peak, a crest factor, a wave factor, a pulse factor, a margin factor, and a root mean square each of the four groups of vibration signals as features.

5. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 2, wherein the performing dimensionality reduction on extracted features using kernel principal component analysis specifically comprises:SI, selecting a radial basis function as a kernel function, and calculating a kernel matrix using a formula KfXpXy) = exp (--);S2, centralizing the kernel matrix to obtain a new matrix Kc;S3, calculating eigenvalues of Kc and corresponding eigenvectors;S4, determining a number p of principal components, selecting eigenvectors corresponding to largest p eigenvalues, and constructing a dimensionality reduction space; andS5, projecting an original matrix into the dimensionality reduction space to obtain feature data subjected to dimensionality reduction.

6. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 2, wherein the machine learning model specifically comprises:the machine learning model being composed of a plurality of support vector machinemodels, and a Gaussian function being selected as a kernel function of the machine learning model, wherein a first-level support vector machine model is to distinguish between no load and other cases, a second-level support vector machine model is to distinguish between normal load and other cases, and a third-level support vector machine model is to distinguish between low load and high load.

7. The method for identifying the load state of the drum of the chili pepper harvester based on vibration signals according to claim 2, wherein the adjusting a penalty coefficient c and a kernel function parameter y of the model using a particle swarm optimization algorithm until a model error reaches a minimum specifically comprises:SI, initializing parameters of the particle swarm optimization algorithm;S2, encoding the penalty coefficient c and the kernel function parameter y as a position of a particle;S3, training the model using the training set;1                         2S4, calculating a fitness value using a formula / = —S™i[y(0 -y^ ] > wherein f is the fitness value, y(i) is a predicted species, and yi is an actual species;S5, updating the position and velocity of the particle; andS6, determining whether the model error reaches a number of iterations; if so, substituting the current penalty coefficient c and kernel function parameter y into the model, and verifying the model using the test set; and if not, continuing to iterative calculation.

Images