Spring tooth type pepper harvester roller fault diagnosis method based on multi-source information fusion

Abstract

The embodiment of the invention provides an elastic claw type pepper harvester roller fault diagnosis method based on multi-source information fusion. The method comprises the steps that vibration signals of a roller rotating shaft, a wall plate and a roller machine and torque and rotating speed signals of the roller rotating shaft under different fault conditions of an elastic claw type pepper harvester roller are collected; then, preprocessing the vibration signal data, the torque signal data and the rotating speed signal data under the same roller fault condition through improved lumped average empirical mode decomposition (MEEMD) and Hilbert transform (HT) to obtain time-frequency diagrams of signal components corresponding to the three signals; and inputting the obtained time-frequency diagram into a multi-channel convolutional neural network (CNN)-spring-tooth type pepper harvester roller fault diagnosis model, and diagnosing the fault state of the spring-tooth type pepper harvester roller. By diagnosing the fault condition of the roller of the spring-tooth type pepper harvester, time and economic losses caused by the fact that the fault cannot be solved in time are avoided, and the production efficiency and fruit quality of the spring-tooth type pepper harvester are improved.

Description

Technical Field

[0001] The present invention relates to a spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion, and more specifically, to a multi-source information data fusion model and a method for agricultural machinery equipment fault diagnosis. The present invention belongs to the technical field of agricultural equipment fault diagnosis. Background Art

[0002] Chili pepper is one of the most widely planted crops in my country, and it ranks among the top in the world in terms of yield and planting area. With the continuous expansion of pepper planting area, mechanized harvesting has gradually become popular. The spring-tine pepper harvester is one of the current mainstream harvester models. It can harvest pepper fruits by rotating the drum to drive the spring teeth on the drum to continuously comb the pepper fruits. However, in actual work, due to differences in the growth of pepper plants, different planting patterns, and complex field environments, the spring-tine pepper harvester drum will have various faults during operation, mainly manifested as drum shaft failure and comb tooth failure.

[0003] At present, there are few studies on how to diagnose the roller failure of spring-tine pepper harvesters in China, and it is still in the initial exploration stage. At present, the diagnosis of roller failure of spring-tine pepper harvesters basically relies on the operator's experience and judgment. This method of diagnosing the roller failure of spring-tine pepper harvesters based solely on the operator's experience has poor accuracy and is likely to lead to the failure of spring-tine pepper harvesters to be discovered in time, which has the potential risk of causing greater damage and affecting the working efficiency of spring-tine pepper harvesters and the quality of pepper fruits. In view of this, it is worth exploring in academia to propose a more efficient and intelligent method for diagnosing roller failure of spring-tine pepper harvesters. Classification is an important technical means in the field of artificial intelligence technology. By classifying, clustering, and extracting features, data or information is divided into different categories or groups to facilitate subsequent data analysis and processing and complete corresponding operations. The convolutional neural network model (CNN) is a commonly used artificial intelligence classification method, and its application in intelligent diagnosis of agricultural machinery faults has broad prospects. Summary of the invention

[0004] In order to solve the technical deficiencies in diagnosing roller faults of a spring-tine pepper harvester during the harvesting process, the present invention provides a roller fault diagnosis method for a spring-tine pepper harvester based on multi-source information fusion.

[0005] The present invention achieves the above technical objectives through the following technical means.

[0006] The embodiment of the present invention provides a method for diagnosing a cylinder fault of a spring-tooth pepper harvester based on multi-source information fusion, which mainly includes the following steps:

[0007] Collect vibration signals on the drum shaft, wall plate, and frame, as well as torque and speed signals of the drum shaft of different spring-tooth pepper harvester drum failures;

[0008] The collected vibration, torque and speed signal data are processed by the improved lumped average empirical mode decomposition method (MEEMD) and Hilbert transform (HT), and then input into the improved multi-channel convolutional neural network fault diagnosis model;

[0009] The improved multi-channel convolutional neural network fusion diagnosis model performs convolution and pooling for feature dimension reduction and feature extraction, and connects the softmax classifier through the fully connected layer to obtain the corresponding spring-tine pepper harvester drum fault category probability, and then determine the fault category.

[0010] Furthermore, the spring-tooth pepper harvester drum fault diagnosis model specifically includes:

[0011] The vibration signals of the drum shaft, wall plate, and frame of the spring-tine pepper harvester under different fault conditions and the torque and speed signals of the drum shaft are collected, and the three groups of different signals are divided into training sets and test sets.

[0012] The collected vibration, torque and speed signals are processed using the improved lumped average empirical mode decomposition method (MEEMD) and Hilbert transform (HT) to obtain the time-frequency diagrams of the three signal components;

[0013] The obtained time-frequency graph is input into an improved multi-channel convolutional neural network (CNN), which performs feature dimension reduction and feature extraction on the time-frequency graph corresponding to the three signal data of vibration, torque and speed of the fault sample through two layers of convolution and pooling operations;

[0014] In a multi-channel convolutional neural network (CNN), the convolution layer and pooling layer perform feature dimension reduction and feature extraction. The feature matrix is ​​converted into a one-dimensional vector through a 1×1 convolution layer and a fully connected layer replaced by global average pooling and input into the softmax classifier for matrix calculation to obtain the corresponding fault probability value.

[0015] Use independent test set signal data to verify the accuracy of the model to ensure that it meets the preset performance requirements. If not, perform parameter adjustment training again;

[0016] Complete the training of the fault diagnosis model for the drum of the spring-tine pepper harvester.

[0017] The three groups of fault signal preprocessing methods preprocess the collected vibration, torque and speed signals through the improved lumped average empirical mode decomposition method to obtain several intrinsic mode function components, and ensure that the decomposed intrinsic mode function components are close to the standard intrinsic mode function components, and then use the Hilbert transform to convolve the obtained intrinsic mode function to obtain a time-frequency diagram. Specifically include:

[0018] Add a white noise signal of a certain amplitude to the signal so that its root mean square value is equivalent to the noise of the signal to be decomposed;

[0019] Add two groups of positive and negative white noise signals y(t) with equal absolute values ​​to the signal to be decomposed x(t), and perform EEMD decomposition on them to obtain two groups of results z i+ (t) and z i- (t), and then average the two sets of results:

[0020] z i ( t ) = 0 . 5 [ z i + ( t ) + z i − ( t )]

[0021] Where i=1,2,3,...,m, to eliminate the white noise residue.

[0022] If the intrinsic mode function component z obtained by EEMD decomposition i If (t) is not a standard eigenmode function component, then the empirical mode decomposition (EMD) is performed again:

[0023]

[0024] Where d1(t) is the first intrinsic mode function component obtained by EMD decomposition of the first quasi-intrinsic mode function component z1(t), q1(t) represents the superposition of the remaining residual components, and the final MEEMD can be expressed as:

[0025] x ( t ) → ∑ L = 1 m [ d i ( t ) ] + r ( t )

[0026] Where, d i (t) represents the final intrinsic mode function component, where i=1,2,.......,m, r(t) represents the final residual component.

[0027] The Hilbert transform (HT) performs convolution processing on the obtained intrinsic mode function to obtain a time-frequency diagram and extract corresponding data, specifically including:

[0028] According to the formula

[0029]

[0030] It can be seen that for the eigenmode function d i(t) Hilbert transform is the function and By performing convolution, the instantaneous frequency and instantaneous amplitude of the corresponding intrinsic mode function component can be obtained, thereby obtaining the time-frequency diagram of the corresponding intrinsic mode function component, and the time-frequency diagram is input into the improved multi-channel convolutional neural network fault diagnosis model as the input of the training model.

[0031] The purpose of building a multi-channel convolutional neural network fault diagnosis model is to:

[0032] The basic components of the multi-channel convolutional neural network model are convolutional layer, pooling layer, convolutional layer and fully connected layer replaced by global average pooling, and softmax classifier connected to the fully connected layer;

[0033] The mathematical model of the convolutional layer in the model can be expressed as:

[0034]

[0035] σ is the activation function. In this example, the σ activation function is the Sigmoid function, and its mathematical expression is:

[0036]

[0037] The activation function is used to delinearize and enhance the expressive power.

[0038] The number of feature extraction network groups of the multi-channel convolutional neural network model is set to 2, and each layer consists of 1 convolution and 1 downsampling layer;

[0039] The classification network consists of a 1×1 convolutional layer, a global average pooling layer, and a softmax output layer;

[0040] The training set signal data is input for iterative training and parameter adjustment to obtain the fault diagnosis model of the spring-tooth pepper harvester drum.

[0041] The divided test set signal data is input into the constructed spring-tine pepper harvester drum fault diagnosis model to evaluate the performance of the model.

[0042] Different fault categories are divided into:

[0043] The roller faults of the spring-tooth pepper harvester are divided into two categories, namely, shaft faults and comb tooth faults. The shaft faults can be specifically divided into four types, namely, loose bearing seat, imbalance, misalignment, and friction. The comb tooth faults can be specifically divided into two types, namely, comb tooth deformation and comb tooth missing. The divided fault categories correspond to the softmax classification channels.

[0044] A softmax fault classifier for a spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion is as follows:

[0045] The three sets of time-frequency graphs corresponding to the vibration, speed, and torque signals of the same fault are input into the multi-channel convolutional neural network. After the neural network convolution and pooling extraction, the global average pooling layer converts the output feature map into a one-dimensional feature vector, which is passed into the softmax classifier for matrix calculation to obtain the probability value of the corresponding fault category.

[0046] In the example of the present invention, the softmax classifier maps the input one-dimensional feature vector to the category, and the result is given in the form of probability, and the formula is as follows:

[0047]

[0048] In the example of the present invention, x represents the input one-dimensional feature vector, y represents the corresponding fault category label, and the K value corresponds to the number of drum fault categories, which is the same as the number of softmax channels.

[0049] in the formula θ k = [ θ 1 θ 2 ⋅⋅⋅ θ K ] T is the weight parameter, corresponding to the classifier parameter θ k Obtained by training the softmax classifier, it can be used as a parameter to calculate the probabilities of all possible categories of the items to be classified.

[0050] Parameter θ k Multiply it with the feature vector and output the probability value belonging to each category. The probability value corresponds to the corresponding fault category to complete the fault identification and classification.

[0051] The input time-frequency graph is subjected to feature dimension reduction and feature extraction by a multi-channel convolutional neural network and then input into a softmax classifier for classification. The output is the probability P of the corresponding fault time-frequency graph belonging to each type of roller fault. y , P y It represents the probability value corresponding to the fault category, and its size sum is 1; compare the output probability value P y The roller fault classification channel corresponding to its maximum probability value is the diagnosed roller fault category. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 A schematic diagram of the basic process of a spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion provided in an embodiment of the present invention.

[0053] Figure 2 A schematic structural diagram of a spring-tine pepper harvester provided in an embodiment of the present invention.

[0054] Figure 3 A schematic diagram of the installation positions of acceleration, torque and speed sensors provided in an embodiment of the present invention.

[0055] Figure 4 A model training flow chart of a spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion provided in an embodiment of the present invention.

[0056] Figure 5 A schematic flow chart of a multi-source information data level fusion method provided in an embodiment of the present invention.

[0057] Figure 6 Schematic diagram of a CNN (Convolutional Neural Network) fault diagnosis training model for a spring-tooth pepper harvester drum provided by an embodiment of the present invention DETAILED DESCRIPTION

[0058] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0059] The embodiments of the present invention are aimed at the occurrence of comb tooth shedding, deformation, loose bearing seat, imbalance, misalignment, and friction failures of the spring-tine pepper harvester's drum due to different pepper planting densities, aging of the drum's mechanical structural materials, or contact with other debris during operation. Timely diagnosis of the existing drum failures of the spring-tine pepper harvester is of great significance for ensuring the normal operation of the spring-tine pepper harvester and reducing the harvesting cost. In order to enable the spring-tine pepper harvester to timely diagnose the type of current drum failure, a method for diagnosing drum failures of the spring-tine pepper harvester based on multi-source information fusion is proposed. Figure 1 As shown, the method includes but is not limited to the following steps:

[0060] Collect vibration signals on the drum shaft, wall plate, and frame of the spring-tooth pepper harvester drum under different fault conditions, as well as torque and speed signals of the drum shaft;

[0061] The collected vibration, torque and speed signal data are processed by the improved lumped average empirical mode decomposition method (MEEMD) and Hilbert transform (HT), and then input into the improved multi-channel convolutional neural network diagnostic model;

[0062] The improved multi-channel convolutional neural network diagnosis model is used to perform convolution and pooling to extract features, and the fully connected layer is connected to the softmax classifier to obtain the corresponding probability value of the spring-tine pepper harvester drum fault category, and the fault category is determined according to the probability value.

[0063] The structure of the spring-tine pepper harvester and the location where the sensors are installed and the signal data samples are collected, such as Figure 2 and Figure 3As shown, the vibration signal on the drum shaft is collected by acceleration sensor 1, the vibration signal on the drum wall plate is collected by acceleration sensor 5, the vibration signal on the drum frame is collected by acceleration sensor 3, the torque signal on the drum shaft is collected by torque sensor 2, and the speed signal on the drum shaft is collected by speed sensor 4. The vibration signals at multiple positions, as well as the torque and speed signals of the drum shaft are collected. Since the vibration of the spring-tooth pepper harvester comes from multiple components and is affected by the terrain during the harvesting process, the vibration signal collected by a single acceleration sensor cannot well represent the actual fault condition. The torque and speed signals of the drum shaft are collected by torque sensor 2 and speed sensor 4, which can provide information from multiple dimensions to support us to understand the changing characteristics of the signal when the drum is in different fault conditions. Using multiple acceleration sensors to collect vibration signals on the drum shaft, drum wall plate, and drum frame respectively can provide the vibration conditions of each component under different fault categories, thereby reducing the impact of vibration caused by terrain, material impact, etc.

[0064] Further, the spring-tooth pepper harvester drum fault diagnosis model, wherein the training process of the diagnosis model is as follows Figure 4 As shown, specifically including:

[0065] The vibration signals of the drum shaft, wall plate, and frame of the spring-tine pepper harvester under different fault conditions and the torque and speed signals of the drum shaft are collected, and the three groups of different signal data are divided into training sets and test sets.

[0066] The ratio of the training set data to the test set data is 4:1;

[0067] The collected vibration, torque and speed signals are processed using the improved lumped average empirical mode decomposition method (MEEMD) and Hilbert transform (HT) to obtain the time-frequency diagrams of the three signal components;

[0068] The obtained time-frequency graph is input into an improved multi-channel convolutional neural network (CNN), which extracts features from the time-frequency graphs corresponding to the three signal data of vibration, torque and speed of the fault sample through two layers of convolution and pooling operations;

[0069] In a multi-channel convolutional neural network (CNN), the convolution layer and pooling layer perform feature dimension reduction and feature extraction, and the feature matrix is ​​converted into a one-dimensional vector through a 1×1 convolution layer and a fully connected layer replaced by global average pooling, and then input into the softmax classifier for matrix calculation to obtain the corresponding fault probability value;

[0070] Use independent test set signal data to verify the accuracy of the model to ensure that it meets the preset performance requirements. If not, perform parameter adjustment training again;

[0071] Complete the training of the fault diagnosis model for the drum of the spring-tine pepper harvester.

[0072] The method for diagnosing the cylinder fault of a spring-tooth pepper harvester based on multi-source information fusion, the multi-source information data-level fusion method flow and the training model (CNN) for diagnosing the cylinder fault of a spring-tooth pepper harvester are as follows: Figure 5 and Figure 6 As shown, in:

[0073] The three sets of collected signal data are preprocessed by the improved lumped mean empirical mode decomposition method (MEEMD) to obtain several intrinsic mode function components, and ensure that the decomposed intrinsic mode function components are close to the standard intrinsic mode function components, including:

[0074] A white noise signal of a certain amplitude is added to the signal so that its root mean square value is equivalent to the noise of the signal to be decomposed, and the difference between the root mean square value and the signal to be decomposed does not exceed 10% to eliminate aliasing;

[0075] Add two groups of positive and negative white noise signals y(t) with equal absolute values ​​to the signal to be decomposed x(t), and perform EEMD decomposition on them respectively to obtain two groups of results z i+ (t) and z i- (t), and then average the two groups of results to eliminate the white noise residue;

[0076] If the eigenmode function component z in EEMD i (t) is not a standard intrinsic mode function component, then EMD decomposition is performed again, specifically including: if the intrinsic mode function component z in EEMD i If (t) is not a standard intrinsic mode function component, then EMD decomposition is performed again:

[0077]

[0078] Where d1(t) is the first eigenmode function component obtained by EMD decomposition of the first quasi-eigenmode function (eigenmode function component) component z1(t), q1(t) represents the superposition of the remaining residual components, and the final MEEMD can be expressed as:

[0079] x ( t ) → ∑ L = 1 m [ d i ( t ) ] + r(t)

[0080] Where, d i (t) represents the final intrinsic mode function component, where i=1,2,.......,m, r(t) represents the final residual component. The specific number of decompositions and the number of modal functions obtained, m, are determined by whether they are standard mode function components, which can maximize the preservation of the information characteristics contained in the signal while eliminating white noise.

[0081] The Hilbert transform (HT) processes the obtained intrinsic mode function to obtain a time-frequency diagram, which specifically includes:

[0082] According to the formula

[0083]

[0084] It can be seen that for the eigenmode function d i (t) Hilbert transform is the function and By performing convolution, the instantaneous frequency and instantaneous amplitude of the corresponding intrinsic mode function component can be obtained, thereby obtaining the time-frequency diagram of the corresponding intrinsic mode function component as the input of the training model.

[0085] The construction of a multi-channel convolutional neural network fault diagnosis model specifically includes:

[0086] The basic components of the multi-channel convolutional neural network model are convolutional layer, pooling layer, convolutional layer and global average pooling instead of the softmax classifier connected by the fully connected layer. The fully connected layer is replaced by a 1×1 convolutional layer and global average pooling to reduce overfitting, improve computational efficiency, and avoid the gradient vanishing or exploding problem of the fully connected layer.

[0087] The mathematical model of the convolutional layer in the model can be expressed as:

[0088]

[0089] For the Tier The output of a neuron; For the Tier The input of a neuron; M j is the input feature map; For the Layer network; ω is the weight matrix, i.e., the filter; For the Layer The network bias of each neuron;

[0090] σ is the activation function. In this example, the σ activation function is the Sigmoid function, and its mathematical expression is:

[0091]

[0092] The activation function is used to delinearize and enhance the expressive power.

[0093] The number of feature extraction network groups of the multi-channel convolutional neural network model is set to 2, and each layer consists of 1 convolution and 1 downsampling layer.

[0094] The classification network consists of 1 convolution layer, 1 global average pooling layer and 1 softmax output layer; 1 convolution layer and 1 global average pooling layer are the replaced fully connected layers, which are used to compress the number of channels of the last pooling layer of the model, so that the number of channels is the same as the number of fault categories, while reducing the risk of overfitting and avoiding the gradient vanishing or exploding of the fully connected layer. Global average pooling converts the output feature map into a one-dimensional feature vector.

[0095] The sizes of the front and rear convolution kernels are set to 5 and 4 respectively, and the sizes of the pooling layers are both set to 2.

[0096] The standard intrinsic mode function components d corresponding to the three groups of signals of vibration, torque and speed of the same spring-tooth pepper harvester drum fault sample are i The time-frequency diagram of (t) is extracted by uniform sampling and converted into a 3-channel input with a size of 30×30×3; the training set signal data is input into the model for iterative training, and the dropout technology is used to improve the diagnostic accuracy and prevent overfitting. The coefficient size is set to 0.99; the exponential decay method is used to set the learning rate to improve the training stability, and the decay rate size is 0.9.

[0097] The training set data is input for iterative training and the parameters are optimized to obtain the fault diagnosis model of the drum of the spring-tine pepper harvester.

[0098] The divided test set data is input into the constructed spring-tine pepper harvester drum fault diagnosis model to evaluate the performance of the model.

[0099] Different fault categories are divided according to the location of the fault on the drum:

[0100] The working conditions of the drum of the spring-tooth pepper harvester are divided into normal and faulty. The faults can be specifically divided into shaft faults and comb tooth faults. The shaft faults can be specifically divided into four types, namely loose bearing seat, imbalance, misalignment, and friction. The comb tooth faults can be specifically divided into two types, namely comb tooth deformation and comb tooth missing.

[0101] A softmax fault classifier for a spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion is as follows:

[0102] In the example of the present invention, global average pooling converts the output feature map into a one-dimensional feature vector and passes it into the softmax classifier.

[0103] The softmax classifier maps the input one-dimensional feature vector to the category, and the result is given in the form of probability. The formula is as follows:

[0104]

[0105] In the example of the present invention, x represents the input one-dimensional feature vector, y represents the corresponding fault category label, the maximum value of K is 6, and categories 1, 2, 3, 4, 5, and 6 correspond to the 6 fault categories of the drum, that is, the 6 channels of the softmax classifier.

[0106] In the formula, θ k = [ θ 1 θ 2 ⋅⋅⋅ θ K ] T is the weight parameter, corresponding to the classifier parameter θ k Obtained by training the softmax classifier, it can be used as a parameter to calculate the probabilities of all possible categories of the items to be classified.

[0107] Parameter θ k Multiply it with the one-dimensional feature vector and output the probability values ​​belonging to each category. The probability values ​​correspond to the corresponding fault categories, completing the fault identification and classification.

[0108] Regarding the softmax fault classifier described in the present invention, it should be noted that the number of channels compressed by the fully connected layer replaced by the 1×1 convolutional layer and the pooling layer is 6, and the 6 channels belong to 2 major types of faults, corresponding to 6 fault categories.

[0109] The input time-frequency graph is subjected to feature dimension reduction and feature extraction by a multi-channel convolutional neural network and then input into a softmax classifier for classification. The output is the probability P of the corresponding fault time-frequency graph belonging to each type of roller fault. y , P y It represents the probability value corresponding to the fault category, and its size sum is 1; compare the output probability value P y The roller fault classification channel corresponding to its maximum probability value is the diagnosed roller fault category.

[0110] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for diagnosing drum faults of spring-tooth pepper harvesters based on multi-source information fusion, characterized in that , including the following steps: Step S1: collecting vibration signals on the drum shaft, wall plate, and frame of the spring-tooth pepper harvester drum under different fault conditions and torque and speed signals of the drum shaft; Step S2: The three groups of signal data of vibration, torque and speed collected under the same fault are processed by improved lumped average empirical mode decomposition method (MEEMD) and Hilbert transform (HT) to obtain time-frequency diagrams, which are input into the improved multi-channel convolutional neural network diagnosis model; Step S3: After the improved multi-channel convolutional neural network diagnosis model is subjected to convolution and pooling for feature dimension reduction and feature extraction, the corresponding spring-tooth pepper harvester drum fault category probability is obtained through the softmax classifier to further determine the fault category.

2. The method for diagnosing cylinder faults of a spring-tine pepper harvester based on multi-source information fusion according to claim 1 is characterized in that: The construction and training of the spring-tine pepper harvester drum fault diagnosis model specifically includes: The vibration signals of the drum shaft, wall plate and frame of the spring-tine pepper harvester under different fault conditions and the torque and speed signal data of the drum shaft are divided into training set and test set. The drum vibration, torque and speed signals collected under the same fault condition are processed by improved ensemble mean empirical mode decomposition method (MEEMD) and Hilbert transform (HT) to obtain the time-frequency diagram of the three signal components. The obtained time-frequency diagram is input into an improved multi-channel convolutional neural network (CNN), which extracts features from the time-frequency diagram of the three signal components of vibration, torque and speed of the fault sample through 2-layer convolution and pooling operations. In the multi-channel convolutional neural network (CNN), the convolution layer and the pooling layer perform feature dimension reduction and feature extraction, and the softmax classifier is connected through the fully connected layer replaced by the 1×1 convolution layer and the global average pooling to obtain the probability value corresponding to the fault type. The accuracy of the model is verified by using the independent test set signal data to ensure that it meets the preset performance requirements. If not, the parameter adjustment training is performed again to complete the training of the spring-tine pepper harvester drum fault diagnosis model.

3. The method for diagnosing cylinder faults of a spring-tine pepper harvester based on multi-source information fusion according to claim 1 is characterized in that: The vibration, torque and speed signals collected under the same drum fault condition are preprocessed by an improved lumped mean empirical mode decomposition method (MEEMD) to obtain a number of intrinsic mode function components, and the decomposed intrinsic mode function components are ensured to be close to the standard intrinsic mode function components, and then the obtained intrinsic mode function components are convoluted using Hilbert transform (HT) to obtain a time-frequency diagram, which specifically includes: A white noise signal of a certain amplitude is added to the signal so that its RMS value is equivalent to the noise of the signal to be decomposed, and the difference between the RMS value and the signal to be decomposed does not exceed 10% to eliminate the aliasing phenomenon; two groups of positive and negative white noise signals y(t) with equal absolute values ​​are added to the signal to be decomposed x(t), and the lumped average empirical mode decomposition (EEMD) is performed respectively to obtain two groups of results z i+ (t) and z i- (t), and then the two sets of results are averaged to eliminate the residual white noise; if the intrinsic mode function component z is obtained by lumped average empirical mode decomposition (EEMD) i (t) is not a standard eigenmode function component, then z i (t) Perform empirical mode decomposition (EMD) to obtain the standard intrinsic mode function component d i (t), and superimpose the residual components of the existing modal components to maximize the original information of the signal; finally, the intrinsic mode function d i (t) Perform Hilbert transform (HT) to obtain the time-frequency diagram of the corresponding intrinsic mode function component.

4. According to claim 2, a spring-tine pepper harvester drum fault diagnosis model is constructed and trained, and the basic characteristics of the model are: a multi-channel convolutional neural network model, whose main components are a convolution layer, a pooling layer, a 1×1 convolution layer, a fully connected layer replaced by a global average pooling, and a softmax classifier; the number of feature extraction network groups is set to 2, and each layer consists of 1 convolution layer and 1 downsampling layer; the fully connected layer compresses the number of channels to the same number of fault categories in the softmax fault classifier; the standard intrinsic mode function component d corresponding to the three groups of signals of vibration, torque, and speed of the same spring-tine pepper harvester drum fault sample is converted into the standard intrinsic mode function component d i The time-frequency diagram of (t) is extracted by uniform sampling and converted into a 3-channel input with a size of 30×30×3; finally, the training set signal data is input into the model for iterative training, and the dropout technology is used to improve the diagnostic accuracy and prevent overfitting; the exponential decay method is used to set the learning rate to improve the training stability.

5. According to claim 2, different types of spring-tine pepper harvester drum failures are classified, characterized in that: The failure categories of the drum of the spring-tooth pepper harvester are divided into two categories: drum shaft failure and comb tooth failure. The drum shaft failure can be specifically divided into four types, namely, loose bearing seat, imbalance, misalignment, and friction. The drum comb tooth failure can be specifically divided into two types, namely, comb tooth deformation and comb tooth missing.

6. The softmax classifier in the model of the spring-tine pepper harvester drum fault diagnosis method based on multi-source information fusion according to claim 2 is characterized by: After the three groups of time-frequency graphs of the same fault sample are input, the neural network convolution and pooling features are extracted, and the feature map is converted into a one-dimensional feature vector through a fully connected layer replaced by a convolution layer and a global average pooling layer, and then input into the softmax classifier; the input one-dimensional feature vector is calculated by matrix multiplication, and the probability value of the fault category corresponding to the drum fault of the spring-tine pepper harvester is output.

7. According to claim 6, the input time-frequency diagram is input into the softmax classifier after feature extraction by a multi-channel convolutional neural network, and the probability values ​​corresponding to the six fault categories are output. The probability values ​​are compared, and the fault classification channel corresponding to the maximum probability value is the fault category of the spring-tooth pepper harvester drum determined by it, thereby completing fault identification.

Images