Conveyor belt wear diagnosis method and system based on GoogLeNet neural network

Abstract

The invention discloses a conveyor belt wear diagnosis method and system based on a GoogLeNet neural network, and the method comprises the steps: capturing images of different surfaces of a conveyor belt of a coal preparation plant through employing a CMOS industrial camera, and constructing an image wear data set; the image wear data set is preprocessed, and a new image wear data set is obtained through noise adding, fisheye transformation, filtering and histogram equalization; and constructing a GoogLeNet neural network model, and classifying the images in the new image wear data set by using the model to complete wear diagnosis of the conveyor belt. And evaluating the classification result. According to the method, the diagnosis precision is remarkably improved, the safety of equipment is enhanced, and the labor cost is reduced through automatic detection.

Description

Technical Field

[0001] The present invention relates to the technical field of coal mine conveyor belt wear diagnosis, and in particular to a conveyor belt wear diagnosis method and system based on a GoogLeNet neural network. Background Art

[0002] During the production and operation of coal mining enterprises, a large amount of coal needs to be conveyed for a long time and at a high load. Therefore, the wear and tear of the conveyor belt have become important indicators to measure the safety of the conveyor belt. At present, the research on conveyor belt failure mainly focuses on acute failures such as belt tear protection, deviation and foreign body detection, and there is little research on chronic failures such as conveyor belt wear.

[0003] In this context, neural networks have been introduced into the field of fault diagnosis. However, the traditional methods based on grayscale analysis and image feature extraction have low fault diagnosis accuracy, and some faults often rely on manual experience judgment, which is highly subjective and the diagnosis effect is not ideal. Therefore, a fault diagnosis method with less manual participation and higher accuracy is urgently needed. Summary of the invention

[0004] The purpose of the present invention is to provide a conveyor belt wear diagnosis method and system based on the GoogLeNet neural network, which can more accurately and efficiently diagnose the wear of the conveyor belt in the coal preparation plant, thereby providing strong support for the on-site production of the coal preparation plant and reducing economic losses. It also provides a new, flexible and effective tool for researchers in this field, and promotes scientific research and industrial development in related fields.

[0005] The present invention adopts the following technical solution:

[0006] A conveyor belt wear diagnosis method based on GoogLeNet neural network includes the following steps:

[0007] S1. Use CMOS industrial cameras to capture images of different surfaces of the conveyor belt in a coal preparation plant and construct an image wear dataset.

[0008] S2. Preprocess the image wear dataset by adding noise, fisheye transformation, filtering and histogram equalization to obtain a new image wear dataset.

[0009] S3. Build a GoogLeNet neural network model and use it to classify images in the new image wear dataset to complete the wear diagnosis of the conveyor belt.

[0010] Among them, the GoogLeNet neural network model includes an input layer, 9 Inception v1 modules with dimensionality reduction function, 5 global average pooling layers, a convolution layer, a fully connected layer and 3 output layers; a 1x1 convolution layer is added before the 3x3 convolution layer, before the 5x5 convolution layer and after the 3x3 max pooling of the original Inception v1 module to obtain an Inception v1 module with dimensionality reduction function, which includes a convolution layer and a pooling layer; 2 of the output layers are auxiliary classification layers.

[0011] Furthermore, in step S2, a new image wear dataset is obtained including the following contents:

[0012] S201, adding Gaussian noise to the images in the image wear dataset, the specific expression of Gaussian noise is:

[0013]

[0014] Where G(x) represents Gaussian noise, C represents the image, σ represents the standard deviation of the distribution, μ represents the mean of the distribution, and σ 2 Represents the variance of the distribution.

[0015] S202, performing fisheye transformation on the image obtained in step S201, the fisheye transformation function expression is:

[0016]

[0017] Among them, c x 、c y Both represent the center of the transformation, d represents the distortion factor, and T(x) and T(y) represent the transformation values ​​of image pixels x and y respectively.

[0018] The image pixel positions are normalized to a grid, with the upper left pixel representing the position (-1, -1), the lower right pixel representing the position (1, 1), (0, 0) representing the exact center pixel of the image, and the x-axis and y-axis ranging from -1 to 1.

[0019] S203, using frequency domain filtering to process the image obtained in step S202, using Fourier transform to convert the image from the spatial domain to the frequency domain, and combining signal processing technology to filter different frequency domain points, and then using inverse Fourier transform to convert the image back to the spatial domain.

[0020] S204, performing histogram equalization on the image obtained in step S203 to obtain a new image wear data set. The specific formula is:

[0021]

[0022] Among them, pr () represents the overall probability density function of the grayscale value distribution of pixels in the image, r k Indicates the gray level with gray value k, n k represents the number of pixels with gray value k, and n represents the total number of pixels.

[0023] The calculation formula of the cumulative grayscale histogram is:

[0024]

[0025] Among them, F r () represents the distribution function, r j Represents the gray level with gray value j.

[0026] Let F r (r k ) is s k , will s k Merge to the nearest grayscale value to get a new grayscale value. The specific formula is:

[0027] s ′ k =(L-1)s k

[0028] Among them, s ′ k Represents the new gray value, and L represents the total number of gray values.

[0029] For ′ k Round off to establish r k →s ′ k The mapping relationship is as follows:

[0030] s ′ k =T(r k )

[0031] Wherein, T() represents the mapping function.

[0032] According to the mapping relationship, the image after histogram equalization is drawn. The specific formula is:

[0033]

[0034] Among them, p s () represents the overall grayscale distribution characteristics of the image after histogram equalization, p s (s ′ k ) represents the image after histogram equalization ′ k The marginal probability distribution of r represents the gray level, T-1 (s ′ k ) represents the set of original grayscale values.

[0035] Further, in step S203, the Fourier transform includes the following contents:

[0036] The image function f(x, y) of size A*B performs a discrete Fourier transform of length B on y, and then performs a Fourier transform of length A on x to obtain the Fourier transform result of the image. The specific formula is:

[0037]

[0038] Among them, x and y represent image pixels, x = [0, 1, 2, ···, A-1], y = [0, 1, 2, ···, B-1], F(u, v) represents the Fourier transform result of f(x, y), F(u, v) represents the Fourier transform result of f(x, y), u and v both represent coordinates in the frequency domain, u = [0, 1, 2, ···, A-1], v = [0, 1, 2, ···, B-1].

[0039] The A*B matrix defined by u and v is a frequency domain matrix.

[0040] Further, in step S203, the inverse Fourier transform includes the following contents:

[0041]

[0042] Furthermore, in step S3, completing the wear diagnosis of the conveyor belt includes the following:

[0043] S301 and the auxiliary classification layer are located behind the second and fourth Inception v1 modules with dimensionality reduction function, respectively, and are used to classify the extracted features. The new image wear dataset in step S2 is integrated, and then scaled, normalized and data enhanced. The processed dataset is randomly allocated into a training set and a test set in a ratio of 7:3. The training set is used to train the GoogLeNet neural network model and optimize it using the loss function and the Adam optimizer. The test set is used to verify the classification accuracy of the model.

[0044] S302, using mini-batch technology, divide the training set into a set number of small batches, each small batch includes 10 samples, input the small batches into the GoogLeNet neural network model, obtain the predicted probability of each category, and use the cross entropy loss function to calculate the loss value of the small batch to obtain the difference between the predicted result and the true label; the specific expression of the cross entropy loss function is:

[0045]

[0046] Among them, L represents the cross entropy loss function, N represents the total number of categories, and y i represents the true label of the i-th category, P i Represents the predicted probability of the i-th category.

[0047] S303. Calculate the gradient of the loss function relative to the model parameters and use the Adam optimizer to adjust the weights and biases of the model. The specific formula is:

[0048]

[0049] Among them, w represents the weight of the model, w′ represents the updated model weight, η represents the learning rate, represents the gradient of the loss function.

[0050] S304, repeating steps S302-S303 until the set number of training rounds is reached, and obtaining a trained GoogLeNet neural network model.

[0051] S305, input the test set into the trained GoogLeNet neural network model, transfer it to the Inception v1 module with dimensionality reduction function through the input layer, reduce the depth of the image through a 1x1 convolution kernel, and then extract features through 3x3 and 5x5 convolution kernels to obtain the corresponding feature map, merge the feature maps output by the 9 Inception v1 modules with dimensionality reduction function to obtain a new feature map; the new feature map passes through a global average pooling layer to reduce its spatial dimension to 1×1 and retain the corresponding features to obtain a first feature map; the first feature map passes through a convolution layer to further extract features, and the extracted features are combined through a fully connected layer to obtain the predicted probability of the corresponding category, which is output through the output layer and standardized using the Softmax activation function, and the category with the largest probability is taken as the final classification result.

[0052] The calculation formula of the Softmax activation function is:

[0053]

[0054] Among them, z i represents the linear combination output of the neurons in the output layer corresponding to the i-th category, z j represents the linear combination output of the neurons in the output layer corresponding to the jth category.

[0055] Furthermore, the final classification results are evaluated to calculate their accuracy, precision and recall. The specific formula is:

[0056]

[0057] Among them, Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive examples, TN represents the total number of samples correctly predicted as negative examples, FP represents the total number of samples incorrectly predicted as positive examples, FN represents the total number of samples incorrectly predicted as negative examples, Precision represents the precision rate, and Recall represents the recall rate.

[0058] When the accuracy, precision and recall rates are all above 80%, it means that the final classification results meet the requirements; otherwise, increase the samples of the training set, adjust the hyperparameters or optimize the GoogLeNet neural network model, and use the training set and test set to process the model again.

[0059] Furthermore, the present invention also proposes a conveyor belt wear diagnosis system based on the GoogLeNet neural network, comprising:

[0060] The dataset construction module is used to capture images of different surfaces of the conveyor belt in a coal preparation plant using a CMOS industrial camera to construct an image wear dataset.

[0061] The data set processing module is used to preprocess the image wear data set, and obtain a new image wear data set through adding noise, fisheye transformation, filtering and histogram equalization.

[0062] The data classification module is used to build a GoogLeNet neural network model and use the model to classify images in the new image wear dataset to complete the wear diagnosis of the conveyor belt. The GoogLeNet neural network model includes an input layer, 9 Inception v1 modules with dimensionality reduction function, 5 global average pooling layers, a convolution layer, a fully connected layer, and 3 output layers. A 1x1 convolution layer is added before the 3x3 convolution layer, before the 5x5 convolution layer, and after the 3x3max pooling of the original Inception v1 module to obtain an Inception v1 module with dimensionality reduction function, which includes a convolution layer and a pooling layer. Two of the output layers are auxiliary classification layers.

[0063] The result evaluation module is used to evaluate the final classification results and calculate their accuracy, precision and recall.

[0064] Furthermore, the present invention also proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor implements the steps of the conveyor belt wear diagnosis method based on the GoogLeNet neural network when executing the computer program.

[0065] Furthermore, the present invention also proposes a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and when the computer program is run by a processor, the conveyor belt wear diagnosis method based on the GoogLeNet neural network is executed.

[0066] Compared with the prior art, the present invention adopts the above technical solution and has the following technical effects:

[0067] The method proposed in the present invention realizes automatic and accurate identification of the conveyor belt status through deep learning technology, significantly improves the diagnostic accuracy, supports real-time monitoring, timely discovers potential wear problems, and enhances the safety of the equipment. At the same time, data-driven decision support enables management to scientifically maintain the conveyor belt and extend the service life of the conveyor belt. In addition, automated detection reduces labor costs, has good applicability and broad application prospects, and provides an effective basis for equipment management. BRIEF DESCRIPTION OF THE DRAWINGS

[0068] Figure 1 It is an overall implementation flow chart of the present invention.

[0069] Figure 2 It is a structural diagram of the Inception v1 module of the present invention with the dimensionality reduction function added. DETAILED DESCRIPTION

[0070] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and cannot be used to limit the protection scope of the present invention.

[0071] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments.

[0072] It should be noted that, in the description of the present invention, unless otherwise specified, "plurality" means two or more than two; the terms "upper", "lower", "left", "right", "inner", "outer", "front end", "rear end", "head", "tail" and the like indicate the orientation or position relationship based on the orientation or position relationship shown in the drawings, which is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present invention.

[0073] Furthermore, the terms “first,” “second,” “third,” etc. are used for descriptive purposes only and are not to be construed as indicating or implying relative importance.

[0074] At the same time, in the description of the present invention, unless otherwise clearly specified and limited, the terms "connected" and "connection" should be understood in a broad sense, for example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances.

[0075] To achieve the above object, the present invention proposes a conveyor belt wear diagnosis method based on GoogLeNet neural network, such as Figure 1 As shown, the specific steps are as follows:

[0076] S1. Use CMOS industrial cameras to capture images of different surfaces of the conveyor belt in a coal preparation plant and construct an image wear dataset.

[0077] S2. Since the number of available worn belt images is small, which affects the adequacy of training data, it is necessary to process the existing images to enrich the dataset. Simple translation or flipping alone cannot effectively enhance the persuasiveness of the training set, so a more complex nonlinear method is used to process the images to improve the effectiveness of the dataset and obtain a new image wear dataset; the specific content is:

[0078] S201. Although noise may interfere with the assessment of conveyor belt image wear, most noise will not affect the judgment of the degree of belt wear. Therefore, in order to enrich the data set and improve the adaptability and scalability of the training network, Gaussian noise is added to the images in the image wear data set. Gaussian noise refers to noise whose probability density function obeys a normal distribution. The specific expression is:

[0079]

[0080] Where G(x) represents Gaussian noise; C represents the image; σ represents the standard deviation of the distribution, and the larger the value of σ, the more obvious the Gaussian noise; μ represents the mean of the distribution; σ 2 Represents the variance of the distribution.

[0081] S202, perform fisheye transformation on the image obtained in step S201 to change the shape of the wear stripes but not the number of wear layers, thereby enhancing the data set. The fisheye transformation distorts the pixels in the image based on the distance from a given central pixel, expands the middle of the image and compresses the edges, changing the shape of the wear stripes. By adding fisheye lenses to multiple parts of the conveyor belt image, the supplemented data set is significantly different from the original image. The fisheye transformation function expression is:

[0082]

[0083] Among them, cx 、c y Both represent the center of the transformation, d represents the distortion factor, and T(x) and T(y) represent the transformation values ​​of image pixels x and y respectively.

[0084] All image pixel positions are normalized to a grid, with the upper left pixel representing position (-1, -1), the lower right pixel representing position (1, 1), (0, 0) representing the exact center pixel of the image, and the x- and y-axes ranging from -1 to 1.

[0085] S203, using frequency domain filtering to process the image obtained in step S202, and using Fourier transform to convert the image from the spatial domain to the frequency domain, the specific content is:

[0086] For the image function f(x, y) of size A*B, perform a discrete Fourier transform of length B on its row variable y, and then perform a Fourier transform of length A on its column variable x to obtain the Fourier transform result of the image. The specific formula is:

[0087]

[0088] Wherein, x=[0,1,2,···,A-1], y=[0,1,2,···,B-1], F(u,v) represents the Fourier transform result of f(x,y), the coordinate system where F(u,v) is located is called the frequency domain, u and v both represent the coordinates in the frequency domain, u=[0,1,2,···,A-1], v=[0,1,2,···,B-1].

[0089] The A*B matrix defined by u=[0, 1, 2, .. . , A-1] and v=[0, 1, 2, .. . , B-1] is a frequency domain matrix.

[0090] By combining signal processing technology to filter different frequency domain points, the images processed in this way better meet the requirements of deep learning for the robustness of image sets.

[0091] The image is converted back to the spatial domain using the inverse Fourier transform. The specific formula is:

[0092]

[0093] S204, the original image may be not clear enough because its grayscale distribution is concentrated in a narrow interval. By changing the grayscale distribution of the original image to obtain a uniform histogram, the dynamic range of the grayscale value difference between pixels is increased, thereby enhancing the overall contrast of the image. The too dark and too bright images will become clear after histogram equalization. Therefore, the image histogram obtained in step S203 is equalized to calculate the grayscale value histogram. The specific formula is:

[0094]

[0095] Among them, p r () represents the overall probability density function of the grayscale value distribution of pixels in the image, r k Indicates the gray level with gray value k, n k represents the number of pixels with gray value k, and n represents the number of all pixels, which is used for normalization.

[0096] When p r (r k ) is 1, it can be regarded as a probability density function, then the cumulative grayscale histogram is to find the probability density function p r (r k ) distribution function F r (r k ); Calculate the cumulative grayscale histogram, the specific formula is:

[0097]

[0098] Among them, F r () represents the distribution function, r j Indicates the gray level with gray value j, n j Represents the number of pixels with gray value j;

[0099] Let F r (r k ) is s k , will s k Merge to the nearest grayscale value to get a new grayscale value. The specific formula is:

[0100] s ′ k =(L-1)s k

[0101] Among them, s ′ k Represents the new gray value, and L represents the total number of gray values.

[0102] If the grayscale value ranges from 0 to 5, then L=6.

[0103] For ′ k Round off to establish r k →s ′ k The mapping relationship is expressed as:

[0104] s ′ k =T(r k )

[0105] Among them, r krepresents the gray level with gray value k, and T() represents the mapping function.

[0106] According to the mapping relationship, the image after histogram equalization is drawn. The specific formula is:

[0107]

[0108] Among them, p s () represents the overall grayscale distribution characteristics of the image after histogram equalization; p s (s ′ k ) represents the gray value s in the image after histogram equalization ′ k The marginal probability distribution of reflects the gray value s in the equalized image ′ k Total frequency of occurrence; T -1 (s ′ k ) represents the set of original grayscale values, which will be mapped to new grayscale values ​​s under the action of the mapping function. ′ k .

[0109] r∈T -1 (s ′ k ) is because the new gray value s ′ k The mapping to raw grayscale values ​​may be a one-to-many relationship.

[0110] S3. Build a GoogLeNet neural network model and use it to classify images in the new image wear dataset to complete the wear diagnosis of the conveyor belt. The specific contents are as follows:

[0111] S301, GoogLeNet neural network model includes input layer, 9 Inception v1 modules with dimensionality reduction function, 5 global average pooling layers, several convolutional layers, fully connected layers and 3 output layers, totaling 22 layers; the Inception v1 module with dimensionality reduction function includes multiple convolutional layers and pooling layers; 2 of the output layers are auxiliary classification layers, which calculate the loss from their output to the true label and update the network parameters through back propagation, which helps to solve the problem of gradient disappearance and ensure information transmission during training. At the same time, these auxiliary layers can better guide the learning of the model and improve the overall recognition accuracy.

[0112] Among them Figure 2The figure shows the structure of the Inception v1 module with dimensionality reduction. The top Filterconcatenation indicates that the feature maps from different convolutional layers and pooling layers are spliced ​​together in this step. Multiple 1x1convolutions represent 1x1 convolution layers for dimensionality reduction, which helps reduce the computational burden of subsequent convolutions. 3x3convolutions and 5x5convolutions, these layers are used to extract features of different scales. 3x3maxpooling shows the pooling operation, which helps further reduce the dimension of the feature map and enhance the stability of the feature. The improvement of the Inception structure diagram is that 1x1 convolution kernels are added before 3x3convolutions, before 5x5convolutions, and after 3x3maxpooling, respectively, to reduce the thickness of the feature map.

[0113] The auxiliary classification layers are placed after the 2nd and 4th Inception v1 modules, which allows the network to introduce additional classification tasks at deeper levels to enhance the transfer of information. Each auxiliary classification layer is responsible for classifying the features extracted by a specific layer.

[0114] The GoogLeNet neural network model has a deeper structure. Through the parallel connection of the Inception v1 module, it can extract multiple features at the same time and expand in depth and width. At the same time, based on the original Inception structure, a structure with dimensionality reduction function is added. A convolution layer with a convolution kernel size of 1x1 is added to the branch, which can significantly reduce the dimensionality and reduce the model training parameters.

[0115] S302, integrate the new image wear dataset in step S2, and then perform scaling, normalization and data enhancement to improve the generalization ability of the model; randomly allocate the processed dataset into a training set and a test set in a ratio of 7:3, the training set is used for model learning, and is optimized using a loss function and an Adam optimizer; the test set is used to verify the classification accuracy of the model.

[0116] S303, using mini-batch technology, divide the training set into multiple small batches, each small batch contains 10 samples. This method can not only speed up the training process, but also effectively reduce the consumption of computing resources. Input the small batch into the GoogLeNet neural network model to obtain the predicted probability of each category, and use the cross entropy loss function to calculate the loss value of the small batch to obtain the difference between the predicted result and the true label; the learning rate is set to 0.0001, and the specific expression of the cross entropy loss function is:

[0117]

[0118] Among them, L represents the cross entropy loss function, N represents the total number of categories, and y i represents the true label of the i-th category, P i Represents the predicted probability of the i-th category.

[0119] S304: After the loss calculation is completed, back propagation is performed to calculate the gradient of the loss function relative to the model parameters, and the Adam optimizer is used to adjust the weights and biases of the model to reduce the loss. The specific formula is:

[0120]

[0121] Among them, w represents the weight of the model, w′ represents the updated model weight, η represents the learning rate, represents the gradient of the loss function.

[0122] S305, repeat iteration: repeat steps S303-S304 until the model converges or reaches the set number of training rounds to obtain a trained GoogLeNet neural network model.

[0123] S306, input the test set into the trained GoogLeNet neural network model, transmit it to the Inception v1 module with dimensionality reduction function through the input layer, reduce the depth of the image through the 1x1 convolution kernel, reduce the calculation amount and the number of parameters of the subsequent convolution layer, and retain important feature information, and then extract features through 3x3 and 5x5 convolution kernels to obtain corresponding feature maps, merge the feature maps output by the 9 Inception v1 modules with dimensionality reduction function to obtain a new feature map; the new feature map passes through the global average pooling layer to reduce its spatial dimension to 1×1, further reduce the number of parameters, prevent overfitting, and retain important feature information to obtain the first feature map; the first feature map passes through the convolution layer to extract higher-level features, and the extracted features are combined through the fully connected layer for higher-level learning to obtain the predicted probability of the corresponding category, which is output through the output layer and standardized using the Softmax activation function, and the category with the largest probability is taken as the final classification result.

[0124] In this embodiment, the wear degree is divided into 6 categories, namely, wear layer 0, layer 1, layer 2, layer 3, layer 4, and layer 5, and these images are stored in different folders.

[0125] The calculation formula of the Softmax activation function is:

[0126]

[0127] Among them, z irepresents the linear combination output of the neurons in the output layer corresponding to the i-th category, z j represents the linear combination output of the neurons in the output layer corresponding to the jth category.

[0128] It is the process of summing the index values ​​of all categories, ensuring that the sum of the probabilities of all categories is 1.

[0129] S4. Evaluate the final classification results and calculate their accuracy, precision and recall, which can help judge the performance of the model and adjust and optimize the model if necessary. Accuracy refers to the ratio of the number of samples with the correct class to the total number of samples, precision refers to the proportion of true positive examples in the samples predicted by the model as positive examples, and recall refers to the proportion of true positive examples in all actual positive examples. The specific formula is:

[0130]

[0131] Among them, Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive examples, TN represents the total number of samples correctly predicted as negative examples, FP represents the total number of samples incorrectly predicted as positive examples, FN represents the total number of samples incorrectly predicted as negative examples, Precision represents the precision rate, and Recall represents the recall rate.

[0132] Accuracy is used to measure the overall predictive ability of the model. A high accuracy means that the model can correctly classify most of the input samples. However, accuracy can be misleading in the case of class imbalance. For example, if a certain class occupies a large number of samples, the model can achieve high accuracy even if it only makes correct predictions for this class. Therefore, relying solely on accuracy to evaluate the model is not comprehensive enough.

[0133] High precision means that the model is highly reliable when predicting positive examples, which can reach more than 80%, that is, there are fewer misjudgments (false positives).

[0134] A high recall rate means that the model can identify more than 80% of positive examples, which means that there are fewer missed positives (false negatives).

[0135] If the model performs poorly in the test, take the following actions:

[0136] (1) Data augmentation: Supplement the training set by adding more samples or using more complex image processing methods to improve the generalization ability of the model.

[0137] (2) Hyperparameter adjustment: Modify hyperparameters such as learning rate and batch size to find a more suitable training strategy.

[0138] (3) Model optimization: Adjust the number of layers and neurons of the existing architecture to improve the learning ability of the model.

[0139] (4) Cross-validation: Use cross-validation techniques to evaluate the stability of the model to ensure that the model performs well on different subsets.

[0140] In order to intuitively demonstrate the classification effect of the model, a confusion matrix is ​​generated, in which each row of the matrix represents the actual category and each column represents the predicted category. At the same time, the recognition of the degree of damage is displayed through a visualization tool to provide a reference for subsequent practical applications.

[0141] After the training is completed, use the model saving function to save the trained model parameters for subsequent loading and use. Use the following function to save the model:

[0142] model.save('model.h5')#Save model

[0143] from keras.models import load_model

[0144] model=load_model('model.h5')#load model

[0145] The method proposed in the present invention can effectively identify different degrees of wear and tear, and provide a scientific basis for the daily maintenance and management of a coal preparation plant, thereby reducing economic losses and improving production efficiency.

[0146] The embodiment of the present invention also proposes a conveyor belt wear diagnosis system based on the GoogLeNet neural network, including a data set construction module, a data set processing module, a data classification module, a result evaluation module and a computer program that can be run on a processor. It should be noted that each module in the above system corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0147] The embodiment of the present invention further provides an electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor. It should be noted that when the processor executes the computer program, the specific steps of the method provided in the embodiment of the present invention are corresponding to the specific steps of the method provided in the embodiment of the present invention, and the processor has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0148] The embodiment of the present invention further provides a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program. It should be noted that when the computer program is executed by the processor, it corresponds to the specific steps of the method provided in the embodiment of the present invention, and has the functional modules and beneficial effects corresponding to the execution method. For technical details not described in detail in this embodiment, please refer to the method provided in the embodiment of the present invention.

[0149] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A conveyor belt wear diagnosis method based on GoogLeNet neural network, characterized in that: include: S1. Use CMOS industrial cameras to capture images of different surfaces of the conveyor belt in the coal preparation plant and construct an image wear dataset; S2, preprocessing the image wear dataset by adding noise, fisheye transformation, filtering and histogram equalization to obtain a new image wear dataset; S3, construct a GoogLeNet neural network model, and use the model to classify images in the new image wear dataset to complete the wear diagnosis of the conveyor belt; Among them, the GoogLeNet neural network model includes an input layer, an Inception v1 module with dimensionality reduction function, a global average pooling layer, a convolution layer, a fully connected layer, and an output layer; a 1x1 convolution layer is added before the 3x3 convolution layer, before the 5x5 convolution layer, and after the 3x3 max pooling of the original Inception v1 module to obtain an Inceptionv1 module with dimensionality reduction function, which includes a convolution layer and a pooling layer; there are two auxiliary classification layers in the output layer.

2. The conveyor belt wear diagnosis method based on GoogLeNet neural network according to claim 1 is characterized in that: In step S2, a new image wear dataset is obtained including the following contents: S201, adding Gaussian noise to the images in the image wear dataset, the specific expression of Gaussian noise is: Where G(x) represents Gaussian noise, C represents the image, σ represents the standard deviation of the distribution, μ represents the mean of the distribution, and σ 2 represents the variance of the distribution; S202, performing fisheye transformation on the image obtained in step S201, the fisheye transformation function expression is: Among them, c x 、c y Both represent the center of the transformation, d represents the distortion factor, T(x) and T(y) represent the transformation values ​​of image pixels x and y respectively; Normalize the image pixel positions into a grid, where the upper left pixel represents the position (-1, -1), the lower right pixel represents the position (1, 1), (0, 0) represents the exact center pixel of the image, and the x-axis and y-axis range from -1 to 1; S203, using frequency domain filtering to process the image obtained in step S202, using Fourier transform to convert the image from the spatial domain to the frequency domain, and combining signal processing technology to filter different frequency domain points, and then using inverse Fourier transform to convert the image back to the spatial domain; S204, performing histogram equalization on the image obtained in step S203 to obtain a new image wear data set. The specific formula is: Among them, p r () represents the overall probability density function of the grayscale value distribution of pixels in the image, r k Indicates the gray level with gray value k, n k represents the number of pixels with gray value k, and n represents the total number of pixels; The calculation formula of the cumulative grayscale histogram is: Among them, F r () represents the distribution function, r j represents the gray level with gray value j; Let F r (r k ) is s k , will s k Merge to the nearest grayscale value to get a new grayscale value. The specific formula is: s ′ k =(L-1)s k Among them, s ′ k Represents the new gray value, and L represents the total number of gray values; For ′ k Round off to establish r k →s ′ k The mapping relationship is as follows: s ′ k =T(r k ) Wherein, T() represents the mapping function; According to the mapping relationship, the image after histogram equalization is drawn. The specific formula is: Among them, p s () represents the overall grayscale distribution characteristics of the image after histogram equalization, p s (s ′ k ) represents the image after histogram equalization ′ k The marginal probability distribution of r represents the gray level, T -1 (s ′ k ) represents the set of original grayscale values.

3. The conveyor belt wear diagnosis method based on GoogLeNet neural network according to claim 2 is characterized in that: In step S203, the Fourier transform includes the following contents: The image function f(x, y) of size A*B performs a discrete Fourier transform of length B on y, and then performs a Fourier transform of length A on x to obtain the Fourier transform result of the image. The specific formula is: Wherein, x and y represent image pixels, x=[0,1,2,···,A-1], y=[0,1,2,···,B-1], F(u,v) represents the Fourier transform result of f(x,y), F(u,v) represents the Fourier transform result of f(x,y), u and v both represent coordinates in the frequency domain, u=[0,1,2,···,A-1], v=[0,1,2,···,B-1]; The A*B matrix defined by u and v is a frequency domain matrix.

4. The conveyor belt wear diagnosis method based on GoogLeNet neural network according to claim 2 is characterized in that: In step S203, the inverse Fourier transform includes the following contents: Among them, f(x, y) represents the image function, F(u, v) represents the inverse Fourier transform result of f(x, y), the size of the image is A*B, u and v both represent coordinates in the frequency domain, and x and y represent image pixels.

5. The conveyor belt wear diagnosis method based on GoogLeNet neural network according to claim 1 is characterized in that: In step S3, completing the wear diagnosis of the conveyor belt includes the following: S301,GoogLeNet neural network model includes input layer, 9 Inception v1 modules with dimensionality reduction, 5 global average pooling layers, convolutional layer, fully connected layer and 3 output layers; The auxiliary classification layers are located after the second and fourth Inception v1 modules with dimensionality reduction function, respectively, to classify the extracted features; The new image wear dataset in step S2 is integrated, scaled, normalized and data enhanced, and the processed dataset is randomly divided into a training set and a test set in a ratio of 7:

3. The training set is used to train the GoogLeNet neural network model and optimize it using the loss function and the Adam optimizer; the test set is used to verify the classification accuracy of the model; S302, using mini-batch technology, divide the training set into a set number of small batches, each small batch includes 10 samples, input the small batches into the GoogLeNet neural network model, obtain the predicted probability of each category, and use the cross entropy loss function to calculate the loss value of the small batch to obtain the difference between the predicted result and the true label; the specific expression of the cross entropy loss function is: Among them, L represents the cross entropy loss function, N represents the total number of categories, and y i represents the true label of the i-th category, P i Represents the predicted probability of the i-th category; S303. Calculate the gradient of the loss function relative to the model parameters and use the Adam optimizer to adjust the weights and biases of the model. The specific formula is: Among them, w represents the weight of the model, w′ represents the updated model weight, η represents the learning rate, represents the gradient of the loss function; S304, repeating steps S302-S303 until the set number of training rounds is reached, and obtaining a trained GoogLeNet neural network model; S305, input the test set into the trained GoogLeNet neural network model, transmit it to the Inception v1 module with dimensionality reduction function through the input layer, reduce the depth of the image through a 1x1 convolution kernel, and then extract features through 3x3 and 5x5 convolution kernels to obtain a corresponding feature map, merge the feature maps output by the 9 Inception v1 modules with dimensionality reduction function to obtain a new feature map; the new feature map passes through a global average pooling layer to reduce its spatial dimension to 1×1 and retain the corresponding features to obtain a first feature map; the first feature map passes through a convolution layer to further extract features, and the extracted features are combined through a fully connected layer to obtain a predicted probability of the corresponding category, which is output through the output layer and normalized using a Softmax activation function, and the category with the largest probability is taken as the final classification result; The calculation formula of the Softmax activation function is: Among them, z i represents the linear combination output of the neurons in the output layer corresponding to the i-th category, z j represents the linear combination output of the neurons in the output layer corresponding to the jth category.

6. The conveyor belt wear diagnosis method based on GoogLeNet neural network according to claim 5 is characterized in that: It also includes evaluating the final classification results and calculating its accuracy, precision and recall. The specific formula is: Among them, Accuracy represents the accuracy rate, TP represents the total number of samples correctly predicted as positive examples, TN represents the total number of samples correctly predicted as negative examples, FP represents the total number of samples incorrectly predicted as positive examples, FN represents the total number of samples incorrectly predicted as negative examples, Precision represents the precision rate, and Recall represents the recall rate; When the accuracy, precision and recall rates are all above 80%, it means that the final classification results meet the requirements; otherwise, increase the samples of the training set, adjust the hyperparameters or optimize the GoogLeNet neural network model, and use the training set and test set to process the model again.

7. A system for the conveyor belt wear diagnosis method based on the GoogLeNet neural network according to claim 1, characterized in that: include: The dataset construction module is used to capture images of different surfaces of the conveyor belt in the coal preparation plant using a CMOS industrial camera to construct an image wear dataset; The data set processing module is used to pre-process the image wear data set, and obtain a new image wear data set by adding noise, fisheye transformation, filtering and histogram equalization; The data classification module is used to build a GoogLeNet neural network model and use the model to classify images in the new image wear data set to complete the wear diagnosis of the conveyor belt. The GoogLeNet neural network model includes an input layer, 9 Inception v1 modules with dimensionality reduction function, 5 global average pooling layers, a convolution layer, a fully connected layer, and 3 output layers. A 1x1 convolution layer is added before the 3x3 convolution layer, before the 5x5 convolution layer, and after the 3x3max pooling of the original Inception v1 module to obtain an Inception v1 module with dimensionality reduction function, which includes a convolution layer and a pooling layer. Two of the output layers are auxiliary classification layers. The result evaluation module is used to evaluate the final classification results and calculate their accuracy, precision and recall.

8. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, the steps of the conveyor belt wear diagnosis method based on the GoogLeNet neural network described in any one of claims 1 to 6 are implemented.

9. A computer-readable storage medium storing a computer program, characterized in that: When the computer program is executed by a processor, the conveyor belt wear diagnosis method based on the GoogLeNet neural network described in any one of claims 1 to 6 is executed.

Images