A jujube machine blocking material plate dynamic adjustment method based on a neural network

Abstract

The application discloses a kind of based on neural network's red jujube machine baffle plate dynamic adjustment method, it is related to sorting equipment technical field, steps are as follows: the image that red jujube machine is shot in operation is collected, the background segmentation algorithm processing is carried out to the color image collected, and background area is removed;The image collected is classified, and different label numbers are given to each class, and good training data samples are constructed, neural network training is carried out;Design is used for the multi-branch convolutional neural network structure of classification, training is carried out using data set, and neural network model is obtained after training is completed;The image that camera real-time collection is input into neural network model, corresponding label value is given, and label value is stored to same position, and real-time update is carried out;Length fixed label value is calculated after real-time update, the proportion of each class is counted, and the direction of baffle plate adjustment is determined;According to the determined baffle plate adjustment scheme, motor is controlled, and adjustment is completed.The application can improve the efficiency of red jujube machine feeding sorting.

Description

Technical Field

[0001] This invention belongs to the field of sorting equipment technology, specifically relating to a method for dynamically adjusting the baffle of a jujube machine based on neural networks, thereby improving the efficiency of feeding and sorting jujubes. Background Technology

[0002] With the rise of artificial intelligence, machine sorting is gradually replacing manual sorting. Manual sorting standards are inconsistent and highly subjective, failing to meet market demands. Due to intense market competition in jujube sorting equipment, improving sorting efficiency and ease of use is crucial to enhancing the equipment's competitiveness. Existing equipment has pre-set baffle angles, making dynamic adjustments based on roller coverage and material accumulation in real time. Summary of the Invention

[0003] (1) The technical problem to be solved by the present invention

[0004] During operation, the angle of the baffle plate of the jujube sorting machine cannot be automatically adjusted to adapt to material sorting; and the problem of low roller coverage or material accumulation and return caused by the angle of the baffle plate should be solved.

[0005] (2) Technical solution adopted in this invention

[0006] To address the aforementioned technical problems, the present invention aims to provide a method for dynamically adjusting the angle of the baffle plate in a jujube machine based on a neural network, thereby improving the sorting efficiency and stability of the jujube machine and reducing the need for secondary sorting of materials.

[0007] The method described in this invention specifically includes the following steps:

[0008] Step S1: Use a camera to capture RGB images taken during the operation of the jujube machine, and perform background segmentation algorithm processing on the captured color images to remove the background area;

[0009] Step S2: Classify the acquired images, assign different labels to each category, construct training data samples, and train the neural network;

[0010] Step S3: Design a multi-branch convolutional neural network structure for classification, train it using the dataset from step S2, and obtain the neural network model after training is complete;

[0011] Step S4: Input the images captured by the camera in real time into the neural network model through step S1, give the corresponding label values, store the label values ​​in the same location, and update them in real time;

[0012] Step S5: Calculate the label values ​​that are updated in real time and have a fixed length, and count the proportion of each category to determine the direction of the baffle adjustment in order to improve the effective filling rate of materials on the jujube machine.

[0013] Step S6: Based on the baffle adjustment scheme determined in step S5, the control unit sends the control to the servo motor to complete the adjustment of the baffle.

[0014] Furthermore, in step S1, after removing the background from the acquired original RGB image, the image is then divided equally to ensure that each image is the same size and has only one groove for storing materials.

[0015] Furthermore, step S2 includes the following steps:

[0016] Step S201: Adjust the image size to a uniform 224*224 pixels;

[0017] Step S202: Manually classify the images and assign different labels, where 0 represents no dates, 1 represents a single date, and 2 represents multiple dates;

[0018] Step S203: Divide the labeled images into a training set, a validation set, and a test set. The division is random, with the training set accounting for 80%, and the validation set and test set each accounting for 10%.

[0019] Furthermore, the design of the convolutional neural network in step S3 includes the following steps:

[0020] Step S301: Design a convolutional neural network consisting of 8 convolutional layers, 4 batch normalization layers, 8 ReLU activation function layers, 1 global mean pooling layer and 1 Softmax multi-classification layer. The convolutional neural network includes 8 convolutional unit blocks, and each unit block contains a feature extraction layer and a feature stacking layer.

[0021] Step S302: Input the training samples into the input layer of the dual-branch convolutional neural network, use the AdamW optimization algorithm instead of the traditional SGD algorithm and Adam algorithm to train the dual-branch convolutional neural network, use cross-entropy to calculate the loss function, and train the network until the loss function of the multi-branch convolutional neural network reaches the minimum value.

[0022] Step S303: The model converges and the trained network weight coefficients are obtained. These network weight coefficients can be used for subsequent predictions.

[0023] Further, in step S301, in each of the eight convolutional unit blocks, features are first extracted by convolution using a 3*3 convolutional kernel and a 1*1 convolutional kernel, and then fed into a feature stacking layer for information stacking. In even-numbered unit blocks, a batch normalization layer is added to directly stack with the feature stacking layer for output. After passing through a ReLU activation function layer, the output is sent to the next unit block. After passing through eight convolutional unit blocks, the data is flattened by a global mean pooling layer, and finally the corresponding result is output through a Softmax classification layer.

[0024] Furthermore, the storage of the prediction results in step S4 includes the following steps:

[0025] Step S401: Scale the image to 224*224 pixels and input it into the trained convolutional neural network model. The model will output the label value corresponding to each image in turn.

[0026] Step S402: Store the label values ​​sequentially in the storage unit, and store them once every k images, where k is the number of images evenly divided in step S1. Then continuously update and increase the length of the label values.

[0027] Step S403: When the tag value length matches the number of rollers corresponding to the machine chain, the subsequent prediction results will continue to be stored, while the tag data stored at the beginning of the storage unit will be deleted.

[0028] Furthermore, step S5, which calculates the proportion of different label values, includes the following steps:

[0029] Step S501: The data stored in the storage unit are A1, A2, A3, ..., An, where n is the number of machine rollers, and A1, A2, etc. represent the predicted label values. The number of data with a label value of 0 is accumulated and recorded as S0, the number of data with a label value of 1 is accumulated and recorded as S1, and the number of data with a label value of 2 is accumulated and recorded as S2.

[0030] Step S502: Calculate the probability of each class. In these three types of probabilities, first compare whether P2 exceeds the set threshold M2, where the threshold M2 is set in advance. If the set threshold M2 is reached, a signal to increase the angle of the baffle plate will be sent. If not, compare whether the value of P0 reaches the set threshold M0. If it exceeds the threshold M0, a signal to decrease the angle of the baffle plate will be sent. If not, no adjustment is needed.

[0031] Step S503: During the dynamic update of the stored tag value data, P0, P1 and P2 are continuously updated, and the baffle plate is judged in real time according to the order of step S502 to determine whether it needs to be adjusted.

[0032] The beneficial effects of this invention are:

[0033] (1) The method of the present invention can automatically identify the distribution of materials in an image without manual observation and adjustment;

[0034] (2) The method of the present invention can automatically adjust according to the image recognition results to ensure the effective full coverage of red dates on the machine. Attached Figure Description

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0036] Figure 1 This is a flowchart of a method according to an embodiment of the present invention;

[0037] Figure 2 This is a flowchart illustrating the working structure of an embodiment of the present invention;

[0038] Figure 3 This is a diagram of the convolutional neural network structure in an embodiment of the present invention. Detailed Implementation

[0039] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0040] Example

[0041] In the existing technology, because the angle of the baffle plate in the jujube sorting equipment cannot be automatically adjusted, it is impossible to get feedback based on the fullness or accumulation of materials on the rollers, so as to adjust the angle of the baffle plate automatically.

[0042] Based on this, this embodiment proposes a method for dynamic adjustment of the baffle plate of a jujube machine based on neural networks, so as to improve the effective filling rate of the jujube machine and improve its production efficiency.

[0043] Please see Figures 1-3 As shown, the method in this embodiment includes the following steps:

[0044] Step S1: Use a camera to capture RGB images taken while the jujube machine is running. Perform background segmentation on the captured color images to remove the background area.

[0045] In step S1, after removing the background from the acquired original RGB image, the image is divided into equal parts to ensure that each image is the same size and has only one groove for storing materials.

[0046] Step S2: Classify the acquired images, assign different labels to each category, construct training data samples, and train the neural network;

[0047] Step S2 specifically includes the following steps:

[0048] S201: Adjust the image size to a uniform 224*224 pixels;

[0049] S202: Manually classify the images and assign different labels, where 0 represents no dates, 1 represents a single date, and 2 represents multiple dates;

[0050] S203: Divide the labeled images into training, validation, and test sets. The division is random, with the training set comprising 80%, and the validation and test sets each comprising 10%.

[0051] Step S3: Design a multi-branch convolutional neural network structure for classification, train it using the dataset from step S2, and obtain the neural network model after training is complete;

[0052] Step S3, designing the convolutional neural network, specifically includes the following steps:

[0053] S301: Design a convolutional neural network consisting of 8 convolutional layers, 4 batch normalization layers, 8 ReLU activation function layers, 1 global mean pooling layer and 1 Softmax multi-classification layer. The convolutional neural network mainly includes 8 convolutional unit blocks, each of which contains a feature extraction layer and a feature stacking layer.

[0054] S302: Input the training samples into the input layer of the dual-branch convolutional neural network, use the AdamW optimization algorithm instead of the traditional SGD algorithm and Adam algorithm to train the dual-branch convolutional neural network, use cross-entropy to calculate the loss function, and train the network until the loss function of the multi-branch convolutional neural network reaches the minimum value.

[0055] S303: The model has converged, and the trained network weight coefficients have been obtained. These network weight coefficients can be used for subsequent predictions.

[0056] In S301, each of the eight convolutional unit blocks first uses a 3*3 convolutional kernel and a 1*1 convolutional kernel to extract features, and then sends them to the feature stacking layer for information stacking. In even-numbered unit blocks, a batch normalization layer is added to directly stack with the feature stacking layer for output. After passing through the ReLU activation function layer, the output is sent to the next unit block. After passing through eight convolutional unit blocks, the data is flattened by the global mean pooling layer, and finally the corresponding result is output through the Softmax classification layer.

[0057] In step S3 of this embodiment, the accuracy of material distribution identification is improved by using convolutional neural networks for training and recognition. Unlike traditional algorithms that use parameters such as area to set separate parameters for different grades of jujubes, this improves the ease of use of the machine.

[0058] Step S4: Input the images captured by the camera in real time into the neural network model through step S1, give the corresponding label values, store the label values ​​in the same location, and update them in real time;

[0059] The storage of the prediction results in step S4 mainly includes the following steps:

[0060] S401: Scale the image to 224*224 pixels and input it into the trained convolutional neural network model. The model will output the label value corresponding to each image in turn.

[0061] S402: Store the label values ​​sequentially in the storage unit, and store them once every k images (k is the number of images evenly divided in step S1), continuously updating and increasing the length of the label values;

[0062] S403: When the tag value length matches the number of rollers corresponding to the machine chain, the results of subsequent predictions will continue to be stored, while the tag data stored at the beginning of the storage unit will be deleted.

[0063] Step S5: Calculate the label values ​​that are updated in real time and have a fixed length, and count the proportion of each category to determine the direction of the baffle adjustment in order to improve the effective filling rate of materials on the jujube machine.

[0064] Step S5, which calculates the proportion of different label values, mainly includes the following steps:

[0065] S501: The data stored in the storage unit are A1, A2, A3, ..., An, (n is the number of machine rollers, and A1, A2, etc. represent the predicted label values). The number of data with a label value of 0 is accumulated and recorded as S0, the number of data with a label value of 1 is accumulated and recorded as S1, and the number of data with a label value of 2 is accumulated and recorded as S2.

[0066] S502: Calculate the probability of each class. In these three probability categories, first compare whether P2 exceeds the set threshold M2 (pre-set manually). If it reaches the set threshold M2, a signal to increase the angle of the baffle plate will be sent. If it does not reach the threshold M2, compare whether the value of P0 reaches the set threshold M0. If it exceeds the threshold M0, a signal to decrease the angle of the baffle plate will be sent. If it does not exceed the threshold M0, no adjustment is required.

[0067] S503: During the dynamic update of the stored tag value data, P0, P1 and P2 are continuously updated, and the baffle plate is judged in real time according to the order of S502 to determine whether it needs to be adjusted.

[0068] In this embodiment, step S5 can update the stored recognition results in real time and perform calculations in real time, ensuring the effective filling rate of the roller material in the production of the jujube machine, improving the efficiency of the production process, reducing the probability of material return, and avoiding damage to the material caused by multiple re-selection.

[0069] Step S6: Based on the baffle adjustment scheme determined in step S5, the control unit sends the information to the servo motor to complete the adjustment of the baffle.

[0070] The feature of the neural network-based dynamic adjustment method for the baffle of the jujube machine in this embodiment is that it proposes an algorithm for real-time automatic adjustment of the baffle angle and an algorithm for automatic detection of the material distribution on the rollers of the jujube machine.

[0071] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic adjustment of the baffle plate of a jujube machine based on neural networks, characterized in that, Includes the following steps: Step S1: Use a camera to capture RGB images taken during the operation of the jujube machine, and perform background segmentation algorithm processing on the captured color images to remove the background area; Step S2: Classify the acquired images, assign different labels to each category, construct training data samples, and train the neural network; Step S3: Design a multi-branch convolutional neural network structure for classification, train it using the dataset from step S2, and obtain the neural network model after training is complete; Step S4: Input the images captured by the camera in real time into the neural network model through step S1, give the corresponding label values, store the label values ​​in the same location, and update them in real time; Step S5: Calculate the label values ​​that are updated in real time and have a fixed length, and count the proportion of each category to determine the direction of the baffle adjustment in order to improve the effective filling rate of materials on the jujube machine. Step S6: Based on the baffle adjustment scheme determined in step S5, the control unit sends the control to the servo motor to complete the adjustment of the baffle. In step S1, after removing the background from the acquired original RGB image, the image is then divided into equal parts to ensure that each image is the same size and has only one groove for storing materials. Step S5, which calculates the proportion of different label values, includes the following steps: Step S501: The data stored in the storage unit are A1, A2, A3, ..., An, where n is the number of machine rollers, and A1, A2, A3, ..., An represent the predicted label values. The number of data with a label value of 0 is accumulated and recorded as S0, the number of data with a label value of 1 is accumulated and recorded as S1, and the number of data with a label value of 2 is accumulated and recorded as S2. Step S502: Calculate the probability of each class. , , In these three types of probabilities, first compare whether P2 exceeds the set threshold M2, where the threshold M2 is set in advance. If the set threshold M2 is reached, a signal to increase the angle of the baffle plate will be sent. If not, compare whether the value of P0 has reached the set threshold M0. If it exceeds the threshold M0, a signal to decrease the angle of the baffle plate will be sent. If it does not exceed the threshold M0, no adjustment is required. Step S503: During the dynamic update of the stored tag value data, P0, P1 and P2 are continuously updated, and the baffle plate is judged in real time according to the order of step S502 to determine whether it needs to be adjusted.

2. The method according to claim 1, characterized in that, Step S2 includes the following steps: Step S201: Adjust the image size to a uniform 224*224 pixels; Step S202: Manually classify the images and assign different labels, where 0 represents no dates, 1 represents a single date, and 2 represents multiple dates; Step S203: Divide the labeled images into a training set, a validation set, and a test set. The division is random, with the training set accounting for 80%, and the validation set and test set each accounting for 10%.

3. The method according to claim 1, characterized in that, The design of the convolutional neural network in step S3 includes the following steps: Step S301: Design a convolutional neural network consisting of 8 convolutional layers, 4 batch normalization layers, 8 ReLU activation function layers, 1 global mean pooling layer and 1 Softmax multi-classification layer. The convolutional neural network includes 8 convolutional unit blocks, and each unit block contains a feature extraction layer and a feature stacking layer. Step S302: Input the training samples into the input layer of the dual-branch convolutional neural network, use the AdamW optimization algorithm instead of the traditional SGD algorithm and Adam algorithm to train the dual-branch convolutional neural network, use cross-entropy to calculate the loss function, and train the network until the loss function of the multi-branch convolutional neural network reaches the minimum value. Step S303: The model converges and the trained network weight coefficients are obtained. These network weight coefficients can be used for subsequent predictions.

4. The method according to claim 3, characterized in that, In step S301, each of the eight convolutional unit blocks first uses a 3*3 convolutional kernel and a 1*1 convolutional kernel to extract features, and then sends them to the feature stacking layer for information stacking. In even-numbered unit blocks, a batch normalization layer is added to directly stack with the feature stacking layer for output. After passing through the ReLU activation function layer, the output is sent to the next unit block. After passing through eight convolutional unit blocks, the data is flattened by the global mean pooling layer, and finally the corresponding result is output through the Softmax classification layer.

5. The method according to claim 1, characterized in that, The storage of the prediction results in step S4 includes the following steps: Step S401: Scale the image to 224*224 pixels and input it into the trained convolutional neural network model. The model will output the label value corresponding to each image in turn. Step S402: Store the label values ​​sequentially in the storage unit, and store them once every k images, where k is the number of images evenly divided in step S1. Then continuously update and increase the length of the label values. Step S403: When the tag value length matches the number of rollers corresponding to the machine chain, the subsequent prediction results will continue to be stored, while the tag data stored at the beginning of the storage unit will be deleted.

Images