Film mulching device for maize film side sowing and intelligent control method

Abstract

The application relates to the technical field of intelligent control, and particularly discloses a film mulching equipment for corn side seeding and an intelligent control method thereof. The film mulching equipment uses artificial intelligence technology based on deep learning to analyze soil state images and the height of the whole machine, excavates virtual and real state feature information of the soil, and intelligently recommends the compacting strength of a compacting roller in combination with the height feature of the whole machine. In this way, intelligent recommendation of the compacting strength can be realized, so that the soil tightness is optimized, the growth and development of crops are promoted, and the economic benefits of corn side seeding with film mulching are improved.

Description

Technical Field

[0001] This application relates to the field of intelligent control technology, and more specifically, to a mulching device for corn film-side sowing and its intelligent control method. Background Technology

[0002] Corn cultivation is an important part of agricultural production, and improving the efficiency and yield of corn cultivation is a key goal of agricultural technology. Mulched corn planting is a new planting technique that has developed in recent years. It improves the soil environment by mulching, increasing soil temperature and humidity, promoting corn growth, and thus contributing to increased crop yield and quality. Current mulching methods mainly involve trenching, laying the film, and planting through holes in the film. However, these traditional mulching methods suffer from problems such as difficulty in recycling residual film and the need for significant manual labor for seedling removal and placement.

[0003] In response, patent CN103766050A discloses a corn mulch-covered ridge-planting and side-seeding machine. This machine employs a non-furrowing mulch-laying method, using a roller to compact the soil into an arched mulch strip that is higher in the middle and lower on both sides. Then, a pressure roller and tensioning shaft are used to tension the mulch film and lay it flat on the mulch strip. The pressure roller further compacts the soil on both sides of the mulch film to secure it. This mulch-laying method facilitates rainwater collection and mulch film recycling; moreover, the side-seeding method eliminates the need for manual film breaking and seedling placement after sowing, saving labor. However, the compaction intensity needs to be adjusted according to the soil density and the overall height of the machine when using the roller to compact the soil.

[0004] In traditional methods, the adjustment of compaction intensity is usually based on farmers' experience and intuition. However, this method has some limitations. First, relying solely on farmers' experience and intuition makes it difficult to comprehensively and accurately assess soil firmness and overall machine height. Second, experience and intuition are personal knowledge and skills, difficult to replicate and pass on between different regions and different farmers. With the development of agricultural technology and changes in the farmer population, traditional adjustment methods based on experience and intuition may gradually become inapplicable. Therefore, there is a need for a mulching device for corn side-planting and its intelligent control method. Summary of the Invention

[0005] To address the aforementioned technical problems, this application is proposed. Embodiments of this application provide a mulching device and its intelligent control method for corn side-planting under mulch. This method utilizes deep learning-based artificial intelligence technology to analyze soil condition images and the overall machine height, extracting information on the soil's virtual and solid state characteristics, and intelligently recommending the compaction intensity of the compaction rollers based on the overall machine height characteristics. This enables intelligent recommendation of compaction intensity, thereby optimizing soil compaction, promoting crop growth and development, and improving the economic benefits of corn side-planting under mulch.

[0006] Accordingly, according to one aspect of this application, a mulching device for corn film-side sowing is provided, comprising: a compaction intensity adjustment module for adjusting the compaction intensity of the compaction roller based on soil density and overall machine height; a soil compaction module for compacting the soil according to the compaction intensity of the compaction roller to obtain an arched film-laying belt that is higher in the middle and lower on both sides; a film-laying module for using a film-pressing wheel and a tensioning shaft to tension and spread the mulch film mounted on the film-laying roller flat onto the arched film-laying belt to obtain mulch film covering the film-laying belt; and a film-fixing module for using a soil-covering shovel behind the film-pressing wheel to shovel the surface soil on both sides of the mulch film covering the film-laying belt toward both sides of the mulch film covering the film-laying belt, and using the pressing wheel to compact the soil covering the mulch film covering the film-laying belt to obtain mulch film fixed on the film-laying belt; wherein, the compaction intensity adjustment module includes:

[0007] Soil condition monitoring unit, used to acquire soil condition images captured by camera;

[0008] The overall height information acquisition unit is used to acquire the overall height value collected by the sensor;

[0009] A soil state feature extraction unit is used to extract soil state features from the soil state image to obtain a sequence of context soil local state feature vectors.

[0010] The whole machine height one-hot encoding unit is used to perform one-hot encoding on the whole machine height value to obtain the whole machine height one-hot encoded feature vector;

[0011] The compaction intensity recommendation unit is used to determine the recommended value of compaction intensity based on the semantic fusion features of the sequence of local soil state feature vectors in the context and the one-hot encoded feature vector of the overall machine height.

[0012] In the aforementioned mulching equipment for corn film-side sowing, the soil state feature extraction unit includes: an image segmentation subunit, used to segment the soil state image into blocks to obtain a sequence of local soil state image blocks; a local soil state feature extraction subunit, used to extract soil state features from each local soil state image block in the sequence of local soil state image blocks to obtain a sequence of local soil state feature vectors; and an inter-regional association coding subunit, used to extract inter-regional contextual association features from the sequence of local soil state feature vectors to obtain a sequence of contextual local soil state feature vectors.

[0013] In the above-mentioned mulching equipment for corn film-side sowing, the soil local state feature extraction subunit is used to: pass the sequence of soil local state image blocks through a soil state feature extractor based on a convolutional neural network model to obtain the sequence of soil local state feature vectors.

[0014] In the aforementioned mulching equipment for corn film-side sowing, the soil state feature extractor based on the convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a flattening layer, and an output layer.

[0015] In the above-mentioned mulching equipment for corn film-side sowing, the inter-regional association coding subunit is used to: pass the sequence of the soil local state feature vectors through a soil state inter-regional association encoder based on a converter module to obtain the sequence of the context soil local state feature vectors.

[0016] In the aforementioned mulching equipment for corn film-side sowing, the compaction intensity recommendation unit includes: a feature fusion subunit, used to fuse the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the whole machine height to obtain a multimodal semantic fusion feature vector; an optimization fusion subunit, used to optimize the fusion of the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the whole machine height based on feature value granularity to obtain an optimized feature vector; an optimization application subunit, used to fuse the optimized feature vector and the multimodal semantic fusion feature vector to obtain a multimodal semantic fusion feature vector; and a decoding regression subunit, used to pass the multimodal semantic fusion feature vector through a decoder to obtain a decoded value, the decoded value being used to represent the recommended compaction intensity value.

[0017] In the above-mentioned mulching equipment for corn film-side sowing, the feature fusion subunit is used to: use a cross-modal interactive fusion device to process the sequence of the context soil local state feature vector and the whole machine height one-hot encoded feature vector to obtain the multimodal semantic fusion feature vector.

[0018] In the above-mentioned mulching equipment for corn film-side sowing, the feature fusion subunit is used to: use a cross-modal interactive fusion device to semantically fuse the sequence of the context soil local state feature vector and the whole machine height one-hot encoded feature vector with the following cross-modal interactive fusion formula to obtain the multimodal semantic fusion feature vector;

[0019] The cross-modal interaction fusion formula is as follows:

[0020]

[0021] Where V is the one-hot encoded feature vector of the overall machine height, and A is a 1×N... w The matrix, N w h is the dimension of the one-hot encoded feature vector of the entire machine. i B is the i-th context soil local state feature vector in the sequence of context soil local state feature vectors, and B is 1×N hThe matrix, N h is the dimension of each of the aforementioned context soil local state feature vectors, N is the number of the aforementioned context soil local state feature vectors, σ is the Sigmoid function, V' is the multimodal semantic fusion feature vector, and M w and M h This represents a convolution operation with a 1×1 kernel, where S is the weight value.

[0022] The aforementioned mulching equipment for corn film-side sowing further includes: a training module for training the soil state feature extractor based on a convolutional neural network model, the soil state inter-region association encoder based on a converter module, the cross-modal interactive fusion device, and the decoder. The training module includes: a training data acquisition unit for acquiring training data, which includes: a training soil state image and a training overall machine height value; a training image segmentation unit for segmenting the training soil state image into blocks to obtain a sequence of training soil local state image blocks; a training soil local state feature extraction unit for passing the sequence of training soil local state image blocks through the soil state feature extractor based on the convolutional neural network model to obtain a sequence of training soil local state feature vectors; and a training inter-region association encoding unit for passing the sequence of training soil local state feature vectors through the soil state inter-region association encoder based on the converter module to obtain a training context. The system comprises: a sequence of soil local state feature vectors; a training unit for one-hot encoding of the whole-machine height, used to perform one-hot encoding on the training whole-machine height values ​​to obtain a training whole-machine height one-hot encoded feature vector; a training feature fusion unit, used to process the sequence of soil local state feature vectors and the training whole-machine height one-hot encoded feature vector using a cross-modal interactive fusion generator to obtain a training multimodal semantic fusion feature vector; a feature aggregation optimization unit, used to perform feature aggregation optimization on the training multimodal semantic fusion feature vector to obtain an optimized training multimodal semantic fusion feature vector; a decoding loss calculation unit, used to pass the optimized training multimodal semantic fusion feature vector through the decoder to obtain a decoding loss function value; and a training unit, used to train the soil state feature extractor based on the convolutional neural network model, the soil state inter-regional association encoder based on the converter module, the cross-modal interactive fusion generator, and the decoder based on the decoding loss function value and through gradient descent directional propagation.

[0023] According to another aspect of this application, an intelligent control method for a mulching device used for corn film-side planting is provided, comprising:

[0024] Acquire soil condition images captured by a camera;

[0025] Obtain the overall height value of the machine collected by the sensors;

[0026] Extract soil state features from the soil state image to obtain a sequence of contextual soil local state feature vectors;

[0027] The overall height value is one-hot encoded to obtain the overall height one-hot encoded feature vector;

[0028] Based on the semantic fusion features of the sequence of local soil state feature vectors in the context and the uniquely encoded feature vector of the overall machine height, a recommended value for compaction intensity is determined.

[0029] Compared with existing technologies, the mulching equipment and its intelligent control method for corn mulching side planting provided in this application utilize deep learning-based artificial intelligence technology to analyze soil condition images and overall machine height, extracting information on the soil's virtual and solid state characteristics, and intelligently recommending the compaction intensity of the compaction rollers based on the overall machine height characteristics. This enables intelligent recommendation of compaction intensity, thereby optimizing soil compaction, promoting crop growth and development, and improving the economic benefits of corn mulching side planting. Attached Figure Description

[0030] The above and other objects, features, and advantages of this application will become more apparent from the more detailed description of the embodiments of this application in conjunction with the accompanying drawings. The drawings are provided to further illustrate the embodiments of this application and form part of the specification. They are used together with the embodiments of this application to explain this application and do not constitute a limitation thereof. In the drawings, the same reference numerals generally represent the same components or steps.

[0031] Figure 1 This is a block diagram of the compaction intensity adjustment module in a mulching device for corn film-side sowing according to an embodiment of this application.

[0032] Figure 2 This is a schematic diagram of the compaction intensity adjustment module in a mulching device for corn film-side sowing according to an embodiment of this application.

[0033] Figure 3 This is a block diagram of a soil condition feature extraction unit in a mulching device for corn film-side sowing according to an embodiment of this application.

[0034] Figure 4 This is a block diagram of a compaction strength recommendation unit in a mulching device for corn film-side sowing according to an embodiment of this application.

[0035] Figure 5 This is a flowchart of an intelligent control method for a mulching device for corn film-side sowing according to an embodiment of this application. Detailed Implementation

[0036] Hereinafter, exemplary embodiments according to this application will be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of this application, and not all embodiments of this application. It should be understood that this application is not limited to the exemplary embodiments described herein.

[0037] As mentioned in the background section above, patent CN103766050A discloses a corn mulch ridge-laying and side-planting seeder, which specifically includes: a compaction intensity adjustment module for adjusting the compaction intensity of the compaction roller based on the soil density and the overall height of the machine; a soil compaction module for compacting the soil according to the compaction intensity of the compaction roller to obtain an arched mulch belt that is higher in the middle and lower on both sides; a mulch laying module for using a pressing wheel and a tensioning shaft to tension the mulch film installed on the mulch film installation roller and lay it flat on the arched mulch belt to obtain a mulch film covering the mulch belt; and a mulch fixing module for using a soil covering shovel behind the pressing wheel to shovel the surface soil on both sides of the mulch film covering the mulch belt towards both sides of the mulch film covering the mulch belt, and using the pressing wheel to compact the soil covering the mulch film covering the mulch belt to obtain a mulch film fixed on the mulch belt.

[0038] In the actual mulching process, specifically when using rollers to compact the soil, the compaction intensity directly affects the soil's density, thus impacting corn growth and development. Excessive compaction leads to over-compaction, resulting in high soil density, a hardened surface, reduced water permeability, and restricted aeration and permeability, hindering root growth and development. Impaired root growth reduces the plant's ability to absorb nutrients and water, further limiting corn plant growth and development. Conversely, insufficient compaction results in loose soil, preventing adequate root support and stability, potentially causing plant tilting, breakage, or unstable growth. Furthermore, a loose soil surface allows for easy water evaporation and loss, affecting corn's water supply and consequently impacting plant growth and yield. Additionally, insufficient compaction may not effectively suppress weed growth. Weeds compete for water, nutrients, and sunlight, negatively impacting corn plant growth. Therefore, it is necessary to adjust the compaction intensity of the roller to control the degree of soil compaction, thereby achieving appropriate soil compaction and planting depth.

[0039] Traditional methods of adjusting compaction intensity typically rely on farmers' experience and intuition. However, this method, which depends solely on farmer experience, has limitations. It is difficult to comprehensively and accurately assess soil texture and overall machine height. Furthermore, this method is difficult to replicate and pass on between different regions and farmers. With the development of agricultural technology and changes in the farming population, traditional experience- and intuition-based adjustment methods may gradually become less applicable.

[0040] To address the aforementioned technical problems, the technical concept of this application is to utilize deep learning-based artificial intelligence technology to analyze soil condition images and overall machine height, extract virtual and solid soil condition characteristics, and intelligently recommend the compaction intensity of the compaction rollers based on the overall machine height characteristics. This enables intelligent recommendation of compaction intensity, thereby optimizing soil compaction, promoting crop growth and development, and improving the economic benefits of corn mulched side planting.

[0041] Figure 1 This is a block diagram of the compaction intensity adjustment module in a mulching device for corn film-side sowing according to an embodiment of this application. Figure 2 This is a schematic diagram of the compaction intensity adjustment module in a mulching device for corn film-side sowing according to an embodiment of this application. Figure 1 and Figure 2 As shown, the mulching device 100 for corn film-side sowing according to an embodiment of this application includes: a soil state monitoring unit 110 for acquiring soil state images captured by a camera; a machine height information acquisition unit 120 for acquiring machine height values ​​collected by sensors; a soil state feature extraction unit 130 for extracting soil state features from the soil state images to obtain a sequence of context soil local state feature vectors; a machine height one-hot encoding unit 140 for performing one-hot encoding on the machine height values ​​to obtain a machine height one-hot encoded feature vector; and a compaction intensity recommendation unit 150 for determining a recommended value of compaction intensity based on the sequence of context soil local state feature vectors and the semantic fusion features of the machine height one-hot encoded feature vector.

[0042] In the aforementioned mulching equipment 100 for corn film-side sowing, the soil condition monitoring unit 110 is used to acquire soil condition images captured by a camera. It should be understood that the soil's density refers to its looseness and compaction. In the actual mulching process, different soil conditions require different compaction intensities to achieve the optimal mulching effect. If the soil is too loose and the compaction intensity is insufficient, the mulching strip may not form completely, and the soil may not adhere tightly to the film material, affecting the mulching effect. Conversely, if the soil is too compact and the compaction intensity is too high, the soil may be over-compacted, affecting corn growth and root development. Therefore, when adjusting the compaction intensity of the compaction roller, the soil density should be fully considered to achieve the best mulching effect. In the technical solution of this application, soil density can be characterized by soil condition images. Therefore, by acquiring soil condition images captured by a camera, visual features such as the texture, color, and shape of the soil surface can be obtained, thereby determining the soil's looseness and compaction, and guiding the adjustment of the compaction intensity.

[0043] In the aforementioned mulching equipment 100 for corn film-side sowing, the overall machine height information acquisition unit 120 is used to acquire the overall machine height value collected by sensors. It should be understood that the overall machine height refers to the vertical distance between the working part of the mulching equipment and the ground. The overall machine height directly affects the compaction effect of the compaction roller. When the overall machine height is low, the contact area between the compaction roller and the soil increases, and the force applied by the compaction roller to the soil is greater. Conversely, when the overall machine height is high, the contact area decreases, and the force applied by the compaction roller to the soil decreases. Therefore, adjusting the compaction intensity of the compaction roller based on the overall machine height information can optimize the degree of soil compaction by the compaction roller to adapt to different soil conditions and mulching requirements.

[0044] In the aforementioned mulching device 100 for corn film-side sowing, the soil state feature extraction unit 130 is used to extract soil state features from the soil state image to obtain a sequence of contextual soil local state feature vectors. Specifically, Figure 3 This is a block diagram of a soil condition feature extraction unit in a mulching device for corn film-side sowing according to an embodiment of this application. Figure 3 As shown, the soil state feature extraction unit 130 includes: an image segmentation subunit 131, used to segment the soil state image into blocks to obtain a sequence of local soil state image blocks; a local soil state feature extraction subunit 132, used to extract soil state features from each local soil state image block in the sequence of local soil state image blocks to obtain a sequence of local soil state feature vectors; and an inter-regional association coding subunit 133, used to extract inter-regional contextual association features from the sequence of local soil state feature vectors to obtain a sequence of contextual local soil state feature vectors.

[0045] Specifically, the image segmentation subunit 131 is used to segment the soil state image into blocks to obtain a sequence of local soil state image blocks. In the technical solution of this application, considering that the soil state image may contain a large amount of detailed information, directly processing the entire image may result in the loss of subtle features. Therefore, in order to capture the local detailed features in the soil state image more precisely, the soil state image is further segmented into multiple small blocks, each representing a local area of ​​the soil. This allows attention to be focused on each image block, enabling independent feature mining of each block, thereby improving the accuracy and reliability of feature extraction and allowing for a more detailed analysis of soil characteristics. Furthermore, segmenting the soil state image into blocks reduces the influence of irrelevant information in the entire image, avoiding interference from noise or background in the entire image on soil feature extraction.

[0046] Specifically, the soil local state feature extraction subunit 132 is used to extract soil state features from each soil local state image patch in the sequence of soil local state image patches to obtain a sequence of soil local state feature vectors. In a specific example of this application, the encoding method for extracting the soil state features from each soil local state image patch in the sequence of soil local state image patches to obtain the sequence of soil local state feature vectors is to pass the sequence of soil local state image patches through a soil state feature extractor based on a convolutional neural network model to obtain the sequence of soil local state feature vectors. Those skilled in the art should know that a convolutional neural network (CNN) is a deep learning model widely used in computer vision, which can effectively and automatically learn meaningful feature representations from images. In the technical solution of this application, the sequence of soil local state image patches is input into a soil state feature extractor based on a convolutional neural network model. The convolutional neural network can apply the convolution kernels of the convolutional layers to perform sliding convolution operations on the soil local state image patches to capture the spatial relationships and local patterns of the image, thereby extracting the soil local state features and encoding the soil state features in each soil local state image patch into a feature vector representation, providing accurate input for subsequent soil state analysis and processing. More specifically, the soil state feature extractor based on the convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a flattening layer, and an output layer.

[0047] Specifically, the inter-regional association encoding subunit 133 is used to extract inter-regional contextual association features of the sequence of soil local state feature vectors to obtain the sequence of context soil local state feature vectors. It should be understood that, considering the usual correlations and dependencies between soil states in different regions, in order to capture the global correlation information between various soil local states, inter-regional contextual association features of each soil local state feature vector are further extracted. In a specific example of this application, the encoding method for extracting the inter-regional contextual association features of the sequence of soil local state feature vectors to obtain the sequence of context soil local state feature vectors is to pass the sequence of soil local state feature vectors through a soil state inter-regional association encoder based on a converter module to obtain the sequence of context soil local state feature vectors. It should be understood that the converter module is a module based on a self-attention mechanism, commonly used to process sequence data and capture the contextual relationships in the sequence. It can adaptively weight each element in the sequence, thereby better capturing the dependencies and contextual information between elements. In the technical solution of this application, the soil state inter-regional association encoder uses the converter module to encode the sequence of soil local state feature vectors. Here, the converter module can perform weighted encoding on each soil local state feature vector based on the self-attention mechanism and the relationship between each soil local state feature vector in the sequence of soil local state feature vectors and other soil local state feature vectors. This allows each soil local state feature vector to contain contextual information from other soil local state feature vectors, thereby comprehensively considering global information in the sequence, capturing the correlation and dependency between soil local states, and improving the ability to analyze and predict soil states.

[0048] In the aforementioned mulching equipment 100 for corn film-side sowing, the one-hot encoding unit 140 for the overall height is used to perform one-hot encoding on the overall height value to obtain a one-hot encoded feature vector for the overall height. It should be understood that the overall height is a discrete feature with a finite range of values. In machine learning tasks, if the original overall height value is directly used as the feature input, the model may consider larger overall height values ​​to have a greater impact on the result. Therefore, to avoid the influence of numerical magnitude on the model, the overall height value is further one-hot encoded to convert it into a sparse binary feature vector, allowing the model to treat different overall height values ​​more fairly. In one-hot encoding, for a discrete feature with n different values, it is represented as a binary vector of length n, where each value corresponds to one dimension of the vector, with only one dimension having a value of 1 and the remaining dimensions having values ​​of 0. This encoding method ensures that each value is completely independent, without introducing any bias or size relationship. This enables the model to better handle discrete features and make better use of information about the overall height of the machine, thereby increasing the feature representation capability and improving the model's ability to model the overall height of the machine.

[0049] In the aforementioned mulching equipment 100 for corn film-side sowing, the compaction intensity recommendation unit 150 is used to determine a recommended value for compaction intensity based on the semantic fusion features of the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the overall machine height. Specifically, Figure 4 This is a block diagram of a compaction strength recommendation unit in a mulching device for corn film-side sowing according to an embodiment of this application. Figure 4 As shown, the compaction intensity recommendation unit 150 includes: a feature fusion subunit 151, used to fuse the sequence of the context soil local state feature vector and the whole machine height one-hot encoded feature vector to obtain a multimodal semantic fusion feature vector; and a decoding regression subunit 152, used to pass the multimodal semantic fusion feature vector through a decoder to obtain a decoded value, the decoded value being used to represent the recommended compaction intensity value.

[0050] Specifically, the feature fusion subunit 151 is used to fuse the sequence of the context soil local state feature vector and the one-hot encoded feature vector of the overall machine height to obtain a multimodal semantic fusion feature vector. In a specific example of this application, the encoding method for fusing the sequence of the context soil local state feature vector and the one-hot encoded feature vector of the overall machine height to obtain the multimodal semantic fusion feature vector is to use a cross-modal interactive fusion processor to process the sequence of the context soil local state feature vector and the one-hot encoded feature vector of the overall machine height to obtain the multimodal semantic fusion feature vector. It should be understood that soil state and overall machine height are both important factors affecting the planting effect on the corn mulch side. By comprehensively considering soil state features and overall machine height features, a more comprehensive and accurate input can be provided for the intelligent recommendation of compaction intensity values. Since the context soil local state feature vector and the one-hot encoded feature vector of the overall machine height belong to different modalities, cross-modal information fusion is required to comprehensively utilize the two types of information. Here, the cross-modal interaction fusion engine introduces an attention mechanism to adaptively select and weight information from different domains or modalities, achieving targeted fusion of cross-domain data. Specifically, it calculates attention weights using the correlation information between the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the overall machine height. Based on these attention weights, it weights and fuses the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the overall machine height, achieving effective information exchange and integration. This improves the expressive power of the features, enabling the model to better capture the correlation and complementarity between features from two different modalities, thereby enhancing the model's performance and generalization ability.

[0051] Specifically, the multimodal semantic fusion feature vector is obtained by processing the sequence of local soil state feature vectors and the one-hot encoded feature vector of the overall height using a cross-modal interactive fusion device. This includes: using the cross-modal interactive fusion device to perform semantic fusion on the sequence of local soil state feature vectors and the one-hot encoded feature vector of the overall height using the following cross-modal interactive fusion formula; wherein the cross-modal interactive fusion formula is:

[0052]

[0053]

[0054] Where V is the one-hot encoded feature vector of the overall machine height, and A is a 1×N... w The matrix, N w h is the dimension of the one-hot encoded feature vector of the entire machine. iB is the i-th context soil local state feature vector in the sequence of context soil local state feature vectors, and B is 1×N h The matrix, N h is the dimension of each of the aforementioned context soil local state feature vectors, N is the number of the aforementioned context soil local state feature vectors, σ is the Sigmoid function, V' is the multimodal semantic fusion feature vector, and M w and M h This represents a convolution operation with a 1×1 kernel, where S is the weight value.

[0055] Specifically, the decoding regression subunit 152 is used to pass the multimodal semantic fusion feature vector through a decoder to obtain a decoded value, which represents the recommended compaction intensity value. It should be understood that the decoder is a machine learning model that can convert the input encoded representation into a target numerical output. In the technical solution of this application, the multimodal semantic fusion feature vector contains a comprehensive feature representation of soil condition and overall machine height information. Further inputting the multimodal semantic fusion feature vector into the decoder allows the decoder's regression mapping capability to learn the mapping relationship between this comprehensive feature representation and the compaction intensity. This transforms the multimodal semantic fusion feature vector into a specific recommended compaction intensity value, which serves as the basis for adjusting the compaction intensity of the compaction roller, thereby achieving intelligent compaction intensity control and improving the effect of corn mulching side planting.

[0056] In a specific example of this application, the decoding regression subunit 152 is configured to: use the decoder to perform decoding regression on the multimodal semantic fusion feature vector using the following decoding formula to obtain the suppression intensity value, wherein the decoding formula is: Where X is the multimodal semantic fusion feature vector, Y is the suppression intensity value, and W is the weight matrix. This represents matrix multiplication.

[0057] In a specific example of this application, the mulching device for corn film-side sowing further includes: a training module for training the soil state feature extractor based on a convolutional neural network model, the soil state inter-regional association encoder based on a converter module, the cross-modal interactive fusion device, and the decoder. The training module includes: a training data acquisition unit for acquiring training data, which includes: a training soil state image and a training overall machine height value; a training image segmentation unit for segmenting the training soil state image into blocks to obtain a sequence of training soil local state image blocks; a training soil local state feature extraction unit for passing the sequence of training soil local state image blocks through the soil state feature extractor based on the convolutional neural network model to obtain a sequence of training soil local state feature vectors; and a training inter-regional association encoding unit for passing the sequence of training soil local state feature vectors through the soil state inter-regional association encoder based on the converter module to obtain... The training process includes: a sequence of local soil state feature vectors for training context; a training unit for one-hot encoding of whole-machine height values ​​to obtain a training whole-machine height one-hot encoded feature vector; a training feature fusion unit for processing the sequence of local soil state feature vectors and the training whole-machine height one-hot encoded feature vector using a cross-modal interactive fusion generator to obtain a training multimodal semantic fusion feature vector; a feature aggregation optimization unit for performing feature aggregation optimization on the training multimodal semantic fusion feature vector to obtain an optimized training multimodal semantic fusion feature vector; a decoding loss calculation unit for passing the optimized training multimodal semantic fusion feature vector through the decoder to obtain a decoding loss function value; and a training unit for training the soil state feature extractor based on the convolutional neural network model, the soil state inter-regional association encoder based on the converter module, the cross-modal interactive fusion generator, and the decoder based on the decoding loss function value and through gradient descent directional propagation.

[0058] Specifically, in the above technical solution, the sequence of training soil local state feature vectors expresses the image semantic features in the local image semantic space domain of the training soil state image based on the global semantic space domain association representation of the feature context between the local image semantic space domains, while the training whole machine height one-hot encoded feature vector expresses the one-hot encoded representation of the whole machine height value. Therefore, there is a significant feature expression variability between the sequence of training context soil local state feature vectors and the training whole machine height one-hot encoded feature vector based on the feature calculation dimension of different modal data. Therefore, when using a cross-modal interactive fusion device to process the sequence of training context soil local state feature vectors and the training whole machine height one-hot encoded feature vector, it is necessary to solve the problem of insufficient aggregation of the cross-modal interactive fusion distribution of the obtained training multimodal semantic fusion feature vectors in order to improve the decoding and regression effect of the training multimodal semantic fusion feature vectors.

[0059] Therefore, this application preferably performs feature aggregation optimization on the training multimodal semantic fusion feature vector, for example denoted as V, during each decoding regression iteration training of the trained multimodal semantic fusion feature vector through the decoder. Specifically, the feature aggregation optimization unit is used to perform feature aggregation optimization on the trained multimodal semantic fusion feature vector using the following feature aggregation optimization formula to obtain the optimized trained multimodal semantic fusion feature vector; wherein, the feature aggregation optimization formula is:

[0060]

[0061] Where V is the trained multimodal semantic fusion feature vector, v i It is the feature value at the i-th position of the trained multimodal semantic fusion feature vector. It is the square of the 1-norm of the trained multimodal semantic fusion feature vector V, ||V||2 -1 / 2 v' is the reciprocal of the square root of the norm 2 of the trained multimodal semantic fusion feature vector V, L is the length of the trained multimodal semantic fusion feature vector V, ε is the scaling hyperparameter, log represents the logarithmic function value to the base 2, and v' i It is the feature value at the i-th position of the optimized training multimodal semantic fusion feature vector.

[0062] Here, the low-rank structure based on the norm of the trained multimodal semantic fusion feature vector V is represented as a voting cluster of the feature value set of the trained multimodal semantic fusion feature vector V. For each feature value of the trained multimodal semantic fusion feature vector V, a canonical voting process is performed relative to the information perception framework. By aggregating the regression representation of the direction and scale of the feature distribution, feature values ​​belonging to the same regression class are mapped to similar locally canonical coordinate sets, thereby improving the aggregation effect of the feature set of the trained multimodal semantic fusion feature vector V and improving the accuracy of the decoded values ​​obtained by the decoder from the trained multimodal semantic fusion feature vector V.

[0063] In summary, the mulching equipment for corn side-planting according to the embodiments of this application is explained. It utilizes deep learning-based artificial intelligence technology to analyze soil condition images and overall machine height, extracting information on the soil's virtual and solid state characteristics, and intelligently recommending the compaction intensity of the compaction rollers based on the overall machine height characteristics. This enables intelligent recommendation of compaction intensity, thereby optimizing soil compaction, promoting crop growth and development, and improving the economic benefits of corn side-planting under mulch.

[0064] Figure 5 This is a flowchart of an intelligent control method for a mulching device used for corn film-side sowing according to an embodiment of this application. Figure 5 As shown, the intelligent control method for a mulching device for corn film-side sowing according to an embodiment of this application includes the following steps: S1, acquiring a soil state image captured by a camera; S2, acquiring the overall height value of the machine captured by a sensor; S3, extracting soil state features from the soil state image to obtain a sequence of context soil local state feature vectors; S4, performing one-hot encoding on the overall height value to obtain an overall height one-hot encoded feature vector; S5, determining a recommended value for compaction intensity based on the semantic fusion features of the sequence of context soil local state feature vectors and the overall height one-hot encoded feature vector.

[0065] Here, those skilled in the art will understand that the specific operations of each step in the above-described intelligent control method for mulching equipment used for corn film-side sowing have been referenced above. Figures 1 to 4 The description of the mulching equipment for corn film-side planting is detailed therein, and therefore, its repeated description will be omitted.

[0066] In the above embodiments, the descriptions of each embodiment have their own emphasis. Parts not detailed or described in a particular embodiment can be referred to in the relevant descriptions of other embodiments. In the several embodiments provided by this invention, it should be understood that the disclosed devices, systems, and methods can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for example, the module division is only a logical functional division, and other division methods may exist in actual implementation. The modules described as separate components may or may not be physically separated. The components shown as modules may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0067] Furthermore, the functional modules in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or in the form of hardware plus software functional modules.

[0068] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be embraced within the present invention. No appended diagram markings in the claims should be construed as limiting the scope of the claims.

[0069] Furthermore, it is clear that the word "comprising" does not exclude other units or steps, and the singular does not exclude the plural. Multiple units stated in a system claim may also be implemented by a single unit through software or hardware.

[0070] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A film-covering device for side-planting of corn, comprising: The compaction intensity adjustment module is used to adjust the compaction intensity of the compaction roller based on the soil compaction and the overall height of the machine; The soil compaction module is used to compact the soil according to the compaction intensity of the compaction roller to obtain an arched film laying belt that is high in the middle and low on both sides; the film laying module is used to use the film pressing wheel and the tensioning shaft to tension the film installed on the film laying roller and lay it flat on the arched film laying belt to obtain the film covering the film laying belt. The membrane-fixing module is used to use a soil-covering shovel behind the pressing wheel to shovel the surface soil on both sides of the mulch film covering the mulch film onto both sides of the mulch film covering the mulch film, and to compact the soil covering the mulch film covering the mulch film onto the mulch film covering ... Soil condition monitoring unit, used to acquire soil condition images captured by camera; The overall height information acquisition unit is used to acquire the overall height value collected by the sensor; A soil state feature extraction unit is used to extract soil state features from the soil state image to obtain a sequence of context soil local state feature vectors. The whole machine height one-hot encoding unit is used to perform one-hot encoding on the whole machine height value to obtain the whole machine height one-hot encoded feature vector; The compaction intensity recommendation unit is used to determine the recommended value of compaction intensity based on the semantic fusion features of the sequence of local soil state feature vectors in the context and the one-hot encoded feature vector of the overall height. The pressing intensity recommendation unit includes: The feature fusion subunit is used to fuse the sequence of the context soil local state feature vector and the whole machine height one-hot encoded feature vector to obtain a multimodal semantic fusion feature vector; The decoding regression subunit is used to pass the multimodal semantic fusion feature vector through a decoder to obtain a decoded value, which is used to represent the recommended suppression intensity value; The feature fusion subunit is used for: A cross-modal interactive fusion engine is used to semantically fuse the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the whole machine height using the following cross-modal interactive fusion formula to obtain the multimodal semantic fusion feature vector. The cross-modal interactive fusion engine calculates attention weights using the correlation information between the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the whole machine height, and performs weighted fusion on the sequence of context soil local state feature vectors and the one-hot encoded feature vector of the whole machine height according to the attention weights. The cross-modal interaction fusion formula is as follows: in, A is the one-hot encoded feature vector of the entire machine height, where A is 1× The matrix, It is the dimension of the one-hot encoded feature vector of the entire machine. B is the i-th context soil local state feature vector in the sequence of context soil local state feature vectors, and B is 1× The matrix, It is the dimension of the feature vector of each local soil state in the aforementioned context. It is the number of feature vectors of the local soil state in the context. It is the Sigmoid function. It is the multimodal semantic fusion feature vector. and A convolution operation representing a 1×1 convolution kernel. It is the weight value; The device also includes a training module for training a soil state feature extractor based on a convolutional neural network model, a soil state inter-regional correlation encoder based on a converter module, a cross-modal interactive fusion unit, and a decoder. The training module includes: A training data acquisition unit is used to acquire training data, which includes: training soil state images and training overall machine height values; The training image segmentation unit is used to perform image segmentation processing on the training soil state image to obtain a sequence of training soil local state image blocks; A training soil local state feature extraction unit is used to pass the sequence of training soil local state image patches through the soil state feature extractor based on the convolutional neural network model to obtain a sequence of training soil local state feature vectors. The training region association coding unit is used to pass the sequence of the training soil local state feature vectors through the soil state region association encoder based on the converter module to obtain the sequence of training context soil local state feature vectors. A training unit for one-hot encoding of overall machine height is used to perform one-hot encoding on the training overall machine height value to obtain a training overall machine height one-hot encoded feature vector. The training feature fusion unit is used to process the sequence of the training context soil local state feature vector and the training whole machine height one-hot encoded feature vector using a cross-modal interactive fusion device to obtain the training multimodal semantic fusion feature vector. The feature aggregation and optimization unit is used to perform feature aggregation and optimization on the trained multimodal semantic fusion feature vector to obtain an optimized trained multimodal semantic fusion feature vector; The decoding loss calculation unit is used to pass the optimized training multimodal semantic fusion feature vector through the decoder to obtain the decoding loss function value; The training unit is used to train the soil state feature extractor based on the convolutional neural network model, the soil state inter-regional association encoder based on the converter module, the cross-modal interactive fusion unit, and the decoder based on the decoding loss function value and through directional propagation of gradient descent. The feature aggregation optimization unit is used to perform feature aggregation optimization on the trained multimodal semantic fusion feature vector using the following feature aggregation optimization formula to obtain the optimized trained multimodal semantic fusion feature vector; wherein, the feature aggregation optimization formula is: in, It is the trained multimodal semantic fusion feature vector. It is the first of the trained multimodal semantic fusion feature vectors Feature values ​​at each position, It is the trained multimodal semantic fusion feature vector The square of the 1 norm, It is the trained multimodal semantic fusion feature vector The reciprocal of the square root of the 2-norm, It is the trained multimodal semantic fusion feature vector The length, and It's a scaling hyperparameter. This represents the value of the logarithmic function with base 2. It is the first of the optimized training multimodal semantic fusion feature vectors Feature values ​​at each position; The low-rank structure of the norm of the trained multimodal semantic fusion feature vector is represented as a voting cluster of the feature value set of the trained multimodal semantic fusion feature vector. For each feature value of the trained multimodal semantic fusion feature vector, canonical voting is performed relative to the information perception framework. By aggregating the regression representation of the direction and scale of the feature distribution, feature values ​​belonging to the same regression class are mapped to similar locally canonical coordinate sets, thereby improving the training multimodal semantic fusion feature vector. The aggregation effect of feature sets.

2. The mulching equipment for corn film-side sowing according to claim 1, characterized in that, The soil condition feature extraction unit includes: The image segmentation subunit is used to perform image segmentation processing on the soil state image to obtain a sequence of local soil state image blocks; The soil local state feature extraction subunit is used to extract the soil state features of each soil local state image block in the sequence of soil local state image blocks to obtain a sequence of soil local state feature vectors. The inter-regional association coding subunit is used to extract the inter-regional contextual association features of the sequence of soil local state feature vectors to obtain the sequence of context soil local state feature vectors.

3. The mulching equipment for corn film-side sowing according to claim 2, characterized in that, The soil local state feature extraction subunit is used for: The sequence of soil local state image patches is processed by a soil state feature extractor based on a convolutional neural network model to obtain a sequence of soil local state feature vectors.

4. The mulching equipment for corn film-side sowing according to claim 3, characterized in that, The soil state feature extractor based on the convolutional neural network model includes an input layer, a convolutional layer, an activation layer, a pooling layer, a flattening layer, and an output layer.

5. The mulching equipment for corn film-side sowing according to claim 4, characterized in that, The inter-regional association coding subunit is used for: The sequence of soil local state feature vectors is passed through a soil state inter-regional association encoder based on a converter module to obtain the sequence of context soil local state feature vectors.

6. The mulching equipment for corn film-side sowing according to claim 5, characterized in that, The feature fusion subunit is used for: A cross-modal interactive fusion processor is used to process the sequence of local soil state feature vectors and the one-hot encoded feature vector of the overall height to obtain the multimodal semantic fusion feature vector.

7. An intelligent control method for a mulching device used for corn film-side sowing, executed by the mulching device as described in any one of claims 1 to 6, characterized in that, include: Acquire soil condition images captured by a camera; Obtain the overall height value of the machine collected by the sensors; Extract soil state features from the soil state image to obtain a sequence of contextual soil local state feature vectors; The overall height value is one-hot encoded to obtain the overall height one-hot encoded feature vector; Based on the semantic fusion features of the sequence of local soil state feature vectors in the context and the uniquely encoded feature vector of the overall machine height, a recommended value for compaction intensity is determined.

Images