Water-soluble fertilizer anti-caking detection method and system

Abstract

The invention provides a water-soluble fertilizer anti-caking detection method and system, and relates to the technical field of detection.The method comprises the steps that a detection position is set and pressed, a telescopic assembly is controlled to press the detection position, and resistance data and sound wave information are detected; and controlling the telescopic assembly to retract and move to the next detection position until the resistance data and the sound wave information of all the detection positions are obtained, so as to determine the caking condition. According to the invention, various data can be comprehensively analyzed according to the various data of the contact head, the telescopic assembly and the microphone of the detection table in the process of pressing the water-soluble fertilizer packaging bag, the caking condition of the water-soluble fertilizer in the packaging bag is determined, the water-soluble fertilizer packaging bag is not damaged in the detection process, nondestructive detection is carried out, the detection cost is reduced, and the detection efficiency is improved. In addition, pressing resistance data and sound wave information in the pressing process can be collected, so that the caking condition is comprehensively analyzed through the two kinds of data, and the analysis accuracy can be improved.

Description

Technical Field

[0001] This invention relates to the field of detection technology, and in particular to a method and system for detecting the anti-caking properties of water-soluble fertilizers. Background Technology

[0002] In related technologies, the detection of powder agglomeration usually employs destructive testing, which involves opening or puncturing the powder packaging to obtain samples for testing. However, for water-soluble fertilizers that are easily soluble in water, the packaging bags have a waterproof and sealed design, which is costly, making destructive testing expensive. Furthermore, breaking the packaging bag may cause the previously unagglomerated water-soluble fertilizer to absorb moisture from the air and agglomerate, resulting in waste of the fertilizer.

[0003] The information disclosed in the background section of this application is intended only to enhance the understanding of the general background of this application and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0004] This invention provides a method and system for detecting the anti-caking properties of water-soluble fertilizers, which can solve the technical problems of high cost and waste in related technologies.

[0005] According to a first aspect of the present invention, a method for detecting the anti-caking properties of water-soluble fertilizer is provided, comprising: placing a water-soluble fertilizer packaging bag flat on a testing platform and setting multiple testing positions on the surface of the water-soluble fertilizer packaging bag; aligning a contact head with the testing position and pressing the testing position with a first preset pressure, wherein the contact head is a cylindrical cavity, the lower surface of the contact head is an annular contact surface, a telescopic component is provided at the axial position of the cylindrical cavity of the contact head, the lower contact surface of the telescopic component is a circular contact surface, the centroid of the circular contact surface coincides with the centroid of the annular contact surface, the circular contact surface can extend out of the annular contact surface to press down on the testing position of the water-soluble fertilizer packaging bag; controlling the telescopic component to press down on the testing position of the water-soluble fertilizer packaging bag. The system presses down on the detection position at a first preset speed, and detects the resistance data of the telescopic component at multiple moments during the pressing process. During the pressing process, sound wave information is collected through a microphone in the cylindrical cavity. When the lower contact surface of the telescopic component extends beyond the annular contact surface by a first preset distance, the pressing stops, and the lower contact surface of the telescopic component retracts above the annular contact surface. The contact head stops pressing on the detection position and moves to the next detection position. The process of acquiring resistance data and sound wave information is iteratively executed until resistance data and sound wave information corresponding to all detection positions are acquired. Based on the resistance data and sound wave information corresponding to multiple detection positions, the agglomeration status of the water-soluble fertilizer is determined.

[0006] According to the present invention, determining the agglomeration status of water-soluble fertilizer based on resistance data and acoustic information corresponding to multiple detection locations includes: sampling acoustic information to obtain acoustic sampling data; obtaining multiple time-domain feature information based on the acoustic sampling data; forming a time-domain feature vector from the multiple time-domain feature information; performing a fast Fourier transform on the acoustic sampling data to obtain acoustic frequency-domain data; obtaining multiple frequency-domain feature information based on the acoustic frequency-domain data; forming a frequency-domain feature vector from the multiple frequency-domain feature information; forming a resistance feature vector from the resistance data; obtaining a subgraph structure corresponding to the detection location based on the time-domain feature vector, the frequency-domain feature vector, and the resistance feature vector; processing the subgraph structure using a first graph model to obtain detection state information for each detection location; determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the distance between the detection state information and the detection location; and processing the graph structure using a second graph model to obtain the agglomeration status of the water-soluble fertilizer.

[0007] According to the present invention, obtaining a subgraph structure corresponding to the detection position based on a time-domain feature vector, a frequency-domain feature vector, and a drag feature vector includes: processing the time-domain feature vector through a first fully connected layer to obtain a first input vector of a first subgraph node in the subgraph structure; processing the frequency-domain feature vector through a second fully connected layer to obtain a second input vector of a second subgraph node in the subgraph structure; processing the drag feature vector through a third fully connected layer to obtain a third input vector of a third subgraph node in the subgraph structure; and processing the first and second input vectors according to the attention mechanism of a first graph model to obtain a first... The connection weights of the subgraph nodes and the second subgraph nodes are obtained by processing the first and third input vectors according to the attention mechanism of the first graph model, and by processing the second and third input vectors according to the attention mechanism of the first graph model, and obtaining the connection weights of the second and third subgraph nodes. Based on the first input vector, the second input vector, the third input vector, the connection weights of the first and second subgraph nodes, the connection weights of the first and third subgraph nodes, and the connection weights of the second and third subgraph nodes, the subgraph structure is obtained.

[0008] According to the present invention, the detection state information of each detection position is obtained by processing the subgraph structure through a first graph model, including: processing the subgraph structure through the first graph model to obtain a first output vector of a first subgraph node, a second output vector of a second subgraph node, and a third output vector of a third subgraph node; concatenating the first output vector, the second output vector, and the third output vector to obtain a detection position feature vector; and inputting the detection position feature vector into a fourth fully connected layer and a first activation layer for processing to obtain detection state information.

[0009] According to the present invention, determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the detection status information and the distance between the detection positions includes: mapping multiple detection positions to multiple nodes in the graph structure; using the detection status information of each detection position as the node input vector of each node; determining an adjacency matrix based on the detection status information and the distance between the detection positions; and determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the adjacency matrix and the node input vector of each node.

[0010] According to the present invention, determining the adjacency matrix based on the distance between detection state information and detection positions includes: according to the formula Determine the data in the i-th row and j-th column of the adjacency matrix. and the data in the j-th row and i-th column ,in, This represents the detection status information of the i-th node. This refers to the detection status information of the j-th node. for and Cosine similarity between them Let be the distance between the i-th detection position and the j-th detection position. The distance between multiple detection locations is the average distance, and "and" indicates a logical AND operation. Indicates except as well as In all other cases, min is the function that takes the minimum value.

[0011] According to the present invention, the clumping status of water-soluble fertilizer is obtained by processing the graph structure through a second graph model, including: determining the degree matrix based on the adjacency matrix; processing the adjacency matrix, degree matrix, and node input vector of each node through the second graph model to obtain the node output vector of each node; concatenating the node output vector of each node to obtain the water-soluble fertilizer state vector; and inputting the water-soluble fertilizer state vector into the fifth fully connected layer and the second activation layer for processing to obtain the clumping status of the water-soluble fertilizer, wherein the clumping status is the probability of clumping of water-soluble fertilizer in the water-soluble fertilizer packaging bag.

[0012] According to the present invention, the method further includes: acquiring the sample subgraph structure of each detection position of the sample water-soluble fertilizer packaging bag; processing the sample subgraph structure through a first graph model to acquire sample detection state information of each detection position; inputting the sample detection state information of each detection position into a sixth fully connected layer and a third activation layer for processing to obtain the clumping probability of each detection position; obtaining a clumping state loss function of the detection position based on the annotation information of each detection position and the clumping probability; acquiring the sample graph structure based on the sample detection state information of each detection position and the distance between each detection position; processing the sample graph structure through a second graph model to obtain the sample clumping status; acquiring the clumping annotation of the sample water-soluble fertilizer packaging bag based on the annotation information of each detection position; acquiring a clumping status loss function based on the clumping probability, the sample clumping status, and the clumping annotation; and training the first graph model and the second graph model based on the detection position clumping status loss function and the clumping status loss function to obtain the trained first graph model and the trained second graph model.

[0013] According to the present invention, a block status loss function is obtained based on the block probability, the sample block status, and the block label, including: according to the formula Obtain the loss function for the clumping status. ,in, This is the probability of sample water-soluble fertilizer clumping determined based on clumping labeling. For sample clumping status, Let be the probability of a block at the i-th detection position. Let N be the probability of a block forming at the i-th detection position, determined based on the annotation information of the i-th detection position. Let N be the number of detection positions, i ≤ N, and both i and N are positive integers. Let max be the function that takes the maximum value.

[0014] According to a second aspect of the present invention, a water-soluble fertilizer anti-caking detection system is provided, comprising: a detection position module for placing a water-soluble fertilizer packaging bag flat on a detection table and setting multiple detection positions on the surface of the water-soluble fertilizer packaging bag; a pressing module for aligning a contact head with the detection position and pressing the detection position according to a first preset pressure, wherein the contact head is a cylindrical cavity, the lower surface of the contact head is an annular contact surface, a telescopic component is provided at the axial position inside the cylindrical cavity of the contact head, the lower contact surface of the telescopic component is a circular contact surface, the centroid of the circular contact surface coincides with the centroid of the annular contact surface, and the circular contact surface can extend out of the annular contact surface to press down on the detection position of the water-soluble fertilizer packaging bag; and a resistance data module for controlling the telescopic component to press down on the detection position of the water-soluble fertilizer packaging bag according to a first preset speed. The system includes a pressure-pressing detection position and a resistance data detection module that monitors the resistance of the telescopic component at multiple points during the pressing process. A sound wave information module collects sound wave information via a microphone in a cylindrical cavity during pressing. A retraction module stops pressing when the lower contact surface of the telescopic component extends a first preset distance beyond the annular contact surface and retracts the lower contact surface of the telescopic component above the annular contact surface. A movement module stops the contact head from pressing the detection position and moves to the next detection position. An iteration module iteratively processes the acquisition of resistance data and sound wave information until all detection positions have acquired their corresponding resistance data and sound wave information. A clumping status module determines the clumping status of the water-soluble fertilizer based on the resistance data and sound wave information from multiple detection positions.

[0015] Technical Effects: According to the present invention, multiple data points can be collected during the pressing of water-soluble fertilizer packaging bags via the contact head, telescopic component, and microphone of the testing station. These data points can then be comprehensively analyzed to determine the clumping status of the water-soluble fertilizer inside the packaging bag. The testing process does not damage the packaging bag, enabling non-destructive testing, reducing testing costs and minimizing fertilizer waste. Furthermore, it can collect pressing resistance data and acoustic wave information during the pressing process, thereby improving the accuracy of the analysis by comprehensively analyzing the clumping status using both types of data. When acquiring detection state information, an attention mechanism can be used to establish the correlation between the high-dimensional vector representations of the acoustic wave's time-domain characteristics, frequency-domain characteristics, and resistance characteristics. This allows for the fusion of these high-dimensional vector representations using a first graph model, obtaining the output vectors of each subgraph node. The resulting vectors can also represent multiple aspects of the detection location's characteristics. A fourth fully connected layer further strengthens the influence between different high-dimensional vector representations, obtaining detection state information that comprehensively represents the state of the detection location, thus enhancing the ability of the detection state information to describe the state of the detection location. When determining the adjacency matrix, the cosine similarity of the detection state information of nodes can be used to represent the similarity of the clustering status of the detection positions corresponding to the nodes. Furthermore, based on the similarity of the clustering status and the distance between detection positions, the similarity of the clustering status and the environmental influences at each detection position can be comprehensively represented, thereby obtaining various data in the adjacency matrix and improving the accuracy of the association relationships between nodes. During model training, a sixth fully connected layer and a third activation layer can be introduced to assist training, thereby obtaining the clustering status loss function for each detection position to enhance the training strength of the first graph model. Moreover, when obtaining the clustering status loss function, the maximum error between the annotation information and the clustering probability of each detection position can be used to enhance the training strength, which not only improves the accuracy of the sample clustering status but also further improves the accuracy of the clustering probability at each detection position, thus improving the overall training efficiency and model accuracy.

[0016] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Other features and aspects of the invention will become clearer from the following detailed description of exemplary embodiments with reference to the accompanying drawings. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other embodiments can be obtained based on these drawings without creative effort.

[0018] Figure 1A flowchart of a method for detecting the anti-caking of water-soluble fertilizers according to an embodiment of the present invention is shown as an example;

[0019] Figure 2 A schematic diagram illustrating the caking state of water-soluble fertilizer according to an embodiment of the present invention is shown.

[0020] Figure 3 A block diagram of a water-soluble fertilizer anti-caking detection system according to an embodiment of the present invention is shown as an example. Detailed Implementation

[0021] 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. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] The technical solution of the present invention will be described in detail below with reference to specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments.

[0023] Figure 1A flowchart of a method for detecting anti-caking of water-soluble fertilizer according to an embodiment of the present invention is shown as an example. The method includes: step S1, placing a water-soluble fertilizer packaging bag flat on a testing platform and setting multiple testing positions on the surface of the water-soluble fertilizer packaging bag; step S2, aligning a contact head with a testing position and pressing the testing position with a first preset pressure, wherein the contact head is a cylindrical cavity, the lower surface of the contact head is an annular contact surface, a telescopic component is provided at the axial position of the cylindrical cavity of the contact head, the lower contact surface of the telescopic component is a circular contact surface, the centroid of the circular contact surface coincides with the centroid of the annular contact surface, and the circular contact surface can extend out of the annular contact surface to press down on the testing position of the water-soluble fertilizer packaging bag; step S3, controlling the telescopic component to press down on the testing position at a first preset speed. The process involves detecting resistance data of the telescopic component at multiple moments during the pressing process; step S4: during the pressing process, sound wave information is collected through a microphone in the cylindrical cavity; step S5: when the distance from the lower contact surface of the telescopic component extending beyond the annular contact surface reaches a first preset distance, the pressing is stopped, and the lower contact surface of the telescopic component is controlled to retract above the annular contact surface; step S6: the contact head stops pressing the detection position and moves to the next detection position; step S7: the process of acquiring resistance data and sound wave information is iteratively executed until resistance data and sound wave information corresponding to all detection positions are acquired; step S8: based on the resistance data and sound wave information corresponding to multiple detection positions, the agglomeration status of the water-soluble fertilizer is determined.

[0024] The water-soluble fertilizer anti-caking detection method according to an embodiment of the present invention can comprehensively analyze various data collected during the pressing process of the water-soluble fertilizer packaging bag by the contact head, telescopic component and microphone of the detection platform, thereby determining the caking status of the water-soluble fertilizer inside the packaging bag. The water-soluble fertilizer packaging bag is not damaged during the detection process, so as to carry out non-destructive testing, reduce detection costs, reduce waste of water-soluble fertilizer, and collect pressing resistance data and sound wave information during the pressing process, thereby improving the accuracy of analysis by comprehensively analyzing the caking status through the two types of data.

[0025] According to one embodiment of the present invention, the water-soluble fertilizer packaging bag is completely sealed and contains a predetermined weight of water-soluble fertilizer. The testing platform may include a platform for placing the water-soluble fertilizer packaging bag, and may also include a contact head capable of longitudinal lifting and planar displacement. The contact head can be connected to an operating lever of the testing platform, allowing the contact head to be operated for longitudinal lifting and planar displacement via the operating lever. The upper surface of the contact head is a closed circular plane with a connecting structure (e.g., a screw) to connect the contact head to the operating lever. The contact head is shaped as a cylindrical cavity with a circular annular radial section, and the axial direction of the cylindrical cavity is vertical. The lower surface of the contact head is an annular contact surface (the lower surface is not closed). The contact head can move longitudinally downwards, pressing the water-soluble fertilizer packaging bag with its lower surface, and testing the clumping of the water-soluble fertilizer inside the packaging bag via a telescopic component and microphone within its central cavity. Because the lower surface is not closed, the lower contact surface of the telescopic component, in its extended state, can be lower than the lower surface of the contact head. The microphone can be housed within a cylindrical cavity, which not only provides some protection for the microphone but also acts as a sound insulator. For example, after the water-soluble fertilizer packaging bag is pressed against the lower surface of the contact head, the space inside and outside the cavity becomes relatively independent, reducing external noise interference and allowing the microphone to capture the sound of the telescopic component pressing the fertilizer packaging bag. The telescopic component can be an electrically operated telescopic rod, allowing control over the speed and distance of extension. The axis of the telescopic component coincides with the axis within the cylindrical cavity of the contact head, and the lower contact surface of the telescopic component is a circular contact surface, the centroid of which coincides with the centroid of the annular contact surface. The microphone's power and data cables, as well as the power cable of the telescopic component, can pass through the upper surface or side wall of the contact head to connect to an external power or data receiving port. That is, the upper surface or side wall of the contact head can be perforated, and after the power and data cables are passed through these holes, they can be sealed with adhesive.

[0026] According to an embodiment of the present invention, in step S1, the packaging bag of water-soluble fertilizer can be laid flat on the testing platform, and multiple testing positions can be set on the surface of the packaging bag. For example, testing positions can be set at the center of the upper surface of the packaging bag and near the corners, or testing positions can be randomly set on the upper surface. The present invention does not limit the way the testing positions are set.

[0027] According to one embodiment of the present invention, in step S2, the operating lever is controlled to perform planar displacement, causing the centroid of the lower surface of the contact head to rotate relative to the detection position, and the operating lever is controlled to descend, so that the cylindrical cavity of the contact head covers the detection position. Furthermore, the operating lever can be controlled to press the contact head against the detection position with a first preset pressure (e.g., 2N), thus isolating the inside and outside of the cylindrical cavity to reduce external noise interference.

[0028] According to one embodiment of the present invention, in step S3, the control extension component presses down on the detection position at a first preset speed (e.g., 1 mm / s) for 20 seconds. The maximum distance the circular contact surface can extend beyond the annular contact surface is 2 cm, i.e., the detection position is pressed down 2 cm. During the pressing process (i.e., from the moment the lower contact surface of the extension component contacts the detection position until the downward pressing stops), resistance data experienced by the extension component can be monitored at multiple moments, for example, resistance data can be detected once per second. The present invention does not limit the duration of the pressing process, the value of the first preset speed, or the time interval between adjacent moments. The contact head can contact the surface of the water-soluble fertilizer packaging bag in the initial state, and can press down when the contact head extends. Furthermore, when the water-soluble fertilizer clumps, it clumps from the surface inwards. Therefore, when the water-soluble fertilizer clumps, a shell-like structure is formed inside the packaging bag. The resistance generated by the shell-like structure when the contact head presses down is significantly different from the resistance generated when the water-soluble fertilizer is not clumped. During the pressing process, although the water-soluble fertilizer packaging bag itself will also generate resistance, since the downward pressing distance is the same, the resistance generated by the packaging bag is consistent regardless of whether the water-soluble fertilizer clumps. That is, the difference in resistance during pressing is due to whether the water-soluble fertilizer clumps, and is unrelated to the resistance of the packaging bag. Therefore, the resistance during pressing can be used to predict whether the water-soluble fertilizer will clump.

[0029] According to one embodiment of the present invention, in step S4, during the pressing process, sound wave information can be collected by a microphone, that is, sound wave information of the sound generated at the pressing detection position.

[0030] According to one embodiment of the present invention, in step S5, when the lower contact surface of the telescopic component extends beyond the annular contact surface by a first preset distance, the pressing stops; that is, the telescopic component no longer extends but retracts above the annular contact surface. In step S6, instead of pressing the detection position, the operating lever is controlled vertically upward and subjected to planar displacement, aligning the centroid of the lower surface of the contact head with the next detection position. In step S7, the above operations can be iteratively performed until all detection positions have been detected, obtaining the resistance data and acoustic information corresponding to all detection positions.

[0031] According to one embodiment of the present invention, in step S8, the resistance and sound generated when pressing the packaging bag of water-soluble fertilizer are different depending on whether it is in a clumped state or not. For example, when water-soluble fertilizer clumps due to moisture, it clumps from the surface towards the center, increasing surface hardness and local density. Therefore, the resistance when pressing the packaging bag is greater than the resistance when pressing unclumped water-soluble fertilizer. Furthermore, when pressed, the clumped water-soluble fertilizer may crack, producing a crisp sound with a louder volume and narrower bandwidth. Conversely, the sound of unclumped water-soluble fertilizer when pressed is similar to pressing a sandbag, with a lower volume and a wider bandwidth. Therefore, based on these differences, the resistance data and sound wave information collected during pressing at each detection position can be analyzed to obtain the overall clumping status of the water-soluble fertilizer inside the packaging bag.

[0032] According to an embodiment of the present invention, in step S8, determining the agglomeration status of water-soluble fertilizer based on resistance data and acoustic information corresponding to multiple detection locations includes: sampling acoustic information to obtain acoustic sampling data; obtaining multiple time-domain feature information based on the acoustic sampling data; forming a time-domain feature vector from the multiple time-domain feature information; performing a fast Fourier transform on the acoustic sampling data to obtain acoustic frequency-domain data; obtaining multiple frequency-domain feature information based on the acoustic frequency-domain data; forming a frequency-domain feature vector from the multiple frequency-domain feature information; forming a resistance feature vector from the resistance data; obtaining a subgraph structure corresponding to the detection location based on the time-domain feature vector, the frequency-domain feature vector, and the resistance feature vector; processing the subgraph model using a first graph model to obtain detection state information for each detection location; determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the distance between the detection state information and the detection location; and processing the graph structure using a second graph model to obtain the agglomeration status of the water-soluble fertilizer.

[0033] Figure 2 A schematic diagram illustrating the process of obtaining the clumping condition of water-soluble fertilizer according to an embodiment of the present invention is shown.

[0034] According to one embodiment of the present invention, sound wave information can be sampled at a preset sampling rate (e.g., 8000Hz, 11025Hz, 22050Hz, etc.) to obtain sound wave sampling data. The sound wave sampling data can then be analyzed to obtain various time-domain feature information, such as root mean square, peak-to-peak value, kurtosis, skewness, impulse factor, etc. of the sound wave sampling data. These various time-domain feature information can be combined into a time-domain feature vector, that is, each data in the time-domain feature vector is a kind of time-domain feature information.

[0035] According to one embodiment of the present invention, a fast Fourier transform can be performed on the acoustic wave sampling data to obtain acoustic wave frequency domain data, and the acoustic wave frequency domain data can be analyzed to obtain a variety of frequency domain feature information, such as fundamental frequency, spectral centroid, spectral bandwidth, signal-to-noise ratio, etc. The various frequency domain feature information can be combined into a frequency domain feature vector, that is, each data in the frequency domain feature vector is a kind of frequency domain feature information.

[0036] According to one embodiment of the present invention, resistance data can be composed into a resistance feature vector. For example, resistance data can be composed into a resistance feature vector in chronological order. That is, the x-th data in the resistance feature vector is the resistance data at the x-th moment in the pressing process, where x is a positive integer and x is less than or equal to the total number of moments included in the pressing process.

[0037] According to one embodiment of the present invention, both the subgraph structure and the graph structure are graph data structures, which may include nodes and edges. Nodes are participants in the graph data structure, and each node may have its own descriptive information to describe its own characteristics. Edges may indicate whether there is an association between two nodes. If there is an edge connecting two nodes, then there is an association between the two nodes. The descriptive information of each node and the connecting edges can be processed by a graph model (e.g., a graph convolutional neural network model) to obtain richer feature information of the node through its own descriptive information and the descriptive information of other nodes connected to it. The richer feature information of each node can also be combined to obtain the feature information of the entire graph data structure. For example, if multiple nodes in the graph data structure are regarded as a group, then the feature information of the entire graph data structure can be used to describe the feature information of the group.

[0038] According to an embodiment of the present invention, obtaining a subgraph structure corresponding to the detection position based on a time-domain feature vector, a frequency-domain feature vector, and a drag feature vector includes: processing the time-domain feature vector through a first fully connected layer to obtain a first input vector of a first subgraph node in the subgraph structure; processing the frequency-domain feature vector through a second fully connected layer to obtain a second input vector of a second subgraph node in the subgraph structure; processing the drag feature vector through a third fully connected layer to obtain a third input vector of a third subgraph node in the subgraph structure; and processing the first and second input vectors according to the attention mechanism of a first graph model to obtain... The connection weights of the first and second subgraph nodes are obtained. The first and third input vectors are processed according to the attention mechanism of the first graph model to obtain the connection weights of the first and third subgraph nodes. The second and third input vectors are processed according to the attention mechanism of the first graph model to obtain the connection weights of the second and third subgraph nodes. Based on the first, second, and third input vectors, the connection weights of the first and second subgraph nodes, the connection weights of the first and third subgraph nodes, and the connection weights of the second and third subgraph nodes, the subgraph structure is obtained.

[0039] According to one embodiment of the present invention, the dimensions of the time-domain feature vector, the frequency-domain feature vector, and the drag feature vector may be different. To facilitate subsequent calculations, they can be processed through a first fully connected layer, a second fully connected layer, and a third fully connected layer, respectively, to obtain a first input vector, a second input vector, and a third input vector with consistent dimensions. Furthermore, the dimensions of the first input vector, the second input vector, and the third input vector can be higher than the maximum value among the time-domain feature vector, the frequency-domain feature vector, and the drag feature vector. That is, the time-domain feature vector, the frequency-domain feature vector, and the drag feature vector are upgraded through the first fully connected layer, the second fully connected layer, and the third fully connected layer, thereby obtaining information that can more comprehensively describe the features of each detection position, namely, the first input vector, the second input vector, and the third input vector.

[0040] According to one embodiment of the present invention, the first input vector is a high-dimensional vector representation of the temporal characteristics of the sound wave, the second input vector is a high-dimensional vector representation of the frequency characteristics of the sound wave, and the third input vector is a high-dimensional vector representation of the resistance data. There is a certain correlation between the temporal characteristics, frequency characteristics, and resistance data of the sound wave. For example, when water-soluble fertilizer is not clumped, the pressing resistance data is low, the sound volume is low, and the bandwidth is large. Therefore, in the temporal characteristics of the sound wave, the peak-to-peak value is small, and in the frequency characteristics, the spectral bandwidth is large. Conversely, when water-soluble fertilizer is clumped, the pressing resistance data is large, the sound volume is high, and the bandwidth is narrow. Therefore, in the temporal characteristics of the sound wave, the peak-to-peak value is large, and in the frequency characteristics, the spectral bandwidth is narrow. That is, there is a certain correlation between the temporal characteristics, frequency characteristics, and resistance data of the sound wave, but this correlation is difficult to identify through simple numerical values. Therefore, an attention mechanism can be used to automatically calculate the relationship between the first, second, and third input vectors, thereby obtaining the connection weights of the edges between each node in the subgraph structure. The parameters of the attention mechanism can be obtained through training.

[0041] In the example, the first and second input vectors can be concatenated, and the concatenated vector can be processed using an attention mechanism. For example, a weighted summation can be performed on the data of each dimension of the concatenated vector using a set of weight parameters to obtain a weighted summation result. Similarly, a weighted summation result can be obtained for the vector obtained by concatenating the first and third input vectors, as well as the vector obtained by concatenating the second and third input vectors. Each weighted summation result is then processed using an activation function (e.g., the softmax activation function) to obtain the connection weights corresponding to each weighted summation result, namely the connection weights between the first and second subgraph nodes, the connection weights between the first and third subgraph nodes, and the connection weights between the second and third subgraph nodes. After obtaining the input vector of each subgraph node (i.e., the first input vector of the first subgraph node, the second input vector of the second subgraph node, and the third input vector of the third subgraph node) and the connection weights between each subgraph node, the subgraph structure is obtained.

[0042] According to one embodiment of the present invention, processing the subgraph structure through a first graph model to obtain detection state information for each detection position includes: processing the subgraph structure through the first graph model to obtain a first output vector of a first subgraph node, a second output vector of a second subgraph node, and a third output vector of a third subgraph node; concatenating the first output vector, the second output vector, and the third output vector to obtain a detection position feature vector; and inputting the detection position feature vector into a fourth fully connected layer and a first activation layer for processing to obtain detection state information.

[0043] According to an embodiment of the present invention, the first graph model is a graph convolutional neural network model that can obtain the output vector of each subgraph node based on the input vector of each subgraph node and the connection weights between subgraph nodes. For example, the input vector of the subgraph node itself can be weighted and summed with the input vectors of other subgraph nodes through the connection weights to obtain the output vector of the subgraph node. That is, the first output vector of the first subgraph node, the second output vector of the second subgraph node, and the third output vector of the third subgraph node are obtained.

[0044] According to an embodiment of the present invention, as described above, the input vectors of each subgraph node are respectively high-dimensional vector representations of the time-domain features of the sound wave, high-dimensional vector representations of the frequency-domain features of the sound wave, and high-dimensional vector representations of the drag data. Therefore, after the above processing, the output vector of each subgraph node is a vector that fuses its own high-dimensional vector representation with the relevant high-dimensional vector representations, which can have richer representation information. That is, it not only has its own high-dimensional vector representation, but also can reflect the influence of other high-dimensional vector representations on itself, and can be used as the output vector of the subgraph node.

[0045] According to one embodiment of the present invention, the first output vector, the second output vector, and the third output vector can be concatenated to obtain an overall vector representation of a group (subgraph structure) formed by multiple subgraph nodes. Furthermore, the subgraph structure has a correspondence with the detection position. Therefore, the overall vector representation is the feature vector of the detection position, which can be used to represent multiple aspects of the features of the detection position.

[0046] According to one embodiment of the present invention, the detection position feature vector can be reduced in dimensionality through a fourth fully connected layer. For example, the first, second, and third output vectors are all 64-dimensional vectors. After concatenation, the detection position feature vector is a 192-dimensional vector. To facilitate subsequent calculations, it can be reduced to 64 dimensions through the fourth fully connected layer. Furthermore, during the fully connected layer operation, different output vectors can be weighted and summed, allowing the data from different output vectors to be fused, further strengthening the influence between different high-dimensional vector representations. Further, the output vector of the fourth fully connected layer can be processed through a first activation layer. For example, the ReLU activation function can be used to process each data point in the output vector of the fourth fully connected layer to obtain detection state information, which can be used to represent the state of the detection position.

[0047] In this way, the correlation between the high-dimensional vector representations of the temporal, frequency, and drag characteristics of sound waves can be established through an attention mechanism. This allows the first graph model to fuse the various high-dimensional vector representations, resulting in the output vectors of each subgraph node. Furthermore, the concatenated vectors can represent multiple aspects of the detection location's features. The fourth fully connected layer further enhances the influence between different high-dimensional vector representations, obtaining detection state information that comprehensively represents the state of the detection location and improving the ability of the detection state information to describe the state of the detection location.

[0048] According to one embodiment of the present invention, determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the detection status information and the distance between detection positions includes: mapping multiple detection positions one-to-one with multiple nodes in the graph structure; using the detection status information of each detection position as the node input vector of each node; determining an adjacency matrix based on the detection status information and the distance between detection positions; and determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the adjacency matrix and the node input vector of each node.

[0049] According to one embodiment of the present invention, each detection location can be considered as a node in a graph structure, and the detection state information of each detection location can be considered as the node input vector. The data in the adjacency matrix can be used to represent the connection relationships between nodes. Specifically, the data in the i-th row and j-th column can represent the connection relationship between the i-th node and the j-th node. Furthermore, since there is a one-to-one correspondence between nodes in the graph structure and detection locations, the actual physical relationship between detection locations can influence the connection relationships between nodes. For example, two detection locations that are closer in distance theoretically have a higher similarity. Therefore, the higher the correlation between the nodes corresponding to the two detection locations (i.e., the more similar their block conditions are theoretically), the larger the value in the adjacency matrix corresponding to the two nodes. Further, the connection relationship between nodes corresponding to detection locations can be related to the similarity of their block conditions. The higher the similarity of their block conditions, the more similar the environmental influences on the two detection locations, and the larger the value in the adjacency matrix corresponding to the two nodes.

[0050] According to one embodiment of the present invention, determining the adjacency matrix based on the distance between the detection state information and the detection position includes: determining the data in the i-th row and j-th column of the adjacency matrix according to formula (1). and the data in the j-th row and i-th column ,

[0051] (1)

[0052] in, This represents the detection status information of the i-th node. This refers to the detection status information of the j-th node. for and Cosine similarity between them Let be the distance between the i-th detection position and the j-th detection position. The distance between multiple detection locations is the average distance, and "and" indicates a logical AND operation. Indicates except as well as In all other cases, min is the function that takes the minimum value.

[0053] According to one embodiment of the present invention, for and The cosine similarity between the detection positions of the i-th and j-th nodes can represent the similarity of the block conditions between the detection positions of the i-th and j-th nodes. This indicates that the block status of the detection position corresponding to the i-th node is highly similar to that of the detection position corresponding to the j-th node. This indicates that the distance between the detection positions corresponding to the i-th node and the j-th node is less than or equal to the average distance between all detection positions. In other words, the detection positions corresponding to the i-th node and the j-th node are relatively close, indicating a high correlation between the two nodes. Therefore, in and If both conditions are met, the relationship between the i-th node and the j-th node can be represented as follows: and Setting both values ​​to 1 indicates a high degree of similarity between the block formation status and environmental influences at the detection locations corresponding to the two nodes. Conversely, setting them to 0 indicates a lower value. and If both conditions are met, it indicates that the similarity between the block status and environmental influence of the detection positions corresponding to the two nodes is low, which can represent the association between the i-th node and the j-th node. and The value is also set to 0.

[0054] According to one embodiment of the present invention, in other cases besides the two cases described above, for example, in and If both conditions are met, or and If both conditions are met, then... and The value is set to ,express The maximum value between 1 and 2. This can represent the relative relationship between the similarity and distance of the block conditions at the detection locations corresponding to two nodes. The higher the similarity of the block conditions at the detection locations corresponding to two nodes, the higher the value; the greater the distance between the detection locations corresponding to two nodes, the lower the value. Furthermore, in and If both conditions are met, although the block conditions at the detection locations of the two nodes are highly similar, they are far apart, and a ratio greater than 1 can be used to determine their similarity. As the denominator, the value of the above fraction is reduced to decrease the correlation between the two nodes. and When both conditions are met, although the block conditions at the detection locations corresponding to the two nodes are not very similar, their distance is relatively small. Therefore, a ratio less than 1 can be used to determine their similarity. As the denominator, the value of the above fraction is increased to indicate that the environmental influences on the two detection locations are more similar, thereby improving the correlation between the two nodes. Furthermore, by using a minimum value function, the value is... and The value is limited to 1.

[0055] According to one embodiment of the present invention, after obtaining all the data in the adjacency matrix, the adjacency matrix can be determined, thereby obtaining the association relationship between each node. Combined with the node input vector of each node, the graph structure corresponding to the water-soluble fertilizer packaging bag can be obtained.

[0056] In this way, the cosine similarity of the detection state information of nodes can be used to represent the similarity of the block status of the corresponding detection position. Based on the similarity of the block status and the distance between detection positions, the similarity of the block status of the detection position and the environmental influence can be comprehensively represented, thereby obtaining the data of the adjacency matrix and improving the accuracy of the association between nodes.

[0057] According to one embodiment of the present invention, the clumping status of water-soluble fertilizer is obtained by processing the graph structure through a second graph model, including: determining the degree matrix based on the adjacency matrix; processing the adjacency matrix, the degree matrix, and the node input vector of each node through the second graph model to obtain the node output vector of each node; concatenating the node output vectors of each node to obtain the water-soluble fertilizer state vector; and inputting the water-soluble fertilizer state vector into the fifth fully connected layer and the second activation layer for processing to obtain the clumping status of the water-soluble fertilizer, wherein the clumping status is the probability of clumping of water-soluble fertilizer in the water-soluble fertilizer packaging bag.

[0058] According to one embodiment of the present invention, the degree matrix has only non-zero elements on its diagonal, while all other elements are zero. Furthermore, the i-th data point on the diagonal of the degree matrix is ​​equal to the sum of all data points in the i-th row of the adjacency matrix. The second graph model is a graph convolutional neural network model, which processes the adjacency matrix, the degree matrix, and the node input vector of each node to obtain the node output vector of each node. That is, through the adjacency matrix and the degree matrix, the node input vector of each node is fused with the node input vectors of other nodes with related relationships to obtain the node output vector. This allows the node output vector to not only represent the block status at the detection location corresponding to the node, but also the correlation between the block status at other detection locations and the block status at that detection location, thereby describing the block status of a region near the detection location.

[0059] According to one embodiment of the present invention, the node output vectors of each node can be concatenated to obtain a water-soluble fertilizer state vector, representing the overall clumping status of the entire water-soluble fertilizer packaging bag. Further, the water-soluble fertilizer state vector can be input into a fifth fully connected layer to obtain a one-dimensional vector, which is then input into a second activation layer for processing. For example, the second activation layer can process the values ​​within this one-dimensional vector using a sigmoid activation function to obtain the clumping status of the water-soluble fertilizer. The clumping status of the water-soluble fertilizer is a probabilistic result. If the clumping status of the water-soluble fertilizer is higher than a preset probability threshold (e.g., 0.5), it indicates that the overall clumping status within the water-soluble fertilizer packaging bag is relatively severe. For water-soluble fertilizers with severe clumping, it can be determined that its quality inspection is unqualified, its chemical components have become ineffective, and it can be discarded or reprocessed.

[0060] According to an embodiment of the present invention, the first graph model and the second graph model, as well as each fully connected layer, can be trained before use. The method further includes: obtaining the sample subgraph structure of each detection position of the sample water-soluble fertilizer packaging bag; processing the sample subgraph structure through the first graph model to obtain the sample detection state information of each detection position; inputting the sample detection state information of each detection position into the sixth fully connected layer and the third activation layer for processing to obtain the clumping probability of each detection position; obtaining the detection position clumping state loss function according to the annotation information of each detection position and the clumping probability; obtaining the sample graph structure according to the sample detection state information of each detection position and the distance between each detection position; processing the sample graph structure through the second graph model to obtain the sample clumping status; obtaining the clumping annotation of the sample water-soluble fertilizer packaging bag according to the annotation information of each detection position; obtaining the clumping status loss function according to the clumping probability, the sample clumping status, and the clumping annotation; training the first graph model and the second graph model according to the detection position clumping state loss function and the clumping status loss function to obtain the trained first graph model and the trained second graph model.

[0061] According to one embodiment of the present invention, each detection location on the sample water-soluble fertilizer packaging bag can be manually marked, that is, whether the detection location is clumped.

[0062] According to one embodiment of the present invention, the sample subgraph structure can be obtained in a manner similar to that used to obtain the subgraph structure, which will not be described in detail here. The sample subgraph structure is then processed using a first graph model to obtain sample detection state information for each detection location. The processing method is similar to that used to obtain the detection state information described above, and will not be described in detail here.

[0063] According to one embodiment of the present invention, auxiliary training can be performed through a sixth fully connected layer and a third activation layer. For example, the sixth fully connected layer can process the sample detection state information to obtain a one-dimensional vector, which is then input into the third activation layer and processed by the sigmoid function to obtain the block probability of the detection position. This block probability can then be compared with the manually labeled information indicating whether the detection position is blocked (i.e., the labeling information) to obtain the block state loss function for the detection position. For example, the labeling information can also be represented as a probability; if the detection position is blocked, the probability is 1; otherwise, the probability is 0. The cross-entropy loss function can then be constructed using the labeling information and the block probability to obtain the block state loss function for the detection position. Similarly, the block state loss function for each detection position can be obtained.

[0064] According to one embodiment of the present invention, the method for obtaining the sample graph structure and the method for processing the sample graph structure by the second graph model are similar to the methods for obtaining the graph structure and processing the graph structure by the second graph model described above, and will not be repeated here. Furthermore, the probabilistic form of sample block status can be obtained.

[0065] According to one embodiment of the present invention, the clumping label of the sample water-soluble fertilizer packaging bag can be obtained based on the labeling information of each detection location. For example, if the number of detection locations with clumping labeling information exceeds half of the total number of detection locations, the clumping label is set to 1; otherwise, the clumping label is set to 0.

[0066] According to an embodiment of the present invention, obtaining a block status loss function based on the block probability, the sample block status, and the block label includes: obtaining the block status loss function according to formula (2). ,

[0067] (2)

[0068] in, This is the probability of sample water-soluble fertilizer clumping determined based on clumping labeling. For sample clumping status, Let be the probability of a block at the i-th detection position. Let N be the probability of a block forming at the i-th detection position, determined based on the annotation information of the i-th detection position. Let N be the number of detection positions, i ≤ N, and both i and N are positive integers. Let max be the function that takes the maximum value.

[0069] According to one embodiment of the present invention, in formula (2), if the block is labeled as 1, then If it is 1, otherwise, It is 0. If If the error is small, it means that the prediction of the sample clumping status obtained by the model is correct. In this case, the cross-entropy loss function composed of the probability of sample water-soluble fertilizer clumping determined by clumping label and the sample clumping status can be used as the clumping status loss function.

[0070] According to one embodiment of the present invention, if This indicates that the model's prediction of sample block conditions has a large error, and the prediction results for blocks are incorrect. In this case, it can be improved by using a value less than 1. As the denominator, the value of the cross-entropy loss function, which consists of the probability of sample water-soluble fertilizer clumping determined by clumping labeling and the sample clumping status, is amplified, thereby improving training strength and efficiency. To minimize the maximum error between the labeled information and the block probability at each detection location, the block status loss function is reduced during training. This not only reduces the value of the aforementioned cross-entropy loss function but also... The value of decreases, thereby reducing the error between the annotation information and the block probability at each detection location, and further improving the accuracy of the block probability at each detection location.

[0071] According to one embodiment of the present invention, after obtaining the clumping state loss function and the clumping condition loss function at the detection location, the first graph model, the second graph model, and each fully connected layer can be trained through backpropagation. This continuously reduces the clumping state loss function and the clumping condition loss function at the detection location, making the clumping probability at each detection location closer to the labeled information, and making the probability of clumping of the sample water-soluble fertilizer closer to the clumping label, thereby improving the accuracy of the first graph model and the second graph model. After the accuracy of the first graph model and the second graph model meets the requirements (e.g., the accuracy rate of judging whether the sample water-soluble fertilizer packaging bag in the validation set is clumped is higher than a preset accuracy threshold), the trained first graph model and the second graph model are obtained.

[0072] In this way, a sixth fully connected layer and a third activation layer can be introduced during training to assist in training, thereby obtaining the block state loss function for each detection location and improving the training efficiency of the first graph model. Furthermore, when obtaining the block state loss function, the training efficiency can be enhanced by using the maximum error between the annotation information and the block probability at each detection location. This not only improves the accuracy of the block state of the samples but also further improves the accuracy of the block probability at each detection location, thus improving overall training efficiency and model accuracy.

[0073] The water-soluble fertilizer anti-caking detection method according to an embodiment of the present invention can comprehensively analyze various data collected during the pressing of the water-soluble fertilizer packaging bag by the contact head, telescopic component, and microphone of the detection platform to determine the clumping status of the water-soluble fertilizer inside the packaging bag. This method does not damage the water-soluble fertilizer packaging bag during the detection process, enabling non-destructive testing, reducing detection costs, and minimizing waste of water-soluble fertilizer. Furthermore, it can collect pressing resistance data and acoustic wave information during the pressing process, thereby improving the accuracy of the analysis by comprehensively analyzing the clumping status using both types of data. When acquiring detection state information, an attention mechanism can be used to establish the correlation between the high-dimensional vector representations of the acoustic wave's time-domain characteristics, frequency-domain characteristics, and resistance characteristics. This allows for the fusion of various high-dimensional vector representations using a first graph model to obtain the output vectors of each subgraph node. The spliced ​​vectors can also represent multiple aspects of the detection location's characteristics. A fourth fully connected layer further strengthens the influence between different high-dimensional vector representations, obtaining detection state information that comprehensively represents the state of the detection location, thus improving the ability of the detection state information to describe the state of the detection location. When determining the adjacency matrix, the cosine similarity of the detection state information of nodes can be used to represent the similarity of the clustering status of the detection positions corresponding to the nodes. Furthermore, based on the similarity of the clustering status and the distance between detection positions, the similarity of the clustering status and the environmental influences at each detection position can be comprehensively represented, thereby obtaining various data in the adjacency matrix and improving the accuracy of the association relationships between nodes. During model training, a sixth fully connected layer and a third activation layer can be introduced to assist training, thereby obtaining the clustering status loss function for each detection position to enhance the training strength of the first graph model. Moreover, when obtaining the clustering status loss function, the maximum error between the annotation information and the clustering probability of each detection position can be used to enhance the training strength, which not only improves the accuracy of the sample clustering status but also further improves the accuracy of the clustering probability at each detection position, thus improving the overall training efficiency and model accuracy.

[0074] Figure 3An exemplary block diagram of a water-soluble fertilizer anti-caking detection system according to an embodiment of the present invention is shown. The system includes: a detection position module for placing a water-soluble fertilizer packaging bag flat on a detection table and setting multiple detection positions on the surface of the water-soluble fertilizer packaging bag; a pressing module for aligning a contact head with the detection position and pressing the detection position according to a first preset pressure, wherein the contact head is a cylindrical cavity, the lower surface of the contact head is an annular contact surface, a telescopic component is provided at the axial position within the cylindrical cavity of the contact head, the lower contact surface of the telescopic component is a circular contact surface, the centroid of the circular contact surface coincides with the centroid of the annular contact surface, and the circular contact surface can extend out of the annular contact surface to press downwards on the detection position of the water-soluble fertilizer packaging bag; and a resistance data module for controlling the telescopic component to press downwards on the detection position at a first preset speed. The system includes: a resistance data detection module for the telescopic component at multiple moments during the pressing process; an acoustic information module for collecting acoustic information via a microphone in a cylindrical cavity during pressing; a retraction module for stopping pressing and retracting the lower contact surface of the telescopic component above the annular contact surface when the lower contact surface extends beyond the annular contact surface by a first preset distance; a movement module for stopping the contact head from pressing the detection position and moving it to the next detection position; an iteration module for iteratively executing the processing of acquiring resistance data and acoustic information until resistance data and acoustic information corresponding to all detection positions are acquired; and a clumping status module for determining the clumping status of the water-soluble fertilizer based on the resistance data and acoustic information corresponding to multiple detection positions.

[0075] This invention can be a method, apparatus, system, and / or computer program product. The computer program product may include a computer-readable storage medium having computer-readable program instructions loaded thereon for performing various aspects of the invention.

[0076] Those skilled in the art should understand that the embodiments of the present invention described above and shown in the accompanying drawings are merely examples and do not limit the present invention. The objectives of the present invention have been fully and effectively achieved. The functions and structural principles of the present invention have been shown and explained in the embodiments, and any modifications or variations of the embodiments of the present invention may be made without departing from the stated principles.

Claims

1. A method for detecting the anti-caking properties of water-soluble fertilizer, characterized in that, include: Place the water-soluble fertilizer packaging bag flat on the testing table and set multiple testing positions on the surface of the water-soluble fertilizer packaging bag; The contact head is aligned with the detection position and pressed against the detection position according to a first preset pressure. The contact head is a cylindrical cavity, and its lower surface is an annular contact surface. A telescopic component is located at the axial position within the cylindrical cavity of the contact head. The lower contact surface of the telescopic component is a circular contact surface, the centroid of which coincides with the centroid of the annular contact surface. The circular contact surface can extend beyond the annular contact surface to press downwards onto the detection position of the water-soluble fertilizer packaging bag. The telescopic component is controlled to press downwards onto the detection position at a first preset speed, and the telescopic movement is detected at multiple moments during the pressing process. The system collects resistance data from the component; during pressing, it acquires sound wave information through a microphone in a cylindrical cavity; when the lower contact surface of the telescopic component extends beyond the annular contact surface by a first preset distance, it stops pressing and controls the lower contact surface of the telescopic component to retract above the annular contact surface; the contact head stops pressing the detection position and moves to the next detection position; iteratively executes the processing of acquiring resistance data and sound wave information until resistance data and sound wave information corresponding to all detection positions are acquired; based on the resistance data and sound wave information corresponding to multiple detection positions, it determines the agglomeration status of the water-soluble fertilizer.

2. The method for detecting anti-caking of water-soluble fertilizer according to claim 1, characterized in that, Based on resistance data and acoustic information corresponding to multiple detection locations, the agglomeration status of water-soluble fertilizer is determined, including: sampling acoustic information to obtain acoustic sampling data; obtaining various time-domain feature information based on the acoustic sampling data; forming a time-domain feature vector from the various time-domain feature information; performing a fast Fourier transform on the acoustic sampling data to obtain acoustic frequency-domain data; obtaining various frequency-domain feature information based on the acoustic frequency-domain data; forming a frequency-domain feature vector from the various frequency-domain feature information; forming a resistance feature vector from the resistance data; obtaining a subgraph structure corresponding to the detection location based on the time-domain feature vector, frequency-domain feature vector, and resistance feature vector; processing the subgraph structure using a first graph model to obtain detection status information for each detection location; determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the distance between the detection status information and the detection location; and processing the graph structure using a second graph model to obtain the agglomeration status of the water-soluble fertilizer.

3. The method for detecting anti-caking of water-soluble fertilizer according to claim 2, characterized in that, Based on the time-domain feature vector, frequency-domain feature vector, and drag feature vector, a subgraph structure corresponding to the detection position is obtained, including: processing the time-domain feature vector through a first fully connected layer to obtain a first input vector for a first subgraph node in the subgraph structure; processing the frequency-domain feature vector through a second fully connected layer to obtain a second input vector for a second subgraph node in the subgraph structure; processing the drag feature vector through a third fully connected layer to obtain a third input vector for a third subgraph node in the subgraph structure; and processing the first and second input vectors according to the attention mechanism of the first graph model to obtain a first subgraph node. The connection weights of the nodes in the first and second subgraphs are obtained by processing the first and third input vectors according to the attention mechanism of the first graph model, and by processing the second and third input vectors according to the attention mechanism of the first graph model, and obtaining the connection weights of the nodes in the second and third subgraphs. Based on the first input vector, the second input vector, the third input vector, the connection weights of the first and second subgraph nodes, the connection weights of the first and third subgraph nodes, and the connection weights of the second and third subgraph nodes, the subgraph structure is obtained.

4. The method for detecting anti-caking of water-soluble fertilizer according to claim 3, characterized in that, The first graph model is used to process the subgraph structure to obtain the detection state information of each detection location. This includes: processing the subgraph structure using the first graph model to obtain the first output vector of the first subgraph node, the second output vector of the second subgraph node, and the third output vector of the third subgraph node; concatenating the first, second, and third output vectors to obtain the detection location feature vector; and inputting the detection location feature vector into the fourth fully connected layer and the first activation layer for processing to obtain the detection state information.

5. The method for detecting anti-caking of water-soluble fertilizer according to claim 2, characterized in that, The graph structure corresponding to the water-soluble fertilizer packaging bag is determined based on the distance between the detection status information and the detection location, including: mapping multiple detection locations to multiple nodes in the graph structure; using the detection status information of each detection location as the node input vector of each node; determining the adjacency matrix based on the distance between the detection status information and the detection location; and determining the graph structure corresponding to the water-soluble fertilizer packaging bag based on the adjacency matrix and the node input vector of each node.

6. The method for detecting anti-caking of water-soluble fertilizer according to claim 5, characterized in that, Based on the detection status information and the distance between the detection positions, the adjacency matrix is ​​determined, including: according to the formula Determine the data in the i-th row and j-th column of the adjacency matrix. and the data in the j-th row and i-th column ,in, This represents the detection status information of the i-th node. This refers to the detection status information of the j-th node. for and Cosine similarity between them Let be the distance between the i-th detection position and the j-th detection position. This represents the average distance between multiple detection locations, where AND indicates a logical AND operation. Indicates except as well as In all other cases, min is the function that takes the minimum value.

7. The method for detecting anti-caking of water-soluble fertilizer according to claim 5, characterized in that, The second graph model is used to process the graph structure to obtain the clumping status of the water-soluble fertilizer, including: determining the degree matrix based on the adjacency matrix; processing the adjacency matrix, degree matrix, and node input vector of each node using the second graph model to obtain the node output vector of each node; concatenating the node output vectors of each node to obtain the water-soluble fertilizer state vector; and inputting the water-soluble fertilizer state vector into the fifth fully connected layer and the second activation layer for processing to obtain the clumping status of the water-soluble fertilizer, wherein the clumping status is the probability of clumping of the water-soluble fertilizer in the water-soluble fertilizer packaging bag.

8. The method for detecting anti-caking of water-soluble fertilizer according to claim 2, characterized in that, The method further includes: acquiring the sample subgraph structure of each detection location of the sample water-soluble fertilizer packaging bag; processing the sample subgraph structure using a first graph model to acquire sample detection state information for each detection location; inputting the sample detection state information for each detection location into a sixth fully connected layer and a third activation layer for processing to obtain the clumping probability for each detection location; obtaining a clumping state loss function for each detection location based on the annotation information of each detection location and the clumping probability; acquiring the sample graph structure based on the sample detection state information of each detection location and the distance between each detection location; processing the sample graph structure using a second graph model to acquire the sample clumping status; acquiring the clumping annotation of the sample water-soluble fertilizer packaging bag based on the annotation information of each detection location; acquiring a clumping status loss function based on the clumping probability, the sample clumping status, and the clumping annotation; and training the first graph model and the second graph model based on the detection location clumping status loss function and the clumping status loss function to obtain the trained first graph model and the trained second graph model.

9. The method for detecting anti-caking of water-soluble fertilizer according to claim 8, characterized in that, Based on the block probability, the sample block status, and the block label, a block status loss function is obtained, including: according to the formula Obtain the loss function for the clumping status. ,in, This is the probability of sample water-soluble fertilizer clumping determined based on clumping labeling. For sample clumping status, Let be the probability of a block at the i-th detection position. Let N be the probability of a block forming at the i-th detection position, determined based on the annotation information of the i-th detection position. Let N be the number of detection positions, i ≤ N, and both i and N are positive integers. Let max be the function that takes the maximum value.

10. A water-soluble fertilizer anti-caking detection system, used to perform the method as described in any one of claims 1-9, characterized in that, include: The detection position module is used to place the water-soluble fertilizer packaging bag flat on the detection table and set multiple detection positions on the surface of the water-soluble fertilizer packaging bag; A pressing module is used to align the contact head with the detection position and press the detection position according to a first preset pressure. The contact head is a cylindrical cavity, and its lower surface is an annular contact surface. A telescopic component is provided at the axial position within the cylindrical cavity of the contact head. The lower contact surface of the telescopic component is a circular contact surface, the centroid of which coincides with the centroid of the annular contact surface. The circular contact surface can extend beyond the annular contact surface to press downwards onto the detection position of the water-soluble fertilizer packaging bag. A resistance data module is used to control the telescopic component to press downwards onto the detection position at a first preset speed, and during the pressing process... The system detects resistance data of the telescopic component at multiple moments; an acoustic information module collects acoustic information through a microphone in a cylindrical cavity during pressing; a retraction module stops pressing and controls the lower contact surface of the telescopic component to retract above the annular contact surface when the distance the lower contact surface of the telescopic component extends beyond the annular contact surface reaches a first preset distance; a moving module stops the contact head from pressing the detection position and moves to the next detection position; and an iteration module iteratively executes the processing of acquiring resistance data and acoustic information until resistance data and acoustic information corresponding to all detection positions are acquired. Agglomeration status is used to determine the agglomeration status of water-soluble fertilizer based on resistance data and acoustic information corresponding to multiple detection locations.

Images