Intelligent detection method and system for hardened area of grain pile

Abstract

The invention relates to the technical field of intelligent warehousing, in particular to an intelligent detection method and system for a hardened area of a grain pile, and the method comprises the following steps: constructing a data collection device of the grain pile, and obtaining the reflection intensity, three-dimensional point cloud data and curvature of the grain pile; carrying out feature fusion on the reflection intensity and the curvature to obtain a hardened area sensitive feature of the grain pile, and constructing a hardened area detection model of the grain pile; establishing an abnormal surface identification network architecture of the grain pile to obtain abnormal surface features of the grain pile; according to the hardened area detection model, performing intelligent detection on the grain pile in combination with the abnormal surface features to obtain a preliminary detection hardened area; and carrying out boundary positioning on the preliminarily detected hardened area to obtain a hardened area boundary so as to obtain the hardened area of the grain pile. According to the method, the hardened area of the grain pile is effectively obtained, a quantitative decision basis is provided for grain storage safety management, and the method has remarkable economic value and practical significance.

Description

Technical Field

[0001] The present invention relates to the field of intelligent warehousing technology, and in particular to an intelligent detection method and system for compacted grain pile areas. Background Art

[0002] In grain storage, grain compaction is a core issue affecting grain storage safety. Grain compaction occurs when grains adhere to each other, forming lumps due to factors such as moisture and heat migration, microbial activity, or mechanical pressure. This condition not only hinders ventilation and heat dissipation within the grain pile but can also lead to secondary risks such as mold and insect infestation. Traditional detection methods rely primarily on manual or simple sensor monitoring, which has limitations such as low efficiency, limited coverage, and high subjectivity, making it difficult to meet the needs of intelligent management in large grain warehouses.

[0003] In recent years, with the development of three-dimensional vision technology and machine learning, grain pile status analysis based on point cloud data has become a research hotspot. Existing technologies obtain the three-dimensional morphology of the grain pile surface and use geometric features such as curvature and normal direction to identify abnormal areas. However, relying solely on geometric features is easily affected by normal deformations such as natural settlement and surface undulations of the grain pile, resulting in a high error rate. At the same time, the surface reflection intensity of the grain pile is an important parameter reflecting material properties, and its correlation with the compacted area has not been fully explored.

[0004] In view of this, the present invention constructs a data acquisition device integrating multispectral lidar to synchronously obtain high-precision reflection intensity and three-dimensional point cloud data, extracts curvature features based on local surface fitting, adopts a dynamic weighting strategy to obtain compaction-sensitive features, and improves the accuracy of abnormal area recognition by combining the detection model with physical constraint correction. A multi-layer grouping strategy and a dual-branch attention mechanism are designed to effectively capture the material variation and morphological anomalies of the compacted area, and finally realize the precise positioning of the boundary of the compacted area of ​​the grain pile, providing technical support for the intelligent evaluation and decision-making of the compacted state of the grain pile. Summary of the Invention

[0005] In view of the defects in the prior art, the present invention provides an intelligent detection method and system for compacted grain pile areas.

[0006] In order to achieve the above-mentioned purpose, in a first aspect, the present invention provides an intelligent detection method for compacted areas of grain piles, the method comprising the following steps: constructing a data acquisition device for the grain pile, and obtaining the reflection intensity, three-dimensional point cloud data and curvature of the grain pile according to the data acquisition device; performing feature fusion of the reflection intensity and the curvature to obtain sensitive features of the compacted area of ​​the grain pile, and constructing a compacted area detection model of the grain pile based on the sensitive features of the compacted area; establishing an abnormal surface recognition network architecture for the grain pile, and obtaining abnormal surface features of the grain pile based on the abnormal surface recognition network architecture; performing intelligent detection on the grain pile according to the compacted area detection model and in combination with the abnormal surface features to obtain a preliminary detection of the compacted area; performing boundary positioning on the preliminary detection of the compacted area to obtain the boundary of the compacted area, so as to obtain the compacted area of ​​the grain pile. The present invention provides multi-dimensional data support for compaction identification by synchronously acquiring the physical properties and spatial structure information of grain piles; establishes compaction-sensitive features through feature fusion technology, effectively breaking through the limitations of single parameter detection, and significantly improving the accuracy and generalization ability of abnormal area identification; introduces an abnormal surface recognition network architecture, and forms a dual verification mechanism with the compaction detection model, effectively reducing the missed detection rate and false alarm rate; uses the boundary positioning algorithm to accurately outline the contour of the compaction area, realizing a hierarchical detection strategy from coarse to fine; the final compaction area provides a quantitative decision-making basis for grain storage safety management, and has significant economic value and practical significance.

[0007] Optionally, the data acquisition device for constructing the grain pile obtains the reflection intensity, three-dimensional point cloud data and curvature of the grain pile according to the data acquisition device, including: constructing the data acquisition device based on a multispectral laser radar, and using the multispectral laser radar to obtain the reflection intensity; obtaining the surface three-dimensional coordinates of the grain pile according to the multispectral laser radar, and obtaining the three-dimensional point cloud data of the grain pile based on the surface three-dimensional coordinates; performing local neighborhood fitting on the three-dimensional point cloud data to obtain a quadratic surface of the grain pile, and obtaining the curvature based on the quadratic surface. The present invention constructs a grain pile data acquisition device through multispectral lidar, realizes the efficient acquisition and quantitative analysis of multi-dimensional physical characteristics, and simultaneously captures the difference in reflection intensity on the surface of the grain pile, providing key parameters for identifying density variation areas; the three-dimensional point cloud data reconstructs the surface morphology of the grain pile through high-precision coordinates, laying the foundation for spatial structure analysis; the curvature features extracted by local surface fitting technology can quantitatively characterize the tiny deformation of the grain pile surface, effectively amplifying the geometric differences between the compacted area and the normal area; the fusion of the three constructs a full-scale detection system from microscopic reflection characteristics to macroscopic morphological characteristics, significantly improving the sensitivity and positioning accuracy of identifying compacted areas, and providing high-fidelity data input for subsequent intelligent detection models.

[0008] Optionally, the feature fusion of the reflection intensity and the curvature to obtain the sensitive features of the compacted area of ​​the grain pile includes: preprocessing the reflection intensity and the curvature to obtain the standardized parameters of the grain pile, and constructing a spatial index of the standardized parameters; spatially registering the reflection intensity and the curvature based on the spatial index to obtain the registration parameters of the grain pile; based on the registration parameters, performing feature analysis on the reflection intensity and the curvature to obtain the reflection intensity features and curvature features of the grain pile; and weighted feature fusion of the reflection intensity features and the curvature features to obtain the sensitive features of the compacted area. The present invention significantly improves the characterization capability of grain pile compaction features through a multi-parameter fusion mechanism; standardized preprocessing eliminates the dimensional difference between reflection intensity and curvature data, and spatial indexing realizes spatial consistency mapping of multi-source features, ensuring the geometric alignment accuracy of subsequent analysis; spatial registration effectively solves the spatial offset problem between multimodal data of lidar, and provides accurate geometric correspondence for feature fusion; the feature weighted fusion strategy based on registration parameters retains the sensitivity of reflection intensity to density changes, and incorporates the characterization capability of curvature to morphological mutations, forming complementary and enhanced composite features, which significantly improves the anti-interference capability and feature discrimination of compaction area detection.

[0009] Optionally, the weighted feature fusion of the reflection intensity feature and the curvature feature to obtain the sensitive feature of the compacted area includes: obtaining a first dynamic weight coefficient of the reflection intensity feature and a second dynamic weight coefficient of the curvature feature; based on the standardized parameters, combining the first dynamic weight coefficient and the second dynamic weight coefficient to perform weighted feature fusion on the reflection intensity feature and the curvature feature to obtain the sensitive feature of the compacted area. The present invention significantly improves the detection accuracy of compacted areas through a multimodal feature collaborative enhancement mechanism; the dynamic weight coefficient can be adaptively adjusted according to the grain variety or storage stage, strengthening the curvature-dominated morphological analysis in the early stage of compaction, and highlighting the strength-dominated density recognition in the hardening stage, forming a feature-complementary detection mechanism, and significantly improving the feature characterization capability under complex working conditions.

[0010] Optionally, the compaction area detection model of the grain pile is constructed based on the sensitive features of the compaction area, including: mapping the sensitive features of the compaction area to a two-dimensional grid space, and obtaining feature statistics in each grid of the two-dimensional grid space, and obtaining the compaction area feature space of the grain pile based on the feature statistics; combining the compaction area feature space, constructing a hybrid classification decision framework to obtain the compaction probability distribution of the sensitive features of the compaction area; establishing physical constraints of the grain pile, and correcting the compaction probability distribution based on the physical constraints to obtain a corrected compaction probability distribution; constructing an adaptive detection threshold, and detecting the corrected compaction probability distribution according to the adaptive detection threshold to construct the compaction area detection model. The present invention significantly improves the accuracy and robustness of grain pile compaction detection through a multi-dimensional modeling strategy; two-dimensional raster mapping transforms discrete features into structured spatial representations, and feature statistics extraction effectively quantifies the degree of abnormality in local areas, constructing a feature analysis framework with spatial correlation; the hybrid classification decision framework realizes soft division of compaction areas through probability distribution modeling, which is more noise-resistant than the hard threshold method; the physical constraint correction mechanism introduces the mechanical properties and stacking laws of grain storage, effectively filtering out false detection results that do not conform to actual working conditions; the adaptive threshold design can dynamically adjust the discrimination criteria according to the type of grain pile to ensure the sensitivity of detection.

[0011] Optionally, the construction of the adaptive detection threshold includes: obtaining the ambient temperature and ambient humidity of the grain pile, and obtaining a first compensation coefficient for the ambient temperature and a second compensation coefficient for the ambient humidity; based on the ambient temperature, the ambient humidity, the first compensation coefficient and the second compensation coefficient, the adaptive detection threshold is obtained in combination with a baseline detection threshold. The present invention significantly improves the adaptability of grain pile compaction detection through a dynamic environmental compensation mechanism; the adaptive detection threshold is based on the baseline threshold, quantifies the environmental offset through the temperature compensation coefficient and the humidity compensation coefficient, and corrects the detection threshold in real time, effectively solving the problem of feature drift caused by temperature and humidity changes in grain pile detection, so that the detection model has environmental adaptability, can still maintain detection accuracy in scenarios with seasonal temperature and humidity fluctuations or regional climate differences, and significantly enhances the environmental robustness and engineering practicality of the detection system.

[0012] Optionally, the abnormal surface recognition network architecture of the grain pile is established, and the abnormal surface features of the grain pile are obtained based on the abnormal surface recognition network architecture, including: constructing a multi-layer grouping strategy, the multi-layer grouping strategy includes a shallow layer and a deep layer, obtaining the texture features of the grain pile in the shallow layer, and obtaining the morphological features of the grain pile in the deep layer; establishing a dual-branch attention mechanism, the dual-branch attention mechanism includes channel attention and spatial attention, obtaining the abnormal area of ​​the reflection intensity according to the channel attention, and obtaining the sudden change area of ​​the curvature according to the spatial attention; constructing the abnormal surface recognition network architecture in combination with the multi-layer grouping strategy and the dual-branch attention mechanism; performing feature recognition on the surface of the grain pile based on the abnormal surface recognition network architecture to obtain the abnormal surface features. The present invention significantly improves the ability to identify abnormal surfaces of grain piles through multi-level feature extraction and attention enhancement mechanism; the multi-layer grouping strategy constructs a complete feature spectrum from micro to macro through shallow texture feature encoding and deep morphological feature analysis, effectively capturing the surface roughness changes and structural heterogeneity unique to compacted areas; the dual-branch attention mechanism strengthens the characteristic response of areas with abnormal reflection intensity through channel attention, and at the same time uses spatial attention to accurately locate areas with curvature mutations, thereby achieving active enhancement of abnormal features and interference suppression; the synergistic effect of the two significantly improves the sensitivity and positioning accuracy of abnormal surface detection.

[0013] Optionally, the grain pile is intelligently detected based on the compaction area detection model and combined with the abnormal surface features to obtain a preliminary detection of the compaction area, including: feature splicing of the sensitive features of the compaction area and the abnormal surface features to construct a joint feature tensor of the grain pile; constructing the compaction spatiotemporal constraints of the grain pile, and performing time-series tracking of the joint feature tensor based on the compaction spatiotemporal constraints to obtain the compaction evolution characteristics of the grain pile; and intelligently detecting the compaction evolution characteristics based on the compaction area detection model to obtain the preliminary detection of the compaction area. The present invention realizes accurate identification of compacted areas of grain piles through multimodal feature fusion and spatiotemporal evolution analysis; the feature splicing strategy fuses compaction-sensitive features with abnormal surface features in a high-dimensional manner to form a composite feature representation including density variation, morphological mutation and texture anomaly, which significantly improves the feature discrimination of abnormal areas; spatiotemporal constraint modeling introduces time dimension analysis, reveals the dynamic evolution law of compacted areas through time series tracking, and effectively suppresses the random error of single-frame detection; the intelligent detection model combined with spatiotemporal evolution characteristics retains the refined analysis capability of spatial heterogeneity and has the trend prediction capability of temporal continuity. The final output preliminary detection results are both real-time and forward-looking, providing high-confidence candidate areas for subsequent boundary positioning.

[0014] Optionally, the boundary positioning of the preliminary detected compacted area is performed to obtain the boundary of the compacted area to obtain the compacted area of ​​the grain pile, including: selecting candidate seeds based on the preliminary detected compacted area, clustering the candidate seeds to construct an effective seed set; constructing a multimodal feature constraint, and constructing a growth criterion for the effective seed set according to the multimodal feature constraint; obtaining the growth result of the effective seed set according to the growth criterion, and obtaining the boundary of the compacted area according to the growth result to obtain the compacted area. The present invention achieves precise positioning of the compacted boundary of the grain pile through a seed-driven regional growth strategy; the candidate seed selection and clustering mechanism effectively screens high-confidence initial points to avoid noise interference; the multimodal feature constraint fuses the reflection intensity, curvature and spatial structure information to construct a growth criterion that conforms to the physical characteristics of compaction, ensuring that the boundary expansion direction is consistent with the actual morphological changes; the regional growth process is combined with dynamic threshold adjustment to achieve progressive boundary optimization from the inside to the outside, generate a closed and continuous contour of the compacted area, and significantly improve the geometric integrity and spatial positioning accuracy of the detection results.

[0015] In a second aspect, the present invention provides an intelligent detection system for compacted grain pile areas. The system implements the intelligent detection method for compacted grain pile areas provided by the present invention. The system includes an input device, an output device, a processor, and a memory. Its benefits are: the hardware facilities integrated by the present invention have excellent performance, the input device, output device, processor, and memory are interconnected, information transmission between each component is smooth, and an efficient information processing system is constructed through the interaction of multiple hardware facilities. The present invention achieves efficient operation of the entire process through hardware collaborative optimization, significantly improves data processing efficiency, and combines modular design to facilitate function expansion and algorithm iteration, providing reliable technical support for intelligent management of grain storage. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 This is a flow chart of an intelligent detection method for grain pile compaction areas according to an embodiment of the present invention; Figure 2 This is a framework diagram of an intelligent detection system for compacted grain pile areas according to an embodiment of the present invention. DETAILED DESCRIPTION

[0017] Specific embodiments of the present invention will be described in detail below. It should be noted that the embodiments described herein are for illustrative purposes only and are not intended to limit the present invention. In the following description, numerous specific details are set forth to provide a thorough understanding of the present invention. However, it will be apparent to one of ordinary skill in the art that these specific details are not necessarily required to practice the present invention. In other instances, well-known circuits, software, or methods are not specifically described to avoid obscuring the present invention.

[0018] Throughout this specification, references to "one embodiment," "an embodiment," "an example," or "an example" mean that a particular feature, structure, or characteristic described in connection with the embodiment or example is included in at least one embodiment of the present invention. Therefore, appearances of the phrases "in one embodiment," "in an embodiment," "an example," or "an example" in various places throughout this specification are not necessarily all referring to the same embodiment or example. Furthermore, the particular features, structures, or characteristics may be combined in any suitable combinations and / or subcombinations in one or more embodiments or examples. Furthermore, those of ordinary skill in the art will appreciate that the figures provided herein are for illustrative purposes only and are not necessarily drawn to scale.

[0019] See Figure 1 One embodiment of the present invention provides an intelligent method for detecting compacted grain pile areas, the method comprising the following steps: S1. Construct a data acquisition device for a grain pile, and obtain the reflection intensity, three-dimensional point cloud data, and curvature of the grain pile according to the data acquisition device.

[0020] In this embodiment, a data acquisition device is constructed based on a multispectral lidar, and the multispectral lidar is used to obtain the reflection intensity; the surface three-dimensional coordinates of the grain pile are obtained according to the multispectral lidar, and the three-dimensional point cloud data of the grain pile is obtained based on the surface three-dimensional coordinates; the three-dimensional point cloud data is localized and fitted to obtain the quadratic surface of the grain pile, and the curvature is obtained based on the quadratic surface.

[0021] Specifically, in response to the need for grain pile compaction detection to be sensitive to material properties, a multispectral lidar is used as the core sensing unit, and a targeted data acquisition device is designed; the multispectral lidar uses three-band laser sources of 1550nm (near infrared), 1064nm (visible light) and 850nm (short-wave infrared), which correspond to the characteristic absorption spectrum bands of different structures in the grain pile, respectively, and enhances the ability to identify compacted areas through the difference in multi-band reflection intensity; the lidar integrates a high-precision rotating pan-tilt head and a ranging module to achieve full coverage scanning of the entire grain pile.

[0022] During the reflection intensity acquisition phase, ambient light compensation is performed on the grain pile. The built-in photosensor is used to collect the ambient light intensity of the grain pile in real time. A dynamic calibration mechanism is then used to correct the original reflection intensity to satisfy the following relationship: in, is the corrected reflection intensity, is the original reflection intensity, is the distance value of the current point, To calibrate the reference distance, is the ambient light attenuation coefficient, is the ambient light intensity.

[0023] Furthermore, through laboratory calibration, a mapping relationship between the reflection intensity of each band and the density and moisture content of the grain pile is established to make the physical meaning of the reflection intensity interpretable. The data acquisition device has a built-in edge computing unit to execute the above-mentioned dynamic calibration mechanism in real time, output high-quality reflection intensity, and provide high signal-to-noise ratio input for subsequent multimodal fusion.

[0024] During the three-dimensional point cloud data acquisition stage, the three-dimensional coordinate data of the grain pile is synchronously acquired based on the multispectral lidar; an adaptive scanning strategy is adopted to perform an initial rapid scan of the entire warehouse grain pile, and then a fine scanning mode is started for areas with high reflection intensity to balance efficiency and accuracy.

[0025] Specifically, point cloud preprocessing includes: noise filtering, using a statistical outlier removal algorithm to remove outliers in the neighborhood; multi-view registration, using an iterative closest point algorithm to align point clouds from different scanning perspectives and control registration errors; and data interpolation, using radial basis functions to generate virtual point clouds for occluded areas (such as depressions in grain piles). The final output is 3D point cloud data, meeting the requirements of millimeter-level surface topography reconstruction. To verify data validity, the device is equipped with a positioning module, which regularly collects reference point coordinates for accuracy verification.

[0026] The above three-dimensional point cloud data satisfies the following relationship: in, The three-dimensional point cloud data of the grain pile, is the three-dimensional coordinate data of the grain pile, is the total amount of 3D point cloud data, The index variable of the 3D point cloud data.

[0027] During the curvature acquisition phase, the curvature of the grain pile surface is calculated based on the 3D point cloud data, and the geometric anomaly characteristics of the compacted area are quantified. A local neighborhood search mechanism is established. With each 3D point cloud data point as the center, 3D point cloud data within a radius of 10 cm are searched to construct a neighborhood point cloud set. It is necessary to ensure that the neighborhood point cloud set contains at least 20 3D point cloud data points. A quadratic surface fitting is performed on the neighborhood point cloud set to satisfy the following relationship: in, is the height of a point on the grain pile surface, is the horizontal coordinate of a point on the grain pile surface, is the ordinate of a point on the grain pile surface, is the coefficient of the quadratic term, is the cross-term coefficient, is the coefficient of the first-order term, is a constant term.

[0028] Specifically, the least squares method is used to determine the quadratic term coefficients, cross term coefficients, and linear term coefficients in the quadratic surface, and then the curvature is obtained, which satisfies the following relationship: in, is the curvature, is the coefficient of the quadratic term, is the cross-term coefficient, is the coefficient of the first-order term.

[0029] Furthermore, to suppress noise from quadratic surface fitting, two levels of filtering were performed: a goodness-of-fit test to remove neighborhoods with abnormal residuals, and curvature smoothing to filter the curvature. Physical consistency was verified through density inversion, and a multi-dimensional validation mechanism combining spectroscopy, geometry, and physics was employed to ensure that the curvature data truly reflected the compaction state of the grain pile.

[0030] S2. Perform feature fusion of the reflection intensity and the curvature to obtain sensitive features of the compacted area of ​​the grain pile, and construct a detection model for the compacted area of ​​the grain pile based on the sensitive features of the compacted area.

[0031] Among them, S2 specifically includes the following steps: S21. Perform feature fusion on the reflection intensity and the curvature to obtain sensitive features of the compacted area of ​​the grain pile.

[0032] In this embodiment, the reflection intensity and curvature are preprocessed to obtain standardized parameters of the grain pile, and a spatial index of the standardized parameters is constructed; based on the spatial index, the reflection intensity and curvature are spatially aligned to obtain the alignment parameters of the grain pile; based on the alignment parameters, the reflection intensity and curvature are feature analyzed to obtain the reflection intensity characteristics and curvature characteristics of the grain pile; the reflection intensity characteristics and curvature characteristics are weightedly fused to obtain sensitive features of the compacted area.

[0033] In the detection of compacted areas in grain piles, the reflection intensity and curvature data come from different working modes of the multispectral lidar. First, the reflection intensity needs to be radiometrically corrected to eliminate sensor noise and map the original data to a standardized range of 0-1. For the curvature data, adaptive threshold filtering is performed based on the geometric characteristics of the grain pile surface. The outlier removal threshold is automatically determined by statistically analyzing the neighborhood curvature distribution, retaining effective deformation information. Subsequently, a three-dimensional spatial index structure is constructed, dividing the grain pile surface into a cubic grid. Each grid cell stores the mean reflection intensity, curvature standard deviation, and data quality label of the corresponding spatial position. The index system uses an octree structure for spatial segmentation to achieve multi-scale data organization from global to local, providing a basic framework for subsequent spatial registration.

[0034] Furthermore, in order to achieve spatial consistency between reflection intensity and curvature data, an accurate geometric correspondence needs to be established. The common view area between the two is extracted in the spatial index, the stable feature points in the two data are identified, and the initial spatial transformation matrix is ​​constructed. In view of the deformation differences caused by the dynamic settlement of the grain pile, spatial registration is performed to obtain the registration parameters: in each iteration, a significant area with a curvature gradient greater than the threshold is selected as the control point, and the transformation parameters are optimized by minimizing the spatial distribution difference between the reflection intensity and the curvature; the registration accuracy is verified by a cross-validation strategy, the grain pile surface is divided into verification blocks, and the data similarity index of the corresponding verification blocks after registration is calculated to achieve sub-centimeter spatial alignment accuracy, ensure the strict correspondence between the two features in spatial position, and lay a geometric foundation for subsequent feature fusion.

[0035] Based on the registration parameters, feature expression systems for reflection intensity and curvature are constructed separately. For reflection intensity features, a filter bank is used for texture analysis. Multi-directional and multi-scale filter responses are used to capture the periodic variation patterns of the grain pile surface, generating feature vectors containing attributes such as directionality and roughness. The statistical characteristics of the neighborhood intensity distribution are also quantified. For curvature data, a three-dimensional morphological operator is designed for feature extraction. The concave and convex morphology of the grain pile surface is described using Gaussian curvature features, with particular attention paid to the correlation between areas of curvature mutation and compaction boundaries. Ultimately, a multidimensional feature space is formed that encompasses spatial distribution, texture characteristics, and geometric morphology, providing a complementary information source for subsequent fusion.

[0036] Furthermore, a feature fusion network is established to dynamically determine the contribution weights of reflection intensity and curvature features. A dual-channel feature encoder is constructed to perform nonlinear mapping on the standardized reflection intensity and curvature features, respectively, and enhance feature expression capabilities through a residual learning module. At the feature fusion layer, a parallel structure of a spatial attention module and a channel attention module is designed: the spatial attention module generates a spatial weight map of the feature map through convolution operations to enhance the spatial consistency of the hardened area; the channel attention module learns the importance coefficients of each feature channel through global average pooling and a multi-layer perceptron. The two attention weights are adaptively fused through element-wise multiplication to generate comprehensive features that are both sensitive to intensity changes and capable of perceiving geometric deformations. Finally, multi-scale feature aggregation is achieved through a feature pyramid network, which enhances global semantic consistency while maintaining detailed features, and outputs hardened area-sensitive features that are discriminative of hardened areas.

[0037] In an optional embodiment, a first dynamic weight coefficient of the reflection intensity feature and a second dynamic weight coefficient of the curvature feature are obtained; based on the normalization parameter, the reflection intensity feature and the curvature feature are weightedly fused in combination with the first dynamic weight coefficient and the second dynamic weight coefficient to obtain the hardened area sensitive feature. The hardened area sensitive feature satisfies the following relationship: in, is a sensitive feature of the hardened area. is 3D point cloud data, is the first dynamic weight coefficient, Point Cloud The reflection intensity, is the mean value of the reflection intensity, is the standard deviation of the reflection intensity, is the second dynamic weight coefficient, Point Cloud curvature.

[0038] S22. Constructing a compaction area detection model for the grain pile based on the sensitive features of the compaction area.

[0039] In this embodiment, the sensitive features of the compacted area are mapped to the two-dimensional grid space, and the feature statistics in each grid of the two-dimensional grid space are obtained. The feature space of the compacted area of ​​the grain pile is obtained based on the feature statistics; combined with the feature space of the compacted area, a hybrid classification decision framework is constructed to obtain the compaction probability distribution of the sensitive features of the compacted area; physical constraints of the grain pile are established, and the compaction probability distribution is corrected based on the physical constraints to obtain the compaction corrected probability distribution; an adaptive detection threshold is constructed, and the compaction corrected probability distribution is detected according to the adaptive detection threshold to construct a compaction area detection model.

[0040] Specifically, the orthographic projection algorithm is used to expand the three-dimensional point cloud data into a two-dimensional plane, and the geometric consistency of the grain pile surface deformation is maintained through conformal projection transformation. During the projection process, the grain pile surface is divided into square grid cells, and each grid is used as an independent analysis primitive. For each grid cell, its internal feature statistics are calculated, including key indicators such as the mean reflection intensity, curvature standard deviation, and eigenvector variance. An adaptive neighborhood weighted algorithm is designed to dynamically adjust the statistical weight according to the distance from the grid to the center of the grain pile. A smaller kernel function is used in the edge area to suppress noise interference. A feature space of the compacted area is constructed, which contains spatial position coding and multidimensional feature statistics, providing standardized data for subsequent classification decisions.

[0041] Furthermore, a hybrid classification decision framework based on multi-classifier fusion was established by integrating support vector machines and random forest algorithms. A support vector machine was used to initially classify the feature space of the hardened area, while a radial basis function kernel was employed to capture the nonlinear relationships between features. The penalty coefficient and kernel parameter combination were optimized through cross-validation. Random forest classifiers were deployed in parallel, with each tree trained on a random subset of feature statistics. Out-of-bag error estimation was used to control model complexity. A dynamic weight allocation mechanism was designed to automatically adjust decision weights. By fusing the output probabilities of the two classifiers, a hardening probability distribution with confidence metrics was generated.

[0042] Specifically, the hybrid classification decision framework satisfies the following relationship: in, is the probability distribution of hardening, is the mixing weight coefficient, represents the output of the support vector machine, represents the output of the random forest algorithm, is the characteristic statistic of the grain pile, is the two-dimensional grid space coordinate.

[0043] In this embodiment, the physical constraints of the grain pile are introduced to regularize and correct the initial probability distribution. The threshold range of physical parameters for compacted areas and non-compacted areas is established through experimental calibration. When the classification probability exceeds 0.7 in a two-dimensional grid space, it is verified whether it meets the physical constraints of the compacted area: the density of the grain pile is greater than 750kg / m³ and the curvature is greater than 0.1. For abnormal probability values ​​that do not meet the physical constraints, a Markov random field is designed for spatial regularization, and local consistency optimization is achieved through probability transfer between adjacent grids. Especially for the edge area of ​​the grain pile, an edge preservation term is introduced to retain the boundary sharpness while correcting the probability distribution, and finally generate a compaction correction probability distribution that conforms to physical reality. The compaction correction probability distribution satisfies the following relationship: in, Corrected probability distribution for hardening, is the probability distribution of hardening, is the correction value of the probability of compaction, is the density of the grain pile, is the curvature.

[0044] Furthermore, an adaptive detection threshold based on statistical distribution is constructed, and the global characteristics of the probability distribution are corrected through kernel density estimation analysis. The probability value range is divided into intervals of equal width, the probability density of each interval is calculated and the bimodal distribution characteristics are identified, and the probability dividing point between the compacted and non-compacted classes is automatically determined. A dynamic threshold adjustment strategy is designed. For example, when the ambient humidity of the grain pile exceeds 14%, the threshold is lowered by 0.05 to compensate for the feature offset caused by humidity. A sliding window detection mechanism is adopted to achieve full coverage of the grain pile through step-by-step scanning. For the detected compacted areas, morphological post-processing is implemented: first, the closed operation of the neighborhood is performed to fill the holes, and then the isolated areas are removed, and the final output is the compacted area detection result.

[0045] In an optional embodiment, the ambient temperature and humidity of the grain pile are obtained, and a first compensation coefficient for the ambient temperature and a second compensation coefficient for the ambient humidity are obtained; based on the ambient temperature, ambient humidity, the first compensation coefficient and the second compensation coefficient, an adaptive detection threshold is obtained in combination with a baseline detection threshold. The adaptive detection threshold satisfies the following relationship: in, is the adaptive detection threshold, is the baseline detection threshold, is the first compensation coefficient, is the ambient temperature of the grain pile, is the second compensation coefficient, The ambient humidity of the grain pile.

[0046] The above-mentioned hardened area detection results satisfy the following relationship: in, is the detection result of the hardened area, Indicates compaction, Corrected probability distribution for hardening, is the adaptive detection threshold, Indicates non-compacted.

[0047] S3. Establishing an abnormal surface recognition network architecture for the grain pile, and obtaining abnormal surface features of the grain pile based on the abnormal surface recognition network architecture.

[0048] In this embodiment, a multi-layer grouping strategy is constructed, including shallow and deep layers, to obtain the texture features of the grain pile in the shallow layer, and the morphological features of the grain pile in the deep layer; a dual-branch attention mechanism is established, including channel attention and spatial attention, to obtain the abnormal area of ​​reflection intensity based on channel attention, and to obtain the mutation area of ​​curvature based on spatial attention; the multi-layer grouping strategy and the dual-branch attention mechanism are combined to construct an abnormal surface recognition network architecture; based on the abnormal surface recognition network architecture, the surface of the grain pile is feature recognized to obtain abnormal surface features.

[0049] A multi-layer grouped feature extraction architecture based on convolutional neural networks was established, and a progressive feature encoding method was used to achieve feature decoupling from shallow texture to deep morphology. In the shallow network branch, a convolution kernel group was deployed to capture the micro-texture features of the grain pile surface at multiple scales. A residual connection structure was introduced to alleviate the gradient vanishing problem while enhancing information interaction between features at different scales. In the deep branch, a three-dimensional convolution module was constructed to process the spatial topological relationships of point cloud data and generate morphological features that encompass the entire grain pile. A cross-layer feature fusion pathway was designed to splice shallow texture features with deep morphological features in the channel dimension to form a composite feature representation that contains multi-scale contextual information.

[0050] A channel-space collaborative attention mechanism is constructed to achieve adaptive weighting of feature channel dimensions and spatial dimensions. In the channel attention branch, global average pooling is used to compress the spatial dimension and generate a channel description vector. The importance weight of each channel is learned through a shared network containing two fully connected layers, and higher weights are given to channels corresponding to areas with abnormal reflection intensity. In the spatial attention branch, spatial weights are generated through convolution, and the weight values ​​are mapped to the range of 0-1 using an activation function, focusing on strengthening the spatial response of areas with curvature mutations. A dual-branch feature calibration unit is designed to cascade the channel attention weights with the spatial attention map, which can enhance the feature saliency of abnormal surfaces and focus on key areas such as hardening boundaries.

[0051] Furthermore, a network architecture for abnormal surface recognition was constructed, which includes a feature extraction layer, an attention module, a feature fusion layer, and a decision layer. The feature extraction layer reduces the number of parameters through grouped convolution operations while maintaining feature expression capabilities. A feature fusion unit is deployed after the attention module, and a channel attention weighted fusion strategy is adopted to dynamically fuse the original features with the attention-calibrated features. The decision layer consists of parallel classification and regression branches. The classification branch outputs the abnormality probability of each spatial position, and the regression branch predicts the continuous value of the abnormality degree. A multi-task loss function is used during training, and cross-entropy loss and smoothing loss are combined for joint optimization. To enhance the robustness to grain pile deformation, a data augmentation strategy is designed, including random rotation, affine transformation, and noise injection, to simulate surface deformation characteristics under different stacking conditions. Label smoothing technology is combined to prevent overfitting.

[0052] To identify abnormal surface features, input data passes through a multi-layer feature extraction module and a dual-branch attention module. The channel attention branch automatically identifies abnormal reflection intensity patterns associated with compaction, such as areas of sudden reflectivity changes caused by surface hardening. The spatial attention branch accurately locates curvature anomalies and captures subtle surface deformations of the grain pile. The two types of weighted attention features are concatenated through a feature fusion layer to generate a comprehensive feature containing information about the location and type of anomalies. At the decision layer, the boundaries of the abnormal regions are dynamically determined based on local feature statistics. Morphological post-processing is performed to remove isolated noise points and segment adhesion areas to generate a connected abnormal surface contour. The final output of the abnormal surface features includes spatial coordinates, anomaly type, and confidence score, providing accurate anomaly location information for subsequent intelligent detection of compacted areas.

[0053] S4. Based on the compaction area detection model and in combination with the abnormal surface features, an intelligent detection is performed on the grain pile to obtain a preliminary detection of the compaction area.

[0054] In this embodiment, the sensitive features of the compacted area and the abnormal surface features are spliced ​​to construct a joint feature tensor of the grain pile; the spatiotemporal constraints of the compaction of the grain pile are constructed, and the joint feature tensor is tracked in time series based on the spatiotemporal constraints to obtain the compaction evolution characteristics of the grain pile; the compaction evolution characteristics are intelligently detected based on the compaction area detection model to obtain a preliminary detection of the compaction area.

[0055] Specifically, the core prerequisite of intelligent detection is to deeply integrate the sensitive features of the compacted area with the abnormal surface features; establish a feature alignment mechanism, and align the two types of features through a spatial transformation network to eliminate the spatial offset caused by differences in data acquisition perspectives; adopt a channel splicing strategy to connect the sensitive features reflecting the physical properties of compaction (such as reflection intensity gradient and curvature mutation value) with the surface features representing abnormal morphology (such as texture roughness and deformation displacement) in the feature channel dimension to generate a dimensionally expanded joint feature tensor; to improve the efficiency of feature expression, deploy a feature selection module, use the random forest algorithm to evaluate the importance score of each channel feature, and eliminate redundant channels with contributions below the threshold; construct a joint feature tensor containing spatial coordinates, feature types and feature values ​​to form a multi-dimensional grain pile surface state description system.

[0056] Based on the joint feature tensor, spatiotemporal constraints on grain pile compaction are constructed to capture dynamic changes. These spatiotemporal constraints incorporate both physical and statistical rules. At the physical level, an upper limit is set on the expansion rate of the compacted area to prevent false alarms caused by noise. At the statistical level, the correlation between compaction formation and environmental parameters in historical data is analyzed, and a probability transition matrix is ​​established to predict evolutionary trends. Time series tracking utilizes a hybrid architecture of Kalman filtering and long-short-term memory networks: the Kalman filter processes linear dynamics (such as the translation of the center of gravity in the compacted area), while the long-short-term memory network captures nonlinear evolution (such as the acceleration of local compaction caused by sudden changes in humidity). During this evolutionary process, environmental sensor data is accessed in real time, prediction weights are dynamically adjusted, and compaction risks are rapidly responded to. Ultimately, the evolutionary characteristics of grain pile compaction are derived, providing temporal context support for intelligent detection of grain pile compaction.

[0057] Furthermore, the evolution characteristics of compaction are combined with the compaction area detection model to make hierarchical judgments: the first layer marks the strongly compacted areas based on the static threshold; the second layer initiates spatiotemporal continuity verification for the suspected areas. If the area continues to expand in the continuous data or the environmental parameters exceed the standard, it is upgraded to a compaction area; the third layer combines the manual review interface to mark the low-confidence areas as "to be observed" and include them in the active learning sample library. In the post-processing stage of the detection results, morphological closing operations are used to fill the holes, and density clustering is used to merge the fragmented areas to ensure that the compaction boundaries are smooth and continuous. At the same time, a confidence assessment system is designed: the three indicators of static probability, temporal stability, and environmental matching are integrated to generate confidence scores, which are marked with color gradients in the visual interface to guide the clearance machine to prioritize high-confidence areas.

[0058] S5. Perform boundary positioning on the preliminarily detected compacted area to obtain a boundary of the compacted area, so as to obtain the compacted area of ​​the grain pile.

[0059] In this embodiment, candidate seeds are selected based on the preliminary detection of compacted areas, and the candidate seeds are clustered to construct a valid seed set; multimodal feature constraints are constructed, and growth criteria for the valid seed set are constructed based on the multimodal feature constraints; the growth result of the valid seed set is obtained based on the growth criteria, and the boundary of the compacted area is obtained based on the growth result to obtain the compacted area.

[0060] In an optional embodiment, a region growing algorithm is used to locate the compaction boundary. Screening high-quality candidate seeds from the initially detected compaction area serves as the starting point for boundary localization. A seed selection strategy based on feature similarity is employed. Within the initial detection area, a sliding window scan is performed with a step size of 5 cm. Within each window, characteristic metrics such as the mean reflection intensity, curvature standard deviation, and texture complexity are calculated. A dual-threshold screening mechanism is implemented: the reflection intensity threshold is set at ±15% of the local mean, and the curvature mutation threshold is set at 0.03 mm⁻¹. Only window centers that meet both criteria are retained as candidate seeds. To eliminate redundant seeds, a density peak clustering optimization algorithm is deployed. Cluster centers are automatically determined by calculating local density and relative distance. Seeds with a spatial distance of less than 10 cm and high feature similarity are merged into the same cluster. The resulting valid seed set includes three elements: spatial location, feature vector, and confidence score. The confidence score is determined by the degree of match between the seed features and the typical compaction pattern, providing priority guidance for subsequent region growing.

[0061] Specifically, a multimodal feature constraint system encompassing physical properties, geometric morphology, and temporal evolution was established: The physical property constraint, based on the physical mechanism of grain pile compaction, sets a reflection intensity gradient threshold and a curvature change threshold to suppress erroneous expansion of non-compacted areas; the geometric morphology constraint utilizes three-dimensional Gaussian surface fitting, requiring that the angle between the main curvature direction of the growth area and the direction of gravity of the grain pile be less than 30° to ensure that the boundary conforms to the compaction deformation law; the temporal evolution constraint incorporates historical detection results, giving growth priority to areas that remain stable for three consecutive scanning cycles. An adaptive growth criterion was designed based on multimodal constraints, dynamically adjusting the growth threshold: Merging is allowed when the reflection intensity difference and curvature difference of neighboring points are less than a threshold. The threshold is linearly adjusted based on the confidence of the current seed, and the growth order follows the confidence priority. High-confidence seeds are expanded first, forming a pattern of progressive growth from the core area to the edge.

[0062] Furthermore, the valid seed set is sorted in descending order of confidence to form a growth sequence. During initialization, the highest-confidence seed is pushed into the growth queue, and its neighboring points are subsequently ejected and tested. For each neighboring point, it is verified whether it meets the multimodal growth criteria. If so, it is merged into the current region and the newly merged neighboring point is pushed into the growth queue. A dynamic boundary detection mechanism is used during the growth process. When encountering an edge with a sudden change in reflection intensity or a point with a reverse change in curvature, an edge preservation strategy is initiated: local edges are detected, and if the edge strength exceeds a preset threshold, growth in the current direction is terminated. For areas of stagnant growth, a hole-filling strategy is implemented, detecting isolated holes with a diameter less than 5 cm and filling them using a morphological reconstruction algorithm. The final growth result includes attributes such as the spatial outline, area, and center of mass coordinates of the compacted area. To verify the accuracy of boundary positioning, a cross-validation strategy is used, comparing the detection results with manually annotated samples. Based on the growth results, the boundary is smoothed using polynomial curve fitting to eliminate jagged edges. The final output is a closed boundary contour that conforms to the grain pile compaction law, resulting in the compacted area of ​​the grain pile, meeting the requirements for accurate grain condition monitoring.

[0063] S6. Obtain the priority coordinates of the grain pile for cleaning based on the compacted area, and automatically generate a set of operation instructions for the grain cleaning machine.

[0064] In an optional embodiment, first, priority evaluation is performed based on the compacted area, and priority is sorted based on the area, thickness, curvature mutation value and spatial position of the compacted area: the volume of the compacted area is calculated through three-dimensional point cloud analysis, and the degree of compaction hardening is quantified in combination with the reflection intensity gradient. The hierarchical analysis method is used to assign differentiated weight coefficients to attributes such as physical properties of different areas and the influence of grain pile stability to obtain priority scores, and then the disposal level is obtained; secondly, a spatial topology optimization algorithm is designed to generate a cleaning path, and the boundary coordinates of the compacted area are projected to the working coordinate system of the bin clearing machine, and an improved genetic algorithm is used for path planning: the chromosome encoding adopts a mixed encoding method of regional sequence and working point coordinates, and the fitness function integrates the path length, The three indicators of operation overlap rate and equipment energy consumption are used to generate the optimal path sequence covering all compacted areas through selection, crossover and mutation operations; finally, the operation instruction set of the cleaning machine is generated based on the optimal path sequence, and the optimized path sequence is converted into executable operation instructions for the cleaning machine: the drilling depth of the drill bit of the cleaning machine is dynamically adjusted according to the compaction thickness, a speed transition section is set at the junction of the areas, and secondary crushing instructions are inserted into the core areas with too high priority scores. A standardized instruction package containing coordinates, operation parameters and steering instructions is generated, and control instructions are sent to the cleaning machine through wireless transmission. The status data of the cleaning machine is received synchronously to form a closed-loop control, which controls the offset between the actual operation trajectory and the planned path, and realizes accurate and efficient disposal of the compacted areas of the grain pile.

[0065] The above priority evaluation satisfies the following relationship: in, is the priority index, is the area of ​​the hardened region, is the average thickness of the hardened area, is the distance between the compacted area and the edge of the grain pile, The maximum span distance of the grain pile edge.

[0066] See Figure 2 In an optional embodiment, the present invention provides an intelligent detection system for grain pile compaction areas, the system comprising an input device, an output device, a processor, and a memory, wherein the hardware facilities are interconnected, wherein the memory is used to store a computer program, the computer program comprising program instructions, and the processor is configured to call the program instructions and execute the specific steps of the embodiment of the intelligent detection method for grain pile compaction areas provided by the present invention. The intelligent detection system for grain pile compaction areas provided by the present invention has a complete structure, is objective and stable, and improves the overall applicability and practical application capabilities of the present invention.

[0067] In summary, the method of the present invention provides an intelligent detection method and system for compacted grain pile areas. By constructing a data acquisition device integrated with a multispectral lidar, high-precision reflection intensity and three-dimensional point cloud data are simultaneously acquired, and curvature features are extracted based on local surface fitting. A dynamic weighting strategy is adopted to obtain compaction-sensitive features. The detection model corrected by physical constraints is combined to improve the accuracy of abnormal area recognition. A multi-layer grouping strategy and a dual-branch attention mechanism are designed to effectively capture material variations and morphological anomalies in compacted grain piles, and ultimately achieve precise positioning of the boundaries of compacted grain pile areas. The method of the present invention is easy to understand, simple to calculate, and has a small workload. It is convenient for engineering application and provides a theoretical basis and technical support for the further development of intelligent warehousing technology.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the above embodiments, or make equivalent replacements for some or all of the technical features therein. These modifications or replacements do not deviate the essence of the corresponding technical solutions from the scope of the technical solutions of the embodiments of the present invention, and they should all be included in the scope of the claims and description of the present invention.

Claims

1. An intelligent detection method for grain pile compaction area, characterized in that: The steps include: Constructing a data acquisition device for a grain pile, and obtaining the reflection intensity, three-dimensional point cloud data, and curvature of the grain pile according to the data acquisition device; Performing feature fusion on the reflection intensity and the curvature to obtain sensitive features of the compacted area of ​​the grain pile, and constructing a compacted area detection model of the grain pile based on the sensitive features of the compacted area; Establishing an abnormal surface recognition network architecture for the grain pile, and obtaining abnormal surface features of the grain pile based on the abnormal surface recognition network architecture; Based on the compaction area detection model, the grain pile is intelligently detected in combination with the abnormal surface features to obtain a preliminary detection of the compaction area; The boundary of the compacted area detected in the preliminary inspection is located to obtain the boundary of the compacted area, so as to obtain the compacted area of ​​the grain pile.

2. The intelligent detection method for grain pile compaction area according to claim 1 is characterized in that: The data acquisition device for constructing the grain pile, obtaining the reflection intensity, three-dimensional point cloud data and curvature of the grain pile according to the data acquisition device, includes: Constructing the data acquisition device based on a multispectral laser radar, and using the multispectral laser radar to obtain the reflection intensity; Obtaining the three-dimensional coordinates of the surface of the grain pile according to the multispectral laser radar, and acquiring three-dimensional point cloud data of the grain pile based on the three-dimensional surface coordinates; A quadratic surface of the grain pile is obtained by performing local neighborhood fitting on the three-dimensional point cloud data, and the curvature is obtained according to the quadratic surface.

3. The intelligent detection method for grain pile compaction area according to claim 1 is characterized in that: The step of fusing the reflection intensity and the curvature to obtain sensitive features of the compacted area of ​​the grain pile includes: Preprocessing the reflection intensity and the curvature to obtain standardized parameters of the grain pile, and constructing a spatial index of the standardized parameters; Performing spatial registration on the reflection intensity and the curvature based on the spatial index to obtain a registration parameter of the grain pile; Performing feature analysis on the reflection intensity and the curvature according to the registration parameters to obtain a reflection intensity feature and a curvature feature of the grain pile; The reflection intensity feature and the curvature feature are weightedly fused to obtain the sensitive feature of the hardening area.

4. The intelligent detection method for grain pile compaction area according to claim 3 is characterized in that: The step of performing weighted feature fusion on the reflection intensity feature and the curvature feature to obtain the hardening area sensitive feature includes: Obtaining a first dynamic weight coefficient of the reflection intensity feature and a second dynamic weight coefficient of the curvature feature; Based on the standardized parameters, the first dynamic weight coefficient and the second dynamic weight coefficient are combined to perform weighted feature fusion on the reflection intensity feature and the curvature feature to obtain the hardening area sensitive feature.

5. The intelligent detection method for grain pile compaction area according to claim 1 is characterized in that: The constructing of the compaction area detection model of the grain pile based on the sensitive features of the compaction area includes: Mapping the sensitive features of the compacted area to a two-dimensional grid space, obtaining feature statistics within each grid of the two-dimensional grid space, and obtaining a feature space of the compacted area of ​​the grain pile based on the feature statistics; Combining the feature space of the hardened area, a hybrid classification decision framework is constructed to obtain the hardening probability distribution of the sensitive features of the hardened area; Establishing physical constraints for the grain pile, and correcting the compaction probability distribution based on the physical constraints to obtain a corrected compaction probability distribution; An adaptive detection threshold is constructed, and the compaction correction probability distribution is detected based on the adaptive detection threshold to construct the compaction area detection model.

6. The intelligent detection method for grain pile compaction area according to claim 5 is characterized in that: The step of constructing an adaptive detection threshold comprises: Acquiring the ambient temperature and ambient humidity of the grain pile, and acquiring a first compensation coefficient for the ambient temperature and a second compensation coefficient for the ambient humidity; The adaptive detection threshold is obtained based on the ambient temperature, the ambient humidity, the first compensation coefficient, and the second compensation coefficient in combination with a baseline detection threshold.

7. The intelligent detection method for grain pile compaction area according to claim 1 is characterized in that: The step of establishing an abnormal surface recognition network architecture for the grain pile and obtaining abnormal surface features of the grain pile based on the abnormal surface recognition network architecture includes: Constructing a multi-layer grouping strategy, wherein the multi-layer grouping strategy includes a shallow layer and a deep layer, obtaining texture features of the grain pile in the shallow layer and obtaining morphological features of the grain pile in the deep layer; Establishing a dual-branch attention mechanism, the dual-branch attention mechanism includes channel attention and spatial attention, obtaining the abnormal area of ​​the reflection intensity according to the channel attention, and obtaining the sudden change area of ​​the curvature according to the spatial attention; Combining the multi-layer grouping strategy and the dual-branch attention mechanism to construct the abnormal surface recognition network architecture; The abnormal surface features are obtained by performing feature recognition on the surface of the grain pile based on the abnormal surface recognition network architecture.

8. The intelligent detection method for grain pile compaction area according to claim 1 is characterized in that: The method of performing intelligent detection on the grain pile based on the compaction area detection model and in combination with the abnormal surface features to obtain a preliminary detection of the compaction area includes: Performing feature splicing on the sensitive features of the compacted area and the abnormal surface features to construct a joint feature tensor of the grain pile; Constructing a spatiotemporal constraint for the grain pile to form compaction, and performing time-series tracking on the joint feature tensor based on the spatiotemporal constraint to obtain a compaction evolution feature of the grain pile; The compaction area is initially detected by performing intelligent detection on the compaction evolution characteristics based on the compaction area detection model.

9. The intelligent detection method for grain pile compaction area according to claim 1, characterized in that: The performing boundary positioning on the preliminarily detected compacted area to obtain a boundary of the compacted area to obtain the compacted area of ​​the grain pile includes: Select candidate seeds based on the preliminary detected hardened area, and cluster the candidate seeds to construct a valid seed set; Constructing a multimodal feature constraint, and constructing a growth criterion for the effective seed set according to the multimodal feature constraint; The growth result of the effective seed set is obtained according to the growth criterion, and the boundary of the hardened area is obtained according to the growth result to obtain the hardened area.

10. An intelligent detection system for grain pile compaction areas, characterized in that: The system includes an input device, an output device, a processor and a memory, and the input device, output device, processor and memory are interconnected, wherein the memory is used to store a computer program, the computer program includes program instructions, and the processor is configured to call the program instructions to execute the intelligent detection method for compacted areas of grain piles as described in any one of claims 1 to 9.

Images