Method and system for multi-dimensional monitoring of grain stores

Abstract

The application discloses a granary multi-dimensional monitoring method and system, and belongs to the technical field of granary monitoring. The method comprises the following steps: constructing a reference coordinate system, acquiring a grain pile surface profile based on a set update interval, and performing coordinate offset compensation; aligning the surface profile with the reference coordinate system to eliminate monitoring position deviation caused by grain pile compaction or loosening; collecting multi-source data in real time, setting multi-dimensional identification conditions to identify scene modes, and performing differential monitoring; when a dynamic scene is determined, triggering a layered monitoring mechanism; when a static scene is determined, triggering a static update mechanism; ensuring that effective monitoring data can be acquired in dynamic and static scenes; continuously monitoring the dynamic scene and performing dynamic termination judgment; once the dynamic scene is determined to end, triggering a smooth detection strategy to improve the long-term adaptability to dynamic changes of the grain pile and avoid sudden changes or faults in the monitoring data.

Description

Technical Field

[0001] This invention belongs to the field of grain storage monitoring technology, and relates to a multi-dimensional monitoring method and system for grain storage. Background Technology

[0002] As the core facility for grain storage, the development of ventilation technology in grain warehouses is closely related to the maintenance of grain quality and the control of losses. Early grain storage relied on natural ventilation, achieving air convection through the opening and closing of warehouse doors and windows. However, due to climatic limitations, it was difficult to effectively regulate the temperature and humidity inside the grain pile. With the application of mechanical ventilation technology, equipment such as fans and ducts forcibly introduced external airflow, achieving preliminary intervention in the temperature and moisture of the grain pile. As the scale of grain reserves expands and different grain varieties are stored in a concentrated manner, the storage environment places higher demands on the uniformity, energy efficiency, and adaptability of the ventilation system. The integration of sensor technology has promoted the improvement of monitoring capabilities, and the deployment of distributed temperature and humidity sensors has upgraded grain pile environmental monitoring from single-point detection to regional coverage.

[0003] However, existing technologies do not consider the impact of real-time changes in the physical state of grain piles during dynamic grain entry and exit from storage on the accuracy of monitoring systems. Especially in dynamic operation scenarios such as simultaneous entry and ventilation, rapid changes in the height, density, and particle distribution of grain piles can lead to blind spots or data distortion in traditional fixed-point monitoring equipment. For example, temperature and humidity sensors with preset depths cannot capture in real time local temperature and humidity anomalies caused by resistance fluctuations inside the grain pile, such as a sudden increase in humidity in ventilation dead corners. This results in the ventilation system lacking effective dynamic data support, ultimately leading to local condensation and mold growth or unreasonable energy consumption, making it difficult to meet the real-time monitoring needs of efficient modern grain storage operations. Summary of the Invention

[0004] In view of the shortcomings of existing technologies, the purpose of this invention is to provide a multi-dimensional monitoring method and system for grain storage, which achieves accurate capture of physical parameters of grain piles and dynamic adaptation of monitoring resources through dynamic contour tracking, multi-dimensional state discrimination, and non-steady-state transition monitoring.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] Multidimensional monitoring methods for grain warehouses include:

[0007] A reference image of the grain pile in an empty state is acquired, a reference coordinate system is constructed, and multi-view sequence images of the grain pile surface are obtained based on the set update interval to generate a point cloud image of the grain pile surface. Coordinate offset compensation is then performed to align the point cloud image of the grain pile surface with the reference coordinate system.

[0008] Real-time acquisition of multi-source images, setting multi-dimensional recognition conditions to identify scene patterns and perform differentiated monitoring. When a scene is determined to be dynamic, a hierarchical monitoring mechanism is triggered; when a scene is determined to be static, a static update mechanism is triggered.

[0009] Continuously monitor dynamic scenes and make dynamic termination judgments. Once the dynamic scene is determined to be over, a smooth detection strategy is triggered.

[0010] Specifically, the steps for identifying scene patterns include:

[0011] The acquired sound pressure spectrum image is segmented into frames, and the sound pressure level of each frame is calculated to obtain a sound pressure level sequence. The standard deviation of the fluctuation of the sound pressure level sequence is then calculated. Perform a Fast Fourier Transform on the sound pressure spectrum image to calculate the energy entropy S. e ;

[0012] Set the sound field fluctuation threshold to σ. TH The energy threshold is S TH To determine the sound field; if And S e >S TH If the sound field detection is successful, the scene will proceed to thermal field detection; otherwise, it will be classified as a dynamic scene.

[0013] The infrared temperature image is preprocessed and the temperature gradient is calculated. The heat flux density is calculated by combining the thermal conductivity of the grain and the heat dissipation is calculated by the central difference scheme.

[0014] If the absolute value of the heat dissipation does not exceed the heat dissipation threshold, the thermal field determination passes and the stress field detection is initiated; otherwise, it is determined to be a dynamic scene.

[0015] The stress field image was generated using Kriging interpolation, and the stress divergence was calculated.

[0016] If the absolute value of the stress divergence does not exceed the stress divergence threshold, the stress field determination passes and proceeds to the multi-field coupling comprehensive determination; otherwise, it is determined to be a dynamic scene.

[0017] Based on the weight of the grain varieties stored in the granary, a comprehensive judgment parameter is calculated, the granary is divided into virtual grids, and the proportion of grids whose comprehensive judgment parameter is less than the comprehensive judgment threshold is counted to obtain the spatial coverage rate.

[0018] If the spatial coverage exceeds the preset coverage threshold, it is determined to be a static scene; otherwise, a re-inspection mechanism is initiated; if the re-inspection finds no abnormalities, it is determined to be a static scene; otherwise, it is determined to be a dynamic scene.

[0019] Specifically, the hierarchical monitoring mechanism includes:

[0020] Construct a reference image set, record the reference time and reference amplitude of multiple reference images, take the average value, and establish an environmental noise database;

[0021] When the grain warehouse is in a dynamic scenario, the propagation time and actual amplitude are collected based on the sound wave device.

[0022] The ratio of propagation speed to empty warehouse speed is calculated, and the grain type calibration coefficient and energy attenuation factor are introduced. Through product coupling, a grain pile density inversion formula is established.

[0023] According to the density inversion formula, the density of each point in the grain pile is calculated, a density point cloud image is generated, and a compaction threshold and an overhang threshold are set to divide the density point cloud image into density regions.

[0024] Gaussian filtering and smoothing are applied to the density point cloud image, and the density gradient is calculated using the central difference method. Different precision grids are set according to the gradient.

[0025] Gravity direction projection correction is performed on the inclined surface grid by incorporating the slope of the grain pile surface.

[0026] Based on Kriging interpolation and coordinate offset compensation, a dynamic grid image adapted to the grain pile morphology is generated.

[0027] Specifically, the hierarchical monitoring mechanism also includes:

[0028] Different sampling frequencies are set for regions with different densities to collect multi-parameter collaborative data;

[0029] A permeability expression is constructed and discretely solved to calculate the pressure gradient and local resistance coefficient, determine the ventilation dead zone, and judge the risk level by combining the density field.

[0030] The risk index is obtained by calculating the difference between actual temperature and dew point temperature, the ratio of carbon dioxide concentration, and the ratio of wind speed, and then constructing a dynamic risk model.

[0031] When the risk index exceeds the risk threshold, an early warning is triggered, generating a three-dimensional visualized risk heat map.

[0032] Specifically, the steps for constructing the permeability expression include:

[0033] Acquire images of grain particles, extract the three-dimensional morphological features of the grain particles, and calculate the sphericity of the particles based on the ratio of the surface area of ​​a sphere of the same volume to the actual surface area of ​​the particles.

[0034] By combining the exponential calculation of the particle sphericity, the original particle diameter is corrected to an equivalent spherical diameter;

[0035] Permeability is set as a tensor. In the calculation of the principal components of the tensor, the density gradient influence coefficient is introduced to calculate the density gradient term. This term is then multiplied by the baseline permeability to obtain the principal component expression.

[0036] The real-time density field of the grain pile is obtained, and the dynamic porosity is calculated based on the ratio of the real-time density field to the actual density of the grain.

[0037] By combining the equivalent spherical diameter and the dynamic porosity, the baseline permeability is updated, and the principal component is updated synchronously.

[0038] Specifically, the static update mechanism includes:

[0039] A three-dimensional monitoring network was constructed to regularly collect carbon dioxide concentration images and infrared temperature images. The metabolic entropy was calculated by correlating the rate of concentration change with temperature using the Boltzmann constant.

[0040] A baseline static scan frequency is set, and when the metabolic entropy exceeds a preset metabolic threshold, the scan frequency is increased based on the enhancement coefficient.

[0041] Calculate the thermal inertia index, mark the area where the thermal inertia index exceeds the thermal inertia threshold as a high inertia area, emit low-power microwaves, receive the temperature rise signal generated by the dielectric loss of the grain pile, invert the deep moisture distribution, and superimpose the moisture distribution with the thermal inertia index to generate a moisture-heat coupling risk map.

[0042] The infrasound array is used to periodically scan the grain pile, acquire the infrasound time domain signal and perform spectrum analysis. Based on the preset baseline, the frequency band energy anomaly area is marked. When the infrasound is continuously abnormal or the moisture exceeds the corresponding threshold, near-infrared imaging is triggered to identify pest characteristics. The infrasound and near-infrared detection results are integrated to generate a risk area mask image.

[0043] Build a predictive model to output the risk index of each region in the future period, optimize the monitoring based on the prediction results, integrate the prediction results with the current monitoring data, and generate a static monitoring report.

[0044] Specifically, the steps for dynamic termination judgment include:

[0045] The initial assessment is conducted by collecting equipment signals. If the equipment operating parameters are lower than the preset percentage of the rated value and the duration exceeds the time limit, and the sound pressure level does not exceed the ambient background noise level within the preset time period, the initial assessment is deemed successful, and the process proceeds to the secondary assessment. Otherwise, the tiered monitoring mechanism continues to be implemented.

[0046] In the secondary judgment, the dynamic termination of the grain pile is decomposed into three major attenuation processes: particles, airflow, and damp heat. The obtained relaxation time is normalized and the relaxation completion degree is calculated. If the relaxation completion degree exceeds the completion threshold and continues for 2 cycles, the secondary judgment is deemed to have passed and the final judgment is entered. Otherwise, the hierarchical monitoring mechanism continues to be executed.

[0047] In the final judgment, the heat flow hysteresis angle, pressure fluctuation entropy and density change acceleration are calculated, and judgment thresholds are set for each. If there are no abnormalities in the judgments, a consistency judgment is made.

[0048] Initiate vortex frequency sweep and calculate echo frequency shift. If the echo frequency shift does not exceed the preset frequency shift threshold, determine dynamic termination and trigger smooth detection strategy; otherwise, continue to execute the hierarchical monitoring mechanism.

[0049] Specifically, the coordinate offset compensation steps include:

[0050] The point cloud image of the grain pile surface is acquired, and after point cloud denoising and contour extraction, a real-time point cloud is generated.

[0051] Based on the reference image, a reference point cloud is constructed. The real-time point cloud is registered with the reference point cloud using the iterative nearest point algorithm to generate a set of point pairs and calculate the global settlement.

[0052] The grain pile is divided into upper, middle and lower layers according to its height. Different settlement correction coefficients are set for each layer. The settlement of each layer is calculated and a layered settlement matrix is ​​generated.

[0053] Based on the theoretical depth of the fixed sensor and the grain layer it belongs to, the actual depth coordinates of the sensor are calculated using a stratified settlement matrix.

[0054] Specifically, the coordinate offset compensation steps also include:

[0055] Set a slope correction threshold. When the slope of the grain pile surface is greater than the slope correction threshold, calculate the half-angle tangent value based on the surface slope, combine it with the actual depth, calculate the overall horizontal correction amount, and allocate it to the X / Y axis of the reference coordinate system according to different scenarios to correct the horizontal offset.

[0056] Set an adjacent threshold, define sensors with a spacing smaller than the adjacent threshold as adjacent sensors, calculate the depth compensation difference between adjacent sensors, and if the depth compensation difference is greater than a preset difference threshold, trigger the lidar retest and take the median of the three compensation calculation results.

[0057] By combining the depth and horizontal offset correction results, the actual three-dimensional coordinates of the sensor are generated.

[0058] The multi-dimensional monitoring system for grain warehouses includes: a correction module, a scene recognition module, and a dynamic judgment module;

[0059] The correction module is used to acquire a reference image, construct a reference coordinate system for the storage body, collect multi-view sequence images of the grain pile surface in real time, generate a real-time point cloud, and perform coordinate offset compensation.

[0060] The scene recognition module includes a data acquisition unit, a recognition unit, and a monitoring unit.

[0061] The data acquisition unit is used to acquire multi-source images in real time;

[0062] The identification unit identifies dynamic or static scenes by using triple physical field coupling judgment, combining multi-field coupling weight to calculate comprehensive judgment parameters, and statistically analyzing the spatial coverage of the virtual grid.

[0063] The monitoring unit is configured with a hierarchical monitoring strategy and a static update strategy;

[0064] The hierarchical monitoring strategy is used to invert the density field and generate a dynamic grid image, calculate the permeability tensor and risk index, and generate a three-dimensional visualized risk heat map.

[0065] The static update strategy is used to construct a three-dimensional monitoring network for carbon dioxide concentration, calculate metabolic entropy, combine thermal inertia index and microwave dielectric loss to invert deep moisture, and identify pest characteristics.

[0066] The dynamic determination module includes multi-level determination units and transition units;

[0067] The multi-level determination unit determines the end of the dynamic scene by making three-level determinations and combining infrasound sweep frequency echo frequency shift analysis.

[0068] The transition unit is used to dynamically adjust the sampling frequency, track relaxation residue and damp heat inertia, optimize the allocation of monitoring resources based on the risk index, and switch to a static low-power mode.

[0069] The beneficial effects of this invention are:

[0070] To address dynamic changes in grain pile morphology, a baseline coordinate system is constructed using lidar and point cloud registration. Combined with layered settlement correction and slope compensation, dynamic calibration of monitoring coordinates is achieved, eliminating positional deviations caused by compaction or accumulation and ensuring data spatial positioning accuracy. Furthermore, the system employs a triple physical mechanism of sound field, thermal field, and stress field interlocking verification to avoid misjudgment based on a single signal, significantly improving scene recognition accuracy. Layered monitoring and static update mechanisms are designed for dynamic and static scenes respectively. In dynamic scenes, acoustic CT and adaptive meshing enable density field inversion. In static scenes, metabolic entropy and thermal inertia tracking provide early warnings of mold risk, reducing energy consumption. For the lag process after dynamic termination, relaxation time spectrum and multi-scale inertial monitoring address the lag in signal judgment by traditional equipment. The transition period strategy, through staged sampling and risk-driven scheduling, captures slow-variable risks such as continuous compaction and vortex attenuation, achieving a smooth transition from dynamic to static and risk management. This provides a highly efficient, low-consumption, and accurate intelligent monitoring solution for grain storage. Attached Figure Description

[0071] Figure 1 A schematic diagram of a multi-dimensional monitoring method for grain warehouses;

[0072] Figure 2 This is a flowchart illustrating the scene pattern identification process in this invention;

[0073] Figure 3 This is a flowchart of the hierarchical monitoring mechanism in this invention;

[0074] Figure 4 This is a flowchart of the static update mechanism in this invention;

[0075] Figure 5 This is a structural diagram of a multi-dimensional monitoring method for grain warehouses. Detailed Implementation

[0076] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments of the present invention and the specific features in the embodiments are detailed descriptions of the technical solution of the present invention, rather than limitations thereof. In the absence of conflict, the embodiments of the present invention and the technical features in the embodiments can be combined with each other.

[0077] Example 1

[0078] refer to Figures 1 to 4 As shown in the figure, this embodiment introduces a multi-dimensional monitoring method for grain warehouses, including the following steps:

[0079] Step S1: In the empty warehouse state, a baseline image is acquired using a deployed linear array camera. Combined with image features of a fixed baseline target in the baseline image, including target corner coordinates and texture pattern distribution, and using camera intrinsic and extrinsic parameters obtained beforehand via Zhang Zhengyou's calibration method, the pixel coordinates of the baseline target are converted into three-dimensional spatial coordinates. This constructs a baseline coordinate system including the warehouse's geometric dimensions and air duct positions. After the grain pile is put into use, a multi-view sequence of images of the grain pile surface is acquired in real time using the linear array camera. Since the obtained image sequence includes images taken from different angles of the grain pile, a stereo matching algorithm, such as window matching based on SAD, is used to calculate image parallax. This, combined with the camera calibration parameters, generates the grain pile surface. The surface cloud image is used to construct the surface contour of the grain pile. At the same time, the update interval is set to dynamically update the surface contour to solve the problem of monitoring coordinate offset caused by dynamic changes in the shape of the grain pile. Since the surface shape of the grain pile changes dynamically due to particle flow and gravity compaction during the process of entering and leaving the warehouse, the actual monitoring position of the sensor will deviate from the theoretical coordinate due to the settlement or lifting of the grain pile. At this time, coordinate offset compensation is performed to align the surface contour with the reference coordinate system, automatically correct the theoretical depth coordinate of the sensor caused by the settlement of the grain pile, eliminate the monitoring position deviation caused by the compaction or loosening of the grain pile, achieve accurate calibration of monitoring coordinates under dynamic changes in the surface shape of the grain pile, and ensure the consistency between the sensor monitoring position and the theoretical coordinate.

[0080] Specifically, the steps for coordinate offset compensation include:

[0081] Real-time acquisition of point cloud images of grain pile surface; removal of image noise by Gaussian filtering; removal of outliers by statistical filtering to achieve point cloud noise reduction; extraction of edge features of grain pile surface image by Canny edge detection algorithm; fitting of grain pile surface plane by RANSAC algorithm; selection of effective contour points located within edge features; generation of real-time point cloud.

[0082] Simultaneously acquire a reference image, extract a reference point cloud, and use the Iterative Closest Point (ICP) algorithm to register the real-time point cloud with the reference point cloud. This includes finding the point with the closest Euclidean distance in the reference point cloud for each point in the real-time point cloud to form a corresponding point pair. Then, calculate the optimal rotation matrix and translation vector using the least squares method to minimize the root mean square error of the corresponding point pair. When the change in root mean square error is less than the mean square change threshold or the number of iterations exceeds the upper limit of iteration, the iteration is stopped, thereby generating a set of point pairs.

[0083] Based on the height difference between each pair of points in the point pair set, the arithmetic mean is calculated to obtain the global settlement ΔZ, thus obtaining the translation vector ΔΘ=(ΔX,ΔY,ΔZ); where, the global settlement reflects the settlement of the grain pile surface relative to the reference plane. When the global settlement is positive, it represents settlement, and when it is negative, it represents accumulation. ΔX and ΔY reflect the horizontal offset of the grain pile surface centroid relative to the empty silo reference centroid, which is directly solved by the ICP algorithm.

[0084] Considering the compaction characteristics of a grain pile with a looser top and a tighter bottom, the grain pile is divided into upper, middle, and lower layers according to height, and a different settlement correction coefficient is set for each layer. The upper layer contains newly piled grain with large particle gaps and a large settlement potential; therefore, the settlement correction coefficient for the upper layer is set to k. upper The middle layer is the reference layer, and its settlement tends to be stable. Therefore, the settlement correction factor for the middle layer is set to k. middle =1.0, at this point no additional correction is needed. The lower layer is compacted by the pressure from the upper and middle layers, resulting in minimal settlement. Set the settlement correction factor for the lower layer to k. lower , and k upper >k middle >k lower The calculated global settlement is multiplied by the settlement correction coefficient for each layer to obtain the layered settlement, thus generating a layered settlement matrix; in this embodiment, k is set respectively. upper =1.2, k lower =0.8;

[0085] For each fixed sensor, the theoretical depth is obtained, and the grain layer to which the sensor belongs is determined by image segmentation algorithm, such as determining the grain layer by threshold-based region segmentation. The corresponding stratified settlement amount is obtained from the stratified settlement matrix, and the actual depth coordinates of the sensor are calculated by adding the theoretical depth. The global settlement amount is then corrected by stratification.

[0086] A slope correction threshold is set, and the slope of the grain pile surface is obtained in real time by calculating the gradient magnitude and direction of the point cloud image of the grain pile surface. When the surface slope is greater than the slope correction threshold, it indicates that there is obvious tilted accumulation. The horizontal offset is corrected by slope correction to correct the horizontal displacement caused by the slope of the grain pile surface. Based on the surface slope, the half angle tangent value is calculated and multiplied by the actual depth to calculate the total horizontal correction amount. The total horizontal correction amount is then distributed to the X / Y axes to obtain the corrected horizontal offset. If the grain pile is a symmetrical cone, such as the central feed inlet, the total horizontal correction amount is distributed to the X / Y axes according to the feed inlet azimuth angle. If there is a main slope direction, it is determined by the horizontal component of the normal vector and distributed according to the projection ratio of the normal vector on the X / Y axes to ensure that the vector sum is the total horizontal correction amount.

[0087] Set an adjacent threshold, convert the actual distance between sensors into the distance between image pixels, define sensors whose distance is less than the adjacent threshold as adjacent sensors, and calculate the absolute value of the difference in the actual depth of adjacent sensors to obtain the depth compensation difference; if the depth compensation difference is greater than the preset difference threshold, trigger the LiDAR retest, perform two more compensation calculations, and take the median as the final value.

[0088] By combining the depth and horizontal offset correction results, the actual three-dimensional coordinates of the sensor are generated to correct the spatial positioning of the monitoring data and ensure that parameters such as temperature and humidity correspond to the actual location of the grain pile.

[0089] Step S2: Real-time acquisition of multi-source images, including but not limited to point cloud images of the grain pile surface, infrared temperature images, and sound pressure spectrum images, and setting multi-dimensional recognition conditions to determine the state of the grain pile in real time to identify scene modes, including dynamic scenes and static scenes, and to perform differentiated monitoring. When a dynamic scene is determined, a hierarchical monitoring mechanism is triggered; when a static scene is determined, a static update mechanism is triggered to capture the dynamic changes in the state of the grain pile in a timely and accurate manner. Differentiated monitoring strategies are implemented according to different scenes to improve monitoring efficiency and targeting, and to ensure that effective monitoring data can be obtained in both dynamic and static scenes.

[0090] In this embodiment, a single signal is prone to misjudgment. For example, changes in lateral pressure caused by grain feeding speed or clumping cannot distinguish between grain pile collapse and normal mechanical vibration. It is easy to misjudge a grain pile that is actually still in a dynamic state as static. At the same time, in the scenario of ventilating while feeding grain, the increase in the height of the grain pile is easily accompanied by the reverse disturbance of the ventilation airflow. For example, the grain pile squeezes the air duct, causing changes in the airflow distribution. At this time, there is no obvious grain flow at the inlet and outlet, but the real-time change in the density of the grain pile has caused fluctuations in ventilation resistance. Multi-dimensional recognition can further distinguish between static mode and pseudo-static mode.

[0091] Specifically, the steps for identifying scene patterns include:

[0092] The acquired sound pressure spectrum image is processed by framing, with a frame length of 1 second and 50% overlap between frames to ensure no signal features are lost. Each frame undergoes grayscale normalization preprocessing. The sound pressure level is obtained by calculating the statistical characteristics of the grayscale values ​​within the frame, thus generating the sound pressure level sequence of the sound pressure spectrum image. The standard deviation of the sound pressure level sequence is then calculated. To characterize the stability of the sound pressure level, a fast Fourier transform is performed on the sound pressure spectrum image to obtain the distribution characteristics of the power spectral density, and then the energy entropy S is calculated. e ;

[0093] Set the sound field fluctuation threshold to σ. TH The energy threshold is S TH To determine the sound field; if And Se >S TH If the sound field test is successful, it indicates that there is no significant friction or collision between the grain particles and the energy is dissipated evenly. If so, the sound field test is successful and the test proceeds to the thermal field test. Otherwise, it indicates that there is micro-movement or flow of the particles. If so, the sound field test is unsuccessful and the test is directly classified as a dynamic scene.

[0094] Once the sound field determination is passed, the infrared temperature image is preprocessed. The Sobel operator is used to calculate the temperature gradient of the infrared temperature image. According to Fourier's law, the heat flux density is proportional to the temperature gradient. The proportionality coefficient is set to the thermal conductivity of the grain, and the heat flux density is calculated. The heat dissipation of the heat flux density is calculated through the central difference scheme, which is the sum of the partial derivatives of the heat flux density in the X, Y, and Z directions, to characterize the spatial convergence of the heat flux.

[0095] A heat dissipation threshold is set for thermal field determination. If the absolute value of heat dissipation does not exceed the heat dissipation threshold, it indicates that there is no heat accumulation or thermal short circuit, and the thermodynamic state of the grain pile is stable. In this case, the thermal field determination is passed and the stress field detection is initiated. Otherwise, it indicates that there is local condensation and heat release or heat absorption in ventilation dead zones. In this case, the thermal field determination is not passed and the scene is directly determined to be dynamic.

[0096] Once the thermal field determination is passed, the discrete pressure sensor data is converted into a pressure distribution image. A continuous pressure field image is generated using the Kriging interpolation algorithm. Since the pressure change in the height direction of the grain pile has been partially offset by the layered deployment of sensors, only the divergence of the pressure field image in the two-dimensional plane is calculated here, which is the sum of the partial derivatives of the pressure in the X and Y directions, and then the stress divergence is obtained to reflect the uniformity of pressure transmission.

[0097] A stress divergence threshold is set to determine the stress field. If the absolute value of the stress divergence does not exceed the stress divergence threshold, it indicates that the stress distribution inside the grain pile is stable and there is no stress mutation caused by local compaction or collapse. In this case, the stress field determination is passed and the multi-field coupling comprehensive determination is entered. Otherwise, it indicates that the grain pile structure has changed, such as deep hysteretic compaction. In this case, the stress field determination is not passed and the scene is directly determined to be dynamic.

[0098] Since single-field judgment is easily affected by noise interference, multi-field coupling is verified by the interlocking of energy dissipation, thermodynamics and mechanics, which improves the reliability of judgment. According to the grain varieties stored in the granary, the field coupling weights are fitted by historical data, and different weights are assigned to the sound field, heat field and stress field. The absolute value of the time change rate of the sound field energy entropy is calculated simultaneously, and the weighted sum is obtained by combining the absolute values ​​of heat dissipation and stress dissipation to obtain the comprehensive judgment parameters.

[0099] The grain warehouse is divided into virtual grids of a preset size, such as 1 cubic meter grids, and a comprehensive statistical threshold is set. The proportion of grids whose comprehensive judgment parameters are less than the comprehensive statistical threshold is counted to obtain the spatial coverage rate, thus ensuring that no local disturbance affects the global judgment.

[0100] If the spatial coverage exceeds the preset coverage threshold, it is determined to be a static scene; otherwise, a re-inspection mechanism is initiated, including infrasound scanning and microwave hygrometer detection. For example, if the peak infrasound energy in a certain area exceeds the environmental background by 10dB and the duration of the excess exceeds 30 seconds, it is determined that there is particle sliding in the grain pile, and the re-inspection shows an anomaly. Microwaves are emitted into the grain pile, and the change in dielectric constant after the interaction between the microwaves and the grain pile material is used to invert the local humidity. If the humidity in a certain area is detected to be greater than 14%RH (the safety threshold here is 13%RH) and the humidity difference with the surrounding area is greater than 2%, it is determined that there is a risk of local condensation in the area, and the re-inspection shows an anomaly.

[0101] If no abnormalities are found during the re-inspection, it is determined to be a static scenario; if any re-inspection indicator exceeds the limit, it is immediately corrected to a dynamic scenario.

[0102] Specifically, the steps of the tiered monitoring mechanism include:

[0103] During non-operational periods in the grain warehouse, such as 2 a.m., the scanning device is activated to scan the empty warehouse to perform background noise scanning, thereby generating a reference image set. The reference time t0 and reference amplitude A0 of the sound wave propagation in the empty warehouse at this time are recorded. The scan is repeated 10 times and the average image is taken to reduce the influence of random noise and establish an environmental noise database, which includes correction coefficients for sound velocity under different temperature and humidity conditions for subsequent data calibration.

[0104] When the grain warehouse is in a dynamic scene, the sound wave transmitter transmits Chirp-coded signals at a preset frequency. Each group of transmission modules is activated in sequence to avoid mutual interference between signals. The sound wave receiver locates the sound source through the TDOA algorithm and synchronously records the time difference of the sound wave from the transmitter to the receiver, which is recorded as the propagation time t, as well as the actual amplitude of the sound wave, to generate a two-parameter image.

[0105] The distance from the transmitter to the receiver of the sound wave transmitter is set to d. Based on the propagation time and the reference time, expressions for the propagation speed v and the empty velocity v0 of the sound wave in the grain pile are established, and the speed ratio of the propagation speed to the empty velocity is calculated. This speed ratio is the ratio of the reference time to the propagation time. Simultaneously, a grain type calibration coefficient and an energy attenuation factor are introduced. Through product coupling, a grain pile density inversion formula containing both propagation time and amplitude attenuation parameters is established. This effectively solves the nonlinear interference problem existing when using sound wave attenuation alone to invert density under high humidity conditions. The energy attenuation factor is obtained by calculating the natural logarithm of the reference amplitude and the actual amplitude, and the grain type calibration coefficient is calibrated through grain pile sample experiments.

[0106] Based on real-time collected data, the density of each point in the grain warehouse under the current dynamic scenario is calculated to accurately obtain the density information of the grain pile, thereby generating a density point cloud image. Compaction threshold and overhang threshold are set, with the compaction threshold being greater than the overhang threshold. The Otsu threshold segmentation algorithm is used to divide the density point cloud image of the grain warehouse into high-density areas, medium-density areas, and normal areas. Marking abnormal areas helps to detect potential problems in the grain pile in a timely manner. Among them, high-density areas are highlighted in red, which are prone to blocking air ducts; medium-density areas are highlighted in yellow; and normal areas are highlighted in blue, which are prone to causing ventilation short circuits.

[0107] The density point cloud image is smoothed by Gaussian filtering to remove data noise. The density gradient in three-dimensional space is calculated by the central difference method. Grids of different precision are set according to the gradient magnitude. Regions with density gradients greater than the encryption threshold are selected and defined as encryption regions, which are equipped with ultrafine grids for monitoring. Regions that do not exceed the encryption threshold are defined as uniform regions and are equipped with regular grids. This reduces the amount of computation while ensuring monitoring accuracy.

[0108] By introducing the slope of the grain pile surface obtained from real-time measurement, the grid on the inclined surface is corrected by gravity direction to ensure that the grid spacing does not exceed 0.3m in the vertical direction, so that the grid better fits the actual shape of the grain pile.

[0109] Based on the Kriging interpolation method, the data from fixed-installation sensors, combined with coordinate offset compensation, is mapped onto dynamically generated grid nodes, ultimately generating a dynamic grid image that adapts to the current shape of the grain pile, reducing monitoring blind spots and providing accurate grid positioning for subsequent data collection.

[0110] Set differentiated sampling frequencies. For high-density areas, activate the high-frequency sampling mode to capture subtle changes. For regular areas, activate the regular sampling mode to reduce data transmission volume and processing pressure. For medium-density areas, activate the medium-frequency adjustable sampling mode and collect multi-parameter collaborative data based on the sampling frequency to provide a basis for grain pile status analysis.

[0111] Considering that grain particles are not standard spheres, a permeability expression is constructed to calculate the permeability. The finite volume method is used to discretize the permeability expression. Combining with the measured wind speed in the air duct, the pressure gradient is solved in real time. During the solution process, the local resistance coefficients at the inlet and outlet of the air duct are calculated. The area where the wind speed is less than the preset ventilation threshold is defined as the ventilation dead zone, and the risk level is judged in combination with the density field situation. For the high-density area, it is marked as the high-warning area, and for the medium-density area, it is marked as the medium-warning area;

[0112] Calculate the difference between the actual temperature and the dew point temperature to reflect the condensation risk. Take the logarithm of the carbon dioxide concentration and calculate the ratio with the standard value to reflect the biological activity. Calculate the ratio of the wind speed to the critical wind speed to reflect the ventilation deficiency situation. Multiply these three factors to obtain the risk index to construct a dynamic risk model; the model reflects the condensation risk by calculating the difference between the actual temperature and the dew point temperature; among them, the dew point temperature is calculated in real time through the Magnus formula;

[0113] When the risk index is greater than the risk threshold, an alarm is triggered, and a three-dimensional visualized risk heat map is generated. The report is transmitted to the grain warehouse monitoring platform through the OPC UA protocol to realize the linkage display with the ventilation system and the equipment for entering and leaving the warehouse.

[0114] Specifically, the construction steps of the permeability expression include:

[0115] The traditional Ergun equation is based on spherical particles, but the actual grain particles are mostly irregular in shape. The grain particle images are obtained by scanning electron microscopy imaging or laser particle size analyzer, and the three-dimensional morphological characteristics of the grain particles are extracted. Based on the ratio of the surface area of a sphere with the same volume to the actual surface area of the particle, the sphericity of the grains stored in the grain warehouse is calculated. Combining with the exponential operation of the sphericity, the original particle diameter is corrected to the equivalent spherical diameter, so that the fluid resistance characteristics of non-spherical particles are closer to the calculation model of spherical particles, thereby reducing the calculation error of permeability caused by particle shape differences. Among them, the exponent is obtained by fitting through the hydrodynamics experiment during particle sedimentation. In this embodiment, the original particle diameter is multiplied by the two-thirds power of the sphericity to obtain the equivalent spherical diameter;

[0116] In actual storage, grain piles will form density gradients due to factors such as settling and agglomeration, resulting in anisotropic particle arrangement. Traditional scalar permeability cannot accurately describe this situation. Therefore, the permeability is upgraded from a single scalar to a tensor form to reflect the resistance differences in different directions. In the calculation of the principal components of the tensor, a density gradient influence coefficient is introduced. The density gradient influence coefficient is multiplied by the density gradient and then 1 is added to obtain the density gradient term. This term is then multiplied by the baseline permeability to obtain the principal component expression, which describes the resistance differences between agglomerated and loose areas. When the density gradient of a certain area is larger, the corresponding permeability will decrease accordingly, thereby simulating the particle compression and increased airflow resistance caused by uneven density.

[0117] Traditional equations using static porosity cannot adapt to porosity changes caused by grain pile settling. By using sonic CT scanning to acquire the density at various locations within the grain pile in real time, a real-time density field is obtained. Porosity is then inverted, and the ratio of the real-time density field to the actual grain density is calculated. Subtracting this calculated ratio from 1 yields the dynamic porosity. Combined with a corrected equivalent spherical diameter, the dynamic porosity is substituted into the calculation of the permeability benchmark value, achieving real-time calibration to update the benchmark permeability. This results in a principal component expression optimized by dynamic porosity, allowing for directional adjustments to the benchmark permeability. When grain pile settling leads to increased density, the porosity decreases accordingly, thus reducing the benchmark permeability value and achieving dynamic calibration of the permeability.

[0118] Specifically, the setting rules for the intermediate frequency adjustable sampling mode are as follows:

[0119] Initialize the basic sampling frequency to f base And set the density gradient sensitivity threshold, gradient enhancement factor, and abnormal event triggering threshold;

[0120] All sensors in the medium-density region are arranged according to f base The frequency acquisition data is used to calculate the density difference with the adjacent grid for each medium-density grid to obtain the local density gradient. The density gradient data is then smoothed by a Gaussian filtering algorithm to reduce random noise interference.

[0121] When the local density gradient of any grid exceeds the density gradient sensitivity threshold, it indicates a significant density change and a tendency to transition to a compacted or loose state, triggering a gradient enhancement mode. The sampling frequency is updated based on the adjustment coefficient, and the mode only applies to grids with gradient exceeding the limit and their adjacent grids, forming a local enhancement region with a radius of R; where R is a configurable spatial influence range parameter.

[0122] Real-time monitoring for abnormal events. Once the collected parameter values ​​exceed the abnormal event trigger threshold, an abnormal event is detected, and the overall sampling frequency of the medium-density area is temporarily increased to the emergency frequency.

[0123] Specifically, the steps of the static update mechanism include:

[0124] A three-dimensional monitoring network for carbon dioxide concentration is constructed, and carbon dioxide concentration images and infrared temperature images are collected regularly. The concentration change rate and temperature are correlated through the Boltzmann constant to calculate metabolic entropy, quantify the energy dissipation of biological activity in the grain pile, set a baseline static scanning frequency, and increase the scanning frequency once the metabolic entropy exceeds the preset metabolic threshold. An enhancement coefficient is set, and scanning is performed by multiplying the enhancement coefficient and the baseline static scanning frequency to locate abnormal biological activity areas.

[0125] Historical data from temperature sensors is acquired to generate a cumulative temperature change image. Based on the ratio of cumulative temperature change to thermal conductivity, a thermal inertia index is calculated, and a thermal inertia threshold is set. Regions with thermal inertia indices exceeding the thermal inertia threshold are marked as high-inertia regions. Low-power microwaves are emitted to these high-inertia regions, and the temperature rise signal generated by the dielectric loss of the grain pile is received. The deep moisture distribution is then inverted, and the moisture distribution is superimposed with the thermal inertia index to generate a humidity-heat coupling risk map to clarify the spatial distribution of high-risk regions. The low-power microwaves are those with a transmission power lower than a preset power threshold.

[0126] The grain pile is periodically scanned using an infrasound array to generate an infrasound time-domain signal image, and spectrum analysis is performed. Based on a preset baseline, anomaly areas in the frequency band are marked to preliminarily determine suspected pest areas.

[0127] When infrasound waves are continuously abnormal or moisture exceeds the corresponding threshold, near-infrared imaging is triggered. Near-infrared reflectance spectrum images are acquired through a hyperspectral camera, and pest characteristics are identified using the ResNet-18 model. The infrasound and near-infrared detection results are integrated to mark pest risk areas and generate risk area mask images and confidence scores to provide a basis for subsequent treatment.

[0128] Based on historical metabolic entropy, thermal inertia, and pest data, a predictive model is constructed using the SARIMAX model to output the risk index of each region in the future. According to the prediction results, the monitoring frequency of high-risk areas is increased, and real-time infrasound monitoring is activated. At the same time, the priority of manual inspection is marked. The prediction results are integrated with the current monitoring data to generate a static monitoring report.

[0129] Step S3: Continuously monitor the dynamic scene and make dynamic termination judgments to determine whether the current dynamic scene has ended. This avoids the limitations of judging the end of the scene solely by the equipment shutdown signal, ensuring the continuity and accuracy of grain pile status monitoring. Once the dynamic scene is determined to have ended, a smooth detection strategy is triggered. The latest grain pile height image is collected using lidar, and the depth coordinates of the fixed layer sensor are corrected to avoid positional deviations caused by grain pile settlement. The bottom acoustic density meter is activated, and a grain pile density layer cloud map is generated by analyzing changes in sound velocity. The focus is on monitoring ventilation dead corners or compacted arching areas in the dynamic operation. After each dynamic cycle, the frequency of abnormal occurrences in each area is automatically counted, an inspection priority map is generated, and the monitoring resource allocation strategy under subsequent dynamic scenes is optimized. This improves the long-term adaptability to dynamic changes in the grain pile, avoids abrupt changes or gaps in monitoring data, and improves the stability and reliability of monitoring data, providing strong support for accurate analysis of the grain pile status.

[0130] In this embodiment, after the grain loading and unloading operation is completed, if ventilation is carried out while the grain is being loaded, there is a gradual process from dynamic to static inside the grain pile. The newly piled grain is continuously compacted due to gravity, causing the actual monitoring depth of the fixed layer sensor to deviate from the theoretical coordinates. At the same time, there are still local airflows inside the grain pile that have not been completely diffused, such as eddies near the air duct, which will cause temporary temperature and humidity fluctuations. These processes cannot be directly determined to end by the equipment shutdown signal. It is necessary to capture the lagging physical changes through transition period monitoring. Therefore, there will inevitably be a non-steady-state stage after the dynamic scenario ends, and it is not possible to directly switch to static monitoring.

[0131] Specifically, the steps for dynamic termination judgment include:

[0132] The initial judgment is performed by collecting equipment signals, such as feeder current, outlet gate switch status, and fan power, to confirm the equipment shutdown signal. Sound pressure spectrum images within a specific frequency band are also collected, and the sound pressure level is calculated. If the equipment operating parameters are lower than a preset percentage of the rated value for a duration exceeding a preset time limit, and the sound pressure level does not exceed the ambient background noise level within a preset time period, the equipment stops operating. There is no residual grain flow impact sound in the silo, and this avoids misjudgments caused by short-term shutdowns. The initial judgment is then passed, and the process proceeds to the secondary judgment stage. Otherwise, the tiered monitoring mechanism continues.

[0133] In the secondary judgment, the dynamic termination of the grain pile is decomposed into three major attenuation processes: particles, airflow, and humidity and heat. The completion degree of each process is quantified by characteristic time. The high-frequency signal spectrum image generated by particle friction is captured by an acoustic emission sensor. The signal attenuation edge is extracted by Canny edge detection, and the particle contact relaxation time τ1 is calculated to reflect the attenuation speed of the relative motion of the particles. An airflow velocity vector image is generated by an ultrasonic velocimeter and a pressure sensor. The vortex motion trajectory is tracked by optical flow method, and the vortex relaxation time τ2 is calculated. The temperature and humidity distribution image is obtained based on the temperature and humidity sensor group, the gradient attenuation rate is calculated, and the particle contact relaxation time τ3 is derived to quantify the evolution process of humidity and heat distribution from non-equilibrium state to equilibrium state. The obtained relaxation time is normalized. The relaxation completion degree is calculated by multiplying the exponential attenuation of the three types of relaxation time. The comprehensive judgment is made in the form of product, which solves the one-sidedness of traditional single indicators.

[0134] If the relaxation completion rate exceeds the completion threshold and continues for two consecutive cycles, the secondary judgment is deemed passed and the final judgment is initiated; otherwise, the hierarchical monitoring mechanism continues to be executed.

[0135] In the final judgment, FFT analysis of deep temperature sensor data is used to generate a spectral image and calculate the thermal hysteresis angle. A pressure fluctuation image is generated using a piezoelectric film array at the bottom of the silo, and the pressure fluctuation entropy is calculated. A density change image is generated using an acoustic density tomography system, and the density change acceleration is calculated using the inter-frame density change rate. Judgment thresholds are set for each. If no abnormalities are found in any of the judgments, a consistency judgment is made. Specifically, thermal inertia thresholds, stress inertia thresholds, and mass inertia thresholds are set. If the thermal hysteresis angle is less than the inertia threshold, it indicates that there is no significant delay in heat conduction, no risk of local heat accumulation, and no abnormality in thermal inertia monitoring. If the pressure fluctuation entropy is greater than the stress inertia threshold, it is determined that the stress distribution is uniform, there is no continuous compaction trend, and no abnormality in stress inertia monitoring. If the absolute value of the density change acceleration is less than the mass inertia threshold, it is confirmed that there is no significant settlement of the grain pile and no abnormality in mass inertia monitoring.

[0136] Initiate vortex frequency sweep, activate the directional infrasound transmitter to scan, generate an echo frequency shift image, calculate the echo frequency shift through image registration. If the echo frequency shift does not exceed the preset frequency shift threshold, it is determined that the residual airflow attenuation is complete, at which point the process is dynamically terminated and a smoothing detection strategy is triggered; otherwise, the layered monitoring mechanism continues to be executed.

[0137] Specifically, the steps of the smoothing detection strategy include:

[0138] Based on the time series images, the transition period is divided into a high-sensitivity period, a decay period, and a stable transition period. Each period corresponds to a different sensor sampling frequency, and the sampling frequency will gradually decrease as the stage progresses.

[0139] We continuously tracked the relaxation residue and calculated relevant parameters. We combined infrasound frequency sweep to determine the residual vortex and take corresponding ventilation measures. At the same time, we managed the humid heat inertia, tracked the heat flow hysteresis angle, and started microwave hygrometer scanning when the threshold was exceeded. We also built a model based on relevant laws to predict humid heat diffusion.

[0140] For biological activity, continuous monitoring of metabolic entropy change is performed, and near-infrared spectroscopy scanning is activated when the entropy change rate exceeds the threshold.

[0141] By integrating multi-dimensional data through risk index calculation, dynamically adjusting the sampling frequency according to the risk level, and using historical data to build a database and model, accurate monitoring and prediction of the transition period can be achieved.

[0142] Specific measures were developed to address delayed compaction and abnormal humidity and heat. Energy consumption was optimized by adopting strategies such as time-sharing power supply. Analysis reports were generated, and the accuracy of the judgments was verified periodically. When the static scenario conditions were met, the system automatically switched to a static low-power mode, achieving a smooth transition from dynamic to static and risk management.

[0143] Example 2

[0144] Please see Figure 5 Another embodiment of the present invention provides a multi-dimensional monitoring system for grain warehouses, comprising: a correction module, a scene recognition module, and a dynamic determination module;

[0145] The correction module is used to acquire baseline images of the empty warehouse using deployed linear array cameras, construct a baseline coordinate system for the warehouse using fixed baseline targets, acquire multi-view sequence images of the grain pile surface in real time, generate point cloud images of the grain pile surface using a stereo matching algorithm, generate real-time point clouds through statistical filtering and RANSAC plane fitting, and register the baseline point clouds using the ICP algorithm to generate a set of point pairs; combined with the layered compaction characteristics of the grain pile, calculate the global settlement and layered settlement matrix, and dynamically correct the three-dimensional coordinates of the sensors by combining slope correction and depth compensation of adjacent sensors, achieving millimeter-level calibration of the monitoring coordinates, eliminating coordinate deviations caused by grain pile settlement or uplift, and ensuring that all parameters are consistent with the actual position of the grain pile;

[0146] The scene recognition module includes a data acquisition unit, a recognition unit, and a monitoring unit;

[0147] The data acquisition unit is used to acquire multi-source images in real time, process the sound pressure spectrum image in frames and calculate the sound pressure level, fluctuation standard deviation and energy entropy, calculate the heat flux density by performing finite difference on the infrared temperature image, and generate a continuous pressure field image by Kriging interpolation, providing multi-dimensional data support for scene recognition.

[0148] The recognition unit makes a judgment through the coupling of three physical fields: sound field, thermal field and stress field. It sets a triple judgment threshold, calculates comprehensive judgment parameters by combining multi-field coupling weights, and counts the spatial coverage of the virtual grid. It eliminates pseudo-static false judgments by infrasound scanning and microwave humidity re-examination, and accurately identifies dynamic or static scenes.

[0149] The monitoring unit is configured with a hierarchical monitoring strategy and a static update strategy;

[0150] The layered monitoring strategy is used to initiate the transmission of acoustic Chirp signals in dynamic scenarios, locate the sound source through the TDOA algorithm, invert the density field by combining the Topp model and the grain calibration coefficient, generate a dynamic grid image by combining Gaussian filtering and density gradient analysis, calculate the permeability tensor and risk index, generate a three-dimensional risk heat map by Kriging interpolation, and link the ventilation system for early warning.

[0151] The static update strategy is used to construct a three-dimensional monitoring network for carbon dioxide concentration. The metabolic entropy is calculated by the Boltzmann constant, and the deep moisture is inverted by combining the thermal inertia index and microwave dielectric loss. Infrasound spectrum analysis is used to identify pest characteristics and predict the risk index for future periods.

[0152] The dynamic determination module includes multi-level determination units and transition units;

[0153] The multi-level judgment unit is used to eliminate misjudgment of equipment shutdown signals, capture hysteresis compaction and airflow relaxation, and combine infrasound sweep frequency echo frequency shift analysis to achieve multi-dimensional verification of the end of dynamic scenarios.

[0154] The transition unit is used to divide the dynamic termination transition period into a high-sensitivity period, a decay period, and a stable transition period, dynamically adjust the sampling frequency, track relaxation residue and damp heat inertia, optimize the allocation of monitoring resources based on the risk index, and automatically switch to a static low-power mode.

[0155] Working principle and effects:

[0156] A baseline coordinate system is constructed using lidar and point cloud registration. Point clouds on the grain pile surface are collected in real time and the contour is dynamically updated. Through layered settlement correction and slope compensation, sensor coordinate offsets caused by grain pile compaction or accumulation are eliminated, achieving accurate calibration of the monitoring position. Through multi-dimensional physical quantity detection, various parameters are analyzed, and the dynamic or static state of the grain pile is verified by a triple-mechanism interlock, avoiding misjudgment by a single signal. In dynamic scenarios, acoustic CT scanning and adaptive mesh technology are used, combined with an improved permeability tensor model, to achieve grain pile density inversion and ventilation resistance prediction. At the same time, the sampling frequency is dynamically adjusted according to the density gradient to focus on high-risk areas. In static scenarios, biological activity and heat and moisture transfer are monitored in a low-power mode through metabolic entropy change and thermal inertia tracking, and pest warning is achieved by combining infrasound and near-infrared technology. When dynamic termination is determined, the relaxation time spectrum of particles, airflow, and heat and moisture is quantified and verified by thermo-inertial index to solve the problem of missed detection in the lag process. During the transition period, a smooth transition to static monitoring is achieved through staged sampling and risk-driven scheduling.

[0157] This invention precisely addresses the monitoring coordinate offset caused by dynamic changes in grain pile morphology, improving data spatial positioning accuracy. Multi-dimensional scene recognition significantly reduces the false judgment rate, avoiding missed detections due to pseudo-static risks. Layered monitoring and dynamic grid technology effectively capture blind spots in grain pile deformation and fluctuations in ventilation resistance. A static update mechanism achieves early warning of biological activity and pests in a low-power mode. Dynamic termination judgment and transition period strategies ensure comprehensive monitoring of lagging physical processes, avoiding monitoring gaps. The overall solution, through intelligent linkage across the entire process, increases the lead time for grain pile risk warnings, optimizes monitoring resource allocation and energy consumption management, and provides systematic technical support for the intelligent management of grain storage.

[0158] The above description is merely a preferred embodiment of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should also be considered within the scope of protection of the present invention.

Claims

1. A multi-dimensional monitoring method for grain warehouses, characterized in that, include: A reference image of the grain pile in an empty state is acquired, a reference coordinate system is constructed, and multi-view sequence images of the grain pile surface are obtained based on the set update interval to generate a point cloud image of the grain pile surface. Coordinate offset compensation is then performed to align the point cloud image of the grain pile surface with the reference coordinate system. Real-time acquisition of multi-source images, setting multi-dimensional recognition conditions to identify scene patterns and perform differentiated monitoring. When a scene is determined to be dynamic, a hierarchical monitoring mechanism is triggered; when a scene is determined to be static, a static update mechanism is triggered. The hierarchical monitoring mechanism includes: Based on an acoustic device, propagation time and actual amplitude are collected; the speed ratio of propagation speed to empty warehouse speed is calculated; a grain type calibration coefficient and energy attenuation factor are introduced, and a grain pile density inversion formula is established through product coupling; according to the density inversion formula, the density at each point of the grain pile is calculated, generating a density point cloud image; compaction threshold and slack threshold are set, and the density point cloud image is divided into density regions; the density point cloud image is subjected to Gaussian filtering and smoothing; the density gradient is calculated using the central difference method; different precision grids are set according to the gradient; the surface slope of the grain pile is introduced to adjust the grid on the inclined surface. Gravity direction projection correction; dynamic grid images adapted to grain pile morphology are generated based on Kriging interpolation and coordinate offset compensation; differentiated sampling frequencies are set for different density areas to collect multi-parameter collaborative data; a permeability expression is constructed and discretely solved to calculate pressure gradient and local drag coefficient, determine ventilation dead angles, and judge risk level in combination with density field; the difference between actual temperature and dew point temperature, carbon dioxide concentration ratio, and wind speed ratio are calculated and multiplied to obtain risk index, and a dynamic risk model is constructed; when the risk index is greater than the risk threshold, an early warning is triggered, and a three-dimensional visualized risk heat map is generated; The static update mechanism includes: Carbon dioxide concentration images and infrared temperature images are collected periodically. The concentration change rate and temperature are correlated using the Boltzmann constant to calculate metabolic entropy. A baseline static scanning frequency is set. When the metabolic entropy exceeds a preset metabolic threshold, the scanning frequency is increased based on the enhancement coefficient. A thermal inertia index is calculated. Areas where the thermal inertia index exceeds the thermal inertia threshold are marked as high-inertia areas. Low-power microwaves are emitted, and the temperature rise signal generated by the dielectric loss of the grain pile is received to invert the deep moisture distribution. The moisture distribution is superimposed with the thermal inertia index to generate a moisture-heat coupling risk map. The grain pile is scanned periodically using an infrasound array to acquire infrasound time-domain signal images and perform spectrum analysis. Frequency band energy anomaly areas are marked based on a preset baseline. When the infrasound is continuously abnormal or the moisture exceeds the corresponding threshold, near-infrared imaging is triggered to identify pest characteristics. The infrasound and near-infrared detection results are integrated to generate a risk area mask image. A prediction model is constructed to output the risk index of each area in the future period. The monitoring is optimized based on the prediction results. The prediction results are integrated with the current monitoring data to generate a static monitoring report. Continuously monitor dynamic scenes and make dynamic termination judgments. Once the dynamic scene is determined to be over, a smooth detection strategy is triggered.

2. The multi-dimensional monitoring method for grain warehouses according to claim 1, characterized in that, The steps for identifying scene patterns include: The acquired sound pressure spectrum image is segmented into frames, and the sound pressure level of each frame is calculated to obtain a sound pressure level sequence. The standard deviation of the fluctuation of the sound pressure level sequence is then calculated. ; Set the sound field fluctuation threshold to To determine the sound field; if If the sound field detection is successful, the scene will proceed to thermal field detection; otherwise, it will be classified as a dynamic scene. The infrared temperature image is preprocessed and the temperature gradient is calculated. The heat flux density is calculated by combining the thermal conductivity of the grain and the heat dissipation is calculated by the central difference scheme. If the absolute value of the heat dissipation does not exceed the heat dissipation threshold, the thermal field determination passes and the stress field detection is initiated; otherwise, it is determined to be a dynamic scene. The stress field image was generated using Kriging interpolation, and the stress divergence was calculated. If the absolute value of the stress divergence does not exceed the stress divergence threshold, the stress field determination passes and proceeds to the multi-field coupling comprehensive determination; otherwise, it is determined to be a dynamic scene. Based on the weight of the grain varieties stored in the granary, a comprehensive judgment parameter is calculated, the granary is divided into virtual grids, and the proportion of grids whose comprehensive judgment parameter is less than the comprehensive judgment threshold is counted to obtain the spatial coverage rate. If the spatial coverage exceeds the preset coverage threshold, it is determined to be a static scene; otherwise, a re-inspection mechanism is initiated; if the re-inspection finds no abnormalities, it is determined to be a static scene; otherwise, it is determined to be a dynamic scene.

3. The multi-dimensional monitoring method for grain warehouses according to claim 2, characterized in that, The steps for constructing the permeability expression include: Acquire images of grain particles, extract the three-dimensional morphological features of the grain particles, and calculate the sphericity of the particles based on the ratio of the surface area of ​​a sphere of the same volume to the actual surface area of ​​the particles. By combining the exponential calculation of the particle sphericity, the original particle diameter is corrected to an equivalent spherical diameter; Permeability is set as a tensor. In the calculation of the principal components of the tensor, the density gradient influence coefficient is introduced to calculate the density gradient term. This term is then multiplied by the baseline permeability to obtain the principal component expression. The real-time density field of the grain pile is obtained, and the dynamic porosity is calculated based on the ratio of the real-time density field to the actual density of the grain. By combining the equivalent spherical diameter and the dynamic porosity, the baseline permeability is updated, and the principal component is updated synchronously.

4. The multi-dimensional monitoring method for grain warehouses according to claim 3, characterized in that, The steps for dynamic termination determination include: The initial assessment is conducted by collecting equipment signals. If the equipment operating parameters are lower than the preset percentage of the rated value and the duration exceeds the time limit, and the sound pressure level does not exceed the ambient background noise level within the preset time period, the initial assessment is deemed successful, and the process proceeds to the secondary assessment. Otherwise, the tiered monitoring mechanism continues to be implemented. In the secondary judgment, the dynamic termination of the grain pile is decomposed into three major attenuation processes: particles, airflow, and damp heat. The obtained relaxation time is normalized and the relaxation completion degree is calculated. If the relaxation completion degree exceeds the completion threshold and continues for 2 cycles, the secondary judgment is deemed to have passed and the final judgment is entered. Otherwise, the hierarchical monitoring mechanism continues to be executed. In the final judgment, the heat flow hysteresis angle, pressure fluctuation entropy and density change acceleration are calculated, and judgment thresholds are set for each. If there are no abnormalities in the judgments, a consistency judgment is made. Initiate vortex frequency sweep and calculate echo frequency shift. If the echo frequency shift does not exceed the preset frequency shift threshold, determine dynamic termination and trigger smooth detection strategy; otherwise, continue to execute the hierarchical monitoring mechanism.

5. The multi-dimensional monitoring method for grain warehouses according to claim 4, characterized in that, The steps for coordinate offset compensation include: The point cloud image of the grain pile surface is acquired, and after point cloud denoising and contour extraction, a real-time point cloud is generated. Based on the reference image, a reference point cloud is constructed. The real-time point cloud is registered with the reference point cloud using the iterative nearest point algorithm to generate a set of point pairs and calculate the global settlement. The grain pile is divided into upper, middle and lower layers according to its height. Different settlement correction coefficients are set for each layer. The settlement of each layer is calculated and a layered settlement matrix is ​​generated. Based on the theoretical depth of the fixed sensor and the grain layer it belongs to, the actual depth coordinates of the sensor are calculated using a stratified settlement matrix.

6. The multi-dimensional monitoring method for grain warehouses according to claim 5, characterized in that, The coordinate offset compensation steps also include: Set a slope correction threshold. When the slope of the grain pile surface is greater than the slope correction threshold, calculate the half-angle tangent value based on the surface slope, combine it with the actual depth, calculate the overall horizontal correction amount, and allocate it to the X / Y axis of the reference coordinate system according to different scenarios to correct the horizontal offset. Set an adjacent threshold, define sensors with a spacing smaller than the adjacent threshold as adjacent sensors, calculate the depth compensation difference between adjacent sensors, and if the depth compensation difference is greater than a preset difference threshold, trigger the lidar retest and take the median of the three compensation calculation results. By combining the depth and horizontal offset correction results, the actual three-dimensional coordinates of the sensor are generated.

7. A multi-dimensional monitoring system for grain storage facilities, used to implement the multi-dimensional monitoring method for grain storage facilities as described in any one of claims 1-6, characterized in that, include: The module consists of a correction module, a scene recognition module, and a dynamic determination module. The correction module is used to acquire a reference image, construct a reference coordinate system for the storage body, collect multi-view sequence images of the grain pile surface in real time, generate a real-time point cloud, and perform coordinate offset compensation. The scene recognition module includes a data acquisition unit, a recognition unit, and a monitoring unit. The data acquisition unit is used to acquire multi-source images in real time; The identification unit identifies dynamic or static scenes by using triple physical field coupling judgment, combining multi-field coupling weight to calculate comprehensive judgment parameters, and statistically analyzing the spatial coverage of the virtual grid. The monitoring unit is configured with a hierarchical monitoring strategy and a static update strategy; The hierarchical monitoring strategy is used to invert the density field and generate a dynamic grid image, calculate the permeability tensor and risk index, and generate a three-dimensional visualized risk heat map. The static update strategy is used to construct a three-dimensional monitoring network for carbon dioxide concentration, calculate metabolic entropy, combine thermal inertia index and microwave dielectric loss to invert deep moisture, and identify pest characteristics. The dynamic determination module includes multi-level determination units and transition units; The multi-level determination unit determines the end of the dynamic scene by making three-level determinations and combining infrasound sweep frequency echo frequency shift analysis. The transition unit is used to dynamically adjust the sampling frequency, track relaxation residue and damp heat inertia, optimize the allocation of monitoring resources based on the risk index, and switch to a static low-power mode.

Images