A method for controlling the level of a sand bin with double levels

Abstract

The present application belongs to the technical field of material level height measurement, and particularly relates to a sand bin material level height control method with double material levels. In the method, annular acoustic transducer arrays are arranged in high and low material level areas of the inner wall of the sand bin. Cross detection of multi-path acoustic waves is used to obtain acoustic field slice data. After preprocessing, the data is input into a deep neural network model to calculate two-dimensional material density maps of each height layer, and the maps are fused and reconstructed into a three-dimensional visual model of the sand material accumulation pattern inside the sand bin. Based on the model, the actual filling height of the high and low material level areas is calculated, and after comparison with a preset threshold, corresponding control instructions are triggered. Through non-contact acoustic tomography and intelligent calculation, the present application improves the measurement stability, control accuracy and system robustness under extreme working conditions.

Description

Technical Field

[0001] This invention belongs to the field of material level measurement technology, specifically relating to a method for controlling the material level height of a sand silo with dual material levels. Background Technology

[0002] In the field of modern industrial production and automated warehouse management, real-time monitoring of material levels in bulk material storage systems is a core element in ensuring production continuity and safety. With the advancement of Industry 4.0, higher demands are being placed on the precise metering and inventory allocation of solid particulate materials in industries such as coal, building materials, and chemicals. Material level control not only optimizes resource allocation and reduces operating costs but also prevents production accidents such as overflowing warehouses or shutdowns due to empty warehouses, serving as a crucial foundation for the digital transformation of the warehousing process.

[0003] As a common raw material storage facility in construction and industrial manufacturing, the precise control of sand silos' material level is directly related to the smooth operation of the production process. Traditional sand silo management typically establishes a dual-level early warning mechanism, triggering automatic feeding or shutdown commands through real-time sensing of material accumulation height. This process involves complex physical environment sensing, requiring the measurement system to maintain stable data output and logical judgment even under harsh operating conditions such as high temperature, high dust, and strong mechanical vibration.

[0004] Traditional radar or ultrasonic measurement technologies often employ single-point ranging, making it difficult to detect the uneven accumulation of sand due to its angle of repose. This results in measurement data that fails to represent the actual volume distribution of the material, easily leading to frequent false alarms in the material level detection system. Furthermore, contact sensors are easily damaged by continuous impact and wear from sand, and traditional signal processing algorithms struggle to effectively eliminate nonlinear noise interference from high-concentration dust environments, causing frequent zero-point drift and accuracy degradation in the sensors. Existing technologies lack the ability to reconstruct the three-dimensional morphology of materials within the silo, failing to capture the dynamic characteristics of sand flow and limiting the adaptability of the control system to complex operating conditions. Summary of the Invention

[0005] The purpose of this invention is to provide a method for controlling the material level height of a sand silo with dual material levels, thereby solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: a method for controlling the material level height of a sand silo with dual material levels, comprising the following specific steps:

[0007] Step 1: Arrange at least two sets of annular acoustic transducer arrays along the circumference of the inner wall of the sand silo, one set corresponding to the high material level area and the other set corresponding to the low material level area. Each transducer has the function of transmitting and receiving acoustic waves.

[0008] Step 2: Each transducer is excited sequentially as a sound source, and the remaining transducers synchronously receive the sound wave signals passing through the internal space of the sand chamber, forming a multi-path cross-covering sound field slice dataset.

[0009] Step 3: Preprocess the collected acoustic echo signal to filter out environmental noise and irrelevant reflection components, and extract attenuation characteristics and flight time information on the effective propagation path;

[0010] Step 4: Input the pre-processed acoustic data into a pre-trained deep neural network model. This model calculates a two-dimensional density map of the material distribution on the cross-section of the sand bin based on the difference in reflection intensity of sound waves at the air-sand interface and the attenuation law of the propagation path.

[0011] Step 5: Merge the two-dimensional density maps of multiple height layers and generate a three-dimensional visualization model of the sand accumulation morphology inside the sand bin using a voxelization reconstruction algorithm;

[0012] Step 6: Based on the three-dimensional visualization model, calculate the actual material filling height corresponding to the high material level area and the low material level area respectively, and compare it with the preset threshold to trigger the corresponding feeding or shutdown control command.

[0013] Preferably, in step one, the annular acoustic transducer array consists of no less than 8 uniformly distributed transducer units, and the angle between adjacent transducers is no greater than 45 degrees, to ensure that the sound field coverage has no blind spots.

[0014] Preferably, in step two, the acoustic excitation adopts a polling method. In each round of excitation, only one transducer is activated as the main sound source, while the other transducers are in the receiving state. After completing one round, the next transducer is switched as the new sound source until all transducers have completed one transmission.

[0015] Preferably, the preprocessing in step three includes time-domain filtering, spectrum analysis, and time delay correction, which are used to eliminate the influence of sound speed drift caused by temperature changes and to remove fixed reflection interference from the warehouse wall structure.

[0016] Preferably, in step four, the deep neural network model adopts a multilayer perceptron architecture or a sequence modeling structure. Its input is the normalized echo amplitude and time delay sequence of each receiving channel, and its output is the probability value of material presence at discrete grid points.

[0017] Preferably, in step five, the voxelization reconstruction algorithm maps the two-dimensional density map of each height layer to a unified coordinate system, connects adjacent layer data through interpolation and smoothing, and finally constructs a continuous three-dimensional volume data representation of the sand surface contour.

[0018] Preferably, in step six, the actual filling height of the high-level and low-level material areas is not determined by a single measuring point value, but is determined based on the average or maximum height of the top surface of the material in the corresponding area in the three-dimensional model, thus avoiding local concavity and convexity errors caused by the angle of repose.

[0019] Preferably, the deep neural network model is trained with a large amount of simulation and measurement data before deployment, covering sand samples with different particle sizes, moisture content and packing morphology, so that it has the ability to generalize and recognize acoustic responses under complex working conditions.

[0020] Preferably, the operating frequency of the acoustic transducer is set within a specific range, which can ensure the ability to penetrate high concentrations of dust and obtain sufficient resolution to distinguish the air-sand interface.

[0021] Preferably, the control system further includes an anomaly diagnosis module, which automatically marks sensor failures or material flow anomalies when the reconstruction results fluctuate drastically or do not meet physical constraints, and initiates a backup logic judgment process.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] 1. This invention, by introducing acoustic tomography principles and deep learning computation mechanisms, achieves non-contact three-dimensional reconstruction of the material accumulation morphology inside a sand silo, overcoming the measurement deviation problem caused by the angle of repose effect in traditional single-point ranging methods. Due to the adoption of a fully acoustic, non-invasive sensing architecture, the system does not require direct contact with the material, improving long-term operational stability and reliability in extreme industrial environments such as high wear and high dust levels.

[0024] 2. By establishing independent height determination logic for each of the two material level areas and combining it with three-dimensional morphological information for comprehensive evaluation, the false trigger rate of high and low material level alarms is reduced, improving the accuracy and safety of warehouse automation control. This method has good scalability and can be adapted to various bulk material types and warehouse geometries, providing highly versatile and robust technical support for the digital upgrade of intelligent warehousing systems. Attached Figure Description

[0025] Fig. 1 This is a schematic diagram of the overall technical solution architecture according to the present invention;

[0026] Fig. 2 This is a schematic diagram of data flow according to the present invention;

[0027] Fig. 3 A flowchart illustrating the attenuation characteristics and time-of-flight information according to the present invention;

[0028] Fig. 4A flowchart for generating a three-dimensional visualization model of the sand accumulation morphology inside a sand silo according to the present invention;

[0029] Fig. 5 This is a flowchart of the feeding or shutdown control instructions according to the present invention. Detailed Implementation

[0030] Example 1: To make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to specific embodiments.

[0031] Please refer to Figs. 1 to 5 In this embodiment, the implementation process of a method for controlling the material level height of a sand silo with dual material levels is as follows:

[0032] Specifically, in step one, at least two sets of annular acoustic transducer arrays are arranged circumferentially on the inner wall of the sand silo. One set corresponds to the high-level material area, and the other set corresponds to the low-level material area. Each transducer has both acoustic wave transmission and reception functions. In actual engineering deployments, the sand silo is usually a cylindrical structure, and its internal space is filled with high-concentration dust, with the ambient temperature fluctuating with production conditions. To achieve omnidirectional sound field coverage, the annular acoustic transducer array in this embodiment consists of no fewer than eight evenly distributed transducer units. In a preferred embodiment, to achieve higher reconstruction resolution, a ring array of 16 transducer units is used. The angle between adjacent transducers is precisely controlled at 22.5 degrees. This high-density arrangement ensures that the sound field coverage within the cross-section of the sand silo is free of blind spots.

[0033] The transducer unit is installed using an embedded flange. Mounting holes are pre-drilled in the sand silo wall, ensuring the transducer's radiating surface is flush with the inner wall of the silo, thus preventing direct impact from material flow on the sensor. Each transducer unit integrates a piezoelectric ceramic wafer, with its operating frequency set within a specific range of 25 kHz to 35 kHz. This frequency range was chosen after rigorous engineering considerations: if the frequency is higher than 40 kHz, the sound waves attenuate too quickly in dust-filled air, resulting in insufficient signal-to-noise ratio at the receiver; if the frequency is lower than 20 kHz, the sound wavelength is too long to effectively distinguish subtle undulations on the sand surface and is easily affected by industrial machinery vibration noise.

[0034] To address the dual-level control requirement, the first ring-shaped acoustic transducer array is installed at the top of the sand silo, positioned 0.5 meters below the highest safe material level line designed for the silo, for monitoring high-level alarms. The second ring-shaped acoustic transducer array is installed slightly below the middle of the sand silo, corresponding to the lowest sustaining material level line during production, for monitoring low-level warnings. Both arrays are connected to the central control unit via shielded cables.

[0035] Specifically, in step two, each transducer is sequentially excited as a sound source, while the remaining transducers synchronously receive the sound wave signals passing through the internal space of the sand chamber, forming a multi-path cross-coverage sound field slice dataset. In this embodiment, the sound wave excitation adopts a strict polling method. The timing controller of the main control unit sends a trigger pulse, first activating transducer number 1 to enter the transmission mode, at which time this transducer converts electrical energy into high-power ultrasonic pulses. At the same time, transducers numbered 2 to 16 in the same array all switch to the receiving mode.

[0036] Because sound waves take time to propagate through the air, the transducers at the receiving end capture the waveforms after transmission, reflection, and scattering along different paths. In a complete excitation cycle, after the first transducer completes its transmission, the system waits for a preset aftershock decay time before switching to the second transducer as the main sound source, with the remaining 15 transducers handling reception. This process continues until all 16 transducers have taken turns acting as a transmission source. Since each transmission source corresponds to 15 receiving channels, a single array can acquire 240 independent acoustic propagation paths in a complete acquisition cycle. These paths intertwine into an extremely complex mesh structure within the sand chamber's cross-section, forming a sound field slice dataset encompassing time, amplitude, and spatial location information. To ensure real-time data transmission, the completion time for each polling round is controlled to within 200 milliseconds.

[0037] Specifically, step three involves preprocessing the acquired acoustic echo signal to filter out environmental noise and irrelevant reflections, and extracting attenuation characteristics and time-of-flight information along the effective propagation path. Preprocessing is a crucial step in ensuring reconstruction accuracy and includes several sub-steps:

[0038] A. Time-domain filtering. The acquisition system first performs high-speed sampling of the original analog signal at 1 MHz, converting it into a digital sequence. By setting a bandpass filter that matches the center frequency of the transducer, low-frequency mechanical noise generated by the stirring motor, conveyor belt, etc., as well as random high-frequency thermal noise generated by the electronic circuit, is filtered out.

[0039] B. Spectrum Analysis. The signal is converted to the frequency domain using Discrete Fourier Transform, and the energy distribution at the dominant frequency is observed. A significant shift in the dominant frequency indicates potential aging of the transducer or material buildup on the probe surface.

[0040] C. Time Delay Correction. The speed of sound propagation in air is significantly affected by temperature. In this embodiment, a high-precision temperature sensor is synchronously installed beside the transducer array. The computing unit corrects the ideal flight time according to the linear relationship between the speed of sound and temperature based on the real-time monitored temperature values. For example, the speed of sound is approximately 343 meters per second at 20 degrees Celsius, while it increases by approximately 12 meters per second when the temperature rises to 40 degrees Celsius. This correction eliminates measurement deviations caused by seasonal changes or residual heat in the sand.

[0041] D. Feature Extraction. In the preprocessed waveform, the algorithm automatically identifies the first arriving peak, and the difference between this moment and the transmission moment is the flight time. Simultaneously, the envelope area of ​​the received waveform is calculated and compared with the reference energy in an empty state to calculate the energy attenuation rate of the sound wave along its path.

[0042] Specifically, in step four, the preprocessed acoustic data is input into a pre-trained deep neural network model. This model calculates a two-dimensional density map of the material distribution across the cross-section of the sand silo based on the difference in reflection intensity and propagation path attenuation of sound waves at the air-sand interface.

[0043] The deep neural network model architecture used in this embodiment is an optimized fusion model of a multilayer perceptron and a convolutional structure. The model input is a 240-dimensional vector, which contains the time-of-flight offsets of all propagation paths and the normalized amplitude attenuation coefficients. The model was trained on massive amounts of simulation data before deployment. During the training phase, tens of thousands of different sand accumulation morphologies (including segregation, funnel shape, convex shape, etc.) were generated using Monte Carlo simulation methods, and the corresponding sound field response was calculated using acoustic simulation software.

[0044] In the actual calculation process, the neural network transforms the one-dimensional acoustic feature vector into a two-dimensional probability distribution map through nonlinear mapping logic. The two-dimensional density map divides the cross-section of the sand silo into a 32x32 discrete grid. Each grid point represents the physical state of that location within the sand silo. The neural network's output layer outputs 1024 values, each corresponding to the probability of material presence at a given grid point. If the value approaches 1, it indicates that the location is covered by sand; if the value approaches 0, it indicates that the location is air. Due to the significant difference in acoustic impedance between sand and air, sound waves experience severe reflection and transmission losses when contacting the sand surface. The deep neural network can sensitively capture these subtle nonlinear attenuation characteristics, thus achieving precise boundary delineation.

[0045] Specifically, in step five, the two-dimensional density maps of multiple height layers are fused, and a three-dimensional visualization model of the sand accumulation morphology inside the sand bin is generated through a voxelization reconstruction algorithm. In this embodiment, the two-dimensional density maps obtained from the high-level array and the low-level array are transformed by spatial coordinate transformation and mapped to a unified three-dimensional Cartesian coordinate system.

[0046] The voxel reconstruction algorithm first defines the interior of the sand silo as a spatial set composed of tiny cubes (voxels). For the missing intermediate region between the two sets of ring arrays, the algorithm employs a smoothing technique based on Laplace interpolation. This technique calculates the continuous evolution logic of material distribution between adjacent layers based on the density distribution trend of known height layers. For example, if the density map of the high-level layer shows sand in the central area, while the low-level layer shows a higher density at that location, the interpolation algorithm will construct a cone structure with a certain slope. After interpolation and edge sharpening, the system constructs a continuous three-dimensional volumetric data field. Subsequently, using an isosurface extraction algorithm, points with a probability value of 0.5 are found in the volumetric data field, and these points are connected to form a complete, closed triangular mesh surface. This surface represents the actual accumulation contour of the sand inside the sand silo, visually showing the angle of repose where the material accumulates higher on one side of the silo wall and lower on the other.

[0047] Specifically, in step six, based on the three-dimensional visualization model, the actual material filling height corresponding to the high material level area and the low material level area are calculated respectively, and compared with the preset threshold to trigger the corresponding feeding or shutdown control command.

[0048] Unlike traditional single-point measurements that only obtain a distance value, this embodiment uses surface integrals to determine the actual material level. In the 3D visualization model, the actual filling height of the high-level area is not taken as the value of a single highest point. The system first identifies all feature points on the top surface of the material within the high-level monitoring area. A. The system calculates the weighted average height of all feature points in this area as the average height of the area. B. Simultaneously, it extracts the value of the highest point in this area as an overflow warning reference. C. By calculating the ratio of the projected area of ​​the surface on the horizontal plane to the total area of ​​the region, the flatness of the material distribution is determined.

[0049] When the average height of the high-level material zone exceeds the preset upper threshold (e.g., 2 meters from the top of the silo), the control logic determines that the sand silo is approaching full. To prevent a roof collapse, the control unit immediately outputs a shutdown command, cutting off the power to the front-end feeding conveyor and simultaneously sending a high-level alarm to the central control room. When the actual filling height of the low-level material zone is lower than the preset lower threshold (e.g., 5 meters from the bottom of the silo), it indicates that the material inventory is insufficient to sustain subsequent production processes. At this point, the control unit triggers a feeding command, starting the upstream vibrating feeder to replenish material into the sand silo. During this process, the system continuously monitors changes in the three-dimensional morphology until the material level rises back to a safe range.

[0050] To further enhance the system's robustness, an anomaly diagnosis module is integrated into the control system. This module continuously monitors the logical rationality of the reconstruction results. For example, it is physically impossible for sand to suddenly increase in height by 3 meters in an extremely short time (e.g., within 1 second). If the 3D model calculated by the system exhibits drastic fluctuations exceeding physical constraints between two adjacent acquisition cycles, the anomaly diagnosis module will automatically identify that the current data may be affected by large pieces of material adhering to the sensor surface or strong external electromagnetic interference. In this case, the system will not immediately trigger control actions but will automatically mark the result as a suspected fault and initiate a backup logical judgment process, namely, verifying the validity of the data by extending the observation period and increasing the data smoothing and averaging through multiple rounds. If the fault persists, the system will switch to manual control mode and prompt maintenance personnel to clean the transducer probe.

[0051] Before formal deployment, the deep neural network model in this embodiment underwent an enhanced training process for complex working conditions. By introducing measured samples with different particle sizes (from fine sand to gravel), different moisture contents (5% to 20%), and different packing morphologies, the model acquired extremely strong generalization and recognition capabilities. In humid environments, the scattering characteristics of sound waves on the surface of sand change. Since the training set already covers such data, the neural network can automatically adjust the calculation weights to offset the influence of humidity on density map generation, ensuring the stability of measurement accuracy.

[0052] Example 2: Based on Example 1, this example further optimizes and expands the excitation strategy and data processing flow of the transducer to adapt to larger-scale and more complex irregular sand silos.

[0053] Specifically, in step one, the transducer arrangement is not limited to a horizontal ring. For ultra-large diameter sand silos, this embodiment adopts a cross-spiral arrangement. In addition to the conventional ring arrays at high and low material levels on the inner wall of the sand silo, a set of oblique sensor links connecting the upper and lower arrays is added. This layout greatly enhances the ability to perceive the material flow mechanism in the central area of ​​the sand silo. The protection level of each transducer unit is upgraded to the highest level, using a stainless steel corrugated pipe protective sleeve, which can withstand the enormous lateral stress generated during the fall of the sand.

[0054] Specifically, in step two, this embodiment introduces a concurrent pulse excitation mechanism. Instead of polling each transducer one by one, the array is divided into several relative transmission groups. For example, transducers 1 and 9 simultaneously transmit acoustic pulses with slightly different frequencies. The receiving end uses digital quadrature demodulation technology to separate the acoustic signals from different sources. This method halves the time required for the entire sound field scan, improving the real-time tracking speed during dynamic feeding.

[0055] Specifically, in the preprocessing stage of step three, an acoustic fingerprint identification operation is added. Since the sand silo wall may generate fixed echo interference, the system pre-records a set of wall echo fingerprints in an empty silo state. During production, this fingerprint signal is subtracted from the real-time acquired signal, thereby completely eliminating the interference of silo wall structural components (such as reinforcing ribs and internal struts) on material identification. Furthermore, to address sound velocity drift, this embodiment employs a self-calibration technique: direct sound velocity measurement is performed using two transducers at a fixed distance in the array. Since the geometric distance between these two transducers is a known physical constant, the most accurate sound wave propagation velocity in the current environment can be directly calculated by measuring the sound wave propagation time between them, without relying on an external thermometer. This method eliminates the impact of temperature sensor response lag.

[0056] Specifically, the deep neural network model in step four is replaced with a sequence modeling structure based on the Transformer architecture. Since sound field slice data inherently possesses strong spatial correlation, the Transformer model, through a self-attention mechanism, can automatically learn the coupling relationships between different propagation paths. For example, if several adjacent paths simultaneously exhibit high attenuation, the self-attention mechanism assigns higher weights to these paths, enabling a clearer delineation of the sand's edge slope when calculating the two-dimensional density map. The model's output is no longer merely discrete grid probabilities but also includes evaluation coefficients reflecting prediction confidence. When environmental dust reaches a critical concentration leading to extremely poor signal quality, the confidence coefficient decreases, thus alerting the system control logic to enter a conservative mode.

[0057] Specifically, in step five, the voxelization reconstruction algorithm incorporates dynamic time warping logic. The system considers not only the current two-dimensional density map but also the model evolution trends from the previous five historical moments. This fusion of time dimensions makes the generated three-dimensional visualization model smoother, eliminating instantaneous spikes caused by local material collapse or airflow disturbances. The three-dimensional model is stored as a high-dimensional tensor data structure, facilitating overlay and comparison with the building information model of the sand silo to calculate accurate inventory volume values.

[0058] Specifically, the control command triggering logic in step six has been refined and hierarchically classified.

[0059] A. Level 1 Warning: When the actual filling height reaches 80% of the threshold, the system activates frequency conversion control, reduces the feeding frequency of the feeder, and enters a low-flow, precise feeding state.

[0060] B. Level 2 alarm: When the height reaches the 100% threshold, a shutdown command will be executed immediately.

[0061] C. Safety Redundancy Judgment: If the 3D model shows that the material accumulation has a serious lateral phenomenon (i.e., one side is extremely high and the other side is extremely low), even if the average height does not exceed the standard, the system will trigger a balanced feeding suggestion or stop unilateral feeding to prevent uneven lateral pressure from damaging the silo structure.

[0062] Through the above-described optimized implementation methods, this invention not only solves the measurement deviation caused by the angle of repose, but also achieves a leap from simple material level control to refined material morphology management, greatly improving the safety and automation level of large sand silo operation.

[0063] Example 3: This example further describes a method for controlling the material level height of a sand silo with dual material levels under mixed storage conditions of multiple types of sand. Under such conditions, due to the significant differences in the density, moisture content, and surface roughness of the sand, traditional single models often struggle to maintain constant accuracy.

[0064] Specifically, in step one, the annular acoustic transducer array employs a multi-frequency composite unit. Each mounting node integrates a low-frequency transducer (20 kHz) and a high-frequency transducer (45 kHz). The low-frequency acoustic waves have extremely strong penetrating power, enabling the perception of the overall contours deep within the sand layer; the high-frequency acoustic waves have better surface resolution, enabling the capture of subtle undulations on the top of the sand.

[0065] Specifically, in step two, the data acquisition system is excited in an alternating frequency manner. First, a low-frequency scan of the entire array is performed to acquire large-scale acoustic field slices inside the sand chamber; then, a high-frequency scan is performed. The two datasets are marked with the same timestamp under the drive of a clock synchronization signal, forming a multispectral feature space.

[0066] Specifically, in step three, the preprocessing step includes envelope statistical feature extraction. The algorithm not only extracts the time of flight and amplitude but also calculates the skewness, kurtosis, and energy centroid shift of the echo waveform. These second-order statistical features reflect the roughness of the material surface. For example, the echoes from fine-grained dry sand are more concentrated, while the echo waveforms from large-grained or wet sand exhibit significant broadening and multipath scattering effects.

[0067] Specifically, in step four, the deep neural network model employs a dual-branch architecture. Branch one processes low-frequency acoustic data and outputs a coarse structure of the material distribution inside the sand bin. Branch two processes high-frequency acoustic data and extracts the microscopic contour features of the material surface. Subsequently, a fusion layer nonlinearly concatenates the features from the two branches. This multi-frequency fusion technique enables the model to automatically identify the type of material currently inside the bin. In the model output layer, the system generates a material type probability distribution in real time, thereby automatically adjusting the subsequent voxelization algorithm parameters based on the repose angle characteristics of different sand materials.

[0068] Specifically, in step five, the voxelization reconstruction algorithm introduces a physical constraint-based smoothing operator. This operator uses the natural angle of repose of the sand as a constraint. If the reconstructed surface slope exceeds the physical maximum angle of repose of the sand type, the algorithm will automatically perform local optimization adjustments to eliminate obvious noise interference, making the reconstructed 3D visualization model more consistent with physical reality.

[0069] Specifically, in step six, the control system connects to a cloud database. This database records all infeed and discharge curves and corresponding three-dimensional morphological evolution processes of the sand silo during past operating cycles. The control logic is no longer a simple threshold comparison, but introduces predictive control logic. A. Predictive Judgment: Based on the current feed flow rate and the material growth rate displayed by the reconstructed model, the system calculates the estimated remaining time to reach the high material level threshold. B. Advance Decision-Making: If the upstream production line cannot complete the switchover within the estimated remaining time, the system will reduce the feed rate in advance, thereby achieving seamless connection of the production process. C. Linkage Strategy: The system can also calculate the total mass of the material in the silo in real time based on the sand volume calculated from the three-dimensional model and the known material density. This mass information is fed back to the production management system for material balance accounting and batching ratio control.

[0070] When the anomaly diagnosis module detects data anomalies, this embodiment also provides a self-healing logic based on virtual path reconstruction. If a transducer in the array fails, resulting in data loss, the system utilizes the spatial completion capability of a convolutional neural network to infer the signal value that the damaged transducer should have received based on the sound field slices measured by other normal transducers. This soft fault self-healing technology greatly extends the continuous operation cycle of the system under extreme conditions and reduces the frequency of downtime maintenance.

[0071] Furthermore, the control system in this embodiment also integrates a web-based remote monitoring interface. By rendering the voxel-reconstructed 3D model in a lightweight manner, maintenance personnel can observe the sand accumulation pattern inside the sand silo in real time from a central control room thousands of kilometers away, with a visual effect indistinguishable from having an internal camera installed. However, compared to optical cameras, the acoustic reconstruction method of this invention can still clearly reveal the sand boundary even in dusty environments, providing a reliable data foundation for the construction of digital twins in smart factories.

[0072] In summary, this embodiment further enhances the adaptability and intelligence of the sand silo level control method with dual material levels in complex and variable industrial environments through multi-frequency fusion and predictive control technology.

[0073] Example 4: This example further describes in detail a method for controlling the material level height of a sand silo with dual material levels designed for ultra-deep sand silos. The height of such sand silos is usually more than 30 meters, and it is difficult for a single or double array to cover the entire material level range.

[0074] Specifically, in step one, the number of ring-shaped acoustic transducer arrays is increased to four. In addition to the ultra-high level array at the top and the ultra-low level array at the bottom, an intermediate reference array is added at 33% and 66% of the sand bin height, respectively. This vertically multi-layered architecture constructs a three-dimensional acoustic sensing network.

[0075] Specifically, in step two, a cross-layer excitation mechanism is introduced into the excitation mode. In each scan, not only is polling conducted within each layer, but the transducers of the higher-layer arrays also emit acoustic waves, which are received by adjacent lower-layer arrays. This addition of an oblique cross-layer path effectively compensates for the insufficient observation blind spots between horizontal layer arrays, thereby improving the reconstruction accuracy of the central region of the sand chamber.

[0076] Specifically, in step three, to address signal attenuation caused by the ultra-long propagation distance, the preprocessing module adds a time-varying gain compensation sub-step. As the flight time increases, the system automatically increases the front-end amplification gain of the analog-to-digital converter. Since acoustic wave energy attenuates inversely with increasing propagation distance, the gain compensation function is designed as an exponential growth model, ensuring that weak signals from the bottom of the bin or the far-end bin wall can be completely captured.

[0077] Specifically, in step four, the deep neural network model employs a hierarchical computation strategy. The first layer performs rapid computation on the local data of each array layer to obtain a preliminary two-dimensional map of that layer's cross-section. The second layer acts as a global integrator, receiving the preliminary two-dimensional maps from each layer and cross-layer path features, and performing global feature fusion. This hierarchical architecture reduces the dimensionality pressure of a single computation, enabling the system to maintain a second-level update rate even when processing data from four arrays and thousands of paths.

[0078] Specifically, in step five, the voxel reconstruction algorithm employs an adaptive resolution mesh. In the region near the sand surface, the voxel mesh is refined to the order of 5 cm to capture the funnel-shaped collapse caused by unloading. Inside the material or in pure air regions, the mesh is coarsened to the order of 20 cm. This strategy significantly saves computational resources, ensuring smooth operation of complex 3D reconstruction logic even on industrial embedded computing platforms.

[0079] Specifically, in step six, the control logic incorporates a flow stability assessment index. By analyzing the changes in the three-dimensional model over time, the system calculates the variance of the material surface fluctuations. If the variance increases abnormally, it indicates that dangerous flow phenomena such as bridging or rat holes may have occurred inside the sand bin. In this case, the control system will not only adjust the feed rate but also automatically activate the air cannons or vibrators installed on the bin walls to forcibly eliminate material bridging through external excitation and restore normal fluidized unloading.

[0080] This embodiment also details the emergency response mechanism of the control system. When the 3D visualization model shows that the sand surface is extremely unbalanced and continues to deteriorate, the system will trigger the imbalance risk protection procedure, forcibly stop all feeding and discharging operations, and guide the operation and maintenance personnel to discharge the material through a specific gravity balance path to protect the sand silo structure from buckling due to asymmetric loads.

[0081] Through this comprehensive and three-dimensional monitoring and control architecture, this embodiment successfully extends the method of the present invention to the field of large-scale bulk material storage, solving the long-standing problem of material level monitoring in ultra-deep sand silos.

[0082] Example 5: This example focuses on describing a method for controlling the material level of a sand silo with dual material levels deployed in harsh electromagnetic environments, especially in industrial sites with large frequency converters or high-power inductive loads.

[0083] Specifically, in step one, an all-fiber optic signal transmission architecture is used between the transducer unit and the central control unit. Each transducer node integrates a miniature electro-optic conversion module. The acoustic signal is converted into optical pulses for transmission at the initial stage of generation. Due to the inherent electromagnetic interference resistance of optical fiber, this design completely eliminates signal baseline drift and glitches caused by high-order harmonics of the frequency converter. Furthermore, the transducer's shielding shell uses a special permalloy material, which has high shielding effectiveness against low-frequency magnetic fields.

[0084] Specifically, in step two, the acoustic wave excitation employs spread spectrum coding technology. The transmitted pulse is no longer a single sine wave, but a series of pseudo-random sequence codes. The receiver uses correlation calculation techniques to extract only signals that highly match the transmitted code sequence. This principle, similar to spread spectrum communication, enables the system to accurately identify the arrival time and energy attenuation of acoustic waves even in extreme environments where the noise intensity is higher than the useful signal intensity.

[0085] Specifically, in step three, the preprocessing algorithm introduces nonlocal mean denoising. This algorithm utilizes the redundancy of the acoustic signal in periodic polling, and suppresses sudden impulse noise by searching for similar waveform blocks on the time axis and performing weighted averaging.

[0086] Specifically, in step four, adversarial interference samples are introduced into the deep neural network model during training. During the simulation training phase, various simulated electromagnetic interference waveforms are actively injected into the acoustic data. Through this reinforcement training, the neural network learns to extract robust spatial features from contaminated signals, improving the model's robustness under real-world conditions.

[0087] Specifically, in step six, the control command output module adopts a dual-machine hot-backup redundant architecture. Two completely independent processors simultaneously perform material level calculations and command generation. The control command is only officially issued to the actuator when the judgment results of the two processors are consistent within the allowable error range. If there is a significant ambiguity between the two, the system immediately enters a self-test mode, and a third-party monitoring unit makes a final arbitration by analyzing historical data trends, ensuring extremely high safety of control actions under high-risk operating conditions.

[0088] This embodiment also describes an automatic cleaning and maintenance subsystem for the transducer. Since the abrasive may contain sticky components, scale may form on the transducer's radiating surface when humidity is high. An ultrasonic self-cleaning loop is added to the physical structure of the system in step one. At fixed intervals, the transducer briefly enters a high-frequency self-excited oscillation mode, utilizing the high-intensity ultrasonic cavitation effect to peel off surface deposits, ensuring the acoustic pathway remains in optimal condition.

[0089] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.

Claims

1. A method for controlling the material level height in a sand silo with dual material levels, characterized in that, Includes the following steps: Step 1: Arrange at least two sets of annular acoustic transducer arrays along the circumferential direction on the inner wall of the sand silo, one set corresponding to the high material level area and the other set corresponding to the low material level area. Each transducer in the annular acoustic transducer array has the function of transmitting and receiving sound waves. Step 2: Each transducer is excited sequentially as a sound source, and the remaining transducers synchronously receive the sound wave signals passing through the internal space of the sand chamber, forming a multi-path cross-covering sound field slice dataset. Step 3: Preprocess the collected acoustic echo signal to filter out environmental noise and irrelevant reflection components, and extract attenuation characteristics and flight time information on the effective propagation path; Step 4: Input the pre-processed acoustic data into a pre-trained deep neural network model. The deep neural network model calculates a two-dimensional density map of the material distribution on the cross-section of the sand bin based on the difference in reflection intensity of sound waves at the air-sand interface and the attenuation law of the propagation path. In step four, the deep neural network model adopts a fusion model of multilayer perceptron and convolutional structure, and its input is a feature vector containing the time-of-flight offset and normalized amplitude attenuation coefficient of all propagation paths. The deep neural network model transforms a one-dimensional acoustic feature vector into a two-dimensional probability distribution map through nonlinear mapping logic. The two-dimensional density map divides the cross-section of the sand bin into multi-dimensional discrete grid points. Each discrete grid point corresponds to a material presence probability value. When the material presence probability value approaches a preset maximum value, it indicates that the location is covered by sand. When the material presence probability value approaches a preset minimum value, it indicates that the location is an air space. The deep neural network model utilizes the acoustic impedance difference between sand and air to perform material boundary delineation by identifying the nonlinear attenuation characteristics generated when sound waves contact the sand surface. Step 5: Merge the two-dimensional density maps of multiple height layers and generate a three-dimensional visualization model of the sand accumulation morphology inside the sand bin using a voxelization reconstruction algorithm; In step five, the voxelization reconstruction algorithm includes: transforming the two-dimensional density maps corresponding to the high-level and low-level material regions into spatial coordinates, and mapping them to a unified three-dimensional rectangular coordinate system. The interior of the sand bin is defined as a spatial collection composed of tiny voxels. For the missing intermediate region between the two sets of annular acoustic transducer arrays, an interpolation-based smoothing technique is used to calculate the continuity evolution logic of material distribution between adjacent layers based on the density distribution trend of known height layers. In the constructed continuous three-dimensional volume data field, the isosurface extraction algorithm is used to find the set points whose probability values ​​are equal to the preset intermediate values, and the set points are connected to form a closed triangular mesh surface, which represents the actual accumulation profile of the sand material inside the sand bin. Step 6: Based on the three-dimensional visualization model, calculate the actual material filling height corresponding to the high material level area and the low material level area respectively, and compare it with the preset threshold to trigger the corresponding feeding or shutdown control command.

2. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step one, the annular acoustic transducer array consists of no less than 8 uniformly distributed transducer units, and the included angle between adjacent transducer units is set to no more than 45 degrees to ensure that the coverage of the sound field within the cross-section of the sand chamber eliminates the detection blind zone. The transducer unit is installed by embedded fixing, and the radiation surface of the transducer unit is kept flush with the inner wall of the sand silo to avoid the direct impact of material flow on the sensor. Each transducer unit integrates a piezoelectric ceramic wafer, whose operating frequency is set within a preset frequency range. The lower limit of the preset frequency range is set to be higher than the frequency of industrial mechanical vibration and noise, and the upper limit of the preset frequency range is set to be lower than the high attenuation frequency of sound waves in a dusty environment. This ensures the ability to penetrate high concentrations of dust while obtaining the spatial resolution to distinguish the surface undulations of sand.

3. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step two, the polling method for sound wave excitation includes: The timing controller of the main control unit sends a trigger pulse to activate one of the target transducers in the ring acoustic transducer array to enter the transmission mode and convert electrical energy into ultrasonic pulses. At this time, all the other transducers in the same array switch to the receiving mode. During a complete excitation cycle, after the target transducer completes its transmission, the system executes a preset aftershock decay time to wait, and then switches to the next transducer as the main sound source in sequence, while the remaining transducers are responsible for receiving, until all transducers have taken turns acting as a transmission source once. Each of the aforementioned annular acoustic transducer arrays acquires multiple independent acoustic propagation paths within a complete acquisition cycle. These paths interweave into a grid structure within the cross-section of the sand chamber, forming a sound field slice dataset that encompasses temporal, amplitude, and spatial location information.

4. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step three, the preprocessing includes the following steps: First, perform time-domain filtering to sample the original analog signal at high speed and convert it into a digital sequence. Then, use a bandpass filter to filter out low-frequency mechanical noise from the mixing and conveying equipment, as well as random high-frequency thermal noise generated by the circuit. Second, perform spectrum analysis, use discrete Fourier transform to convert the signal to the frequency domain, monitor the energy distribution of the main frequency point, and determine the performance status or surface material condition of the transducer unit by the offset of the main frequency point. Third, time delay correction is performed by using a temperature sensor installed on the side of the ring-shaped acoustic transducer array to monitor the real-time temperature and correct the ideal flight time according to the linear proportional relationship between sound speed and temperature, thereby eliminating the influence of sound speed drift caused by seasonal changes or residual material temperature. Fourth, feature extraction is performed to identify the first arriving peak in the preprocessed waveform, and the flight time is determined based on the difference between the peak time and the transmission time. At the same time, the envelope area of ​​the received waveform is calculated, and the energy attenuation rate of the sound wave on the propagation path is obtained by comparing it with the reference energy.

5. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step six, the actual filling height is calculated using the surface integral method. By identifying the feature points of the top surface of the material in the monitoring area in the three-dimensional visualization model, the weighted average height of the feature points is calculated as the regional average height. At the same time, the highest point value in the monitoring area is extracted as an overflow warning reference. The triggering logic for the control command includes: When the average height of the high material level area exceeds the preset upper limit threshold, a stop command is executed and the power supply to the feeding equipment is cut off. When the actual filling height of the low material level area is lower than the preset lower threshold, a feeding command is triggered and the feeding equipment is started to replenish the material; during the control process, the system continuously monitors the morphological changes presented by the three-dimensional visualization model until the material level rises back to the safe setting range.

6. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step one, the annular acoustic transducer array adopts a multi-frequency composite unit, with each installation node integrating a low-frequency transducer and a high-frequency transducer. The penetrating power of low-frequency sound waves is used to perceive the overall contour of the deep sand layer, and the surface resolution of high-frequency sound waves is used to capture the micro-undulations on the top of the sand. In step three, when extracting features on the effective propagation path, an envelope statistical feature extraction process is added to calculate the skewness, kurtosis, and energy centroid shift of the echo waveform, and the statistical features are used to reflect the roughness of the material surface. The deep neural network model adopts a dual-branch architecture, which processes low-frequency acoustic data and high-frequency acoustic data respectively. The features of the two branches are nonlinearly spliced ​​together through a fusion layer, thereby identifying the type of material in the warehouse and automatically adjusting the parameters of the voxelization reconstruction algorithm.

7. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step one, multiple sets of the aforementioned annular acoustic transducer arrays are arranged in the vertical direction for the ultra-deep sand silo to construct a three-dimensional acoustic sensing network. In step two, a cross-layer excitation mechanism is introduced, in which the transducer of the upper layer array emits sound waves and the adjacent lower layer array receives them, thus compensating for the observation blind spots between the horizontal layer arrays through an oblique cross-layer path. In step three, time-varying gain compensation is performed to address signal attenuation caused by long-distance propagation. As the flight time increases, the front-end amplification gain is increased according to an exponential growth model to ensure that the reflected signal from the far-end cabin wall is completely captured. The voxelization reconstruction algorithm uses an adaptive resolution mesh, which refines the voxel mesh in the area near the surface of the sand and coarsens the voxel mesh inside the material or in the pure air area.

8. The method for controlling the material level height of a sand silo with dual material levels according to claim 1, characterized in that: In step one, the transducer unit and the central control unit adopt a full fiber optic signal transmission architecture. The acoustic signal is converted into light pulses for transmission in the early stage of generation to resist electromagnetic interference in the industrial field. In step two, the acoustic excitation employs spread spectrum coding technology, the transmitted pulse is set as a pseudo-random sequence code, and the receiver extracts the useful signal matching the transmitted code sequence from the noisy environment through correlation operations.

Images